Last week we hosted Deepak Pathak, the founder and CEO of SkildAI, building a general brain for robots.

Here are our main takeaways from our dinner:

1/ Intelligence is decision-making. Everything else is clock speed.

Body control runs at 500–1,000 Hz. Vision at ~10 Hz. Speech at 2–5 Hz. Math at ~0.001 Hz. It’s the same machinery all the way down.
Sit with that for a second: the way our entire industry splits itself into vision, language, and controls teams may be an artifact of our tooling rather than the problem itself.

2/ We’re building intelligence top-down. Evolution built it bottom-up.

Evolution spent billions of years on movement and hundreds of millions on vision and hands. Language appeared only 10,000–50,000 years ago, and the brain barely changed to accommodate it.
That makes language a product of intelligence, rather than its source.
Deepak’s question to the room: Where is the evidence that scaling the last sliver of that timeline will fill in everything underneath it?
Nobody had an answer.

3/ “General” only means something if it’s falsifiable.

Anyone can put the word in a deck. His definition: any robot, any task, one brain.
They cut a quadruped in half, and both halves keep walking. An arbitrary humanoid walks within seconds.
That’s a claim you can break in public - which is why I trust it more than the adjective.

4/ Omni-bodied is a data strategy wearing a research costume.

Robotics has no internet to scrape. The only large-scale data will come from deployed fleets, and hardware will never converge on a single shape.
There are dozens of TV makers, phone makers, and car makers. Bet on one form factor, and you quietly cap your data ceiling forever.

5/ There are four data sources, and each one’s weakness is another’s strength.

Teleoperation: pristine, tiny.
Gloves: middling on both.
Simulation: infinitely scalable, with a reality gap.
Human video: diverse, free, and never robot data.
Anyone selling you on one source is selling you their own strength.

6/ Use video and simulation for pre-training. Use teleoperation for post-training.

It’s the same shape as LLM training.
His read on the field is that most teams run this backward, burning their scarce, clean data on pre-training - where scale matters more than polish.

7/ Watching Federer for 100 hours will not make you play like Federer.

Video gives you the what. It never gives you the forces, the momentum, or the forehand-to-backhand transition.
Those have to be practiced - and practice happens millions of times in simulation.
This is the cleanest argument I’ve heard against treating egocentric video as a standalone answer.

8/ Deployment is the third pillar, and it’s the one nobody pays enough attention to.

Data. Architecture. Deployment.
Deepak named the third as the single biggest reason this wave could fail.
On the objection that “deployment isn’t venture-scale,” his response was blunt: “A very convenient argument to give.”
Skip deployment, and you lose the only thing that tells you you’re wrong.

9/ Robotics has no working evaluations, so revenue is the scoreboard.

You cannot find two robotics papers evaluated on the same benchmark. There’s no arena, shared setup, or honest baseline.
That leaves money paid as the only real signal - and across the field, that number still rounds to zero.
He called it a convenient shared incentive to remain unmeasured. Hard to argue.

10/ Customers don’t care about accuracy. Zero.

They assume accuracy. That’s what they’re paying for.
What they buy is cycle time and cost.
The metric that earns you a paper and the metric that earns you a purchase order barely overlap. Closing that gap took his team a year of actual deployment.

11/ Evolution did the pre-training. Childhood is post-training.

His curiosity-driven exploration work solved Atari without rewards - and never worked on a real robot.
His theory now: children reach the same milestones whether they grow up in a mansion in LA or a village in India, so most of that learning was already paid for.
That gives robotics permission to be spectacularly data-inefficient because we’re paying the evolution bill from scratch.

12/ Language stops working at about 30 seconds.

“Pass the sandwich” - perfect.
A three-minute unfamiliar task - useless, because you’d have to dictate joint angles.
Beyond that horizon, the interface becomes demonstration.
As he put it: If words were enough, handymen would be out of work.

One more thought on focus, from his advisor:
The world max-pools you to your single highest achievement.
Decide what your max pool is. Spend your time there, and let everything else fall into place.

He joined Tesla as an engineer in 2014 and rose to lead its AI organization.

Here are a few takeaways from our conversation:

1. Evaluation is the bottleneck.

At the end, I asked him what the community should go build. His answer was evaluation.
“Evaluation is such a pain. That is probably the biggest problem.”
Tesla runs roughly a dozen different kinds of evaluation because no single one tells you enough. Then he shared the detail that made the room go quiet: you are trying to detect a 5% improvement with an evaluation that has 50% variance. Change nothing except the random seed you train with, and your error scores can swing wildly. At that point, how do you know whether your change actually did anything?
Building a low-variance estimate of whether your robot got better is real engineering. Almost nobody wants to fund it. Everybody wants to fund the model.
But if you are building a serious product, evaluation is the part that determines whether you actually have one.

2. The shortcut that gets you to a demo costs you the ceiling.

On LiDAR, he was direct: it gets you to market sooner. Then it slows you down.
His framing was that sensors that hand you early wins can pull you away from the right approach entirely. You take the shortcut, patch around it, and eventually find yourself crawling. Cameras give you none of those early wins. But the information is already there: five megapixels, eight cameras, thirty frames per second, and roughly two gigabytes of data arriving every second. The problem is information processing.
His allocation rule was equally direct: if you have an extra thousand dollars to spend on a self-driving car, spend it on compute, not sensors.

3. You cannot buy your way to a data moat.

Everyone in this space is buying egocentric data right now. I asked whether Tesla does the same.
His view was that purchased egocentric data is a drop in the bucket.
The point was not that data does not matter. It is necessary, but nowhere close to sufficient. How you structure the network, formulate the problem, and build the reward and evaluation around it carries much more of the weight. Having a lot of data and using it badly gets you nowhere.
This is also exactly why the humanoid form factor matters. The world is already built for humans, so you do not have to rebuild the world around the robot. More importantly, human video data transfers to human kinematics. Put a gripper or a wheel on the robot and that data stops applying. Preserve the human form and you begin much closer to the solution.
Tesla’s own factories are full of people performing exactly the tasks the robot needs to learn. That is the loop.

4. Demos are not products, and unit economics decide who survives.

Most robotics demos you see are running on a datacenter GPU bolted to the robot. That is thirty to forty thousand dollars of compute inside a machine that eventually needs to sell for around thirty-five thousand. It is also burning five hundred watts against a power budget closer to one hundred and forty.
You cannot simply offload the work to the cloud. The path from pixels to actuation is about 140 milliseconds. No round trip to a server survives that constraint. For anything safety-critical, inference happens locally or it does not happen at all.
His conclusion was that anyone serious about scale eventually designs their own silicon. The barrier at the edge is much lower than it is in the datacenter because you are not trying to compete with a complete NVIDIA system. Tesla’s chip is designed to be exceptional at exactly one shape of problem: high-dimensional video in, low-dimensional actions out.

5. General beats specialized because of economies of scale.

For any individual task, there is some form factor that is technically optimal. But every specialized robot fragments your volume, your supply chain, and your AI team. The general-purpose robot gets cheaper every year because everyone is pushing down the same cost curve. It is the same reason HBM became affordable: the entire industry standardized around needing it.
Eventually, the supposedly efficient specialized robot becomes the expensive one.
Humans are not optimal at most tasks either. They still perform most of the work in the world. Build one general form factor well enough and you can capture eighty to ninety percent of the market.

6. The problem he would work on if he could work on anything else.

Autoregressive decoding.
His reasoning was that autoregressive decoding makes memory bandwidth, rather than compute, the bottleneck. That is why HBM is the constraint and why HBM prices are heading where they are. Reformulate the problem so that it becomes compute-bound, and you can build dramatically better chips per watt and per dollar.

7. On video, simulation, and the other open gaps.

He named several other problems that remain unsolved and worth building around.
- Video supervision. Frontier training moved from text to images and is only now beginning to reach video. Because of how the objective is formulated, most of the signal contained in video gets squandered. Nobody has cracked internet-video pretraining for robotics in the way text pretraining was cracked for language.
- Simulation. Existing simulation tooling still is not good enough for this problem. He added an important caveat, though: neural simulators can be so expensive that collecting data from real robots is often cheaper. The opportunity is to build one that is cost-effective.
- Grading subjective behavior. A collision is objectively bad and straightforward to score. “Was that sudden brake appropriate?” is a much harder question. No model has developed that intuition yet, and he expects humans to remain in that loop for several more years.
- Failure data. Humans rarely fall, which means there is almost no data showing what falling looks like. Teaching a robot to fall safely, especially in a room with a child, is a whole-body planning problem that nobody has solved.
- Actuators. His warning was that magnets are constrained and actuator prices are going to spike the way memory prices did. If you are not locking down supply now, you may not have actuators when you need them.

8. On how Tesla is actually run.

I asked about the company’s hiring and firing playbook because everyone in the room is trying to build a team.
His answer was that Tesla operates like a high-performing sports team. A lot of the process is self-selection. Tesla is the wrong company for most people, but the people it is right for can do extraordinary work there. It is intense, deadlines are aggressive, and the experience can be painful in the moment.
What people get in exchange is leverage. Almost no other company would hand an engineer that much scope. There is no committee verifying every decision. It is your product. You decide what goes into it, and then you build it.

9. On whether robotics is overhyped.

I compared the current moment to crypto in 2021. He did not accept the comparison.
The models are genuinely beginning to work. The excitement is built on real material. But people badly underestimate how difficult it is to move from a working demo to something economically viable.
Hardware is hard. The long tail is long. Validation is brutal.
His sharpest business observation was to ask whether the model itself is the product. If it is, what exactly are you guaranteeing? Nobody buys unguaranteed software for high-consequence work.
That is why he does not think robotics foundation models will commoditize in the same way open-weight LLMs have. You cannot promise performance across hardware you do not control.

Thanks to Alma and Beta Univesity for supporting Icons.

ZaiNar is truly special. They’ve been building in stealth for 9 years with full conviction that their tech is going to change the world. They gathered one of the greatest teams, and now that time has come. It is one of the hottest and greatest companies in physical AI, and their conviction is paying off. The dinner was off the record, but we were able to confirm a few insights that Eric is comfortable sharing.

Here they are:

What they build. Wireless synchronization precise enough that distance becomes a geometry problem. Light travels about a foot per nanosecond, so once radios are synchronized tightly enough, position falls out. It runs on the Wi-Fi and 5G already installed in a building, software only, no tags and no new hardware. Standard enterprise Wi-Fi location gets you to roughly ten meters. Getting inside one meter is where the use cases start.

The infrastructure bottleneck is radio. The consensus conversation is compute, power and chips. Eric’s argument is that communication networks run out first. If physical AI goes where this room believes it goes, there will not be enough sensors or waveforms to carry it. Dense cities can support it, most of a large country cannot. And it has the shape of the EV charging standoff: everyone wants the infrastructure, nobody wants to fund the first mile.

The unit of the problem is the fleet. A camera can tell one robot where the table is. The commercial question is a hundred robots in a warehouse or a hundred drones in an airspace, coordinated safely alongside people. That is a coordination layer, and it is a different business from building either the model or the machine.

Radio is a sensor layer that is already deployed. You cannot move through a propagating wave without disturbing it. Every Wi-Fi and 5G network in the world is producing this signal today and almost nobody is reading it. The camera story of the last thirty years is about to repeat with radio, and 5G is where the next push happens.

On world models. A blueprint of a space is close to useless on its own. What you need is the blueprint, plus where everything currently is, plus what it is likely to do next, with the lag to reality driven toward zero.

On the company. Nine years of building, no press until this year, a first product line they shut down along the way. Eric was open about the identity crisis that followed and about keeping every dollar in the business rather than taking money off the table. That candor is the part of these dinners I value most.

The question I left with. If a network can assert where every person and machine in a space is, that assertion becomes valuable enough that someone will want to forge it. Verifying physical position to a third party, without handing over the underlying trace, feels like the next solution.

Thanks to Alma and Beta University for supporting Icons.

Here are a few insights from the dinner:

1. The data bottleneck is closing, and almost nobody has updated on this.
Karol’s view is that we are quickly escaping robot data scarcity. Every serious lab now collects at scale internally, and on top of that there is a wave of data companies out of YC and elsewhere doing teleop and egocentric collection. The consequence worth sitting with: if you are building a company whose whole thesis is “we supply the data,” you are entering the part of the stack that is getting solved and getting crowded at the same time.

2. The real bottleneck is the last 5%, and it comes from a misaligned objective.
This was the sharpest thing he said. These models are optimized to reproduce the actions in the dataset. What you actually care about is task success. Those two objectives diverge, so additional data on a given task stops moving you toward 100% reliability. Going from 0% to 20% to 60% zero-shot is a data story. Going from 60% to deployable is a reinforcement learning story. That is where PI is putting real effort, and he was open that nobody has cracked it.

The way they approach it: label large volumes of data with language and scores, train a generalized value function that can judge a trajectory in progress, then use that to score everything and improve the policy. The same way you can watch someone attempt a task and predict whether they will succeed.

3. Hardware reliability is the second bottleneck, and it surprised him.
Despite the enormous volume of robots coming out of China, he says there is still no platform robust enough for real deployment. Robots that stay up, stay safe around people, and survive daily use. This is a genuine gap, and it is a hardware company opportunity sitting in plain sight.

4. Cross-embodiment generalization works, and it’s not fully clear why.
The same PI model runs on humanoids, single-arm and dual-arm manipulators, drones, cars, tractors, excavators, and surgical robots. Training across more embodiments makes the model better at any individual embodiment. Karol called this extremely counterintuitive and said plainly that they cannot fully explain it. Tractors and livestock robots somehow improve manipulation.

The limit: give it a brand new robot it has never seen, and it cannot look at the body and figure out how to move. It still needs a little data from that embodiment. He expects in-context adaptation eventually.

5. Generalists beat specialists, which is the worst startup advice ever given.
His words. Normal startup logic says own a narrow wedge, dominate it, then expand. If OpenAI had followed that, they would have started with translation only. What actually happened is that breadth of data distribution made the model better at every individual task, and he expects the same to hold for robotics. So when people ask him which vertical or which task PI will nail first, his answer is: many tasks at once, embrace diversity, and that is how you get good at any single one.

6. Simulation’s near-term job is evaluation.
This reframed the sim debate for me. Evaluation does not scale: the better your model gets, the more real-world samples you need to tell whether the new one beats the old one. Thousands of physical rollouts is impractical. So sim’s first real value is a closed-loop proxy that tells you if your model improved.

On sim for data, his diagnosis is specific: the blocker is asset creation. Rendering fidelity is fine. Locomotion works in sim because you only have to model your own robot accurately. Manipulation fails because you would have to model every object in the world you touch, including how it deforms, and there is no scalable way to do that. Real-world collection is faster and carries no distribution shift.

7. Tactile sensing is overrated as a blocker.
He has worked with tactile sensors a lot and has been disappointed by how little lift they give relative to the integration pain. PI’s workaround is elegant: put a camera on the gripper, read the deformation of the gripper, and infer applied force. Now it is just another vision signal the model already knows how to consume. Tactile data is also awkward because you only get signal when you are touching something, so most of the time it is empty.

8. The VLA versus world model debate is blown out of proportion.
Companies are building their entire identity on an architectural choice you can swap in about three days. His view: data, data quality, diversity, and deployment matter far more at this stage. Worth remembering when you are diligencing robotics companies that pitch architecture as a moat.

9. Compute will be split, and inference is leaving the robot.
Models are already too large to sit entirely on the robot. The likely shape: a system-one policy on the edge handling low-level action and reflexes, and a system-two reasoner in the cloud, possibly parts of the same network. Where connectivity is bad, local inference boxes on site. Robot-specific silicon is further out, though every chip company wants it.

10. Why they chose the intelligence layer.
Every robotics company historically started with a robot they wanted to build. Everything else became subordinate to that, so the AI team stayed small and the intelligence was always the lopsided part. PI inverted it: solving physical intelligence is priority one, and hardware and software choices serve it. He had never seen a company organized this way, which is exactly why he thought it might work.

On the operating system analogy people keep reaching for: he pushed back. Operating systems work because they abstract the hardware away, and that abstraction breaks for robots, because all the complexity lives in the physical body and how it touches the world. The right abstraction sits higher, at the intelligence level, where you specify a command and any robot executes it.

11. What he would actually tell builders to work on.
Two things. Actuators in the West, because if Chinese actuators get restricted we have no domestic answer, and the supply chain problem goes deeper than most people check: rare earth refining, magnets, copper for coils, batteries, and the manufacturing know-how itself. And deployment: understanding ROI, integrating into a workflow, remote intervention, knowing when the robot should ask for help. His heuristic on where to build was blunt and worth writing down: if it is easy, it is probably the wrong place to enter, because it likely works worse than you think and everyone else is already there.

He also made the point that deployment companies end up owning the best data by accident, which everyone else will want.

12. The barrier to entry has genuinely collapsed.
Download an open model, fine-tune on a little data, get signs of life. One deployment company they work with had robots in a logistics setting three months in. That used to take five to seven years. Hardware is cheaper and intelligence is now available, which is why he thinks this is the best moment in history to start a robotics company.

13. There will be no ChatGPT moment.
He thinks it goes the way of Waymo. There was never a single Waymo moment, and then suddenly they were everywhere. Robots will show up in businesses, then industrial settings, then around people, then simple home applications, and at some point you look up and it happened.

14. First deployment will be shaped by tasks, and tasks cut across verticals.
The framework of picking a vertical may be wrong. His example: one of the tasks PI has focused on is assembling packaging boxes at a chocolate factory. You would file that under logistics, and it sits inside food. The same simple, low-variability tasks recur across industries that look unrelated. Start there, then expand into longer-horizon tasks, more human collaboration, mobility, and higher-consequence failures.

15. The home is the last problem, and privacy is a real part of it.
No two homes are alike, the task variety is enormous, and repeatability is low, so ROI per task is weak. You fold laundry twice a week. A home robot has to do many things to justify existing. On top of that: teleoperation means a stranger walking around your house, and even without teleop there is a robot collecting data in your home. He said outright that people underweight this today, and expects early adopters to opt in the way they did with Alexa.

16. On culture: run the experiment.
Most startups run on the conviction of the founder. Karol said PI would be bottlenecked by his ideas if they operated that way. What they kept from academia is the discipline of running the experiment, looking at the evidence without bias, and doubling down on what worked regardless of what you wanted to be true. His read on the history of robot learning is that most missteps happened because the honest path was hard and people believed the easier story.

17. On his own path: someone gave him a chance at every stage.
He knocked on a lab door in Munich as a student who did not know C, did not know Linux, and did not know robots. They gave him a shot anyway. Later, a visiting postdoc told him his two years of PhD work was worthless and he should be doing deep learning. He ignored him, went to his lecture that afternoon, and dropped everything the next day. That person is Sergey Levine, his co-founder. When someone asked what drives him, he credited the people who kept giving him chances.

18. Moats.
Earn the right to have one by making it work first. The moat appears naturally, because the problem is so hard.

Thanks to Alma for supporting Icons. The best immigration law firm with 99% success rate.

Here are a few insights from the dinner:

1. The demand is coming from plant managers.

Luis described a plant director from large OEMs. He did the math himself. One humanoid per production line. He wants to replace everyone on that floor except himself.
Workers lift parts off a rack below waist height all day and get back injuries. Insurance, complaints, union exposure. He’s buying down a liability.

2. Reliability is the product.

The target he named is 2,000 hours with no maintenance. He does not think anyone gets there soon and said so plainly. These machines fall, hit things, break hands. But four months of 24/7 with occasional calibration is enough. At that point ROI stops being a slide and becomes a purchase order.

3. Training still runs on people.

Robot parks. Hundreds of humanoids in rows, running all day, humans in wearables and VR teaching them to pick up a pillow, a can, a phone. That data feeds the foundation models. Then you fine tune on site. For one automotive customer they built a replica of the actual line and trained on the exact part being manipulated. Very little of this is autonomous today.

His guess at what deployment looks like: the robot arrives, fails a task a few times, and someone shows it five times until it works. Closer to onboarding a new hire.

4. Legs are an economics argument.

A wheeled base tips if it accelerates hard. A humanoid shifts its weight and stays up. In lean manufacturing the robot works one station, turning and reaching in a small radius. Take a five second cycle to four and a half and you beat the entire line. That is the case that survives a CFO.

He was also clear that plenty of jobs go to wheels and arms. Walking stays hard because a fall accelerates exponentially, which is why elderly people die from it and why balance is the expensive part.

5. Coordination is the next constraint.

Plants change output daily. Different volumes, different parts, different spool thicknesses. Reassigning a robot to a new line has to be as simple as telling a worker where to stand. He thinks the pieces exist and nobody has assembled them yet. That gap is a company.

6. The real bottleneck is the supply chain.

American teams wait months for components. A Chinese supplier ships a replacement humanoid in two weeks because the chain sits next door to them. If there is a squeeze, everyone here stops.

Silicon is the other half. An Orin NX runs about 1,500 dollars, an AGX about 3,000, for something that should cost a few hundred. Memory prices are off the charts. His view is that this corrects, because a robot only needs inference. Training stays in data centers you rent. The robot needs a cheap edge chip in the fifteen watt range, and the market is moving there.

He said the person he now values most is the one who can tell you which country the rare earths for your motors come from.

7. Robots will need to communicate before they need to be liked.

He is working on whole body communication. No companionship, no personality. Just a robot that can tell you the food spoiled, or signal to a line worker that something is wrong, through posture and gesture. In a factory that is a safety feature.

8. Coarse manipulation is basically solved.

Sensing has stopped being the wall. High density tactile arrays, capacitive skins, even spectroscopy that can tell you whether a fish is fresh. Now the models have to catch up to the sensing, which is a good problem to have. In hand manipulation is the frontier, and his personal benchmark is a robot that can pull a proper espresso. He is opening a lab for cooking.

The last thing.

He was candid in a way people at his level usually are not. Said engineers get locked into ignoring people and that he appreciates MBAs for exactly that reason. Gave out his phone number to the room.

That is the whole point of the format. The best part of these dinners is never the talk. It is thirty five people asking better questions than any panel would, and a speaker who answers them the way he would at dinner.

Thanks to Alma and Beta University for supporting Icons.

We hosted Martin Casado for an SV Icons dinner last night with a group of Stanford founders, engineers, and investors.

Martin leads a16z’s infrastructure practice and is one of the sharpest thinkers and investors in AI. It was a great conversation across AI, robotics, open source, and how to think about where value will accrue as the market develops.

We shared a few takeaways from the evening on LinkedIn and X!

Our dinners are off the record, but a few ideas from the conversation stayed with me:

1. AI may be the first software wave constrained more by capital than engineering.

2. Capital can now convert directly into compute, capability, and demand.

3. Intelligence has a real marginal cost, fundamentally changing the economics of software.

4. Founders have to account for the next model, not just the one powering their product today.

5. Computers are advancing rapidly at reasoning, but embodied intelligence is a different problem.

6. Navigating an unpredictable 3D world is much harder than operating within a constrained environment.

7. Open source is strategically important, but the economics of keeping it at the frontier remain unresolved.

8. Great investors have conviction about the magnitude of a shift and humility about exactly where the value will accrue.

Grateful to Martin for an incredibly thoughtful conversation and to everyone who joined us around the table.

Last week we kicked off Icons Robotics Series with the first guest Lee Redden, pioneer in robotics space, co-founder of Blue River Technology (acq. by John Deere) and now LastFleet. Robotics has been one of Silicon Valley's most heavily funded sectors over the past year, and we're digging deep into where the industry is headed.

A few insights from the dinner:

On finding the problem. Lee started in a Stanford robotics lab building a new robot every month, each one cooler than the last, each one ending up on a shelf collecting dust. The turning point was deciding he wanted to build things that outlive him rather than things that impress his labmates. Lean startup gave him the method: ten customer conversations a week, treating the business as a hypothesis to be falsified before engineering time gets spent. Blue River pivoted from weeding to lettuce thinning on the back of those interviews.

On the bet that worked. The insight was importing internet-scale computer vision into agriculture at a moment when robotics conferences weren’t discussing those datasets or algorithms at all. Within four years, Blue River was thinning roughly 10% of US lettuce.

On being wrong for the right reasons. Mid-company, Moore’s Law flatlined. Lee kept re-speccing and buying the same processor three years running while the entire industry insisted the curve was intact, and the product roadmap had been built on it improving. His framing: judge a decision by whether you ran the right process on the information available, not by the outcome.

On catching the next curve. Deep learning arrived and unlocked a product they couldn’t previously build. They caught it because the unglamorous scaffolding was already in place: hardened cameras, real-time systems, tractor domain knowledge, data collection pipelines. The lesson is that the durable advantage was the boring infrastructure, not the algorithm.

On where robotics actually is. Perception is solved with a good team. Locomotion got solved about four years ago. Manipulation is the open problem, which is where the capital is going, with genuinely state-of-the-art results but unproven commercial fit. Lee’s read is that manipulation landing is the trigger for a wave of company formation, and that it hasn’t landed yet.

On simulation versus human data. A contrarian position: behavior cloning tops out and may produce commercial products in narrow domains, but not generalized ones. Human data is a bootstrap phase. His precedents are face detection trained on synthetic faces, Go self-play, sim miles in autonomous driving, and legged locomotion learned entirely in simulation. He’s also been repeatedly surprised by how crude a simulator can be and still transfer, citing a quadruped that learned to balance on a soft yoga ball modeled as a rigid sphere.

On explainability. The demand for explainability tends to disappear once the learned method outperforms the interpretable one. The more useful distinction is organizational: hand-tuning features requires PhDs and looks like a research project, while collect data, retrain, redeploy looks like an engineered product and is far more reliable to run.

On what deployed products actually look like. State of the art will look like one generalized model. Shipped products will be mostly edge-case hardening, cost engineering, and safety, with the generalized component a small block inside a much larger system. Safety belongs in its own non-negotiable bucket, even when it costs capability. Everything else is a productivity number.

On the real bottleneck. Not intelligence, manufacturing. A hundred thousand deployments requires a hundred thousand robots, and several of the best-funded companies in the category are producing on the order of one robot a day. The gap between perceived deployment and actual deployment is enormous. Lee’s response is to keep the robot cheap and simple enough to actually produce.

On the component that worries him. Motors, and only motors. Everything else has five suppliers and four-day turnaround on custom parts. He thinks motor design is still a generation or two from where it needs to be, and that the right design could take meaningful weight out of a robot while raising its capacity.

On the verticals worth building in. The boring, high-burden human ones: driving, cleaning, food. His argument is that domestic robots may be the killer application for family formation, because cleaning and laundry burden is what parents actually name when asked why they stopped at three kids.

On security. Home robots are a physical attack vector with no precedent. The uncomfortable part is that the dangerous actions are also legitimate commands, so a compromised robot doing something harmful may not look anomalous at all. His view is that the cybersecurity work here is nowhere near where it needs to be.

On advice to founders in the room. Get out of the building. Ten customers a week. Validate before you build. Building in the garage first is fine, but be honest that it’s for fun rather than for a business.

Lee is actively hiring for lastfeet engineering roles - send your CVs here jobs@LastFeet.com.

One of the most fascinating and insightful conversations we’ve had at Icons.

Here are a few takeaways:

• What was refreshing to hear is that Jonathan isn’t the stereotypical Zuckerberg-style founder who succeeded on the first try. His first startup didn’t work out the way he intended. Right after Stanford CS, Jonathan started a company in Silicon Valley and spent seven years building it before reflecting on what went wrong. The answer wasn’t Jonathan. It was the market. He was attached to an idea that simply didn’t have a large enough market. He was stubborn. He believed it could be huge. But that’s not what the market demanded.

• In situations like that, Jonathan suggests being less stubborn. Give yourself the freedom to think differently. Go talk to 100 ICPs and verify whether they actually care about the problem you’re trying to solve. If the answer is yes, go solve it. If the answer is no, pivot away. Not just pivot slightly, but jump away from what you had before - teleport. All your existing collateral can become a curse when you’re trying to find a truly great startup idea.

• But what about insights? Didn’t we learn at Stanford that we should stick with an “insight,” following Andy Rachleff’s Product-Market Fit framework? Jonathan’s view is: challenge your insight. Most insights only exist within a specific time horizon. Imagine having a brilliant insight around automation before 2023. Then ChatGPT arrives. Do you still hold on to that insight? Probably not. Humble yourself. Your insight may no longer be true. Don’t become attached to the dream.

• Okay, you’ve pivoted and your old insight is no longer valid. What’s next? Go all in. Jump into the new thing that excites you most. Don’t underestimate your ability to develop new insights. If you’re smart and curious, you’ll go deep and find them again, but this time inside a market that’s actually growing fast enough to matter.

• When Jonathan started Turing, OpenAI called and asked how many people he could dedicate to expert-skill labeling. He wanted to say an even bigger number because the demand was so overwhelming. The market signal was impossible to ignore. In just a few years, Turing grew to hundreds of millions ARR. Today, it serves many of the leading AI labs and also helps enterprises adopt AI by connecting them with the best solutions available.

• Is the opportunity around data labeling limited? Eventually, yes. But not anytime soon. Jonathan’s view is that we’re still decades away from fully automating the process. At the same time, Turing has built a second business that leverages the latest AI models and innovations to help enterprises deploy AI directly into their operations.

• How would Jonathan screen for startup ideas? He would look for highly fragmented markets with mostly analog competitors. Real estate is one example: fragmented, less technology-driven, and deeply connected to the physical world.

• Another way to think about opportunities is to become an input to AI companies. What will they need to reach the next level? It could be data. It could be infrastructure. It could be something entirely different.

• Jonathan believes founders need to stay several years ahead of competitors. How do you get ahead? Reading books isn’t enough. You need high-variance learning so you don’t get trapped in a local minimum. That means constantly meeting new people, exposing yourself to new ideas, and learning from what others have built, especially in Silicon Valley, where the density of ambitious and talented people remains incredibly high.

Last week, we hosted Sri Viswanath for an Icons dinner.

Sri was previously CTO of Atlassian, later Head of Ventures at Coatue, where he managed investments across a multi-billion-dollar platform.

He recently left to build Sycamore.

When I asked him why he walked away from one of the top venture positions in the industry, his answer was immediate:

“There’s no better time to build. Generational companies are going to be built now.”

Here are a few takeaways from the conversation:

1. There’s probably never been a better time to build than right now.

Software is changing at record speed. Entire assumptions around how products are built, teams are structured, and value is created are being rewritten in real time.

The old software stack is slowly dissolving underneath us while a completely new one is emerging.

2. Lean companies will disproportionately win in the AI era.

Organizations that can rethink workflows from scratch and rebuild around AI-native operations will move dramatically faster than incumbents defending legacy processes.

Historically, changing the rules inside an existing system has often been harder than building a new one entirely.

3. Incumbents may still benefit massively from AI.

While startups move faster, enterprises already have customers, workflows, distribution, and operational inefficiencies worth billions.

Once AI deployment becomes reliable enough, the ROI for large organizations could materialize almost overnight.

4. Sri’s perspective on layoffs and restructuring was surprisingly calm.

His view was that most of today’s turbulence is transitional rather than existential.

AI will eliminate some functions, but it will also create entirely new categories of work, products, and companies.

Especially in Silicon Valley, ambitious and adaptive people tend to regroup quickly. The long-term impact may ultimately become far more productive than destructive.

5. One of the most interesting experiments Sri runs today is a Slack environment where AI agents communicate autonomously with each other.

Not humans prompting models one by one.

Agents coordinating, delegating, responding, and operating together inside shared workflows.

It hints at software evolving from passive tools into active organizational participants.

6. The future org chart may become dramatically smaller than people expect.

A large portion of middle layers inside organizations primarily exist to move information, summarize context, coordinate teams, and route decisions.

AI increasingly attacks exactly those functions.

7. At Sycamore, Sri’s core bet is that enabling AI at scale inside enterprises requires far more than simply plugging into existing models.

The hard part is orchestration, permissions, workflows, memory, governance, and reliability across large organizations.

The winners won’t just be companies building models.

They’ll also be the companies building the operational layer that allows enterprises to deploy AI safely and effectively at scale.

Thanks to Alma and Informed Ventures for supporting Icons!

A few things I will be carrying with me from the dinner:

1/ On knowing when you’re wrong. Higgsfield burned $10M of a $16M seed on consumer before Alex admitted AI doesn’t fix consumer retention. The courage to change course when you’ve already spent the money is rare. He did it.

2/ On picking your hero. Instead of chasing OpenAI, he studied Canva — 20M paying users, $4B ARR, built quietly out of Australia. That choice of role model says a lot. Not every great company looks like the ones on the magazine covers.

3/ On staying close to the work. Alex still knows which model releases shipped this week and what his customers are paying. When the company has grown so fast, you cannot step away. He isn’t. That proximity is the company.

4/ On what actually wins. Six releases a week. Cash-flow breakeven. No marketing budget. “We had to believe the best product wins.” Not the biggest raise. The best product.

5/ On speed as a tool. The way Alex described it, speed wasn’t the strategy. It was the byproduct of a small, deeply aligned team that knew exactly what it was building and why. Velocity as a function of clarity, not chaos.

What stayed with me wasn’t a tactic. It was Alex himself. The clarity he brought to every question, about the wins, the missteps, the parts of the journey most founders dress up, was the most inspiring thing in the room.

Thank you, Alex. And thank you to everyone who came.

A separate thank you to Alma for sponsoring the dinner and to the Central Asian Innovation Hub for providing us the space to make this happen.

The bottleneck in enterprise AI is organizational, not technical: Enterprise buyers are defensive, uninformed, and incentivized to avoid risk rather than adopt tools. This creates structurally slow adoption regardless of how good the technology becomes.

If thinking becomes a commodity, the next bottleneck shifts to the real world: When intelligence and software get cheap, the constraint moves to physical experimentation and infrastructure. AI can’t commoditize the physical world because it requires real-world feedback loops.

First-principles thinking matters more than a polished pitch: When VCs ask about moats or competitors, they’re testing how deeply founders think, not what they answer. Oftentimes, the best response to “Why can’t Google do this?” is “They can.”

The best founders are defensive of their equity: They ask “Why do I need this money?” before “Should I take it?” Raising only makes sense when there are specific offensive levers to pull. Otherwise, founders dilute themselves for optionality.

Deedy, thank you for your thoughtful insights and exciting stories!

Thank you to Silkroad Innovation Hub and Alma for supporting the event.

Chris Larsen, co-founder and Executive Chairman of Ripple, has spent over 30 years building at the frontier of fintech - E-Loan, Prosper, Ripple. Each time, he built where banks didn't want him to exist. Each time, he felt that fear.

When I worked in fintech, we were constantly looking at Ripple. They were the market. Last week, I got to sit down with Chris for dinner with Icons. What I wanted to understand: how do you keep building for three decades when the system is designed to stop you?

A few insights from our conversation:
1. That fear never goes away, and sometimes it's right to be scared.
All three companies terrified him. But he's also been fired, nearly went bankrupt, and spent eight years at Prosper when he should have walked away after two. That "yeah" feeling doesn't mean you won't fail. It just means you're willing to find out.

2. Don't hang on too long.
Chris spent eight years at Prosper. His biggest regret isn't that it didn't work, it's that he didn't walk away sooner.

"Try to fail sooner, don't hang on too long. This is a very short life. You don't have any time to waste."

VCs can always recapitalize a company. You can't recapitalize your time. This one hit harder as we continue to navigate the fast-changing world.

3. Self-awareness matters as much as capability.
After four years as Ripple's CEO, Chris brought in Brad Garlinghouse. That was 11 years ago. The boundaries are ruthlessly clear: any email Chris sends to Ripple employees copies Brad. Nobody can come to him and undo Brad's decisions.

"The CEO has to be seen as the person making decisions."

Knowing what he's good at gave Chris his life back. He's 65, still building, but now he has time for what matters. He swims. He works on climate. He spends time with his kids.

He's good at chaos. Brad's good at scaling. That partnership works because the boundaries are clear.

4. Culture requires relentless repetition.
"Talk about your core values all the time, until you are blue in the face. You think people can't possibly want to hear this again. Nope, they want to hear it all the time."

Chris and Brad talk about core values in every meeting. They do celebrations. People get jackets with the values on them. It sounds corny, but the discipline is rare.

All three of Chris's companies were built where banks didn't want him to exist. The pattern: find where the system is broken, build the alternative, fight the incumbents.

Success wasn't just about the technology or the strategy. It was about having the conviction to build where everyone said you couldn’t.

Thank you to Chris for the time and the honesty he shared with us. Thank you to everyone who joined us.

Democratization of education

Technology has a long track record of democratizing access, and today’s LLM capabilities can provide students with something that used to be scarce: a personalized tutor and an individualized on-demand curriculum. That can widen opportunities by making education cheaper and more broadly accessible. At the same time, classes and universities still matter because they still serve an important social role as places where people meet and learn together.

Contrarian approach to Lean Startup for deep-tech companies

For deep-tech startups, shipping early can backfire: users cannot validate what is not technically viable yet. They will ask for better “horses,” not a car. Until the technology can deliver a real end-to-end experience (as with Waymo), early releases risk non-adoption and misrepresent the true potential.

Obsession with a problem

One of the most frequent mistakes founders make is giving up too early. Sometimes they are actually working on a meaningful problem, but after a few attempts they quit because their initial solution does not work. A better approach is to be obsessed with the problem, not a specific solution or technology. That mindset keeps you adapting, and staying focused on what you are ultimately trying to solve.

No shortage of inventions

Most of the technologies we rely on in our daily lives were invented within the last 150 years, which is a reminder of how early we still are relative to the age of humanity. The vast majority of useful inventions are still ahead of us, so there is no shortage of meaningful things left to build.

We are grateful to Sebastian for taking the time and providing valuable learnings with the Icons community!

Thanks to Silkroad Innovation Hub and CMT Digital for supporting the event.

Here are a few insights from the dinner:

1. Stablecoins are becoming a new form of money.
Stablecoins are no longer a niche crypto instrument. They are becoming embedded in the global financial system as programmable, internet-native money. Visa’s role is not to displace banks, but to act as connective tissue between banks, PSPs, and new on-chain payment rails.

2. Stablecoins are additive to Visa’s business, not disruptive.
Rather than cannibalizing card networks, stablecoins expand Visa’s surface area. They unlock new revenue opportunities in cross-border settlement, treasury management, and B2B payments where instant finality, lower costs, and 24/7 availability matter more than interchange economics.

3. PSP → merchant settlement in stablecoins will scale by orders of magnitude.
Merchant settlement in stablecoins is still early, but the direction is clear. Stablecoins offer faster settlement, fewer intermediaries, improved liquidity control, and global reach without correspondent banking friction. Once merchants experience this, it becomes hard to go back.

4. Crypto cards are early and will scale 100x.
Crypto cards today are where cards themselves once were: functional but underdeveloped. This is one of the few categories where startups can launch globally from day one, onboard users across geographies instantly, and compete with banks on speed, UX, and programmability.

5. Privacy is one of the biggest unsolved problems in crypto payments.
Privacy is not a “nice to have.” It is a prerequisite for enterprise and merchant adoption. Multiple models are emerging: on-chain privacy, off-chain abstraction, selective disclosure, and ZK-based approaches. Which architecture wins is still an open question, but some form of privacy must exist.

6. Banks adopting stablecoins will benefit long-term.
Bank resistance to stablecoins is softening. Allowing customers to instantly deposit stablecoins and use them for spending or transfers is low-hanging fruit. Over time, stablecoins strengthen banks’ product offerings rather than weakening them.

7. On-chain credit and programmable loans are a massive upcoming market.
Fully on-chain credit, programmable lending, and stablecoin-based financial strategies are poised to scale significantly. Every major bank will need a coherent stablecoin and on-chain strategy.

8. The “vibe-coding” market is real and accelerating.
The cost of building financial products is collapsing. AI-assisted development enables many more builders to ship wallets, payments, and settlement products faster than incumbents can respond.
Crypto retains defensibility today due to audits and the high cost of mistakes, but tooling will continue to close the gap.

9. Agentic commerce will shift from speculation to utility in 2026.
In 2025, protocols like x402 saw ~95% of volume tied to memecoins. Cuy expects that in 2026, this changes. Wallets will act as universal API keys, enabling pay-per-use access to tools inside agentic IDEs. He expects major tech companies to deploy an x402 server and push agentic commerce into real-world use.