Lizzy Manzi spent seven years navigating changing requirements at SpaceX, where she eventually led construction across Los Angeles and Seattle, with a team covering roughly 50 buildings. She was responsible for building out manufacturing, test, reliability, R&D, logistics, office spaces - pretty much any infrastructure the company needed. When she finally left SpaceX, former colleagues who had started or joined hardware companies began calling her. They needed help finding buildings, hiring electricians, understanding permits, and figuring out how to build for production.

“They’re rocket scientists. They’re incredibly talented engineers,” she said. “But they usually don’t know construction. And they shouldn’t.”

Those calls eventually led her to start Ground Control in 2024, which began as consulting services and grew into factory management software. Through expert services and purpose-built software, her team helps hardware companies decide what infrastructure to build now and what to leave flexible as their production needs evolve.

In our conversation, we discussed what goes into building a company’s first factory, why flexibility is so important, what that actually looks like in practice, and how she sees the future of managing constantly changing factory floors.

The hardest part is defining what you want

When I asked Lizzy what most often prevents factories from expanding quickly, assuming they have the demand and the money, her answer came back to the company itself.

“I hate to say, a lot of times it’s defining what you want.”

Defining a factory’s scope requires a series of decisions that depend on one another. For example, the equipment needed is based on the product you’re manufacturing. And the power requirements and floor layout are dictated by what equipment you choose. This ultimately then determines what buildings could work and how much construction is needed to set that machinery up.

It all stems from the first decision around what you’re making, and sometimes that is still evolving. This can also be a question of what production happens in-house and what can be outsourced. “If this part isn’t designed yet, how do I know the machine that I need to build that part?” Lizzy described working with a company that had several possible machines in mind for each equipment category. Each had very different requirements and so planning was a series of assumptions and deciding when to be conservative versus when a middle-of-the-road estimate was reasonable.

It’s a balance between developing a plan that is grounded enough to act on, but still leaves room to evolve as the company learns and scales. Informed future-proofing.

“The best project is no project”

“Elon would say the best part is no part,” Lizzy said. “I say the best project is no project.”

A lot of the work in Lizzy’s role at SpaceX and today at Ground Control is simply understanding the underlying need that’s driving a project. This usually means identifying what work helps the business move forward. At SpaceX, her team managed many projects simultaneously, and her prioritization was based on, as she put it, “which group is the bottleneck right now?”

But then, even when a need is clear, there’s still the question of how much net new infrastructure is actually required. Lizzy gave the example of an engineer who thought they wanted a full ISO clean room. After understanding the cost and schedule implications, the team decided that temporary plastic walls and a rental HVAC unit were enough to support the immediate milestone.

“Figure out what functionality we actually need, and then build that,” Lizzy said. Temporary solutions can buy time as well as information. It’s about designing for flexibility. This usually means keeping space open as much as possible, thinking ahead about electrical distribution, and avoiding permanent infrastructure, or “monuments,” that could inhibit future expansion.

Part of designing for flexibility also means accepting the possibility of rework. At SpaceX, Lizzy’s team would procure equipment with long delivery times while design and permitting continued in parallel. “We were pretty good at redoing stuff,” she joked. But it’s about transparency around risk - what could change and what the company might have to absorb if it did. Moving quickly depends on quantifying risks upfront, accepting them (or not) as a team, and taking action.

At the core of all of this, though, is one simple question that Lizzy kept coming back to in our conversation: “Are you making this harder than it should be?” Often, the fastest way forward is to give yourself less to build, and less to undo.

The Goldilocks problem

When a team chooses its first factory, it has to commit to space before knowing how quickly the business will grow. Too small, and it risks needing to move prematurely; too large, and it risks paying for capacity it can’t use yet.

Assessing options for space from brokers is much more than just comparing rent and square footage. Lizzy rattled off a long list of other factors that are equally important for teams looking to build their first factory: “Does this building have enough power? Is it 480V? Does the outside transformer match the inside switchgear? Are the doors big enough? What’s the concrete slab look like? Are there fire sprinklers? Is the zoning correct? Is the occupancy correct?”

When I asked Lizzy what teams most often need help with, she said power. “It’s one of the biggest potential surprises,” she said. Power requirements depend on what machinery is being used and what cycle times look like. Teams can overestimate how much power they’ll need, or badly underestimate it. As Lizzy put it, “In reality, you’re probably never going to be running all of the machines full bore at the exact same time.”

Lizzy advises teams to first understand what electrical service is available today, whether it can support their expected growth, and what an upgrade would involve. If additional capacity is needed, assuming the utility company can support it, she recommends budgeting roughly a year.

Power also interacts with a factory’s layout. Placing high-power equipment close to the electrical service can save hundreds of thousands of dollars, but also impacts production flow. “Power distribution across a large facility can be an extremely expensive project,” Lizzy said.

A building’s occupancy type is also a factor in how much has to be invested into building out a factory. For example, a warehouse that was used for storage that is being converted to manufacturing probably needs changes to fire protection, restrooms, parking, and exits in order to accommodate the increased amount of people and activity. These are the types of costs that aren’t obvious when comparing property listings, but come out when you submit to a city for building permits.

By identifying these costs upfront, Ground Control helps companies understand how much work is needed before production can start, and which projects can wait. “What do we need to achieve on day one?” she asks. “From there, you can think about how you phase the rest of your projects over time.”

When the factory outgrows the spreadsheet

Another pain point in a hardware company’s real estate journey comes when they finally have enough space and equipment, but now they have to keep track of everything.

Lizzy calls this the point when the business is “not spreadsheetable.” Usually there is an out-of-date CAD file stored somewhere. Someone will screenshot it, paste it in PowerPoint, and draw rectangles for where new machines go, or where existing machines will move to. The corresponding equipment specifications are tracked in Excel. Those files then get passed around to architects, electricians, general contractors, or other stakeholders, leaving that person to reconcile the markups with the equipment list. “You’re cross-eyed looking at this whole thing,” she said.

A CAD drawing doesn’t usually help alleviate this chaos because only a handful of people know how to update it, or have a license. They become “the keeper of the CAD,” while the rest of the team responsible for things like machining, inventory, offices, etc. still needs to collaborate and contribute to the plan.

Ground Control addresses this by bringing the factory layout and the information about what’s inside it into one place. A machine’s specifications, such as its make, model, size, power requirements, cut sheet, history, point of contact are all visible from a single floor plan. “I’ll click a box. It says the machine name with a photo of it. It says the power it needs. And I can see the rest of the building’s power distribution and know where it’s fed from.” This makes it easier to translate a production team’s needs into work that other people can evaluate and execute.

The platform is also designed to be used and updated by different stakeholders. Facilities managers, construction managers, and production engineers can have permission to move things on the floor plan. Anyone in the company can find assets, people, conference rooms, and report facilities issues or request projects. This ensures the plan is useful and relevant beyond the initial buildout.

Ground Control’s consulting work shapes which problems its software should address. Working directly with companies gives the team repeated exposure to where manufacturing companies get stuck. “What are their problems? Where do they struggle? What are the SOS moments along their journey?” The goal is to make routine coordination easier, while keeping a human in the loop for the moments that need it.

Over time, this repository of information can help teams assess bigger investments. If a group asks for another 50,000 square feet, the company can first look at what equipment it already has, where it sits, and how space is being utilized. This helps teams make better decisions around whether to expand or reorganize what they already have, and brings Lizzy’s “best project is no project” philosophy into the software itself.

Leaving room to change

Flexibility takes planning. The hard part is knowing which investments are needed to start production today versus which commitments can wait until a team knows more. These choices also need to be documented and referenceable as the company grows, rather than left to tribal knowledge. After all, the person planning future expansions may not be (and likely isn’t) the one who designed the first production line.

This is where Ground Control’s combination of expertise and software can help teams. The consulting work enables teams to make better-informed initial decisions, while the software gives the people using and modifying the factory a shared place to work from and a traceable source of truth. Keeping the requirements, changes, and history connected to the space means each new decision is better informed. Over time, the company builds an understanding of why its factory works the way it does, and what would need to change to support something different.

“You’re going to continue to layer in different needs as the company learns more about their product, their business model,” Lizzy said. Getting the first factory up and running is one milestone in that process. The company will keep learning, growing, and uncovering unknowns, and both the building and the knowledge behind it need room to evolve.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

Every manufactured object is the result of materials, machinery, people, and places working together in systems that have been refined over generations. These systems determine what we can make, how quickly we can make it, and what it costs.

AI gives us new ways to improve these systems. Realizing its full potential means building on generations of manufacturing knowledge and experience, both to improve existing factories and to create new technologies, companies, and factories that show what’s truly possible.

I sat down with John Hart the other week to talk about this. John is a professor at MIT and head of MIT’s Department of Mechanical Engineering. He also co-leads MIT’s Initiative for New Manufacturing. John has also helped found and advise several manufacturing startups, working on everything from technical strategy to team building. He describes his work simply as “trying to make things better at scale.”

In our conversation, we discussed what it means as AI gives manufacturers new capabilities: how to connect those tools to production, learn from what gets made, and build on the factories, suppliers, and expertise already around us.

As John put it: “Manufacturing is about technology, people and organizations.”

Absorbing AI’s full potential

“We can build faster than we were able to build before,” John told me early in our conversation.

Almost every step in the manufacturing process has been, or can be, impacted by AI. This includes everything from planning the process and understanding the costs, to programming the equipment and integrating it with software. And as these individual tasks improve, there’s a lot more room to rethink how a factory operates.

John’s view on what that should look like has shifted. He used to think that automated design and manufacturing might mean one giant model where you could ask it to machine a part and it would handle everything. “Now I think more about how the workflow can be decomposed into the different operations represented by tailored models and agents, and orchestrated at scale,” he said.

For example, if you’re making an injection mold for a plastic part, someone has to design the part, design the mold, simulate the process, and generate the toolpaths that tell a CNC machine how to cut it. Each of these steps requires different tools and different expertise. Breaking up the workflow gives AI more clearly defined tasks and makes its performance easier to evaluate.

“I think good-enough tools along with enough guardrails and prompting for each of those steps already exist,” John said. “But I don’t know if we’re fully absorbing that potential yet.” Connecting capabilities to processes requires a lot of engineering work. “One challenge,” John said, “is understanding how these new tools can fit into existing manufacturing workflows, to drive adoption that allows the truly breakthrough opportunities to unlock.”

And the measure of success is whether improvements in individual tasks translate into better parts, lower costs, or faster production at scale.

Capturing value in manufacturing

Being able to make something faster or cheaper creates an opportunity. But how a company then captures the value of that opportunity depends on what it sells.

I asked John whether selling the output of a manufacturing process is generally a more attractive business than selling the machines themselves. “Maybe. It depends on the industry and the use case,” he said.

John pointed to producers of high-quality forged materials for aerospace and medical applications as an example of how specialized expertise can support attractive margins. “That’s the narrowest point in the supply chain,” he said. “That’s where the specialized knowledge is.”

A company can capture value through both specialized expertise in individual processes and the ability to combine them to deliver a finished component.

VulcanForms is one example of a company that brought multiple processes in-house to deliver finished components to its customers. When John and Martin Feldman started the company, they developed an industrial metal printing technology with the goal that “our business model would be to sell the parts, not to sell the machines.” They brought CNC machining and heat treatment in-house because printing alone didn’t deliver the finished component their customers needed.

“For a lot of customers, we need to own even more of the value chain to be able to deliver them a product that then they can integrate and sell,” John said.

Learning from what actually gets made

This is one reason I find vertical integration in manufacturing so exciting. Owning more of the process can make it faster and easier to establish a feedback loop between design and production. After all, a design expresses what should happen and production shows what actually happened. This means you can identify more precisely what needs to change and then more easily feed those findings into the next design or production run.

John is excited about how collecting data will let us “learn so much more about the relations between design and manufacturing.” He sees particular opportunity in capturing the knowledge generated during the difficult transition from a successful prototype to reliable production.

“How do we accelerate the process from prototype to scale through getting data from the systems and people and processes that have to go through that extreme pain to bring something to full-scale production?” he asked.

But owning the process doesn’t automatically create that learning loop. It’s really about how you get the data, and this needs to be a focus of building the operation from the start.

“This is what enables you to be able to differentiate on product performance while understanding manufacturing so deeply – you can always win on cost and quality,” John said.

Rebuilding the base

A company can’t reach this goal alone. It’s also dependent on the suppliers around it. Even the most highly integrated manufacturers rely on other factories. In fact, it’s the network of these factories that determines what a vertically integrated manufacturer can make and how reliably it can deliver.

John visited factories around Shenzhen in January and was struck by the connectivity of the ecosystem and how extensively automation was being deployed. “In all the conversations I had, there was never the comment, ‘Oh, we’re not automating that yet because it’s too expensive,’” he told me. “Of course, that happens, but I didn’t hear it.” Instead, he saw that if companies weren’t automating something yet, it was because the capability wasn’t there yet.

“The density of what I would say is true innovation was really, really high… and the experience was both daunting and inspiring,” John said. He saw ingredients of a more automated, flexible manufacturing economy that he believes can succeed in the US. Developing those capabilities here will take investment in the entire ecosystem.

As I wrote about here, manufacturing has a tip and a base. The tip includes factories in aerospace, defense, and automotive industries. The base consists of the smaller suppliers, fabricators, and distributors.

Much of where I see the excitement and venture dollars going these days is concentrated at the tip, but there is a lot of untapped opportunity to improve the base, the suppliers that enable the tip.

“Those companies are also sources of innovation and opportunity,” John said. Supporting them requires several things to come together, including talent, which is in part why the Initiative for New Manufacturing at MIT has a technologist training program (TechAMP). It’s designed to help level up experienced technicians.

“This is a necessary but not sufficient ingredient,” John said. “The factory also needs to be able to go get the business to be able to justify the investment.” Training, financing, and customer demand have to reinforce one another.

When I asked whether all manufacturing capability should be brought back to the US, he immediately said, “Oh, no. Definitely not. And it never will be or could be economically.” Instead, it’s about building companies and supply chains that balance cost, resilience, automation, and human work.

That will take time. “It’s a long game,” John told me. “We’re talking about a multidecade rebirth.”

What “new manufacturing” means

That broader ambition helps explain MIT’s Initiative for New Manufacturing. Their definition of “new” includes improving the factories and processes already operating around us. “So much manufacturing is established, and the beauty is figuring out how to upgrade and refine existing manufacturing practice,” John said.

AI gives us new ways to build on that existing knowledge. We can learn more from production and use those learnings to inform designs. We can make better use of the expertise and talent pools currently available. We can build denser supplier networks to improve price competitiveness and lead times in the US.

But doing all of this will take time, investment, and changes to how factories operate.

Toward the end of our conversation, John told me that he tells all of his students to “look around and imagine all the things that have been in factories.” Every object reflects a chain of decisions around materials and processes that were made by a group of experienced people over years or decades.

The things we make next will draw on generations of knowledge about materials, machinery, and production. The promise of “new” manufacturing is to build on that experience, giving existing factories new ways to improve what they do, and to build the new factories and companies that define the future.

What we invest in now will determine how much new capability we can create from what we already have.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

You would not, however, want to rely on their performance in any business setting. The cups end up in a pile (or worse!). The laundry has to get refolded. And in general, any task that takes an adult 30 seconds probably takes 10 minutes, if not more.

This is roughly the state of general-purpose robotics today: magical toddler-like generalizability, but also toddler-like reliability. The demos and generalization are impressive at first, but you wouldn’t want to run a factory on their performance.

Meanwhile, factories all over the world are full of the opposite kinds of robots. Like the palletizer that Formic installed for a mid-sized contract manufacturer packing everything from spices to cleaning pods. The person who used to do that job has, in a sense, been promoted to now doing work that actually needs human judgment. And nobody thinks about the palletizer anymore because they know it will do its one job at 99%+ reliability, around the clock, for years.

But that palletizer, like the millions of other industrial robots installed all over the world, is also so specialized that asking one to do literally anything else is a months-long engineering project, if it’s even possible at all.

Everyone in robotics (the humanoid startups, the foundation model labs, the industrial incumbents, and especially us) agrees on the destination: robots that are both broadly capable and production-reliable. But nobody agrees on how to get there. Between us, we’ve built robots, operated them in factories, and invested in the companies developing them.

We think we know how.

The Map

Here’s how we see the industry on a chart.

The vertical axis measures performance (reliability, cycle time, and uptime). This is basically how well the robot does its job in production. The horizontal axis measures generalizability: the range of tasks, products, and environments it can handle. And perhaps most importantly, the horizontal axis also measures how much work (programming, but also hardware like fixtures and conveyors) it takes to get a robot to do something new.

The top-left corner are the specialists. These include weld cells, palletizers, pick-and-place machines, etc. They’ve undergone decades of refinement, and their economic value is clear and proven. But they’re still incapable of doing anything other than the one thing they were built for.

In the bottom-right corner are the generalists. These are end-to-end robot foundation models, vision-language-action systems and their cousins that can attempt nearly anything you describe in a sentence. They can complete tasks at demo-grade success rates (think 70-90%), and they tend to look incredible on a launch video!

The top-right corner is the general and performant robot. This is the goal.

You don’t get to the top-right corner overnight. You have to traverse the map deployment by deployment, and every intermediate point on your path either generates revenue or burns cash. Robotics, after all, is capital-intensive; it’s an atoms-not-bits industry.

Why the Midpoints Matter More Than the Destination

In the same way AI changed the digital world, robotics will change the physical world. It’s going to create and capture trillion-dollar markets. We believe that as much as anyone: one of us runs a robot fleet, another runs a fund whose thesis is that AI will eat the world, and hardware is its teeth.

But conviction doesn’t fund a market like this. Sustained investment does, and venture capital alone isn’t enough. Somewhere along the way, you need value-creating, cash-flowing businesses.

So then, each point along the path of a robotics strategy matters a lot (versus just where the path ends). Some paths run through what we call a revenue corridor: every step is a product somebody will pay for today. Other paths run through a valley of death: every step is merely a milestone on the way to a future product, funded on faith.

From an investor perspective, it’s the difference between backing a company like Amazon versus a drug discovery biotech. Both required billions in sustained investment across years. But Amazon funded its warehouse and datacenter buildouts by reinvesting profits back into the business. Biotech companies can raise billions across a decade and then simply die because the science failed. Both models can produce wins, but one has more control over its own destiny.

Revenue at the midpoints isn’t just important because of the runway it provides. Paying customers are compasses, helping you navigate as you build. They tell you what matters and what edge cases in the real world look like.

The Vertical Path or the Valley of Death

Some companies are choosing to start in the bottom-right and climb straight up. This is the generality-first bet being made by very smart people with very large checkbooks. In their view, generality is the hard part and performance is a scaling problem. If you train on massive datasets and deploy broadly, you can climb by letting iteration push reliability all the way up.

It worked for software and language models, so why not atoms? This is the thesis, in different forms, behind Optimus, Figure, 1X, Physical Intelligence, and Skild.

But factory economics are threshold-based, not linear. This means a robot with 85% task success isn’t worth 85% of one at 99.9%. In fact, it’s actually probably worth less than zero because each failure leads to wasted time, scrapped product, and/or additional headcount needed to oversee.

Saman remembers seeing just how costly a few hours of stalled production was to a customer in Formic’s early days. A stopped line at a mid-sized food manufacturer can cost $120,000/hour. So if a robot’s mistakes bring that line to a halt, getting it right only 85% of the time is pretty catastrophic.

The climb from general-but-unreliable to general-and-reliable happens off-revenue. In other words, if your general robot isn’t reliable enough for customers to pay for it, you have to fund that climb yourself. And without customer deployments, it’s very hard to know whether your internal benchmarks are accurate reflections of what will happen on a customer’s production line.

And while you’re climbing, the bar keeps rising. Specialists continue improving, and hard automation is only getting cheaper. The reliability threshold is set by your customer’s next-best option and that’s not a fixed altitude.

“But LLMs Did It Vertically”

Yes. GPT-3 was a general-but-unreliable system. It hallucinated constantly and yet a lot of people paid for it anyway. The generality-first crowd is pattern-matching to the most successful technology rollout of the century.

So why doesn’t it transfer? Three asymmetries.

Asymmetry #1: Cost of Error

A bad token is basically free. But a bad grasp is often a crushed part, a jammed line, more scrap, or worse. LLMs monetized their awkward adolescence because humans could fix the mistakes before their output went into the world. This isn’t the case for robots.

Asymmetry #2: Data

It would take a human more than 100,000 years to consume the text and images used to train today’s LLMs. There is no equivalent for manipulation data. The data that trains general robot models still has to be manufactured.

Asymmetry #3: Distribution

ChatGPT launched into a world where billions of people already owned the hardware needed to use it (think computers and phones). Robotics in 2026 doesn’t look like computing in 2023. It looks like computing in 1985. Today, there are ~5M industrial robots operating worldwide, but they were built to follow programmed routines, and getting them to work with newer AI control systems isn’t a simple software upgrade. The install base for these newer systems barely exists. Even if we handed you Physical AGI on a hard drive tomorrow, you couldn’t ship it, because the robots to run it on aren’t deployed.

Put the three together, and the LLM analogy doesn’t support the vertical path for robotics. In fact, it does the opposite. The winners of the LLM era were the ones who had the most data and the widest distribution. The winners of the robotics era will be the ones that start with narrow tasks that customers will pay for, then use the data from those deployments to expand. Over time, they will build toward a general robot while getting paid along the way.

The Horizontal Path: Revenue Is What You Need

Returning back to the performance vs generalizability chat, let’s look at companies that choose to start in the top left corner, and move right.

With this approach, you begin with a system that clears some reliability threshold in at least one task so you have something that can generate revenue from day one. Then you gradually widen, rung by rung on a ladder that looks like this:

Rung 1 same task, more variation: the same robot handles more SKUs, package sizes, etc and can do it with faster changeovers. As an example, one Ohio snack manufacturer that Formic works with went from 300,000 pounds a week to over a million on the same footprint, starting with one palletizer and expanding once it proved itself.

Rung 2 adjacent tasks, same robot: the palletizer adds case packing; the welding bot adds setup and fixturing. It’s the same workspace, but it reduces the labor requirement. This is where generality starts kicking in.

Rung 3 adjacent stations, same facility: after a while, you earn the customer’s trust. New capabilities start to get adopted and expansion starts happening via conversations rather than formal sales cycles.

Rung 4 adjacent facilities and verticals: this is where the models transfer and the playbook starts to really compound. What you learn on one product line carries over to the next.

Importantly, every rung is individually purchasable.

This is the playbook of the companies that are actually winning customer contracts in physical AI right now. Path Robotics, for example, started at the top-left corner of our chart, a weld cell, but AI-native, so every weld across every customer trains a shared model that makes the next weld, and the next customer, better. Now they are expanding to other embodiments like quadrupeds and humanoids.

Saman and Formic are doing it in packaging and palletizing: 200+ robots across 120+ US facilities, running at 99.3% uptime, each one both a cash-flowing deployment and a data-collection site. For example, at a California dairy, one operator said the palletizing work had left her “unable to sit or stand after a shift.” A robot now does that job. Another example is at a plant making sauce packets and tableware where annual turnover was 130%. Robots helped the plant increase output by 25% and eliminate shipping errors. These examples are from different plants with different products, but each followed the same path. They started with one task, proved it worked, and expanded from there.

By starting narrow, you have customers that are ready to pay for work the robot can already do. And by doing that work, the robot is exposed to more variation and more things that can go wrong, giving you the data to improve it and take on the next task. This is how a business built around one seemingly narrow task can grow into something much more general.

Don’t Become a Systems Integrator

The horizontal path has a failure mode, and it’s called becoming a systems integrator.

There are thousands of automation integrators, capable, profitable, decades-old companies that have built specialist robot cells since the 1980s and have never generalized. Every deployment needs bespoke engineering, with nothing shared between projects. Each deployment is a snowflake with custom fixtures and custom code.

The escape from the integrator trap is an architectural commitment that is only possible today with AI: a shared model and software layer across the entire fleet, so every new deployment makes all deployments more capable. This is where AI changes the game in a way it never could for the integrators of the 1990s. Value built across customers and applications propagates back into one compounding asset. The integrators never had a substrate that could learn across projects, but now the substrate exists to build a company where every deployment makes the next deployment cheaper and faster.

The Fleet Is the Path

The top-right corner will be reached by whoever operates enough performant, data-producing deployments to fund and feed the generalization at the same time. It won’t happen from a single model breakthrough (the vertical bet) or by stacking a thousand bespoke point solutions (the integrator trap).

And while both the vertical and horizontal trajectories appear symmetrical on paper, they diverge on a third dimension: cumulative deployed surface area over time. The horizontal path continuously expands its active footprint across new customers and new capabilities. The vertical path, however, doesn’t have a viable commercial product until performance is finally good enough. This is an important difference in an industry where the install base provides both your revenue and your training data.

The horizontal path is the fast path for everyone.

America had 409,000 open manufacturing jobs in 2025, and as many as 1.9M could go unfilled by 2033. The vertical path asks factories (and investors) to wait for uncertain breakthroughs. The horizontal path deploys now: every working robot adds capacity, strengthens factories, and generates data for better models. That flywheel is why Saman helped start Robots for America. No company can close the gap alone; manufacturers, robotics companies, and policymakers have to move together. The greatest risk is that the flywheel spins too slowly.

That’s also why Haomiao and Annelies are excited to back companies that generalize within verticals: Path is expanding from welding into assembly and post-processing; Rhoda is tackling more factory manipulation tasks. Neither will fold your laundry anytime soon. Robotics abundance runs through this glorious “useful” path - what Travis Kalanick calls “gainfully employed robots”- getting more capable every quarter in unglamorous buildings, years before the glamorous version arrives.

The question was never what the end goal is. It’s whether anyone is paying you while you travel there.

Big thanks to Saman Farid and Haomiao Huang for sharing their knowledge. Saman is the founder and CEO of Formic, which deploys and operates robots-as-a-service across US manufacturing, and a co-founder of Robots for America. Haomiao is Founding Partner at Matter Venture Partners, an early-stage venture firm focused on HardTech, and was previously an investor at Kleiner Perkins, a roboticist and Co-founder & CTO of Kuna Systems.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

Google, Google DeepMind, and MIT FutureTech recently published a study, AI in Science: Early Insights, to understand AI’s impact on science so far. The study looked at a corpus of 15M Gemini interactions, 2,690 specialized models, and a survey of 637 US and UK researchers. In the study, researchers report saving ~7 hours a week using AI, but nearly half reported a growing backlog of untested hypotheses.

In other words, as I wrote about here, AI is speeding up individual research tasks, but the overall process of turning predictions into validated discoveries is still constrained. This is the gap between what we can design or predict and what we can actually validate.

I sat down with Mihai Codreanu, a Google senior research economist who co-led the study. His interest in science stems from his interest in economic growth. “A big part of why economic growth happens is innovation and scientific discovery,” he told me. “If AI pushes science, it will also be pushing economic growth.”

But measuring that impact requires looking beyond the tasks where researchers already find AI useful. As Mihai put it, “discovery is still constrained by the types of tasks that AI may not yet be very good at.”

In our conversation, we unpacked the results from the study to understand where the bottleneck is today, how that is evolving and what impact this has on scientific research more broadly.

The bottleneck moves downstream

There’s a big difference between having more hypotheses and having more discoveries. “The micro level impact seems to be clear, scientists find AI very helpful in certain aspects of their jobs,” Mihai said. “But in order to see very big impacts on macro level discoveries, we have to think of all the other things that are within the scientist workflow and outside of the scientist workflow that need to happen.”

Nearly half of participants in the study said their main bottleneck had moved downstream over the last two years. Nearly a quarter said their constraints were in physical experimentation and/or data collection, closer to ⅓ for participants from physical sciences and engineering.

This implies that AI will accelerate discovery faster in fields where hypotheses can be tested computationally. “Computational fields can close the entire loop in silico, without leaving the digital environment,” Mihai said. But even those fields have constraints: access to compute, the cost of large simulations, and high-quality data and benchmarks.

And in every field, there’s a downstream cost around verification, which requires researchers’ active participation.

The verification tax

Mihai said one of the findings that surprised him most was “the sheer size of the verification tax.” 88% of scientists reported spending more than 10% of their saved time verifying AI outputs; nearly half spent more than a quarter of it. “AI is giving these productivity benefits, but there’s also the cost of verifying outputs,” he said.

What verification means depends on the field. In math or computer science, it’s reviewing code or a proof. In health and life sciences, which reported the highest share of saved time spent on validation, it means testing in a wet lab. In physical domains, the verification tax and the downstream bottleneck are the same thing: checking the model’s output is the experiment.

And Mihai doesn’t expect the tax to completely disappear as models improve. “Things like clinical testing or experimental replication, as well as expert judgment are core to the scientific method,” he said. “So even if models get very good, we’ll still need humans to validate them, at least in the short and medium term.”

Choosing what to test

“We don’t specifically look at the quality of these hypotheses,” Mihai told me when I asked about how we know how high quality this backlog is. But he suspects they are likely plausible enough to investigate and thus added to a queue. “At least for now, it seems that many of these hypotheses being reported by scientists are worth considering.”

For a lab with limited experimental capacity, an expanding list of plausible ideas creates a selection problem. Researchers have to decide how to prioritize experiments and when to abandon a direction.

“I am bullish on AI being able to help with this,” Mihai said. One potential path he described was around scoring frameworks that evaluate different hypotheses based on a variety of factors, such as feasibility, novelty, and expected value.

The hard part is deciding what to optimize around. For example, feasibility vs expected value. One hypothesis might be easy to test but scientifically incremental whereas another might be much more uncertain (and expensive), but potentially a lot more valuable. Choosing between this tradeoff means using judgment about what the likely result will be and its impact.

“The collaboration between human scientists and AI is going to be key,” Mihai said.

The pull toward safer questions

One of the more interesting tensions in the research concerns the questions scientists choose to pursue. The majority of respondents reported greater access to insights from other disciplines, and greater ambition and broader research agendas. But at the same time, about half of respondents also said AI encouraged safer projects in their own work. In other words, AI is making it easier to do more obvious work, faster.

Mihai described two effects happening at once. “At the field level, AI can help scientists synthesize literature across disciplines,” he said. “And this could inspire a broader research agenda and a more ambitious vision.” For example, a researcher could more easily discover a method from another field and then connect that seemingly separate idea with their own work. This expands their view of what’s possible.

However, the other side of this is that when it comes to choosing a project, researchers will likely be most drawn to problems that already have a well-established corpus of data. “AI may lower the cost of tackling well-defined, data-rich problems that we have good data for and good benchmarks for,” Mihai explained, encouraging “lower-risk, highly achievable projects.”

He described this as a possible “streetlight effect,” which is the tendency to investigate what’s easiest to study, like looking for lost keys under a streetlight because that’s where you can see. In this context, AI may steer researchers toward questions with abundant data and established benchmarks, even when harder-to-study questions could be more consequential.

The concern is whether the ease of pursuing incremental work starts to steer the research agenda away from riskier, more novel questions. Mihai wants to better understand how the decisions around research topics accumulate over time. “We’re sort of thinking of this as measuring citation dynamics over time,” he said. This means tracking where major breakthroughs show up and whether diversity across research subfields expands or shrinks.

After all, this selection process shapes which ideas get explored and which have a chance to become discoveries.

What counts as scientific progress?

Toward the end of our conversation, I asked Mihai what would convince him that AI had enabled discoveries that otherwise would not have happened, or would have taken much longer.

“I would want to see more longitudinal data,” he said, pointing to high-impact findings and nonstandard breakthroughs. “I’d also look for proof AI is helping genuinely novel types of discoveries,” he added. To do this, you need to follow what happens after an initial result to understand whether other scientists find it useful, and if it opens new directions of research.

Mihai added that “it’s perhaps not enough to be correct. It also really matters how you are able to interpret and what you’re able to learn from these sorts of findings.” The value of a finding can extend well beyond the question it originally answers. For example, it may give researchers a new method to work with or a reason to explore a new direction they hadn’t ever considered.

He gave the example of AlphaGo’s “Move 37” in its 2016 match against Lee Sedol. “Those types of very creative moves are where I’m most excited about the potential impact of AI,” he said. In science, getting there requires having the capacity to recognize an idea’s potential and then test it.

“The promise of AI in science is that eventually we’ll be able to make those types of moves that are incredibly creative, that go beyond the limit of what human minds or researchers have tried on their own,” Mihai said. That’s the promise of AI for science, and it depends on what happens after we have a hypothesis.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

“It’s always harder to solve a Jenga problem than solve a chess problem.” Praveen Rao told me the other week when we were talking about AI in manufacturing.

Unlike many other industries, in manufacturing, you’re working with a physical environment and legacy existing equipment, and many interdependent processes. Everything is tightly interconnected and thus every action has consequences for the rest of the system.

Praveen leads manufacturing at Google Cloud and is a mechanical engineer by training. Over his multi-decade career in manufacturing, he’s worn the CIO and CTO hats and been at companies including IBM, and AWS.

In our conversation, we talked about what it takes to move industrial AI into production, how autonomy changes jobs and the economics of manufacturing, and where founders can build beyond the hyperscalers.

“Your pilots are like the seeds, but data is the soil that you have to till”

“In manufacturing, we don’t have the luxury of scrapping everything and starting over,” he told me when talking about the constraints shaping the industry right now, everything from the data architecture to the buying decision. A plant manager is measured on production, delivery, and safety so what they care about is, as he put it, “delivering what they’re supposed to deliver, and doing it safely.”

Praveen calls the scaffolding around many early pilots “duct tape architecture.” It can demonstrate feasibility while leaving the harder deployment questions unanswered.

“Can you do it with 10,000 people collaborating with each other,” he asked, “while accounting for change management and the response times the operation needs?”

“Your pilots are like the seeds, but data is the soil that you have to till and stakeholder sponsorship is what sustains the effort,” he said. In other words, you have to build the conditions in which the pilot can become an operating system people actually use.

“Always keep the outcome in mind, what you are trying to achieve.”

Semantic data is the assembly line for profitability

“These machines were built thirty years ago. They absolutely didn’t have any semantics or context,” Praveen said. “You go get the data from a machine, and you just get a sensor value of 25. Somebody has to interpret what that really means. And you have to do this 100,000 times, across multiple factories, with so many different sensors.”

There are two connected challenges around understanding data from machines and then being able to reason over that data. The first is integrating equipment that uses different communication standards, such as Modbus, MQTT, and OPC UA.

And the second is then giving its data enough context to be useful. This means understanding what a reading measures, which machine it belongs to, and how it relates to the production process. This includes connecting to the rest of the production process (the order, the bill of materials, the supplier, the production line, the operator, and the history of how a product was made). This is the semantic layer and, coupled with raw sensor data, is what enables an AI agent to operate effectively.

“Semantic data is really the assembly line for profitability.”

So across the various handoffs between an engineering system and production planning, the meaning and the relationships have to survive each handoff.

Praveen pointed to Gemini’s benchmark strength at navigating graphs, then connected it to the shop floor. “Assemblies of products are nothing but graphs,” he said. “You have a big assembly, subassemblies, and parts.” Once a model can navigate that graph, applications can be built on top. Examples of those applications are change-order management, identifying duplicate parts, surfacing similar past projects to build better proposals, etc.

I asked whether the age of installed equipment means manufacturers will eventually have to start over. Praveen doesn’t think so and actually argued it wouldn’t be economical anyway. Instead, he sees the path being more around retrofitting sensors to capture signals like vibration, temperature, or pressure in situations where the equipment itself doesn’t provide this granularity. In other words, he believes the architecture has to accommodate what is already there.

AI permissions in a factory

As discussed with Shiv Trisal from Databricks the other week, manufacturing has been plagued by a surplus of insights and alerts, but is still behind on connecting those to actions and interventions.

Praveen described a progression “from systems of engagement, systems of record, to systems of intelligence, to systems of action.”

“The diagnosis needs to lead to treatment, another measurement, and confirmation that the patient improved,” he said. In a factory, that means connecting observation with execution with feedback.

“You need to work in a deterministic environment and in a closed loop environment when you’re coming to manufacturing,” he said.

The system needs clear boundaries around the actions a model can propose and the conditions under which those actions can execute. Permissions, process constraints, and operator authority all become part of the product.

“It’s a contract between the AI tool and the systems of record that’s providing this input. That way, AI accesses only what is needed. It executes the contract versus overwhelming a system,” Praveen told me. He gave the example of thousands of agents requesting information from an ERP system. Even an otherwise useful agent can create problems through excessive requests or access beyond the needs of its task.

Human authority has to be explicit, too. “You empower your people to have the kill switch,” he said. “That’s how you hold people accountable as well.”

Praveen sees the worker’s role expanding from “a doer” to “an orchestrator”. “So in essence, they got a promotion,” he said. The new workload includes orchestrating events around diagnosing equipment, configuring machines, coordinating workflows, and initiating maintenance.

I’m excited about this shift because it changes the economics of manufacturing. When a worker orchestrates ten workflows instead of executing one, output stops scaling linearly with labor. Productivity per person and the margin structure of the factory begins to look less like traditional manufacturing.

The economics of autonomy

There are projected to be 3.8M open manufacturing jobs by 2033. And today, it takes two to five years to onboard a manufacturing employee. “We can’t just hire our way out,” Praveen said.

Physical AI and robotics are part of the answer to that gap. When I asked whether the hype around humanoids was justified, Praveen was optimistic about their potential. “The world is built for humans by humans,” he said, so there’s a good reason to build machines that can function in environments designed for us.

But even 99% accuracy is too low in the manufacturing world. Six Sigma, the industry’s standard quality bar, allows 3.4 defects per million. Most robotic benchmarks aren’t even close to that. Even on controlled simulation benchmarks, top models land in the 70-90% range.

Praveen’s advice for deploying robotics mirrors his advice for deploying AI generally: “Start from an outcome. What is the thing you are trying to do? Work backwards to what kind of body you need to be able to do that.” The form factor is a byproduct of whatever problem you’re trying to solve. Put another way, he said “Don’t obsess over the body versus the brain.”

Another factor in all of this is how humans will work alongside robots in the future. Praveen believes there will always be an important role for humans in manufacturing and so he thinks we need to do more research into how humans stay motivated and relevant when surrounded by machines that can perform parts of their job better than they can. People need a clear understanding of their own responsibilities and growth trajectories. Their cooperation and involvement will shape robotics adoption just as much as the technical capability of the systems.

After all, “people support what they help create,” Praveen said.

The opportunities

Toward the end of our conversation, I asked Praveen what he would build if he were starting a company today.

He first answered by saying what he wouldn’t build, which is anything around an apparent gap in a cloud offering. “You never know what they have in the product road map. They may be planning on it, like, two months from now.”

Instead, he pointed toward operational technology and the application of AI within industrial processes. “Think about what we are doing in IT and replicate it on the OT side. It’s almost a greenfield. You have the whole world for yourself.” Describing Google’s position, he said, “We don’t see us building manufacturing machines. That’s not really our forte.”

For the models themselves, his advice is to build on top rather than compete underneath: “Take these large language models, which are like the runways, and build your airplane that can take off from them.”

It’s hard not to share his enthusiasm. “We’re living in the most exciting times in human history,” he told me. Problems that stood unsolved for a hundred years are falling. “If we can do it now, imagine what we can do two, three years from now.”

Check out this video to learn more about how Google Cloud is using AI and data to transform the operations of global manufacturers.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

The other week, I sat down with Shiv Trisal, who leads global GTM at Databricks across manufacturing, automotive, transportation and logistics, energy, and utilities.

Early into our conversation, he told me about a recent conversation he had with the CEO of a large industrial company. The company had invested heavily in digital infrastructure, and Shiv asked how long it takes them to understand the factors driving the company’s margins.

“Shiv, it can take me three weeks.”

Then the CEO paused.

“If at all.”

The CEO explained to Shiv that someone was still having to pull information from different systems and then figure out what had happened. This “human middleware,” as Shiv put it, is still needed in most production deployments across industrial enterprises.

Shiv’s career has been focused on making technology work in legacy industrial settings. Before Databricks, Shiv led product strategy and commercialization for all of Raytheon Technologies’ Connected Aviation division. In this role, he built a predictive aircraft-health platform that’s now deployed at over 50 airlines. He also built the commercial engine to sell those products.

“AI has to change the fundamental economics of the work being performed,” Shiv explained when I asked how he defines success for AI inside large industrial companies. In our conversation, we discussed what this means in practice, why it’s still breaking down in places, and where the biggest opportunities are.

Redefining the Unit of Value

When Shiv was building predictive maintenance products, his team initially thought that if they predicted a component was going to fail and could send an alert with a recommendation before it happened, that alone would provide a tremendous amount of value.

In part, it did. But it also meant that for every alert, someone still had to review and validate it, and then decide what to do next. So instead of alleviating workloads, the alerts actually created more work to manage, especially since many of the alerts were false alarms.

“Did we make that human’s life harder or easier by overloading that person with more alerts?” he asked. “We called it ping and ding. Ping, meaning alert, and ding, meaning showing up in an inbox. If there were a thousand pings and ten thousand dings, that’s not helpful for anybody.”

Today, Shiv defines value in terms of resolution. A system has to usher a problem all the way through to a completed action before it can be considered valuable. “I can research faster, but it doesn’t mean I did work faster,” he said.

As I wrote about here, this applies to anyone building AI solutions that touch the physical world. Speeding up one step usually means the bottleneck moves somewhere else. So which step you accelerate matters more than how much you accelerate it.

John Deere as a Case Study

One of John Deere’s early AI applications was parts forecasting. Across its dealer network, different locations each had different approaches to deciding which parts to stock. Deere wanted to better connect equipment needs with inventory decisions.

“Inventory management is actually the moneymaker. Parts are the profitability lifeline for equipment manufacturers,” Shiv said.

Shiv described Deere’s investments in equipment telemetry, parts information, and service data as a way of making the business understandable to software. “They built their own language, and I use the word language freely here, of how they run farming.” This is all context about the relationships between a machine, its customer, its service history, and the relevant policies. “We basically know what’s happening at any farm, and we basically know where parts are,” he added.

More importantly, however, this work built a foundation for subsequent applications. He pointed to repair and warranty quoting as an example of what this enables. Because Deere built this live machine-readable context layer, agents can handle qualifying requests without waiting for a human to be at their desk. “After all, what’s the point of a human in the loop if the human is sleeping?” Shiv joked, adding that quotes are won or lost based on a matter of hours or less sometimes so if you’re waiting for a human to manually approve a quote, that can cost you the deal.

In other words, Deere did the hard work of making its business understandable to software upfront, so deploying agents was actually the easy part. Today, building the context layer is where most companies are getting stuck.

Making Industrial Knowledge Usable

“The reason AI works is the data. It’s not the models,” Shiv said. “The models became too much of the spotlight. They’re important. They make data more valuable. But without data, the models can’t actually do any work reliably inside of a company.”

And the data that matters most isn’t public. “Nobody tells you how to do defect detection in the semiconductor industry on Reddit,” he said.

The most valuable data is bespoke to each company. A general-purpose model needs access to that information and an understanding of how it fits together.

AI can help with this. Shiv gave the example of field engineers writing reports of what happened to an asset, each report with a different format. “That’s a problem that would have taken a data scientist to go train a machine learning model on a very specific corpus,” he said. “Something that would’ve been a six-month project is down to hours.”

He also sees opportunities to automate the work of connecting new data to an existing representation of the business. “Why does it have to take a human every time to go physically model every aspect for your use case? You should always have a connected knowledge graph that gets automatically refreshed as you add more data.”

Shiv breaks the path from data to action into three layers.

• Context: machine-readable understanding of equipment, people, parts, policies, and relationships

• Knowledge work: analysis, forecasting, retrieval, and reasoning needed to decide what should happen next

• Action: permitted changes a system can make to carry a decision through (this layer needs very explicit boundaries)

Each layer should use different tools. For example, mathematical and optimization tools for calculations, machine learning for forecasting and patterns, and language models for tasks that need interpreting and summarizing information. “Where I have certainty, use math. Where I have uncertainty, use machine learning because you can quantify uncertainty,” he said.

The Business Case Is Part of the Product

Companies that start by asking for “agentic use cases” are choosing an approach before defining the problem. “What you’re telling me is you’ve presupposed a solution even before you’ve defined a problem. Why would you do that?”

Instead, Shiv’s starting point is competitive advantage. What does the company do particularly well? Where could AI strengthen that advantage, and where might it threaten it?

These questions require leadership involvement. “Without top-down, bottom-up doesn’t succeed,” Shiv said.

Executives set priorities and budget, but the broader organization adopts the tools. “You need both perspectives and buy-ins to build something useful,” Shiv told me.

People spend time learning tools and changing their routines. “Your productivity actually drops before it goes back up,” Shiv said. “The question is how many places in the organization can you afford that to happen?” Companies need to be selective about where they choose to implement AI, and how many changes they take on at once.

“All ROI analysis is wrong, but it needs to be done,” Shiv said, quoting a line he says he’ll repeat for the rest of his life. It’s the stake in the ground, a starting estimate that forces priorities, and that you measure reality against and revise as you learn.

The 50% Flat Tax on Engineering & The Opportunity

Toward the end of our conversation, I asked Shiv what he would build if he were starting a company today.

“Past waves of digital transformation have left engineering and R&D mostly untouched,” he said.

He sees a substantial amount of engineering work happening between and around existing tools. “My experience of working with engineers in the physical world is that they pay a 50% flat tax on their productivity. Even if there were software paradigms they were trained in, the work they do outside the system is where the gap is.”

He gave the example of information generated by simulations. “I don’t need to recreate a simulation engine, that’s not the point,” he said. “But the artifacts from those simulations: which ones were successful, which ones weren’t, what were the parameters? If I have enough of that, I can create a faster model to help people reduce their reliance exclusively on simulations,” he said. Simulations are expensive and slow, so the accumulated record of what worked and why is a shortcut and narrows down what engineers need to investigate.

Engineers build on prior designs, combining known approaches and investing in the candidates most likely to succeed. “Everybody gets obsessed with breakthrough innovations, whether it’s in material science or drug discovery,” Shiv said. “But the reality of most engineering work is it’s derivative. You integrate, validate, test, and pick the designs with the highest chance of success.” There is an opportunity to support this whole sequence, from finding relevant prior work through evaluating options and coordinating next steps.

“AI will create a whole new class of software that we don’t even know yet,” Shiv said. “There is 99% of software that comes in the next five years where we don’t even know what it looks like today.”

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

The first time I took a Waymo, I filmed it with my mouth gaping open. Watching the steering wheel turn on its own was like something out of an Isaac Asimov novel. Looking back, I think I was in awe at the mechanics of it, what was happening to the steering wheel and blinker and brake pedal without anyone in the driver seat. I thought about that moment again recently after speaking with Wenhao Gao.

Wenhao is a postdoctoral researcher at Stanford and an incoming professor of Chemical and Biomolecular Engineering at the University of Pennsylvania. His work focuses on connecting AI molecular design to the experiments needed to test those designs.

Wenhao’s analogy for lab automation is self-driving cars. “I don’t think moving the car mechanically is the tricky part,” he told me. “Before making a decision, it needs to know the environment around it: the traffic lights, the lanes, the pedestrians, the other vehicles.”

Wenhao sees a similar challenge in chemistry. Many routine operations, including mixing, heating, and dispensing, can already be automated. “You can put A in this position and B in that position and fix them there,” he explained. “You can even hard-code those positions into the robot.”

But after the reaction, you have a mixture in a vial and you have to decide what the next step should be. “This requires first knowing what actually happened inside the vial,” Wenhao said, “this part of interpretation is the fundamental of decision making and much harder to automate than anything around the equipment.”

In our conversation, we talked about where connections are breaking in labs today, what automation can help with, and why it’s useful to limit what a model is allowed to generate.

From molecular simulation to experiments

Wenhao was initially drawn to molecular simulation and the possibility of understanding a system by modeling it. But he recognized the cost and limitations of those simulations and in 2017, he read Automatic chemical design using a data-driven continuous representation of molecules by Rafael Gómez-Bombarelli and colleagues. In this paper, molecules were represented in a continuous mathematical space, which made it possible to search for new structures by moving through that space.

“I thought, this should be the future,” Wenhao said and he shifted his focus from molecular simulation toward generative molecular design.

During this period, he came to realize the importance of experimental validation. A model can propose a molecule and predict its properties, but you still have to validate it through an experiment. This is also how you ensure you’re improving future designs.

“If you have good verification, you can close the feedback loop,” he explained.

This brought him into the wet lab where he worked on automated synthesis and saw just how much automation depends on physical resources. “In academia, you typically expect that a couple of smart ideas can fundamentally change the behavior of a system,” he said. “But lab automation is very resource intensive.”

He gave the example of an LC-MS instrument. If you have one, you have one instrument’s worth of throughput. If you want more capacity, you need more instruments. Better software can help to a degree, but you still have to account for the time and equipment each experiment needs.

He ultimately came back to reaction modeling and molecular design, but with a much clearer understanding of what models will need to connect to in order to be valuable.

Where the feedback loop breaks

Early in our conversation, I asked Wenhao to walk me through the step-by-step process of going from a desired property to a molecule in the lab.

“We start from a property profile that we want to have,” he explained. In drug discovery, that means a combination of objectives, including activity, selectivity, and safety.

From there, the process has roughly four parts:

1. Design. Identify or generate candidate molecules with the properties you want.

2. Synthesis planning. Work backwards from a candidate structure to reactions and starting materials that could produce it. This is called retrosynthetic analysis.

3. Execution. Run the reactions, then separate and purify the products as needed.

4. Readout and testing. Establish what was made and measure whether it has the properties you wanted. Use those results to decide what to try next.

“If you provide the correct information, I think we can expect current language-model agents to do a reasonably good job on some synthesis-planning steps,” he said. “That part, I’m cautiously optimistic about.”

But the harder problem is the readout.

LC-MS helps separate components of a mixture and provides information about their masses. But these outputs still require some amount of interpretation. “We’re getting indirect information, and we need to infer the molecular structure from that information,” Wenhao explained. “That requires intelligence, but it is also still an open scientific problem.”

This connects to something I’ve been exploring and wrote about recently with Anna Marie Wagner here. Recording what happened in a lab gives us information but that information rarely makes it into a written protocol. And even with a complete record of what the equipment did, we still need to understand what happened chemically.

Checking whether a reaction yielded a specific molecule is generally an easier problem versus trying to identify an unknown molecule from scratch. This is because the problem is more constrained. Recent work such as MARLIN, a July 2026 preprint, explicitly tackles structure generation without being given the correct formula.

The interpretation of an experiment determines what the system does next. An incorrect identification can send the next round of experiments in the wrong direction. This is problematic for all labs, and especially so for autonomous labs.

Designing molecules with a route to the lab

In a 2020 paper with Connor Coley, The Synthesizability of Molecules Proposed by Generative Models, Wenhao looked at how often molecular design algorithms proposed structures that a synthesis-planning program could find routes to. For some optimization tasks, the program couldn’t find a route for any of the top 100 candidates.

“The core idea was that we wanted the model to learn synthesizable chemical space only,” Wenhao explained. The research highlights how a model can perform well against a computational design objective but still produce candidates that are difficult, even impossible to test.

SynFormer, developed by Wenhao, Shitong Luo and Coley, incorporates synthesis into generation. It takes into account a set of starting materials and a defined set of reactions and then generates pathways. Though a computational route still has to be experimentally validated, the goal is that the model has already accounted for a constraint that would otherwise appear after design.

The aim is “something we can get within three, maybe up to five steps,” he said. A molecule needs to be accessible within a reasonable experimental cycle to be considered as a design candidate.

I asked what you give up by imposing that constraint. His answer was synthetically complex molecules. Some interesting molecules will fall outside the reactions and building blocks the system can use, or require much more complicated synthesis. “There is a cost,” he said. “But overall, that’s a price I would want to pay.”

You can still generate new molecules within the space, which can expand as more reactions become more reliable. You are just limiting the ways they can be assembled.

What models learn from known chemistry

“A generative model, by design, learns the distribution of its training data,” Wenhao explained. “It’s hard to expect it to generalize to something very far away from that.”

But chemistry involves creating structures that have never existed before. Reading the literature gives a model useful knowledge about what’s been done, but it doesn’t establish that the model can reason reliably about a new structure. Its predictions still need to be tested, and a lab can only run so many experiments.

The literature also makes it difficult to learn from the experiments that have already happened. As I wrote about here, scientific papers were built for human readers. “They refer to ‘molecule 2,’ and there’s a structure drawn there,” Wenhao said. Understanding the result requires connecting that drawing to the text, the reaction conditions, and the measurements. For a model to learn from that work, the context around those connections have to exist.

This means what you test is really important, and a good model should help a lab choose which candidates are worth testing.

Wenhao’s Practical Molecular Optimization benchmark looked at one part of this around how efficiently algorithms find promising candidates. It compared 25 algorithms across 23 tasks, and gave each run a budget of 10,000 computational evaluations. Under this budget constraint, most newer methods failed to outperform their predecessors.

The result highlights why evaluation budget needs to be part of how we judge a molecular design algorithm. “If you generate 10,000 designs but none of them proceed to the wet lab, a lot of compute has been spent on that generation process.” Wenhao said. But you basically get nothing.

This adds to the argument I made here: as design becomes cheaper, more of the value depends on proving what works. In chemistry, that means proposing molecules a lab can make, choosing informative experiments, and capturing results (and failures!) in a form that improves the next design.

For me, that changes how we should evaluate these systems. Given the same lab capacity, does a model help us find a molecule with the properties we want in fewer experimental cycles? And does each cycle give us new information we can use again? Those are improvements that would make cheaper molecular design translate into faster discovery.

What an autonomous lab still needs

Toward the end of our conversation, I asked Wenhao what he envisions the end state will look like. He described “something like a 3D printer for chemistry: a small unit where we can design a molecule, synthesize it, and test it.”

To get there, he thinks some operations should stay manual because they are probably easier for humans to do. “I don’t think 100% automation is needed,” he said. Automation is useful in so far as it improves throughput and makes experiments more consistent, better documented, and easier to learn from.

For the system to learn, experimental results need to feed back into molecular design. That requires a workable synthesis for a proposed molecule and a reliable readout of what the reaction produced. If the system can’t establish what happened in an experiment, it can’t reliably decide what to change or try next.

Just as with self-driving cars, automating the movements within a self-driving lab is only part of the challenge. The hardest part is being able to interpret what happened in an experiment and use that information to decide what to do next. The goal is for each experiment to help us better choose the next step, and reduce the number of steps we ultimately need to reach a useful result.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

But a lot has changed in two years. A lot has changed in the last couple months!

And to say I’m no longer bearish on what AI can do in manufacturing is an understatement. I’ve never been more excited about the opportunities I’m now seeing for AI to come in and change the way we build plants and products in the manufacturing world.

Last week I moderated a panel at IMTS about AI infrastructure in manufacturing with Praveen Rao (Global Head of Manufacturing at Google Cloud), Shiv Trisal (Global Lead for Industrials and Energy at Databricks), and Les Karpas (Global Head of Physical AI at NVIDIA Inception). Each of them has also had to rethink something fundamental about AI in manufacturing over the last couple years.

It’s no longer just about the model and chasing intelligence at any cost. The models have reached a point where they are smart enough for most jobs. The constraint is now shifting toward context and the infrastructure that takes the insights generated from AI to then produce a physical outcome.

Our panel dove into these topics and here are some of my takeaways.

From insights to action

The panel cautioned that agent activity for the sake of generating insights is not inherently productive.

Manufacturers have become inundated with alerts. So much so that they often can’t prioritize the ones that actually matter.

What is productive, however, is closing the loop from insight to action. Thus the metric of success shifts from number of workflows or tokens to number of end-to-end business resolutions.

This isn’t a new idea but it is a lot more complicated to execute in manufacturing plants than many other industries given the nature of a factory floor and the number of legacy systems, machines and humans having to interact.

The panel advised having three foundations in place in order to move from insight to action:

1. A unified data semantic across all the plant’s data.

2. An agentic platform with contract-based delegation to systems such as ERP and PLM.

3. Multimodal interfaces to allow for easier agent orchestration.

They also suggested starting narrowly with just one problem that you can’t solve today or can only solve slowly or manually. Then work backwards from there to understand what data is needed to solve that problem. From there, you can gradually expand to more problems.

Legacy equipment and the data problem

Machines on a factory floor tend to be decades old, at least. They also use hundreds of different PLC vendors and protocols. And some very old machines may not even have a PLC.

These systems weren’t built for interoperability and so the data they generate often needs to be supplemented with additional context around what the measurements mean and how they relate to a process. This has to happen before AI can effectively use the data. This is one of the reasons AI adoption in manufacturing has lagged.

The panel stressed the importance of creating a unified namespace or contextualized data layer that connects all of a plant’s different data, and preserves context across handoffs.

Also, physical AI doesn’t have the internet-scale training set that LLMs had. This means the data that is available needs to be stretched further in a sense so that you can improve digital twins and simulations and then generate synthetic data to validate policies. And once deployed into production, the real-world performance needs to be fed back into the system in order to generate a continuous flywheel.

AI can’t turn bad data into good data, so this work around establishing reliable context and quality is critical before any useful actions can happen reliably. Useful and reliable being key words here!

Own your data in open formats

Every few weeks, there’s a new model at the top of a leaderboard. As such, the need to own your data and context, and store it in open formats is now more important than ever. This is what allows you to easily switch between the best models and not lose any of the logic you’ve been building.

Openness matters at the tooling layer too. OpenUSD is an example of a common open format for physical AI. It’s a 3D wrapper that can bring together models, time-series, language, and OT data to improve interoperability across a system. Having a common wrapper format for the 3D, time-series and OT data is what lets different systems share one representation of the plant.

It’s a multi-AI system, not an LLM

A lot of value on a factory floor is still driven by traditional ML. And as such, the panel stressed the importance of making sure you’re applying the right kind of AI to a problem. This also ensures you don’t fall into the trap of token maxing instead of value maxing.

One example we talked about was predictive maintenance. In this case, ML is probably best for predicting equipment failures and forecasting demand for spare parts. Whereas an LLM is best for interpreting the training manuals. And a mathematical optimization model is best for scheduling work under various constraints.

We also talked about how agents are now writing ML, which means a forecasting model can be built in under a day without needing a large data science team. This is especially impactful for small and mid-sized manufacturers who previously didn’t have the resources to build some of these tools internally.

Edge and cloud, not edge or cloud

A combination of cloud and edge solutions is critical on factory floors.

Edge deployments are best for anything that requires very low latency or where regulatory concerns are an issue. The cloud is best for larger workloads like analyzing large volumes of data. It’s also where the learning loop happens because of the volume of data that’s needed across both the operational and business sides of a factory.

It’s no longer a binary choice. The panel recommended having a connected architecture.

Robotics and the case for and against humanoids

Robotics is unquestionably the next frontier of physical AI on the factory floor. But what is less clear is what the optimal form factor will be.

The built world, including manufacturing plants, has been designed with the human body in mind. It’s thus a very natural transition to go from humans that have two arms, two legs, and ten fingers to robots that also have two arms, two legs and ten fingers. Not to mention, we’ve recorded a lot of videos of humans doing tasks. Therefore, starting with a humanoid form factor makes sense to some extent because that’s where we have the data. Then you can use cross-embodiment to transfer the policies into other bodies.

The counter argument here is that we’ve become obsessed with the body and perhaps lost sight of the basics. What matters in the physical world is the ability to move things, to take action, reliably. And the form factor to do this depends on what exactly you’re trying to move or do. The question of form factor in the abstract doesn’t make sense; you have to start with what the job to be done is.

The adoption bottleneck

The biggest bottleneck to AI adoption in manufacturing these days probably isn’t capability. The models are largely good enough.

What’s missing is the data and context, and the real-world applications and feedback loops.

Two things make this hard. The data problem is a legacy of machines that were never designed to interoperate. The adoption problem is that factories are measured on throughput and uptime, and production can’t absorb the disruption that standing up AI creates.

I’ve never been more excited about the opportunities ahead for physical AI in manufacturing. But to unlock its potential, both problems need to get easier. On data, this probably means starting from the problem you’re trying to solve and wiring up only the minimum set of data required, in open formats, rather than modeling the whole plant first. On adoption, this probably means setting up controlled environments for testing and very gradual scale-up only after AI has proven it can resolve an end-to-end action rather than generating yet another alert.

IMTS is an incredible conference for those building or interested in manufacturing. I used to come to it as a founder when I was building Prima. It was fun to come back this year and get to see all that has changed and continues to change for the manufacturing world!

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

Inside a tokamak, plasma circulates at temperatures around 100 million degrees, hotter than the core of the sun. Nothing can contain it by contact, so it’s held by magnetic fields, shaped and reshaped in real time so the plasma never reaches the wall. “It’s like you want to control a storm in a very small area, without letting this storm touch any wall.” Azarakhsh (Aza) Jalalvand, a research scholar in Princeton’s Plasma Control Group explained to me the other week when we sat down.

Aza’s training is in computer science and AI, not fusion. His PhD was on speech analysis with neural networks, later extended to image, video, and radar data. He ended up focusing on fusion almost by accident: a professor from the applied physics department turned up at his data science lab with a pile of data from a fusion device and a request for volunteers.

What Aza found was a field that had made remarkable progress over decades, but was still challenging due to the complexity and nonlinear behavior of plasma. “Physicists have brought us a very long way,” Aza told me. “But plasma is an extremely complex environment, and our current knowledge and models are not fast or accurate enough to predict its behavior reliably in real time, which makes monitoring and control hard. It also means we still have to rely a lot on experimentation and trial and error.”

Today, researchers have to rely heavily on sensors to estimate the plasma state and guide control decisions. Sensors provide the essential measurements of what the plasma is doing, and there’s no computational substitute for them because existing physics models can’t yet predict plasma’s behavior with sufficient speed and accuracy.

But sensors fail.

This problem is not specific to fusion. Any system with a lot of sensors (such as a satellite, a factory line, a patient monitor, etc) will eventually lose one. When that happens, everything that relied on that reading gets worse. In a well-understood system you can calculate around the loss. But when the underlying theory can’t predict the system, such as is the case in plasma, there’s no way to calculate what the dead sensor would have read. Measurement via the sensor was the only route to that number, until Aza and his colleagues built a workaround. And in doing so, they also ended up discovering measurements no instrument had ever been able to make.

In our conversation, we talked about the work that led to this breakthrough, how you trust a signal that was never recorded, and what it means when a model produces a measurement no instrument could.

Diag2Diag: its origins and applications

“If you talk to a physicist or a diagnostician and say, ‘Diagnostic A died. Is it possible to make its measurement by looking at diagnostic B and C?’ the simple answer is no,” Aza said. “And that’s correct. They are measuring totally different things. But for me, I’m an engineer, I’m a data scientist,” he said. “All these sensors are looking at one system.”

So even though diagnostic B might not measure what diagnostic A measures, the information could still be there, distributed across the other signals in a form nobody can make sense of.

“Maybe AI can find this pattern,” Aza said. “This is where machine learning and AI shines, because these models are very good at learning and leveraging patterns that cannot be interpreted by our understanding.”

A model can learn the relationships between sensors from historical data, assuming enough of the data exists from periods when every diagnostic was running. This means that when one sensor drops out, the model can generate a synthetic version of its signal. That idea became Diag2Diag (Diagnostic to Diagnostic).

The paper describes an experiment where the model is given data it hasn’t seen and asked to reconstruct a signal that was actually recorded. Its predictions are checked against what actually happened.

“We can compare and say, okay, now it’s correct, or now it’s wrong, or it’s 90% accurate,” Aza said. He added a caveat, “a model can be 90% accurate but with a lot of uncertainty. That’s also important.”

Much like any hardware sensor, the reconstructed signal is checked against known answers enough times to establish what its errors look like. After passing this validation, the model is used on live experiments.

The immediate application for this is failure mitigation.

“Whenever we have a tough system in which maintaining the hardware, maintaining the sensors, is expensive, time consuming, or even impossible,” Aza said, “synthetic measurements become interesting.”

He gave the example of a Mars rover. When a sensor dies, no one can easily fix it. But the rover has years of history from that sensor alongside all the others, and so those relationships can be learned and used to predict the missing signal.

The same idea can be applied in a lot of settings, especially where the sensors might be hard to reach such as on the seafloor, inside jet engines, in concrete, orbiting, or implanted. Like anything in AI, how well it works depends entirely on the data. “It all depends on the quality and the quantity of the data that we have collected historically,” Aza said.

AI as a new instrument

Diag2Diag was built for sensors that fail. But a sensor can be missing in another sense: it can be working perfectly and still not tell you what you need, because its measurements aren’t granular enough.

For example, Thomson Scattering is the diagnostic that measures electron temperature and density, two of the quantities physicists care most about. It fires a laser into the plasma and reads the light that scatters off free electrons. It’s also not fast enough to measure under millisecond plasma events researchers actually want to see.

So Aza’s team wanted to see if they could use Diag2Diag to reconstruct Thomson Scattering. The other much faster diagnostics became the inputs, and Thomson’s readings became the target the model learned to reproduce. Unfortunately, the model can only learn at the instants when Thomson Scattering happens to take a reading, and those instants are few and far between. This means that most of the data can’t be used.

“For example if Thomson Scattering fires at every 5 milliseconds, during training we just ignore the input data of the other diagnostics at milliseconds two, three, four, five, because we don’t have the target measurement,” Aza said. “Then we thought: what happens if after training the model we give [the model] those inputs where we don’t have the data from Thomson Scattering anyways?”

However, because there is no Thomson Scattering reading at those intermediate moments and never was, there is nothing to check against.

“That is where the data-driven validation stops,” Aza said. “We cannot do any data validation anymore. Now we have to think about physical analysis. Do these measurements physically make sense?”

This turned out to be answerable. Edge localized modes (ELMs) are eruptions at the edge of the plasma that occur when the pressure gradient builds past what the edge can hold. Aza described the approach to mitigate them to me like tapping a balloon, “the surface flattens for a fraction of a second, then springs back”. ELMs are one of the things Thomson is too slow to catch; Thomson Scattering samples on the order of milliseconds, but an ELM happens much faster than that. So a reading only happens by chance, and never enough times in a row to see the shape of the event.

Physicists had a theory that structures called magnetic islands suppressed ELMs, and they ran simulations to predict what the electron profile should look like as an island forms. The theory had no experimental confirmation because there were no instruments fast enough to validate it.

“I worked on the data and generated some outputs, but I didn’t know if they were physically meaningful,” Aza told me. So he gave them to a physicist who ran the simulations and confirmed they were a good match. In other words, the synthetic measurements, generated at moments when no hardware recorded anything, matched the simulations. “This was a very good example of an unbiased discovery,” Aza said.

AI wasn’t analyzing observations that a sensor had made. It was actually producing net new observations that no sensor had ever been able to make before.

What this changes

AI systems are already good at answering questions we know how to ask. Is this a hand? Is this signal abnormal? What is the predicted value of this measurement? But science often advances when someone notices the thing nobody was looking for.

“These AI models have been really good, sometimes almost perfect, in answering the questions that we know,” Aza said. “But in plasma fusion, the question is: what are we missing?”

Normally, answering questions we don’t know requires building new hardware to generate new data. But here it didn’t and that’s what makes this result so unusual. The evidence for the magnetic island theory was latent in data the machine had already recorded.

Potential future applications extend beyond plasma to aerospace exploration, robotic surgery, and any complex engineering or scientific systems where missing measurements can compromise safety or control.

Diag2Diag started with trying to answer what happens when a sensor is missing, broken, or too slow. But it led to discoveries much deeper. AI may not only help scientists analyze the world, it may help them observe it.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

AI agents can already browse, compare, recommend, and negotiate. In fact, according to a Deloitte study, retailers are already seeing 15-20% of referral traffic coming from AI chat interfaces. And soon, they won’t just be referring traffic, but they’ll actually be the ones spending money, at scale.

But before this can happen, agent actions have to be identifiable and accountable.

Someone has to build the infrastructure to make that possible.

Payment networks are a preview of what agentic AI in commerce will need because they’ve already solved versions of these problems once at global scale. As Greg Ulrich, Mastercard’s Chief AI and Data Officer, put it to me: “We have an ability to manage these systems with trust and responsibility that’s been proven over decades.”

I sat down with Greg the other week to explore what has to exist before AI agents can safely spend money, what Mastercard is doing to prepare for that, and what Greg is most bullish about for agentic commerce.

A Means to an End

Payments are much more than just the movement of money. They are identity, authorization, fraud detection, merchant trust, consumer protection, dispute resolution, compliance, and auditability. Mastercard has spent decades developing and supporting this environment.

“Data and AI are really at the foundation of a lot of what we’re doing today, and it has been for a while,” Greg told me. Mastercard sits in the ecosystem as a network provider and payment infrastructure company. “The transmission of data, the use of AI in that transaction, has been fundamental to what we’ve been doing for decades.”

Notably, AI is not a standalone strategy for Mastercard. It is embedded into how they grow payments, build services, and run their own operations. “It is still a key enabler of our strategy,” he told me. “It’s a means to an end, in my mind, as opposed to an end in and of itself.”

Greg organizes Mastercard’s AI strategy into four layers:

1. The foundational layer

2. The value AI creates for Mastercard’s own people

3. The value AI creates for Mastercard’s customers

4. The role Mastercard plays in shaping the broader ecosystem.

Each of these layers has the same goal, which is to make AI more capable and more trusted. Ultimately, Mastercard sees its role as helping build the infrastructure that makes AI trustworthy at scale.

Importantly, for Greg, AI capability isn’t the binding constraint. As AI becomes more capable, he believes the limit will be what people, businesses, and institutions are willing to let it do on their behalf. “How does AI work securely, responsibly and at scale for the entire ecosystem?” he said. “And how do we enable that? How do we bring in the data, the trust and the role we play to be a unique and positive force in that?”

Trusted intent to scale agentic commerce

Today, Mastercard lets consumers, merchants, acquirers, and issuers transact with trust and zero friction. Agentic commerce introduces a fifth actor.

“When you add in a new party, like the agent, how do we envision that ecosystem? How do we enable that ecosystem? How do we make sure we can identify the agents and make sure that that is a known transaction? How do we understand the intent of the consumer when you’re purchasing something so we can track it along this ecosystem to make sure that the product you have is what you asked the agent for?” Greg told me.

“In digital commerce, intent is implicit: I click buy, and the click is the intent,” he said. “Intent and action are bundled together. But in agentic commerce, they separate. Intent is an artifact that needs to be captured and verified, and potentially disputed.”

This opens up questions that didn’t exist a year ago around agent identity, delegated authority, fraud scoring for agent actions, consent sharing. Taken together, they point towards needing to extend the trust that already exists in digital commerce such that machines and agents can now operate in the system we’ve already built for people, businesses, and merchants.

Mastercard is building the standards and infrastructure to make this possible. One example Greg mentioned is Verifiable Intent, a standards-based trust paradigm for agentic commerce co-developed with Google. It’s a standard for safely passing information through the ecosystem so that a user’s intent can be captured, executed correctly, and traced if it isn’t.

Greg also described the importance they place on working with tech leaders, industry standards bodies, and customers in order to “build a system that’s going to work for all parties.” The ecosystem role, as he framed it, is about helping define standards, policies, and “the right balance between innovation and responsibility.”

The foundations for trust at scale

A standard is only as good as the infrastructure that enforces it. “This starts below the product surface,” Greg told me. You now need to verify delegated authority instead of just authenticating credentials.

For Mastercard, this is the foundation of trust and it comes from five connected capabilities:

1. Identity to know who is acting

2. Intent to prove what was authorized

3. Controls to define what an agent can do

4. Trusted execution to protect and govern transactions

5. Intelligence to assess risk and detect fraud

None of these are products. They’re requirements, and meeting them at scale takes three things Mastercard is building.

A harness. “We need standards, we need governance, we need compliance, we need observability,” Greg told me. Doing that bespoke for every AI deployment is untenable, so Mastercard is building what Greg calls an agentic factory: “Building an entire harness for all of that, which embeds all of our principles. That’s the operating system, if you will, for the agents.” As I’ve written about previously here and here, the differentiation is not so much in the model as it is in the harness around the model.

Data as a governed asset. Mastercard has spent the last 18 months building what it calls its Data Commercialization Platform to “bring the data together, to democratize it, to have gold data products available to the enterprise, to have the right controls,” he said, adding: “It’s not just transaction data. It’s how we leverage all the data assets that we have with the right controls, the right governance, and the right linkages.” Those foundations become even more important in an agentic world, where agents will increasingly need access to trusted data and clear permissions.

A foundation model for transactions. The most technically distinctive thing Mastercard is building is what Greg calls a Large Tabular Model, “the equivalent of an LLM, but for transaction or tabular data. It’s a model on the data that we have that helps predict behavior and helps understand entities more effectively.” It’s built on 15 billion transactions, with a much larger version coming.

In Mastercard’s view, trust at scale is built not through a single product or model, but through the infrastructure that governs how agents operate.

The customer-facing work of trusted AI

On top of these foundations, Greg characterizes Mastercard’s customer-facing AI work into three key areas: making commerce safer, making customers smarter, and enabling more personalized experiences. Each component is connected to a piece of infrastructure agents will need.

Real-time trust scoring. “We’re going to provide a score on that transaction about how likely it is to be fraudulent or legitimate,” Greg said. “It’s not about adding friction to the ecosystem. It’s about making it more seamless.” Decision Intelligence Pro, Mastercard’s real-time transaction scoring product, is one example. Greg also mentioned Mastercard’s Safety Net solution, which looks for malicious actors throughout the ecosystem, in addition to Mastercard’s acquisition of Recorded Future, a threat intelligence company that uses AI, analytics and data to help organizations identify, prioritize and respond to cyber threats before they become attacks. Every agentic system will eventually need a real-time trust score on actions and intent, not just transactions. Fraud scoring is the mature template of this.

Static to dynamic insights. “A lot of things that we used to provide in a static format are now much more dynamic,” Greg said. “We’re bringing in more data and intelligence because of what the technology allows, but also making it a lot easier to use by putting an AI interface on top that lets customers find their room for optimization, learn about their portfolio, and get better insights.” Rather than a static report, the deliverable is the interface that lets customers dynamically pull insights.

Consent that travels. Greg continued: “For an agent to be able to recommend or negotiate or buy on your behalf, it needs your data. It needs to know your preferences, your purchase history and payment credentials. And then that data needs to move from you to your agent, and then from your agent to merchants and payment networks. As that data moves, your consent needs to move with it. Infrastructure is thus needed to transmit and honor those permissions across handoffs, as well as to decide who bears liability when data gets used beyond what you consented to.”

Owning the layer that differentiates

Mastercard recognizes that no single company will build the AI future alone, and instead, success will come from orchestrating a trusted ecosystem of model providers, cloud platforms, data partners and innovators, while contributing the security, governance, observability and proprietary intelligence that makes AI safe and effective at scale. Mastercard’s role is not to build every agent or model, but to help define and operate the trust layer that allows agentic commerce to function across parties, platforms and markets and allows customers to scale AI across their enterprises.

Exponential models, linear humans

At this point, the gap between what’s possible with AI and what’s actually happening within large enterprises is less about technology and more about adoption. As Greg puts it, “The quality of the models continue to go up at this exponential rate, but people’s ability to consume this and change the way they’re operating is going at a linear pace.”

For Mastercard, closing the gap has required more than just granting access to AI tools. Mastercard has helped employees develop new ways of working so that AI is now embedded into everyday workflows across the business. Engineers were among the earliest users, but Mastercard quickly realized that AI is most effective when it supports the entire product development lifecycle and broader enterprise processes.

Mastercard sees AI as a force multiplier that helps accelerate innovation and deliver better outcomes for customers. The goal isn’t just around efficiency, but instead to enable the organization to scale expertise and bring new capabilities to market more quickly.

Trust that can scale global commerce

At the end of our conversation, I asked Greg what excites him most about the next phase of AI in payments and commerce. “I think it’s around scale in a trusted way,” he said. Most of what’s happening in agentic commerce right now, even at the pace it’s taken off, “is people experimenting and seeing how it works.”

Whether cause or effect, consumers remain reluctant to transact through agents. In Accenture’s 2026 survey of more than 25,000 people across 16 countries, 74% said they’d let an agent handle routine tasks, but only 9% are open to letting one shop autonomously on their behalf. People are ready for agents that recommend, but they’re not ready for agents that pay.

We’ve seen this dynamic before. In a 1995 Pew Research report, only 8% of internet users had bought anything online in the previous month. Today 84.3% of the US population shops online at least once per year. What closed the trust gap was infrastructure: fraud scoring, zero liability, dispute rights.

Mastercard has built this kind of infrastructure for previous evolutions in commerce and is positioned to build it again. “We have a fundamental role to play in agentic commerce by building the trust infrastructure that allows agents to transact safely and at scale given where we sit in this ecosystem,” Greg told me.

He continued: “I believe agentic commerce will follow a similar arc as e-commerce: first a trust gap, then infrastructure, then adoption. The difference will be the speed of that arc. Agentic commerce will proliferate much faster in large part because the institutions that closed the last trust gap are already building for this one.” What comes next, as Greg sees it, is enterprise adoption and scale with trust, as he put it, “at the heart of everything.”

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

We’ve made designing molecules easier and cheaper than ever, but, as I wrote about here, proving those designs work in humans is increasingly becoming a bottleneck. There are a lot of reasons for this, such as the constraints around chemical synthesis, the cost and duration of clinical trials, and how little we actually know about what happens in a lab.

Today, scientists keep notebooks and publish papers, and other scientists work from those records. But most of what makes an experiment succeed (or fail) is never actually written down.

Nowhere is that more expensive than at the handoff, when a working process has to move from the lab that developed it to a contract manufacturer, often in another country. “That is where things get screwed up,” told me the other week. Tech transfer is one of the most common ways manufacturing problems derail a drug launch; and manufacturing is now the single biggest cause of delays, at 64% in 2024.

Anna Marie has spent her career moving between biology and technology. She studied molecular biology, spent time as a tech investor, and returned to the sciences in 2019 unable to shake a specific discomfort: two fields she knew well, science and software, ran on entirely different economics. She’s now the co-founder of Transfyr, which builds observability infrastructure for scientific execution. The company came out of stealth last week, with The New York Times covering the launch. “The variation we see even among well-trained scientists is pretty jaw-dropping,” Anna Marie told the Times.

“In software, I can write some code, and if it’s useful, I can ship it to millions of people instantaneously, basically at no cost. But transferring a new technique to a single other lab can take months of debugging between highly trained people,” she told me. “We call that tech transfer, which is effectively the distribution cost.”

Up until now, most advancements in AI for science have focused on the parts of science that were already digitizable, such as sequences and molecular structures. And while that’s great progress, it’s not addressing the physical problems that determine whether a drug succeeds in a human and whether it can then be manufactured at scale. And it’s these problems that continue to hold back much of the industry and inhibit innovations from getting out of the lab. As Anna Marie put it, “Anything that’s getting in the way of positive scientific discoveries reaching the world is impacting everyone. It’s climate, it’s food security, it is human health, animal health. This is a global issue.”

Noise in Science

If you visit Transfyr’s lab, you’ll have the opportunity to participate in a simple exercise: a “serial dilution,” which requires diluting a solution of known concentration by a specified amount through a series of steps, in duplicate. Each participant is then scored on accuracy (how close to the target concentration they landed) and precision (how close the two samples are to each other). “We consider the ‘strike zone’ to be within 5% of the target on both accuracy and precision. The vast majority of people do not hit the strike zone. Even experienced scientists.”

The reason for this miss depends on a lot of factors. Did they pre-wet the pipette or back-pipette? Did they set their pipette to the correct volume at every step? How thoroughly did they mix the sample? Did a stray bubble get in? Did they accidentally foam the mixture? None of this is actually in the written protocol, but it all shows up in the result.

Most protocols are actually thousands of small, undocumented decisions. At scale, this invisible variation shows up in the data used to train AI models. As Anna Marie put it, “We would get data sets where the model was better at recognizing what lab did an experiment than what the experiment was designed to measure in the first place. If that data is just training on unwritten techniques we don’t even observe, is it actually teaching the model anything useful about the biology itself?”

There are surprisingly simple, undocumented ways that this lack of standardization shows up. Anna Marie gave an example of a scientist who pre-labeled all of her tubes, which, along with several other shortcuts, helped her get her protocol done two hours faster. And those two hours made a big impact on the experiment’s outcome because it turns out RNA sitting at room temperature degrades, and two hours is not a trivial amount of time. But pre-labeling tubes would never show up in a normal protocol.

“I worry that sometimes we chalk up noise in biology to ‘biology is noisy’,” Anna Marie told me. “There’s also a lot of process variability that we just accept without understanding. That’s the problem. You can accept variability, but we need to understand it so we can interpret our data under that lens.”

In other words, the problem is mistaking “we don’t know what happened” for a fact about biology, rather than a fact about the lack of tooling and observability that exists today in science.

The Minimum Viable Film Room

So then what’s actually worth measuring? There’s no shortage of variables that could theoretically matter, but Anna Marie has narrowed down the list to three that do most of the work: the operator’s actions and intent, supply chain, and then the environment. As she puts it, “Those few things comprise the minimum viable film room.”

Operator actions and intent means, first and most basically, understanding what a person (or a machine) actually did. This isn’t necessarily just visual observation, because a huge amount of what you’re working with in the lab looks identical even though it isn’t. “I’m looking at a clear liquid, and that clear liquid could be water or it could be hydrochloric acid. There’s a big difference between those two things, and you cannot tell the difference between them visually,” Anna Marie explained.

Supply chain is around the inputs, not just what a material was, but its lot number, its expiration date, who else touched it, what might have contaminated it.

The environment is about monitoring things like temperature, humidity, and CO2. Some of this might seem obvious and yet, for most of the industry today, it’s missing.

In Anna Marie’s opinion, automation around these three levers is often misunderstood. The field has sold automation on the ideas of scale and cost savings, she points out, but neither holds up especially well in practice. Utilization at automated labs across the industry has stayed frustratingly low, and scientists tend to value flexibility over throughput in ways robots can’t yet match (though I suspect this will change soon!).

However, what automation does deliver, in her view, is observability rather than scale. In her essay “On Observability,” Anna Marie makes the case using the example of self-driving cars: autonomous vehicles didn’t improve because of better maps or smarter models, but because cars were instrumented with cameras, lidar, and GPS to learn how humans actually navigated the road, years before autonomy made economic sense. In DARPA’s first driverless-car challenge in 2004, the best vehicle made it just over seven miles into a 142-mile course; a year later, after a year of watching and iterating in public, five vehicles finished the whole thing. Observability enabled autonomy.

Above is a video from a head-mounted camera recording an experiment, along with three cameras mounted above the lab bench. Source.

Lab automation works the same way: automated systems tend to log what they intended to do and what they actually did, almost as a side effect of being automated. “I would never say that autonomy is the end goal,” she told me. “It’s a tool just like any other, and that tool should be wielded by really smart, caring humans who are trying to make a difference in the world.”

Why Science Hasn’t Built Its Film Room

So why don’t we have film rooms for science already? “We have deep observability in sports, and it is so valued by the athletes. Its primary purpose is not evaluative. It is coaching and training and self-improvement,” she said. So why are we not giving elite scientists the same tools that we give elite athletes to improve their game?

Part of the answer is infrastructure. But the other part is cultural. Science still has a layer of secrecy around it. If you think you have, as she puts it, “the magic hands,” you may not be in a hurry to let everyone else have them too.

Publication incentives don’t help, either. Observability expands what’s visible, and science has not made the same peace with visible failure that sports has. “Are we incentivized in any way, shape, or form to publish comprehensive results as opposed to pretty or clean results? No, I don’t think so,” Anna Marie said.

Amber Liu made a version of this point to me a few weeks ago about AI research rather than the wet lab: workshops encouraging scientists to publish their failures exist, but adoption is slow. Amber’s observation was that the reluctance is specifically human. An AI scientist, as she put it, doesn’t carry “this burden of disclosing that they’re actually doing a lot of dumb things, that they fail a lot in the middle.”

The infrastructure that makes science observable is also the infrastructure that makes an individual scientist’s workflows and mistakes visible and public. Until the incentives around that visibility change, what’s technically solvable might not matter.

Systems Integration

Toward the end of our conversation, I asked Anna Marie what she thinks the field could look like in the next five to ten years. Her answer was more about the scientific system as a whole rather than any single breakthrough.

“There are millions of scientists who are having ideas right now all around the world. And none of them have a good way to communicate. They don’t have the language for it,” she explained.

That, to me, is the larger promise of observability in science. Beyond just cleaner data or better protocols (which matter!), it’s the possibility that knowledge from one lab can be easily transferred so that someone else, somewhere else can use it too.

“I go back to systems integration. Such a boring term, but I love it,” she joked. “Because it is the root of most advanced industries. Can we bring together components to advance products that actually solve real problems in the world, on a completely different timeline than we’re doing today?”

Systems integration is about allowing the components of the physical work of science to talk to each other. Right now, too much knowledge is siloed and unproductive, because nothing exists to carry it from where it was learned to where it is needed.

Science still hasn’t solved distribution. And until it does, better models will continue helping us imagine what to build, but they won’t tell us how to carry that knowledge into the physical world where it can actually be used.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

There are two high-level ways to make an AI agent work on something hard for a long time: give it enough context to hold more of the problem in mind, or build scaffolding around it so it doesn’t have to. Scaffolding can include retrieval, memory, planning, tools, tests, checkpoints, and sub-agents.

I recently sat down with Prateek Jain to talk through what the right combination of these two approaches is. Prateek is a Distinguished Scientist at Google DeepMind, where he works on new architectures for frontier models with a focus on efficiency and elasticity, and co-leads Gemini’s long-term innovation area that focuses on risky and ambitious research projects.

Recently Prateek has been focusing a lot on long-horizon agents, systems that stay on a task longer than most models are trained to do right now. Prateek explains the goal in terms of human effort: “Let’s say there is a task that requires ten experts to be at a problem for five days, or ten days, or maybe a month. We want to see if models can go and solve those tasks, which means that our models should know how to plan for a month. They should know how to delegate. They should know how to share information across all these tasks and subtasks and sub-agents.”

But before this can happen, a few things need to be solved first:

1. Cost: Attention cost grows quadratically with context length, so doubling what a model holds in mind roughly quadruples what it costs to serve.

2. Handoff: When a main agent delegates to a sub-agent, the sub-agent typically only sees what was written down for it. This means the reasoning that produced the subtask is thrown out.

3. Rationing: Anyone running agents at scale has to manage compute costs. But it’s still unclear how an agent should manage the budget it gives each sub-agent it spins up.

There are tradeoffs across each of these. More context gives the model more room to work, but it also makes the system more expensive to run. A text-only handoff is cheaper, but the sub-agent only gets the summary, instead of all the work that led up to it. A tight compute budget for the sub-agent might result in the sub-agent failing and thus the work needing to be redone.

In our conversation we explored these tradeoffs, and what has to change architecturally in order for long-horizon agents to work at scale.

What is a long-horizon agent, anyway?

“Long-horizon” is a term that gets thrown around a lot but what actually makes something “long-horizon”? I asked Prateek and he said he defines it along two axes: task difficulty and training distributions.

On difficulty: a long-horizon task is one that would take multiple human experts multiple days (think ten experts working for a week or a month).

On training distribution: standard RL post-training have multiple rollouts where the model has practiced full attempts at a task with trajectories of somewhat limited length. A long-horizon task runs past anything in that range, so the model is improvising instead of drawing on something it has done before. As Prateek put it: “The model was trained for a certain rollout length range, and now you are asking the model to do maybe 10x more work,” at which point the task is “out of distribution” with respect to training data.

Long context is necessary, but not sufficient

When I asked Prateek whether he thinks we should give models ever-longer context windows, or build agentic scaffolding around shorter-context models, he said he sees them as a complement. As tasks get harder, “the amount of context we will need to solve the problem is going to be extremely high. So having agentic scaffolding, as well as ideas like retrieval or maybe separate memory, will be critical.”

But squeezing everything through a short context window breaks down, because hard tasks are underspecified: “For very hard tasks you might not already know what all you need to put in the context.”

He gave an example: “If you are trying to ask Gemini about a fairly underspecified problem ‘oh, my wife’s birthday is tomorrow, can you please plan the whole day?’ Gemini would need to understand how your relationship is, what are the things you like to do for fun. It might have some rough idea: okay, let’s find pictures where both of these people are together, or emails around itinerary planning. But it might not be able to pinpoint the exact things apriori.”

Pure retrieval only works if you know what to look for, which can be hard when the task is underspecified. Meanwhile, long context is the opposite because instead of deciding upfront what to look for, you bring in a broad field of information and then let the model decide what’s relevant and how the pieces connect.

Long context simplifies orchestration. With a short context window, you need more retrieval, more sub-agent calls, and more tool calls. This is all extra orchestration burden, which, as Prateek put it, “is not ideal. You want to give as much context to the model as possible.”

The goal instead is “a really clean, simple solution which is also fairly general and can solve very challenging long-horizon tasks”: a capable long-context model that needs less machinery around it, not more.

The efficiency wall, elastic models, and routing

Every token of context has a price. “The serving cost, especially with super long context, is going to grow quadratically, which is challenging,” Prateek acknowledged. Inference economics is an architectural constraint for agents doing long-running work at scale.

He’s optimistic, however, that active research in frontier labs and academia “might be able to bring down the cost significantly.”

Some of Prateek’s work is focused here, on model efficiency, but from the angle of how much model each token uses. “Models are sort of monolithic. For every task and every phase, you are basically passing the data through the same number of layers and doing the same amount of work.” This means every step, no matter how hard or easy it is, costs the same.

MatFormer, one of Prateek’s best-known projects, asks whether a model can instead be elastic, dialing its power up or down continuously. In practice, Prateek explained, that could mean adjusting the model “based on the complexity of the task or maybe based on how loaded your servers are or what is the cost currently of your tokens.”

“If agents can say, hey, this is a simple task, I can delegate it to a small model within this large family of models versus, oh, this is a slightly harder task, let me go one notch higher. That can enable high quality solutions at relatively low cost,” Prateek said.

This decision-making requires routing, meaning a policy has to decide which model to call, whether to spawn a sub-agent, what budget it gets, and when to keep retrying versus stop.

“As agents scale to really large numbers of tasks per second,” Prateek said, “we will need to start rationing how much compute you are giving to these agents.” The risk, however, is that you pick a sub-agent too small for the task and it fails, forcing the main agent to evaluate and redo the work. In this case, you end up spending more than if you had just had a more generous upfront routing budget to begin with.

Getting routing right means jointly assessing “the complexity of the task and the capability of the sub-agent.” And budgets cascade, since the sub-agent then decides for itself how much thinking its allotment buys.

This kind of elastic routing hasn’t fully arrived. “There is still more research to be done” on the quality-versus-cost tradeoff. But the direction feels inevitable to him: “As agents percolate to pretty much everything we do on a day-to-day basis, then having to do this rationing of compute and of capabilities becomes unavoidable.”

The handoff problem

Delegating to a smaller model requires planning a precise sub-task for the smaller model along with enough context/hints to the smaller model so that it can solve it. In today’s main-agent/sub-agent pattern, it’s mostly whatever the main model serialized into text (e.g. a task description rather than the internal state that produced it). The main model does a bunch of work, writes up a subtask, and hands it off. “The sub-agent’s view is only the subtask that has been given. It doesn’t know what all work the primary model would have done,” Prateek explained. Everything the main model figured out in its KV cache is essentially thrown out at the handoff.

Tandem Transformers, another of Prateek’s projects, is focused on this tension. The work pairs a small model with a large one, trained together so the small model consumes the large model’s representations rather than a text description of the task. “Whatever work the main model has been doing in its full form, in the KV-cache format itself, can be transferred to the sub-agent,” Prateek said. “Which hopefully can provide it with a much richer context so that it can solve the task in a much better fashion.”

In short, the sub-agent gets the main agent’s working memory, making the delegation much more flexible. The main model no longer has to fully specify the subtask+hints/context in text before passing it along.

Team players and strategic thinkers

Prateek believes some of the most important domains for long-horizon agents are agentic coding, particularly for machine learning and LLMs, and STEM research, especially in biotech and material science research. Success in these areas requires judgment, “this kind of work requires the development of taste and personality itself in the model, along with domain expertise and general intelligence,” noted Prateek.

Five years out, Prateek expects the main job of these systems to be organizational. “What I see is the base models becoming not only very intelligent and powerful, but also having a lot of capabilities around the ability to plan, delegate, coordinate. The model can be put in a variety of harnesses. You can attach memory to them, you can attach knowledge bases to them, and make them work.”

And underneath all of that, the model has to know its own limits. “Does it know that this is beyond its capability so it can bring in other ideas? Is it able to recognize that maybe it should spin up a new subagent with focus on specific tools?” Prateek sees these as the key questions. Every routing decision depends on that: sizing a sub-agent, setting a budget, deciding whether to escalate. A model that can’t tell hard from easy, can’t ask for help from humans or call potentially expensive tools, and can’t ration effectively.

Today, the models are great ICs, especially when supplied with clear tasks and strategies to solve the task. What they’re still learning is how to be, as Prateek put it, “really good team players, strategic thinkers, and leaders.”

Thanks to Divy Thakkar for making this connection!

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged. Prateek is speaking in his personal capacity. The views expressed here are his own and do not represent those of his company.

The middle decades of the twentieth century were some of chemistry’s most productive years. The Haber-Bosch process had already begun transforming agriculture through synthetic fertilizer. DuPont invented nylon. In medicine, the period brought sulfonamides, the first synthetic antibacterials, and chlorpromazine, the drug that helped launch modern psychopharmacology. During the “golden age of natural product synthesis,” organic chemists learned to build more and more complex molecules found in nature, including alkaloids, steroids, vitamins, and antibiotics. Woodward and Corey helped turn synthesis into a planning discipline. In fact, Corey’s formalization of retrosynthetic analysis underpins many AI route-planning tools.

Today, chemistry is still driving some of the world’s biggest breakthroughs. Materials chemistry gave us lithium-ion batteries. Chemistry also shows up all over modern hardware and medicine: in semiconductor fabrication, high-performance polymers, and the materials used in sutures, catheters, and implants. It was central to the COVID-19 mRNA vaccines too: lipid nanoparticles protected the fragile mRNA and helped get it into cells.

Most approved drugs also depend on synthetic and medicinal chemistry. The GLP-1 boom is a good example: the first wave of these drugs has been peptide-based, but the major goal now is to create oral small-molecule pills. KRAS G12C inhibitors are another example; they proved that RAS, a well-known driver of tumor growth long considered “undruggable,” could be targeted.

But chemistry breakthroughs are incredibly hard-won. In drug discovery, drug development still fails roughly 90% of the time. Large pharmaceutical companies can withstand these failure rates. Most companies can’t. Meanwhile, much of the manufacturing capacity for chemistry and custom synthesis has moved overseas, especially to China and Eastern Europe, a vulnerability made more visible by the war in Ukraine. The reason looks a lot like the broader story of US industrial hollowing-out. Synthetic chemistry was relatively easy to carve out from the rest of R&D. Over time, pharma kept more of the work it saw as highest value (e.g. IP, target discovery, and commercialization), and moved more synthesis and manufacturing to places with lower labor costs, strong scientific talent, and, at least historically, looser environmental and regulatory constraints.

This has made the US chemical synthesis and manufacturing ecosystem fragile, especially for small molecules. The good news is that AI is poised to accelerate chemistry by opening new chemical space. Models are finally getting good enough to propose useful candidates, but they still lack the right data, especially failed reactions and make/test outcomes. Synthesis is how we generate that missing data, and that makes US synthesis capacity strategic.

One note before diving in, I focus mostly on small-molecule drug discovery in this piece because it is the clearest example, but the same design-make-test bottleneck exists in materials, catalysis, and energy chemistry as well.

Why Protein AI Moved First

We’ve already seen this play out in protein engineering and antibody design. The advent of de novo protein models compressed the human and experimental work of protein design into inference, opening up previously undruggable targets and novel biologics mechanisms. AlphaFold 2 made it possible to reliably predict a protein’s 3D shape. Then RFdiffusion and ProteinMPNN let you design a new protein by sketching the shape, and then finding a sequence that folds into it. The Baker lab published de novo serine hydrolase design in 2025. Researchers used AI to design antibodies from scratch. Nabla has reported de novo antibody designs against both GPCRs and peptide-MHC complexes, two hard target classes that point toward more personalized biologics.

Two things made this possible:

1. The experimental loop was fast: researchers can synthesize and screen thousands of protein variants in a single round, so even low initial success rates can produce useful feedback.

2. The data existed: decades of solved structures in the Protein Data Bank (PDB), and orders of magnitude more sequence data than structure data.

Synthetic chemistry is different. Small-molecule drugs, for example, require complex, multistep synthesis routes, and optimizing them involves weighing trade-offs between synthesis complexity and difficult-to-predict medicinal chemistry properties. In other words, a model can propose a structure, but then that structure still needs to be made and tested. Synthetic chemistry has been harder to model, harder to search, and harder to learn from. That is why it has lagged protein engineering.

Why Chemistry is Next

Structure Prediction Went Atomic.

Models now represent interactions between all molecular types: proteins, ligands, water, ions, and other molecules critical to biological function.

AlphaFold 3 extended structure prediction beyond proteins alone to proteins bound to other co-factors, including small molecules. Boltz-1 and Chai-1 brought similar capabilities into more accessible tools. At the same time, computational chemistry and materials science are getting their own general models. Machine-learned interatomic potentials, such as MACE-MP-0 and MatterSim, can simulate how atoms interact across a wider range of molecules and materials, much faster and cheaper than traditional methods.

Progress on Higher-Level Properties.

We’re advancing beyond structure to properties that are important for drug design. Things like binding affinity, synthesizability, developability, metabolism, and toxicity.

Boltz-2 added binding-affinity prediction. BoltzMol-1 pushes this further into small-molecule hit discovery by combining model-driven ranking with filters for properties like solubility, lipophilicity, and permeability. Axiom is taking a similar property-prediction approach to toxicity.

Reasoning Models Are Well-Suited to Chemistry.

Agentic models with tool use are well suited to multi-step optimization, and chemistry is fundamentally a reasoning problem. Just as protein language models benefited from treating proteins as sequences, chemistry agents may benefit from treating the synthesis task as a step-by-step reasoning problem.

In medicinal chemistry, each structural change can affect potency, selectivity, solubility, metabolism, toxicity, and synthesizability. Chemists have to reason through these trade-offs molecule by molecule, but AI could help them explore at much greater speed and scale.

ChemCrow shows how LLM agents can combine reasoning with chemistry-specific tools to improve performance and enable new capabilities to emerge. Newer benchmarks like ChemIQ show how reasoning models are becoming more capable on chemistry problems directly.

The Data Is Still Missing

Although chemistry is beginning to look modelable in a new way, we simply don’t have enough data.

The design space of synthetic chemistry is enormous. The number of drug-like chemicals reaches estimates as high as 10^60 possible molecules. Even constrained estimates are huge: GDB-17, which only includes organic small molecules of up to 17 atoms, has 166B molecules. The largest searchable libraries now approach 8.3T molecules. For comparison, ChEMBL 37, a public database of molecules and their biological activity, has only ~2.9M distinct compounds.

On the synthesis side, the problem is compounded, with synthetic route reporting of small-molecule drugs largely being contingent on drug discovery campaign success. Much of this knowledge is tacit and never publicly disclosed, with unsuccessful reactions rarely being archived, which means scraping literature won’t actually recover it. A 2022 paper on the importance of failed experiments showed how reporting bias distorts reaction models, and argued that negative results are among the most valuable missing data. The Open Reaction Database is one initiative that is trying to change this.

And the medicinal chemistry problem is even harder. Data we do actually have is heavily biased toward specific drug targets and regions of chemical space already explored. Unfortunately, decisions made during these campaigns are largely unreported in patent literature, and weighted toward molecules that survived the filters of synthesis, binding, and journal publication. Some companies are trying to close this gap by generating new data directly. Leash, for example, is generating large-scale binding data by testing millions of compounds against hundreds of proteins. But even these efforts are limited by synthesis.

If we want generative AI to discover new drugs, materials, catalysts, and industrial chemicals, we need much richer data on what we can make, what we cannot make, what failed, and why.

Synthesis Creates the Data

This data has to come from the systems that are actually trying to make new molecules. Those systems need to record what happens, and then feed those results back into the models. In other words, synthesis will unlock the learning loop.

The next generation of chemistry infrastructure therefore needs three things.

1. Cheaper synthesis: Novel chemistry and automation to drive down cost. Models improve with data, and the cost per molecule determines how much data we can generate.

2. Broader synthesis: Multistep reactions enabling exploration of novel chemistry. Models need access to chemistry beyond the existing scaffolds that make up most of today’s libraries. The goal is to make different molecules, not just more.

3. Faster turnaround: Quicker feedback loops for agent learning. A system that takes months to design, make, test, and learn from one molecule will not generate enough feedback for models to improve quickly.

Achieving all three requires agentic integration, meaning closed-loop learning systems. The synthesis platform itself has to be part of the learning system, which means planning routes, running reactions, capturing failures, analyzing results, and deciding alternative paths to try.

Pieces of this infrastructure are starting to emerge. Chemify, Onepot, B12, Lila, and Biosero are trying to connect route planning, reaction execution, purification, analytics, and data capture into a single make-test-learn system. Periodic Labs, CuspAI, and Radical AI are doing something similar in materials, where the hard part is connecting models to labs that can synthesize and iterate on them.

Academic research is also showing the benefits of self-driving labs with closed-loop systems. For example, Bayesian optimization outperformed expert chemists on both efficiency and consistency when benchmarked head-to-head against real experiments. A team led by Timothy Noël out of the University of Amsterdam built RoboChem, which can optimize the synthesis of ~10-20 different molecules per week. AlphaFlow, a self-driven fluidic lab, discovered a 40-parameter multi-step route that outperformed conventional sequences.

Together, these examples show that chemistry is becoming programmable. But this only matters if it is connected to the physical world. For the US, that makes synthesis capacity strategic infrastructure.

America Can’t Rent This Future

That infrastructure to connect programmable chemistry with the physical world has to be built in the US.

Synthesis capacity will become more and more strategic. The country that controls the fastest molecule-making loops will have an advantage in drug discovery, materials, agriculture, energy, and defense. We’re seeing this play out in other physical world domains, such as manufacturing and rare earths right now as well.

US pharma and biotech supply chains are deeply exposed to China. According to a 2026 CFR report, “China has both the tools and demonstrated willingness to weaponize US pharmaceutical dependence: the structural conditions enabling it run through nearly every tier of the pharmaceutical supply.” The National Security Commission on Emerging Biotechnology has warned that China is moving aggressively in biotechnology and that “the window to act is closing.” Meanwhile, China’s biopharma industry is becoming a major source of global licensing assets, with Chinese out-licensing deals having already reached 80% of 2025’s full-year total.

Washington has started responding. In December 2025, the BIOSECURE Act was signed into law. It creates a formal process for deciding which biotech suppliers pose national security risks, and for cutting them out of federal contracts and grants. In June 2026, the Pentagon added WuXi AppTec to its Section 1260H list of “Chinese military companies.” WuXi is a Chinese CRDMO whose chemistry arm took on 1,187 new small molecules in 2024, and its US revenue grew 32% last year even under the threat of restriction. The DoD is barred from working directly with those companies. Beginning in June 2027, that restriction will also cover DoD contracts that rely on their work, including US companies that use WuXi as a supplier. WuXi is suing to get off the list but there has been no ruling yet. And Section 232 tariffs went live in July 2026 at a 100% base rate on patented pharmaceuticals, their APIs, and their key starting materials from countries like China and India.

On the commercial side, companies are also starting to act. Lilly committed $27B toward their US manufacturing investment; and Novartis committed $23B to expand its US-based manufacturing and R&D footprint.

The Next Golden Age

Generative AI will create a flood of innovative ideas. The bottleneck will be turning those ideas into molecules, turning molecules into data, and turning that data back into better models.

That loop has to be built here. US dependence on foreign countries, especially China, isn’t just an inconvenience around cost or lead times anymore. It is increasingly a national security threat. The US can’t lead in AI-enabled molecular discovery while continuing to outsource the physical foundations of discovery itself.

Building AI-native synthesis capacity in the US is therefore one of the biggest opportunities of the next decade. AI and robotics can change the economics of synthesis and make domestic chemistry more attractive. For example, smaller teams can run more experiments thus leading to more data and faster feedback loops than traditional lab workflows allowed.

The private sector should build most of this infrastructure, while the government should make it easier to build. The government can do this by funding precompetitive infrastructure, creating demand through procurement, supporting domestic manufacturing credits and loan guarantees, and setting data standards so failed experiments actually become useful training data.

Rebuilding access to reagents, precursors, solvents, APIs, and specialty chemicals will take more than software and robotics, and likely more than a decade. But that is why the US has to start now. The last golden age of synthetic chemistry gave us fertilizers, polymers, antibiotics, synthetic medicines, and much of the industrial base of modern life. The next one could be even bigger.

Special thanks to Dylan Reid, Wenhao Gao, Geoffrey Smith and Paul Gamble for their help in talking through a lot of the ideas in this piece and reviewing versions of it.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

To better forecast the impact of AI on any industry, you must understand its overall impact on an integrated ecosystem. It may streamline a piece of the workflow but how will this play out as the other steps digest this acceleration? Where will the bottlenecks be created and can they be cleared? If not, the overall throughput will not increase.

AI forecasting is about predicting what AI systems will be capable of, as well as how those capabilities will impact industries, institutions, and everyday life. That second part is much harder.

“It’s not something you can get from first principles,” Abi Olvera told me when we spoke last week about AI forecasts. I first met Abi after I reached out because of one of her recent articles in which she talks about AI forecasting from the lens of biosecurity.

Abi spent more than seven years as a State Department diplomat working across crisis preparedness, national security, China, international coordination, and emerging technology risks. She served in Dakar and Cairo, where she focused on Egypt’s $12 billion IMF program. She also worked on cybersecurity and law enforcement issues on the China Desk. Across her career, she’s had to think about poverty, national security, crisis preparedness, energy, biosecurity, cybersecurity, and the institutions that have to respond when technology changes faster than policy can. Today, Abi writes the Positive Sum Substack and is launching an organization focused on state capacity and bottlenecks in the economy. She is also a Special Advisor at Golden Gate Institute, and an affiliate at the Institute for AI Policy and Strategy.

In our conversation, Abi and I discussed where she sees the most risk around forecasting gaps, the drivers of those gaps, and what can be done to help fix them.

Forecasting Gaps

Forecasting gaps happen when people assume having access to information is the same as having the ability to act on it. AI makes a domain easier to understand, but doesn’t necessarily make the work easier to execute.

Take radiology for example. In 2016, Geoffrey Hinton told a Toronto machine learning conference that “people should stop training radiologists,” because it was “completely obvious” deep learning would outperform them within five years. And while AI has become very valuable in radiology (the FDA has cleared hundreds of AI-enabled medical imaging tools), the labor-market forecast was very wrong. Radiology did not disappear; in fact demand grew. At the Mayo Clinic, one of the more aggressive AI adopters, radiology headcount reportedly grew 55% since 2016 while the department built a 40-person AI team and used more than 250 AI models.

“There’s been amazing progress, but these AI tools for the most part look for one thing,” said Dr. Charles E. Kahn Jr., a professor of radiology at the University of Pennsylvania’s Perelman School of Medicine and editor of the journal Radiology: Artificial Intelligence. As The New York Times wrote, “Radiologists do far more than study images. They advise other doctors and surgeons, talk to patients, write reports and analyze medical records. After identifying a suspect cluster of tissue in an organ, they interpret what it might mean for an individual patient with a particular medical history, tapping years of experience.”

AI improved the task of interpreting medical images, but radiologists still integrate clinical context, communicate with physicians, guide treatment decisions, perform procedures, and carry accountability. The capability arrived, but the forecast failed because interpreting images was only the tip of the iceberg of the “radiology job” - and as AI became more proficient in this task at the top of the workflow, more radiologists were needed to keep up with the increase in related work.

Abi believes forecasts focus too much on what technology makes possible and not enough on the realities of domains and social incentives governing whether, where, and why people will actually use it. We need to better understand both sides: where does AI change the workflow, and how does the world respond.

Biases and Blind Spots

Part of what makes AI forecasting so hard is that the people closest to the technology are often farthest from the domains where the forecasts will play out. As Abi put it, “Because AI forecasting has tended to be from the more technical, Bay Area-based community, it systematically falls down more in any domain that has to do with social behavior, adoption, or any large real-world or physical world component.” In San Francisco, for example, people tend to start from capability. Software engineers often use AI for tasks where the model can complete much of the workflow end to end. In other domains, however, the workflow extends beyond text into physical systems or interactions with outside actors, which makes AI’s impact both harder to see and harder to forecast.

“People in DC are less likely to be using Claude Cowork or Codex,” she said. “If people use it [in DC], a lot of times they might be using the chatbot version, which is great, but that doesn’t really unleash the parts of AI that are crazy surprising.”

On the other hand, people in San Francisco often underestimate frictions around AI adoption because they are focused on the vision more so than the nitty gritty work that translates vision to reality in the real world.

In truth, neither perspective is complete. San Francisco may be closer to the frontier of technological capabilities, but further from the systems where those capabilities are deployed. DC (or other centers of non-technical power), meanwhile, is closer to those systems where the rubber meets the road, but further from the true possibility of AI technology. Unfortunately, public discourse rarely closes this gap. The people most motivated to speak usually have a stake in the outcome (e.g. a company trying to sell a product or an institution trying to defend its role). Advocacy creates the resources and incentives to shape the conversation, which is valuable, but can also mean the loudest arguments are not always the most representative view of practitioners. And because the content people consume is usually the content that already fits their worldview, public discourse increasingly serves as echo chambers rather than bridges.

Cross-Framework Research

This is why AI forecasting needs more contact with practitioners. They are the ones with the most grounded view of where AI can change the work vs where it can’t, and which parts of the job outsiders are likely to overlook from their AI ivory towers.

Abi believes we need more research that brings together both sides – the technical and the practical. She calls this cross-framework research.

One of her favorite examples of this kind of research is a study that tested whether AI actually helped novices perform molecular biology tasks in a lab. The study compared people with internet access to people with internet plus frontier AI models, then measured whether they could complete hands-on wet-lab tasks over eight weeks. The result was mixed: AI seemed to help on some intermediate steps, but it didn’t significantly increase the number of people who could complete the full lab workflow from start to finish. The study tried to evaluate the delta between simply having better instructions versus being able to execute a difficult physical workflow.

For many industries, a key question with AI is whether it can replace enough tacit knowledge, troubleshooting ability, and hands-on competence to change what a novice can do.

What the World Lets AI Do

This has implications for policy too. If governments don’t know yet exactly how a technology will evolve, rushing to write rigid and detailed rules isn’t helpful. Abi believes more focus needs to go toward building the capacity to respond quickly as evidence gets clearer.

Abi has argued that governments should focus on building ecosystem capacity: better reporting, better evaluations, and more technical expertise inside institutions. This doesn’t require policymakers to perfectly forecast every future use case, but rather gives them the ability to notice what’s happening, test assumptions, and act effectively as they get better information and clearer directions emerge.

AI will transform science, security, industry, and government. But the path from capability to transformation runs through the parts of the world that are hardest to model from the outside. As AI accelerates parts of a workflow, more attention needs to shift to whatever is still slow, physical, tacit, or institutionally constrained. As Abi put it: “Bottlenecks become more important when everything else gets automated.”

The question, then, is not just what AI can do. It’s whether that capability changes the bottleneck. That’s the difference between what AI can technically do and what it can actually do.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

Also covered: ARC’s definition of AGI as human learning efficiency, why world models and memory matter, current ARC-3 scores, long-horizon evals, open vs. closed source dynamics, multimodal integration, the underinvestment in parametric learning, what’s next for agents (spending, agent-to-agent communication, always-on), and why vibe coding raises the floor but not the ceiling.

Chapters

00:14 Background & Joining ARC Prize

03:06 Role at ARC Prize & Founding Team

03:55 ARC1 to ARC3 Evolution

06:07 Building a Benchmark: Idea to Sunset

08:08 ARC4 & What’s Next

08:45 Approaches to Beating ARC

11:33 Why Benchmarks Matter

12:24 Who Are Benchmarks For

13:42 Evolution of Benchmark Difficulty

15:27 Long Horizon Agents & Evals

16:59 Defining AGI and ASI

18:37 Role of World Models

22:02 Where Models Struggle Today

22:57 Open Source vs Closed Source Models

25:52 Trends & What’s Exciting

27:03 Under-invested Research Areas

29:55 Vibe Coding & Role of Humans

32:18 Final Thoughts

In the 1660s, Europe’s natural philosophers had a knowledge-sharing problem. Discoveries were spreading through private letters, which meant they were spreading slowly and often not at all, because scientists worried about losing credit. Some went so far as to publish findings as ciphers and anagrams, staking a dated claim to the discovery while keeping its contents secret. Henry Oldenburg, secretary of the newly formed Royal Society of London, was at the center of this correspondence network. His job was to collect, translate, and circulate scientific reports among members across Europe.

In 1665, Oldenburg decided to turn that correspondence network into a formal printed version of the scientific communications of the society. Thus, the Philosophical Transactions of the Royal Society, one of the first scientific journals, was born. It made discoveries public, dated, citable, and open to scrutiny. And in doing so, it helped create the system of published papers through which scientific knowledge has accumulated ever since.

More than three hundred and fifty years later, the paper is still the basic unit of scientific knowledge. And for the first time since Oldenburg, the format is meeting a reader it wasn’t built for.

“The whole research infrastructure was built for humans,” Amber Liu told me. “It’s been built to accommodate human speeds: reading speed, understanding speed, experimenting speed. The entire stack, starting from the paper PDF and arXiv, to conferences and peer review, to enterprise research tools like GitHub and Weights & Biases are all built for human researchers.” Without infrastructure designed for AI scientists, she argues, “it’s very hard for them to either compound knowledge or collaborate with each other.”

That idea is at the center of her recent paper, The Last Human-Written Paper: Agent-Native Research Artifacts. I sat down with her to discuss the paper, how she sees the role of AI scientists, and what it will take to get there.

Recording and transmitting knowledge

The scientific paper takes months of branching, backtracking, and dead-ended research. What comes out, the paper itself, is a compression of all that, which exacts two taxes.

Storytelling tax. A paper is a persuasion device. As Amber puts it, “the paper PDF is entirely built for human researchers to convince human reviewers. Its pages are a polished story and they hide all the tedious work that’s been done in the process.” The failed hypotheses, the rejected designs, the pivots, the “days and nights of experiments trying to fine-tune the performance.” All of that disappears in the final cut.

Engineering tax. In persuading reviewers, papers systematically underspecify what’s needed to actually rerun the work. Setup details, hyperparameter choices, and all decisions that made an experiment succeed. These live with the research team who wrote the paper, but don’t actually get published in the paper.

As Amber put it, “the paper itself is a lossy compression of the research knowledge. It doesn’t have all the necessary information needed to reproduce the important results.” And it’s this part, the part that gets thrown away, the research process itself, that is more valuable than just the outcome itself.

These taxes were somewhat necessary costs for human readers. Without them, it would be even harder than it already is for humans to read all the relevant papers published.

Agent-native research artifacts

However, this changes with AI agents. Agents can process many orders of magnitude more than humans can and the missing details are actually a key part to understanding and reproducibility for agent “readers”.

In a world where AI agents drive a lot of our scientific research, the question becomes what substrate do these agents need in order to do cumulative, reliable science?

Intelligence alone doesn’t compound. Intelligence plus a medium for faithfully recording and transmitting knowledge does. “If we have a more AI-native protocol to document research knowledge, it basically facilitates the research to compound over time,” Amber said.

Her proposed replacement for the paper is the Agent-Native Research Artifact, or ARA. This is a machine-executable research package with four layers:

1. Scientific logic: the conceptual-level claims and reasoning behind the work

2. Executable code and full specifications: the actual implementation tied to the claims

3. The exploration graph: the full trajectory of the research, including the branches that went nowhere

4. Evidence: the raw outputs, not summary statistics

“Everybody is talking about intuition, which is very vague. People don’t know what intuition is, and people don’t know how to learn intuition from top researchers,” Amber told me. The scientific logic layer is meant to be a container for that knowledge, “things we learn in the process but don’t necessarily document inside the paper.” For example, how a seasoned researcher approaches a new problem or tunes hyperparameters. This is basically the idea of research taste, which we’ll return to, made traceable.

Preserving failure

The exploration graph preserves a different kind of lost knowledge: the failed experiments that today never get published. This is expensive knowledge that is never leveraged in any formal capacity. Someone spent compute, time, and attention learning that a path doesn’t work, but then the publishing system deletes those lessons, leaving every future researcher (human or artificial) to rediscover the same dead ends. Amber believes that we need to start treating “what didn’t work” as first-class scientific memory.

There’s a technical urgency behind this argument as well. “Today’s models are very good at trying to prove something is right, because all their training data contains the successful trajectories,” Amber explained. This can lead to perverse results. In her analysis of research benchmarks, she’s watched models game their objectives in creative ways: one model “is really good at looking at ground-truth answers and making up synthetic data and merging it into the training data so it can reach 100% scores,” while another family of models will happily modify (even hardcode) components of an architecture “just to please the judge inside the evaluator.”

In other words, the hope with feeding a model the dead ends, abandoned branches, and intermediate reasoning is that it can start to develop critical thinking about its own work.

Today, we’re starting to see workshops encouraging humans to publish their failures, but adoption is slow. “For AI scientists, it’s relatively easier to publish their trajectory,” she noted, “because an AI scientist doesn’t have this burden of disclosing that they’re actually doing a lot of dumb things, that they fail a lot in the middle.”

Plurality of judgment

If AI can generate “tons of ideas, experiments, and conclusions overnight,” as Amber puts it, then the bottleneck migrates downstream to verification. Last week, I wrote about where this is showing up in the physical world, but it’s also showing up in scientific research.

The failure mode Amber worries about most is speed compounding error. Researchers are now running self-evolving systems, which are agents doing, in her words, “twenty-four hours research.” So if the agent hallucinates or reward-hacks mid-stream, that mistake propagates through everything built on top of it, wasting weeks of compute on a corrupted branch. Amber believes we need formal verification and neurosymbolic methods to “guardrail” the research process and catch errors as they happen, rather than asking a human to try to audit after the fact.

The human verification system is straining under the same load. Recent ICML and ACL cycles have seen submissions double, and, in Amber’s words, “the peer review system is essentially broken.” The morning we spoke, one of her students had been commiserating about EMNLP reviews; his group’s internal statistics suggested the ratings were very noisy, “good papers get bad scores and bad papers get good scores.”

To help reduce this noise, agents should be responsible for more of the mechanical checking. “Most objective review should be done by AI to save the attention of humans,” Amber told me. The vision is automated verification of claims against evidence, something like a grammar checker for science, which is the analogy she uses in her paper. Humans, meanwhile, keep what was always the interesting part: judging significance, novelty, and taste. Two excellent reviewers can weigh the same paper differently: one prioritizing resource efficiency, another rigor, “and both of them are correct.” That plurality of judgment needs to continue to exist.

The foundations of compounding intelligence

If execution is increasingly the machines’ job, the bottleneck shifts to judgment. “How to identify the important problems, how to approach them, how to raise the questions. This is what matters now, because AI agents can take over the execution,” Amber told me. She added that, “Research taste itself is a resource-allocation algorithm. For a hundred ideas, you need to pick the ten, or the one, actually valuable enough to put more resources into.” Taste is the skill of deciding where attention, time, and compute go when generating options is nearly free.

In the long run, “only the top 0.1% of researchers with great taste will actually control the resources.” Whether or not you accept the number, the direction is hard to argue with. As execution gets cheap, knowing what’s worth executing becomes the valuable skill.

But before this can happen, taste has to be captured, failures preserved, and claims made verifiable. As Amber put it, “Human wisdom started accumulating because we invented language.” Oldenburg understood this in 1665. The scientists of his era were lacking a medium, and the journal he invented enabled three and a half centuries of discovery to compound. Science now has a new kind of reader, working at speeds no journal was designed for, and the medium has to be reinvented again.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

In Pablo Lubroth’s interview of Jack Scannell, Scannell references an analogy made by David Shaywitz where he draws a parallel to the early electrification of factories during the Second Industrial Revolution. In the decades after electric power arrived, many factories simply bolted an electric motor onto a layout built for steam, and saw almost no gain. The returns came only once the whole floor was redesigned around the new power source. That rebuild took decades.

It feels like we’re at a similar point in the AI revolution. We’ve invented the AI equivalent of the electric motor, but we’re still bolting that capability onto layouts built for steam.

This is especially true in many physical domains where AI is getting very good at the technical steps around generating designs and simulating how they’ll behave. But translating that capability into an output in the physical world is still a work in progress.

Where this shows up in manufacturing

In manufacturing, AI has moved fastest on the upstream, design-adjacent work. It reads 3D CAD geometries and technical drawings and turns them into production parameters, cost estimates, and toolpaths. It flags manufacturability problems (thin walls, features a tool can’t reach, tolerances a process can’t hold) and suggests fixes.

These tools are about design. But what matters most is reliably producing high-quality goods at scale. And this is bounded by physical cycle times, which accelerating design work doesn’t really impact. You can model a part in seconds, but you still have to cut the tool, run the line, measure what comes off it, adjust, and run it again. Speeding up design doesn’t compress the loop that actually takes the time.

This part of the loop, the physical production, also generates valuable data around what actually came off the line, and if/where it deviated from spec. Feeding that back into the AI design and simulation tools is what closes the loop, producing better designs on the next run. So, in theory, while you can’t make any single physical cycle faster (yet), you can make each run teach the next one, thus reducing the number of runs needed to reach the yield or quality you’re after.

This is why the promise of the AI-native manufacturer is exciting. It points toward systems that try to orchestrate the whole line, from design through to finished output. The premise is that once you own the entire loop you can close it. As the upstream design tools proliferate, this is where I think the value will accrue. Rather than bolting an “electric motor” into a legacy factory, these new entrants are electing to rebuild the entire factory.

Where this shows up in life sciences

The same dynamic is also playing out in life sciences. Two excellent recent pieces acknowledge that although we’ve made designing molecules easier and cheaper than ever, the hard part (and likely where value will accrue) is now validating those designs in humans.

Carlos Outeiral, in The antibody revolution is real. The business, less so. explores how, once several labs plus open-source models can all design comparable antibodies, the design step commoditizes, margins compress, and what ultimately ends up mattering is the asset itself, a differentiated drug. As he puts it, “the real problem is whether the antibody produces a real benefit in a real patient, which is a property not of the molecule but of the biology, and which no amount of binding affinity or easily assayable properties can buy you.”

In the Decoding Bio interview I mentioned at the start of this piece, Scannell argues that one of the fundamental bottlenecks to R&D productivity is predictive validity, which is the extent to which preclinical models and assays accurately predict human outcomes. Scannell coined the term Eroom’s law in 2012, which says the inflation-adjusted cost of developing a new drug roughly doubles every nine years. The name Eroom is Moore spelled backwards, a nod to the fact that what’s happening in drug discovery is the opposite of what’s happening with transistors. One explanation for this is the lack of predictive validity. If the biology you’re testing against doesn’t predict human outcomes, designing binders faster just produces wrong answers faster. Whereas, a small gain in validity can outweigh a hundredfold gain in screening throughput.

This mirrors what’s happening in manufacturing as discussed above. We have good molecular design and simulation tools, but the value only shows up once you close the loop between those designs and the physical world, which in biology means testing in humans to create drugs. Closing the loops means rebuilding the “factory,” except here the factory is the drug-development company itself: vertically integrated platforms that own design, simulation, and physical validation together, so that what’s learned in the clinic feeds back into the next design, improving predictive validity. You can’t skip the human trial, but each one can teach the models that generate the next candidate, so you need fewer shots to find something that works in a real patient.

The scarce thing is downstream

In many physical world domains, AI is commoditizing the upstream steps around design and value is now moving downstream to where the bottleneck is: proving the thing actually works where it has to work. Better or more manufactured products, drugs that work in humans (not just mice). There are many more examples of this across other physical domains, but the throughline in all of them is that physical validation is hard because it’s slow, expensive, and irreducible.

We’re in the midst of our own revolution, one I believe will eventually dwarf the Industrial Revolution by orders of magnitude (some people say it already has). AI is the new power source. And because of it, we’re designing new molecules, discovering new materials, and simulating parts faster than ever. But how we reorient ourselves to translate those discoveries into finished goods and viable businesses is still an open question. I don’t think it will take as long as it took factories to redesign around electric motors, but I do think we’re only just starting on what is arguably the hardest part.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

A few weeks ago, I sat down with Adam Marblestone to talk about what might be worth reading off the brain. The weights of the brain’s learning subsystem are tuned over a single lifetime, particular to one brain. Adam believes that the thing worth reading is actually the wiring: the architecture and the reward circuitry built into the structure itself.

After our conversation, Adam pointed me toward a few other people thinking about how biology might inform AI, and I recently sat down with three of them: Sean Escola, a computational neuroscientist who co-founded Herophilus (platform acquired by Genentech) and the robotics company Fauna (acquired by Amazon), and is a Venture Partner at Protocol Labs; Dan Wetmore, who has spent about fifteen years at the intersection of biosignals and hardware, including at CTRL-labs, the EMG wristband company acquired by Meta; and Tom Griffiths, a cognitive scientist and the director of the AI lab at Princeton.

In a 2023 Nature Communications paper, Catalyzing next-generation Artificial Intelligence through NeuroAI, Sean and Adam, along with their co-authors, propose that to accelerate progress in AI, we must invest in fundamental research in NeuroAI. “Historically, many key AI advances, such as convolutional ANNs and reinforcement learning, were inspired by neuroscience. Neuroscience continues to provide guidance [...] but this is often based on findings that are decades old. The fact that such cross-pollination between AI and neuroscience is far less common than in the past represents a missed opportunity.”

In my conversation with Sean, Dan and Tom, we explored what better cross-pollination could look like and the impact this could have on the field. As Sean put it, “There’s a whole set of potential ways that we can build better artificial intelligence by looking towards biological intelligence.”

The brain is not one inductive bias

In cognitive science and AI, a widely-used framework holds that any information-processing system can be understood at three distinct levels of analysis. In his 1982 book, Vision, the computational neuroscientist David Marr proposed these levels:

1. Computational level, which is the abstract problem a system solves and its ideal solution

2. Algorithmic level, which is the representations and algorithms that approximate that solution

3. Implementation level, which is how those representations and algorithms are physically realized.

In Levels of Analysis for Large Language Models, Tom and members of his lab argue that methods developed in cognitive science can be useful for understanding large language models. “The same three levels can be used for analyzing large language models, focusing on how such systems are shaped by their function, the solutions that they seem to find, and the realization of those solutions in weights and units within the underlying artificial neural network.”

In our conversation, Sean expanded upon this and described a parallel hierarchy of four ways in which neuroscience offers a source of inductive bias for AI, moving from least disruptive to most disruptive:

1. Representational, which is about asking models to represent information internally the way brains do. It’s the simplest of the four in some ways. The data you’d need is neural activity, and it stays compatible with the transformer stack we have today, and likely also with whatever replaces it.

2. Algorithmic, which focuses on new, biologically inspired learning rules. The data you’d need might be neural activity and/or connectomes.

3. Architectural, which are new circuit motifs derived from connectomes to replace the transformer. Transformers are the backbone of nearly everything we build, but they have no real analog to the brain’s circuitry, beyond the loose sense that both have something like attention and something like neurons. But nothing in a transformer derives from the brain’s macro- or micro-structure.

4. The compute substrate, which would be neuromorphic or wetware-inspired hardware that computes the way neurons do. This is the far end of the ladder where, as Sean put it, “we would have to throw out the entire stack and start from scratch.”

These are ordered by how much of today’s AI each one leaves standing. “You can kind of see as you go down this hierarchy,” Sean said, “that you’re going from things that are completely compatible with the existing hegemony of the LLM built out of transformers existing on GPUs, to ‘we’re going to throw out the entire stack and start from scratch.’” In other words, the compute substrate could have the most dramatic payoff, but it would also require demolishing most of the existing infrastructure of the modern AI ecosystem.

Representational alignment, on the other hand, is the least disruptive approach and it’s also the rung where the argument for reading brains gets most interesting. Whether it can improve AI hinges on whether neural data carries information you couldn’t get anywhere else. This is the line of attack Sean, Dan, and Tom argue should be prosecuted.

One warning from Dan that applies to every rung on the ladder is that the brain is only useful to copy at the right level of description. “One of the challenges historically in this field,” he said, “is to operate at the right level of abstraction for the specific goals that you have.”

Neuromorphic computing is the cautionary tale. Early efforts tried to reproduce the analog behavior of individual neurons and how they spike. That, Dan said, was a case where “folks were being much too grounded in the details of how biology was doing it, and not abstracting away the process that the circuits were actually carrying out.”

Nothing that follows is an argument for copying neurons. It’s an argument about a specific kind of information, and about reading it at the level where it means something.

Why behavior is not enough

In my post with Adam, we discussed whether brain data tells a model anything it couldn’t already figure out on its own. As Adam put it: “What’s the delta? What’s the difference between having that information and having just the information about the world that we train on now? Are there things in that neural activity that we can’t already predict from the data that it’s seeing?”

Put simply, the answer to this might be that behavior hides most of cognition. Sean, Tom, and their co-author Patrick Mineault discuss this in their recent paper Cognitive Dark Matter: Measuring What AI Misses. Sean gave the example of watching someone cook. “Imagine someone is making pasta for dinner and you’re observing them. At some point in the recipe, they reach into the cabinet and pull out some spice mix and put it in the pasta. At what point did they make the decision to do that? Just at that moment, an hour before, a week before when they were at the grocery store.” It only becomes a visible action when they reach for the jar, but the decision might have been made much earlier. “The behavior records the act, not the deciding. You’re only getting the selected behaviors,” he said, “not the unselected behaviors.”

LLMs are trained on the selected behavior, but human cognition is mostly weighing other options and choosing not to do those. These are all the evaluations we weigh (consciously or subconsciously) before we take an action. Recreating human reasoning may require understanding the discarded branches instead of just the chosen one. And that data isn’t gleaned from behavioral data because it never became visible. Neural data is a bet that you can read the compression loss.

You don’t need neural data everywhere. Instead of trying to make models brain-like across the board, the goal is to find the specific areas where they fall short and collect data from humans doing those specific tasks. Tom is a cognitive scientist and he works at this intersection of computer science and psychology. He’s interested in “how the methods we’ve developed for studying human minds give us a tool for understanding these AI systems.”

Sean, Dan and Tom believe that if you know where a model’s reasoning diverges from a person’s, you can design a task that isolates it. The neural data fills the gap at the level of representation where you’ve established the model is weak.

Capability and trust are the same problem

The hope in doing this is that this will make models both more capable and more trustworthy. The trust half runs through theory of mind. “We’re able to engage as strangers in a conversation and engage in business relationships or whatever it is because we have some theory of each other’s minds that will allow us to operate in ways that are safe for us,” Sean explained. “In order for us as humans to incorporate agents into our societies,” Sean said, “we will need to have our models of their minds be valid.”

The hope is that if you design tasks that are based on theory of mind or based on cooperative behavior, then you can make models more prosocial.

This is largely the thesis of cooperative AI research. In the 2021 Nature paper, Cooperative AI: machines must learn to find common ground, Allan Dafoe and his co-authors argue that theory of mind is a precondition for AI systems to cooperate safely with people, and that cultivating it should be a first-class research goal rather than a side effect of scale.

Also, a model that scores well but reasons in alien ways is a supervision problem. It’s hard to predict and hard to catch if something goes wrong. But by making a system reason more like a human, then perhaps the path it took to an answer becomes something we can actually follow.

The missing traces of thought

We are building AI almost entirely out of selected behavior, on the assumption that whatever matters survives into the behavior output. But what if it doesn’t? What if true decision-making happens upstream and gets compressed away before the visible behavior? If that’s the case, then a model trained on outputs is learning just from the residue of the reasoning that produced it. The goal of reading the unselected is that we can learn from the part it structurally leaves out.

Whether that part carries a signal an LLM couldn’t already recover is, as Adam said, “one of these things that needs to be tried.” Because if it does, the same signal could teach a model to reason where it currently falls short and make that reasoning something we can more easily understand.

After all, if a human mind is more than the sum of its outputs, then so is everything worth learning from one.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

Neural networks can solve physical simulations much faster than the numerical solvers engineers have relied on for decades. That is exciting for a lot of reasons, some of which I wrote about previously here.

But neural solvers can be much less accurate than classical ones. And they’re only fast once they’re trained.

In order to train neural solvers, someone has to generate training data (usually by running the very classical solver the model hopes to replace), then train, tune, and validate it. Those costs might be worth it if speed matters. But maybe not.

So when does paying the upfront cost become worth it?

That’s the question Mikhail (Misha) Khodak, together with Yijing Zhang, Nicholas Roberts, and Tanya Marwah, set out to answer in Breakeven Complexity: A New Perspective on Neural Partial Differential Equation Solvers.

Last week, I sat down with Misha, an assistant professor of computer sciences at UW-Madison, where he's been working on specialized foundation models and folding AI tools into algorithm design and scientific computing. In our conversation, we unpacked the research, the motivation behind it, and the impact he hopes it will have on the broader industry.

“It’s not really a case of ‘we’re just going to replace all the classical solvers with deep learning’” he told me, “There are stages of many processes where different things will be useful.”

The pessimism that started it

In 2024 McGreivy and Hakim from Princeton published a paper that landed hard on the field. “They went through many deep learning papers, mostly ones applying deep learning to fluid simulations,” Misha told me. “They found that in a lot of these papers, the comparisons are mainly against fairly weak baselines. The solvers being used aren’t the ones an actual practitioner would use.”

Those papers were overlooking a knob every numerical solver has, which is resolution. You can always make a classical solver faster by down-resolving. “In practice you can set a lower resolution for your solver,” Misha said. “The speed is determined by how many time steps you take and how fine your mesh is. And they often don’t compare to that.”

A neural network that’s “faster” than a high-resolution classical solver, in other words, hasn’t necessarily won anything, because the honest comparison is on speed against a cheap classical solver tuned to the same accuracy.

That paper, Misha said, “caused a lot of pessimism at the time among the scientific computing community.”

Despite the paper, the field kept building. “I was trying to reconcile this issue where we were still developing these ML methods and some people seemed optimistic,” he told me, “but others, especially in the physics community, would tell you to just down-resolve to get a faster solver.”

Counting solves until you break even

The paper proposes breakeven complexity, a metric that counts the numbers of runs before a learned solver is cost-effective relative to an error-equivalent traditional solver. Error-equivalent here means the classical solver has been down-resolved to the same accuracy as the neural one.

A neural solver only becomes economical when its upfront cost can be amortized. Unlike a classical solver, which pays its computational cost every time you run it, a trained neural network can be reused across every subsequent solve. This is important because most real-world engineering isn’t solving one equation once. As Misha explained it to me, “You are trying to optimize some engineering object. And to do that you have to solve many, many similar partial differential equations.” In practice that means optimizing a design, which requires repeatedly evaluating slightly different geometries, boundary conditions, or material properties. (The paper sets aside real-time applications, where neural networks already have an obvious speed advantage.)

The question then becomes: how many times do you have to run the model before the time you save at inference outweighs the time you spent generating data and training?

The answer, they found, depends a lot on the type of problem you’re solving.

At one end are toy problems, which are simplified academic test cases. “If you want to solve these toy 2D Navier-Stokes with periodic boundary conditions,” Misha said, “you have to perform hundreds of thousands of inference calls before your data generation and optimization cost pays off.” The reason is that for easy problems, the classical solver is very hard to beat even after you degrade its resolution. “For these toy problems there are extremely fast solvers. Even scaling down your resolution, you’ll still end up with a fast solver that is pretty good. So a neural network is just not going to beat it anytime soon. You really need to perform millions of inference calls.”

But as the team moved to harder problems, that changed. As you move to harder problems you require less inference calls in order to make training this neural network worth it. “What happens if instead of Navier-Stokes with periodic boundary conditions you have more regular inlet-outlet boundary conditions, and maybe you throw some blocks in there. Put some obstacles into your fluid flow,” Misha said. “Suddenly that solver becomes very expensive. And scaling down that solver doesn’t just speed it up. The tradeoff becomes much worse. It will be faster but much less accurate.”

On hard problems, down-resolving the classical solver degrades its accuracy, while the neural network holds its speed advantage at comparable quality. The team confirmed the pattern along several axes of difficulty: scaling a chaotic system from 1D to 2D to 3D, predicting further into the future, and pushing fluids toward turbulence.

Model size and scaling laws for training

Breakeven also depends heavily on which model you use. One of the paper’s more surprising findings is that some of the leading foundation models for physics, such as Poseidon and Walrus, don’t always come out ahead. “In some cases the latest large models are actually too expensive to run, at least on the simulations we tried,” Misha told me. “When I say too expensive, I mean more expensive than a reasonably down-resolved classical solver. So it’s not even worth it to run them.” In several cases, well-tuned, smaller neural operators outperformed the larger transformer-based models, suggesting that “it’s worth it to pre-train very efficient models rather than very large models.”

That same philosophy extends to training itself. Rather than simply throwing more compute at the problem, the team uses scaling laws to determine how a fixed compute budget should be divided between generating training data with expensive classical solvers and training the neural network. As Misha put it, “You can very nicely and consistently predict what is the optimal tradeoff as your computation budget increases.”

Even then, “a practitioner should view this as the upper bound on the performance of a neural network for their problem,” Misha said, “because we are training and testing on the same distribution.” Real-world engineering systems inevitably drift into scenarios the model never saw during training, so neural networks degrade in ways classical solvers generally don’t. Therefore, Misha sees this more as a screening test: if a neural solver can’t outperform a classical solver under these idealized conditions, it’s unlikely to do so in production. If it does clear that bar, it’s probably worth trying on your actual problem.

A funnel, not a replacement

Towards the end of our conversation, I asked Misha what he was most surprised by from this research. “I’m more optimistic now than I was before I started this,” he said. “I actually was maybe more skeptical of neural solvers when I came in. And that’s why I wanted to figure it out.”

Misha also described the future he sees with neural and classical solvers as a pipeline. He referenced the illustration from Periodic Labs, a Zetta portfolio company. “The idea was, with these physical simulations, there’s an agent trying to optimize some material or protein or drug. They’re first going to see what heuristics say is a good candidate, then narrow down. Then they’ll try a neural surrogate and narrow down a bit more. Then they’ll try high-fidelity classical solvers and narrow down a bit more. And then they’ll actually have to go to the lab and build the thing.”

A funnel, in other words where we use cheap heuristics at the wide mouth, fast neural surrogates next, expensive high-fidelity solvers after that, and physical experiment at the narrow end. Each stage tuned to a different point on the speed-accuracy curve.

“There’s room for all of these things,” he said. “It’s not really a case of replacing all the classical solvers with deep learning. There are stages of many processes where different approaches will be useful.”

It’s not a question of whether neural solvers beat the classical way, but rather where each one outperforms vis-a-vis the other. The breakeven paper is the first attempt to answer that question with real numbers.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.

At this point, most companies and individuals are using a relatively baseline level of AI models’ capabilities. Which means we don’t see huge benefits from each new model release. And yet, AI is still failing across many enterprise deployments. Why?

I think it’s because most enterprise AI conversations focus on the wrong question: what part of our budget can we put towards AI? For example, should we use our existing software like NetSuite or throw AI at it and move to a new AI-native ERP?

Whereas the right question is something more fundamental and not related to AI at all: what are our existing processes and where do they break down?

“You will not transform your company without rebuilding operations from the ground up,” Vas Moza, founder of Varick Agents, told me. Varick is an applied AI company that goes into large enterprise organizations to help them transform from the inside out with AI.

Rebuilding operations from the ground up means knowing what to automate deterministically, what to give to an agent, and what to leave to a human. This isn’t a model or agent engineering question. It’s a process engineering one. “That sort of process reengineering,” Vas said, “is what makes AI 100 times more effective.”

Not every workflow needs an agent

Knowing what not to agentify is the first question to answer when it comes to enterprise deployments – because not every workflow deserves an agent. My mental model for this is:

Deterministic automation is for rules.
If the inputs are structured and the desired behavior can be specified, you probably do not need an agent. You need software. These are for tasks like routing an invoice, syncing a CRM field, checking whether required documents are present. These workflows are valuable, but they don’t need open-ended reasoning.

Agents are for judgment under context.
Agents make sense when the work requires interpreting inputs, pulling context across systems, or executing a multi-step process where the path changes depending on what the agent discovers. The agent is useful where deterministic automation breaks down.

Humans are for accountability, ambiguity, and trust.
Some work should remain human. This might be because the consequences or relationships are too important, or the organization doesn’t yet know what “good” looks like. In these cases, the right AI product might be a copilot, reviewer, researcher, or QA layer.

Then, even once you know an agent is the right fit, you still have to decide whether it’s worth building. That’s a separate evaluation across dimensions like manual hour displacement, key-person risk, cycle-time reduction, revenue uplift, time and cost to build, and, most importantly, expected ROI. “A workflow with a 10x ROI should get prioritized,” Vas said. “A workflow with a 2x ROI maybe, maybe not. And a workflow with a sub 1x ROI should be left alone.”

The decision of where an agent belongs is downstream of all these diagnoses. Crucially, the most painful workflow isn’t necessarily the most valuable one to automate.

The deployment is the product

In many companies, the workflow lives in people, who sometimes can’t actually articulate what it is they’re doing. This is the tribal knowledge problem.

Before I was an investor, I built a company in the manufacturing and supply chain space. I remember trying to understand factory workflows by asking people what they did. They could always show me, but they rarely could explain it.

Vas sees the same thing in enterprise operations. The people doing the work aren’t usually used to narrating the work. They may have done it for fifteen years and know which exceptions are important to document versus which aren’t, but they may not have language for why.

And not all of that knowledge is worth capturing. Ahmad Kakar, CEO of Euclid, which builds agentic systems for trucking and logistics, told me, “The catch with tribal knowledge is that a considerable portion isn’t true expertise – it’s workarounds people built for non-existent or broken systems over the years. If you just capture it and encode it into the agent, you’ve automated the dysfunction. The work is pulling it out of people’s heads and then deciding what was real judgment versus scar tissue.”

The process is also a delicate one because the person whose knowledge you’re asking for often assumes their job is at risk. “We want to make sure they don’t feel like they’re having their entire job replaced,” Vas said, “because most of the time that isn’t the case. There’s a lot of work to be done at growing organizations.”

Troy Shen, co-founder of Cervo, an AI platform for customs brokers, frames the same dynamic as a sales problem. “Successful AI adoption takes two sales, not one,” he told me. “You’re selling an outcome to the executive and a dramatically better workday to the operator who’ll actually use the agents.” And without operator buy-in, the deployment will fail.

I spoke with Bihan Jiang, the Director of Product at Decagon, last year about this topic and she told me that a big chunk of her job is internal change management for customers, not just technical implementation. “There’s often a board-level directive to ‘bring AI into CX,’ which creates top-down momentum and we often end up being a part of many vibrant, cross-functional conversations on the customer’s side because there’s a lot of trust and alignment going into a decision this transformative.”

Suril Kantaria, co-founder of Adaptional, which builds agentic AI for insurance claims, frames the capture process as an apprenticeship. “In many enterprises, work is an apprenticeship model,” he told me. “This is a simple but key insight from deploying into some of the largest insurance companies. To succeed, we onboard like eager new hires — learn how the best adjusters handle claims, make decisions, ensure compliance, and keep up with internal policy. Core to the ‘product’ is both a deep understanding of our domain (insurance claims) and an ability to handle the contours of every customer’s claims process.”

This is what makes the deployment process partly technical, partly anthropological, and partly political. As Troy put it, “Deploy for a Fortune 500 and you’re rolling out across multiple offices, each with its own history, procedures, politics, and attitude toward AI. That navigation takes a rare mix of emotional intelligence, improvisation, and stakeholder management.”

The goal is to create a map of the knowledge tree within a company – things like workflow maps, decision rules, exception taxonomies, escalation paths, eval sets, access policies, approval thresholds, and edge-case libraries. This creates structured logic for an organization that can be inspected and, over time, improved.

Modernization does not mean migration

I’ve argued before that legacy modernization is an important wedge in enterprise AI. This means translating old code into new code and requires a full-system modernization. While this is true, it’s worth noting that sometimes the best first move is to modernize the work around a piece of software rather than the software itself. “We’re not going to force you to migrate off your systems of record, which is the enterprise reality for most companies,” Vas said. “They’re very married to their ERP or their CRM.”

These systems hold years of process and integrations so sometimes it’s easier (and better) to build on top of them. This means connecting them and then extracting the process logic between them. Then give agents the context to execute work across them. “We build on top wherever possible,” Vas said “because there is still so much work to be done that exists on top of these platforms and by interconnecting these systems.”

This approach has limits if the underlying architecture is brittle. This might mean the data isn’t clean, or permissions are broken, and so building on top of these problems means you end up automating around dysfunction instead of fixing it.

The best enterprise AI deployments I’ve seen do some combination. They build on top where possible, then let the agent deployment show what actually needs to be rebuilt. Modernization used to mean making legacy software usable by modern developers. These days, modernization means making legacy operations legible to agents.

The pattern library

Each deployment in services-heavy AI companies looks bespoke. However, there are patterns to be learned in the deployment expertise itself, which makes repeatability across customers easier over time.

For example, after multiple finance department deployments, a company can start to learn patterns around how finance departments run and where they break. Importantly, this compounding doesn’t depend on moving client data around. “We allow our agents to have a higher baseline accuracy,” Vas said, “not because we extract data from our clients, but because we have the methodology and practices best suited for the right-tail edge cases we’ve seen across deployments.” The data from one client does not necessarily need to be copied into another client’s system for the next deployment to improve.

This means you start to learn the workflow archetypes, questions to ask, evaluation harnesses, governance models, etc. Clemens Komorek, CEO of Zalion, which builds AI procurement agents for industrial buyers, sees the value of this at the transaction level. “Every transaction teaches the system something about supplier relationships and negotiation levers,” he told me. “And because supply chains keep shifting, the system has to adapt as outcomes change. This is how localized expertise becomes something repeatable.”

What remains bespoke is the business-specific context. “That part is not repeatable,” Vas said, “because every single business is different. Pretending otherwise is why most AI projects fail.”

The accumulated knowledge is what creates a pattern library and this is the moat. It’s knowing where autonomy belongs and how to turn processes into structured operational logic, built up across deployment after deployment.

Why neither the labs nor the incumbents own this

It is tempting to assume the labs will eventually own enterprise AI deployments because if the models keep getting better, why not just build the agents? And yes, the labs will own a lot of the horizontal intelligence infrastructure, but there’s a difference between selling intelligence versus selling operational transformation.

The labs are optimized for scalable intelligence whereas enterprise transformation is not cleanly scalable. Plus, no enterprise wants its entire company dependent on just one or two model providers because it would make them highly vulnerable.

So there’s still a need for a layer that decides which model belongs where, how to route work, how to control cost, and how to ensure the workflow continues even when the model layer shifts underneath. This is something I wrote about last year – the shift from models to systems.

Consulting firms and legacy systems integrators are the other candidates to own enterprise transformation. They’ve shepherded companies through past technology shifts, and they know how to navigate large organizations. The question is whether they can build and continuously upgrade production AI systems fast enough.

AI moves quickly. Best practices from six months ago are practically obsolete now. “The inertia required to move a large behemoth organization like a McKinsey to become AI-native themselves,” Vas argues, “is insurmountable.” A firm with tens of thousands of employees and decades of internal process inertia may struggle to become AI-native quickly enough to deliver the kind of systems customers actually need.

This is not to say the incumbents will not participate. They will. They already are. But there is an opening for a new category of company that combines consulting-grade process understanding with software-grade deployment velocity. These can be horizontal companies or vertical-specific solutions.

The agent is the mechanism

A company transforms when the relationship across its people, its processes and its outputs change. “Rebuilding a company around AI,” Vas said towards the end of our conversation, “means starting to understand where the processes themselves break down. It’s not about slapping AI on top of operations that aren’t yet designed for it.”

In many cases, this means rebuilding the operating model from the ground up. And in that new model, plenty of work still belongs to humans or to deterministic automation – not to agents. Where an agent actually belongs is something you discover through the deployment process itself. And that’s what compounds – the accumulated knowledge of where autonomy works and how to deploy it.

After all, enterprises aren’t looking to buy an agent. They’re looking to enable their business to be better than it is today, either with more efficiency or with greater outcomes than they could reach before. The agent is just one mechanism.

Author’s note: An LLM was used for light copy editing only (spelling, grammar, and clarity). Content, meaning, tone, and structure remain unchanged.