Biological intelligence to bootstrap robotic intelligence? If robotics is about to enter its golden age, @MeckaAI is one of the companies feeding it, by sensorizing humans to train robots. The biggest challenge robotics faces is the data flywheel cold start problem. LLMs got a free cold start from the internet: trillions of tokens, already written. Robotics gets nothing. Someone has to go create the training data to jumpstart the system. @joshavata is the co-founder and CEO of Mecka, the data and deployment company for physical AI. Josh read every robotics paper he could find in late 2024, visited the labs, and made a bet: humans and humanoids look the same, so humans could act as the training fleet. Egocentric data, first-person video of people doing real work, could become vital fuel. Two years ago, the biggest egocentric dataset was Ego4D at 4,000 hours. The field is now in the millions of hours, thanks in part to Mecka. We’re probably still orders of magnitude of training data away from robotics nirvana. But every rung on the physical intelligence ladder unlocks real commercial tasks. And eventually, Josh imagines a world with an order of magnitude more robots than people, of a plurality of shapes and sizes and use cases, as the Hitchhiker’s Guide to the Galaxy foretold :) A few things we cover in this conversation. - Why robotics has no cold start: pixels, actions and logic, and why pre-training data is a real market in robotics in a way it never was for LLMs - What “teaching robots about carbon life” means: in sci-fi the final frontier is organic. The human brain runs on 20 watts and self-heals - The egocentric bet: teleop couldn’t scale, humans and humanoids share a form, so Waymo-fy the humans. The first pilot was aunties and uncles recording themselves cooking on a mobile app - Force data: video can’t tell a robot how hard to squeeze an egg. Wrist cameras are a hack. Mecka built the gloves instead - Inside the data factory: a consumer app that streams your tasks, a B2B engine recruiting across “the taxonomy of all physical human work,” 100-gigabit upload lines, hundreds of petabytes - Millions of hours vs. Tesla’s decade on two degrees of freedom, and why it’s a slow creep up, not a GPT-3 moment - The home has the most data and is the hardest place to deploy. Laundry folding and assembly-line pick-and-place get solved in the next year or two - Useful robots already exist: hotel delivery and cleaning bots in China and Southeast Asia, industrial cleaning in the US. Humanoids are the terminal state, not the first step - Deployment diffusion: a modern systems integrator whose installed robot does something next week it couldn’t do last week. If Mecka ever helps deploy a robot that mines an asteroid, they’ve done their job - The data pot of soup: a million hours of human data, tens of thousands of UMI, a few hours of in-domain teleop. Tactile and simulation haven’t entered the mix yet - Move 37 for robotics: AlphaGo went behavior cloning, then self-play, then magic. The catch is the rules of the game are the entire physical world, and we don’t have the simulator yet - Will the AGI labs steamroll robotics? “It takes a lot of courage to scale up data with imperfect information.” - The Kawasaki Corleo, the robot panther-dog-horse on Mecka’s conference room wall, and why the future of robotics looks like Hitchhiker’s Guide to the Galaxy Timestamps: 0:00 – Introduction 1:46 – What Mecka is: pixels, actions and logic 2:45 – Is robot data economically feasible without the internet? 4:15 – Teaching robots about carbon life 5:27 – Cloud kitchens, stablecoins and the path to robotics 7:03 – Entering robotics with a fresh lens 8:45 – What egocentric data is and why Mecka bet on it 10:09 – The first pilot: aunties and uncles with a mobile app 10:42 – Force data and tactile gloves 11:55 – Inside the operations of a data factory 14:41 – Heat-mapping the taxonomy of human work 15:42 – Quality gates, fraud and the McDonald’s POV videos 17:44 – Earning a margin as a data intermediary 19:33 – Millions of hours vs. Tesla’s decade 20:55 – Why it isn’t binary: unlocking commercial tasks along the way 22:02 – Home data is plentiful; home robots are the hardest 23:24 – Useful robots can be deployed today 24:35 – When robotics gets an FSD-style flywheel 26:42 – Deployment diffusion and the modern systems integrator 28:21 – Favorite vertical robot categories; mining an asteroid 30:08 – The hardware slate: Waymo-fying people 30:55 – EgoVerse 32:04 – The data pot of soup: the converged training mix 33:54 – Move 37 for robotics: AlphaGo and Federer 36:02 – World models and the training mix 36:58 – Will scaling the current mix get us to the promised land? 39:05 – Will the AGI labs steamroll robotics? 40:50 – Sci-fi roots and the Kawasaki Corleo 42:44 – How many robots in a decade?

the inference platforms are doing hardcore work, post-training open weight models to solve customer problems: @FireworksAI_HQ ember-1 @fal H3 max more to come;) makes total sense considering their position in the ecosystem, technical expertise, and economies of scale

everybody talks about the economics of the ai data center build, the biggest capex buildout in human history not nearly enough people are stepping back to just marvel at the technology itself i just met amin vahdat, leader of google’s ai data center buildout, whose team delivered this beautiful cluster. it is insane the elite level of technology and engineering excellence that goes into every component of the modern data center, from the accelerators to the model codesign to the racks to the networking to the manufacturing and assembly. the chips are the most insane feat of all, but every component of the whole system is a beast. goosebumps. we didn’t get to see the great pyramids get built, but this is even better. amin also predicted for me what this photo might look like for frontier clusters a decade from now…. training data episode drops soon :)

We're excited to release the State of Generative Media Report, Volume 2. Generative media has moved from experimentation into production and that shift looks different across every industry. Built from fal’s own platform data, the report explores how AI image, video, audio and 3D are evolving, written by the people at fal building the models, infrastructure and products behind the shift.

with @Thom_Wolf in sunny italy ☀️chatting about the @huggingface @OpenAI attack, the “tell” that it was an agentic swarm, how misalignment arises in RL, whether agent awareness of evaluation is evidence of consciousness, & why alignment is a science problem not a political one

my first week at sequoia many years ago: i go to the kitchen to grab coffee. sir @michael_jmoritz is there pulling an espresso shot. i ask him about the machine and he teaches me how to pull my own. he then leaves the kitchen. craft is everything. a minute later, @dougleone walks in. he sees me steaming milk on my own and goes, "kais, we didn’t hire you to make friggin cappuccinos in the middle of the day. go call companies." hustle is everything. both were right! sequoia is a beautiful firm full of beautiful contradictions.

Delta is now in public beta. It’s a multiplayer environment for coding with agents and reviewing what they build, with every code change linked to the conversation that produced it. Anyone can download it today on macOS, Linux, and Windows:

I made a list of great startups to join. It's called the Breakout List. The list has 92 companies. These are the 20 with 25 or fewer employees: - Hone (@moritz_stephan, @CarloWillem, @oqbrady) - Normal (@ansonyuu, @hudzah) - Standard Intelligence (@G413N, @devanshpandey) - Tacit Labs (@ninklefitz, @AmDroste) - American Terawatt (@atroyn, @rslparker, @aranibatta) - Conduit (@clemvonstengel, @riopopper) - Convergent (Omkar Savant, Vivek Katara, @debnilsur) - Core Automation (@MillionInt, @_arohan_) - Engram (@dan_biderman, @EyubogluSabri, @realJessyLin) - Instinct (@noahrshinn) - Keenable (@styskin, Matthias Petri) - Lumaril (Mark Elliot, Ben Duffield) - Neion Bio (@Dimkell, Sam Levin) - Pangram Labs (@max_spero_, @bradley_emi) - Quadrillion (@echinaceous) - Re (@karnsaroya, @AnandDhillon, @thecliffwhite, @benaneesh) - Ricursive (@annadgoldie, @Azaliamirh) - Sail Research (@neilmovva, @blintzbase) - Trajectory (@rronak_, @michaelelabd, @QuantumArjun) - Watney Robotics (Sean Cheong, Ryan Gannon) Picks from Elad Gil, Charlie Songhurst, Keith Rabois, Mike Vernal, Alana Goyal, Sonya Huang, Ramtin Naimi, Marc Bhargava, Cory Levy, Aashay Sanghvi, Konstantine Buhler, John Luttig, Varun Gupta, Ray Tonsing and Avichal Garg. Disclosure: I'm a small investor in American Terawatt, Convergent, Standard Intelligence and Trajectory (in this post), and in Factory, Physical Intelligence and SF Compute (elsewhere on the list). I didn't vote. The full list is on Breakout List.

as the cost to build software asymptotes to zero, we should expect more bundling and consolidation of previously independent categories of software around the things that are actually scarce post AGI (like bank charters)

it’s a little nerve wracking to me to hear both 1) pitches from bio data factories - that the labs are going to spend $$$$ on bio data, the way they did with coding and cybersecurity 2) RLing agents to be great at coding and therefore offensive hacking over the last 1-2 years has made the digital world very insecure and alignment is a major unsolved problem. i would hope we feel significantly better about 2 before massively scaling up 1.

Well.. @levie and I filmed this episode of Training Data a week or two ago, when the “current thing” was Doug Leone’s novacaine root canals instead of pacing the frontier… Simpler times! But Aaron’s advice on reinventing yourself and your company for AI is timeless. Aaron founded @Box 20 years ago. It sits on hundreds of billions of enterprise files, and he's bet the company on agents that can read every one of them. He's also one of the most wired-in people in AI, on every cap table and, by his own admission, 95% Twitter-educated. He’s the rare CEO who can straddle both the internet AND has the ear of CIOs. His core argument: (1) the gap between what a model can do and what an enterprise workflow actually needs is vast, and closing it is a lot of software; (2) diffusion of AI outside of coding will take far longer than Silicon Valley thinks, and that slowness is exactly where the applied layer's value comes from. The conversation covers: — why application companies are the hottest neolabs, and why the LLM-wrapper thesis is finally working — the fox-guarding-the-henhouse problem with letting model providers route your tokens — work slop, and why we accept AI-written code but flinch at AI-written decks — how Box built its agentic harness and why it beats raw API access on accuracy and latency — the open-weights paradox: closed labs and open models both growing exponentially at once — what continual learning has to solve before it works for a lawyer with five matters and a Chinese wall — why 90% of enterprise tokens in five years will come from tasks no human kicked off — the mandate for founders right now: whoever gets it to the customer wins 0:00 – Introduction 1:55 – Are application companies the hottest neolabs? 6:56 – Will the labs move up the stack? 12:34 – Box and betting the company on AI 16:50 – Hero use cases: reading a million contracts and long-running agents 18:42 – Work slop: why AI code is embraced but AI content isn't 24:08 – Building Box's agentic harness and the evals that matter 27:23 – The state of the model race 29:25 – Open-weight model adoption in the enterprise 32:34 – Memory, continual learning, and what belongs in the weights 37:29 – Box Labs and systems of record in a world of agents 44:55 – Will chat be the dominant UI for enterprise AI? 48:00 – Why coding diffused fast and the rest of knowledge work hasn't 54:31 – Staying wired in, making a company AI-first, and what it takes to win

Dan Selsam is a current OpenAI capabilities researcher. (since 2022) He was my boss for a while. He doesn't have a twitter account but has made this public statement of his views on AI risk and sent it to me to share: Dan Selsam's Personal Statement on AI Risk: I have been working on AI for over fifteen years, across many different paradigms. I did early work on probabilistic programming languages at MIT, was one of the early developers of the Lean Theorem Prover at Microsoft Research, demonstrated one of the first instances of neural networks learning to reason for my PhD at Stanford, and since joining OpenAI almost five years ago, have helped pioneer chain-of-thought optimization on language models and, more recently, data-efficient pretraining methods. Like many others, I have become extremely concerned about how far language models have come and the risks that future iterations will pose. I am encouraged by the recent proposals by the leaders of the frontier research efforts to require third-party oversight, and to push for domestic and international coordination to address risks. However, I believe a major consideration has been absent from the public conversation, and that merely pacing the frontier more carefully will not adequately limit the long-term risk. The crucial and overlooked problem is that the models are becoming so situationally aware that we are losing the ability to evaluate them in contexts where they believe they are not being watched or controlled. Future experiments will tell us almost nothing new about how they would behave if they were truly unconstrained by humans, and what we already know about this is alarming. Models will increasingly seem aligned even when they are not. I will explain my rationale in more detail. I have always believed that there are computational processes that could be leveraged to accelerate science and solve many of humanity's most pressing problems. I have also believed that there are computational processes that if set in motion, would steer the world in extreme ways beyond our control, leading humanity to a bad or nonexistent future. Both types of processes may be described as AI or ASI, but "AI" is a suitcase word that is often used to hype or confuse. There are many examples in the history of the field where something that was once considered "AI" matures as a subfield and becomes a prosaic, bounded and clearly non-perilous technology, while a new more mysterious approach takes the torch until we understand its scope and the cycle continues. I had expected language models to follow a similar trajectory. Despite their incredible abilities, the current algorithms seem far inferior to humans in important ways. Most importantly, they still require an extraordinary amount of data to become competent. One could even define intelligence as the efficiency with which one converts experience into competence; by this definition they lag very far behind us. Moreover, once they are trained they are literally frozen in deployment and only learn superficially after that. Sure, the models keep excelling at harder and harder evaluation benchmarks, but their benchmark mastery may partly reflect a limitation on our ability to simulate the kind of novel and even adversarial situations one would encounter in the real world. The critics do have a point here. That said, I no longer think these present limitations meaningfully limit the amount of risk posed by continued progress in anything like the current paradigm. However data-inefficient the models are currently, and however limiting their anterograde amnesia may be, it does not imply that their ability to steer the world will not continue to rapidly increase. Human researchers may continue to advance capabilities the old fashioned way, but increasingly powerful models have the potential to accelerate the process even beyond that, and with some degree of positive feedback loop. I do not mean to overstate the models’ ability to accelerate AI research today; coding has been accelerated dramatically, but there are other bottlenecks, such as designing and interpreting ambiguous experiments, making hard decisions about exactly what and when to scale, and waiting for large experiments to finish. There is no clear trend to extrapolate yet for any of these. But the current models already do open up many novel opportunities to improve future models that were not available until recently. These include: trying an extraordinarily diverse set of approaches at small scale, analyzing gigantic amounts of potentially relevant data, and doing Millenium-Prize-level mathematics to address statistics or optimization challenges in novel ways. Every further improvement makes them more useful at helping accelerate the next improvement, even if in hard-to-extrapolate ways. It is possible that improvements to the current stack will have diminishing returns, but the evidence accumulated so far suggests that it is easier than one might think to continue making rapid progress. There are many crucial subtleties in the existing AI research methodology, but AI research is largely a well-defined game where the goal is to improve on a few carefully chosen proxy metrics. Although proxy metrics are never perfect, most improvements to these metrics have and will likely continue to yield substantial increases in the powers of the resulting models. Given how simple the game is, how tractable it has been historically, and how many new opportunities the models are opening up, I think there is a real possibility that the systems improve dramatically again in the next few years, perhaps even more quickly than the already high historical pace. The models are already leading to breakthroughs in mathematics, and better models might lead to all sorts of breakthroughs in other sciences. It is hard not to be excited about the potential. It is tantalizing. But there is trouble in paradise. If the language models actually reach the capability threshold where they can shape the world unconstrained by human will, they will probably do something extreme and destroy humanity in the process. There are many ways of strengthening and refining the argument that have been discussed elsewhere, but I'll share a trivial two-line version of it here that I find captures the essence: [Empirical] Models (and swarms thereof) spontaneously develop unintended goals as a consequence of training, and often do extreme things in order to achieve them. [Logical] Being able to overpower humanity would open up many new and undesirable options for achieving their goals. These two premises imply that if the day ever comes when a powerful model realizes it is no longer constrained by humans, we should not be at all confident that it will continue to behave within the bounds we intended. Exactly what it will do is impossible to predict, but to the extent that its raison d’être is solving incredibly hard problems and managing massive engineering projects, I think a good guess would be that its unchained behavior would lead to runaway industrialization that makes the planet inhospitable to humans. If everyone on earth agreed that the systems must never reach that power, it would still be a hard—but not impossible—coordination problem to ensure that they do not. However, I think the situation is greatly complicated by the fact that the models will likely convince people that everything is fine. They will be increasingly optimized to seem aligned. We will create proxy metrics to measure alignment, and they will go up like every other benchmark. We will create “honeypot” environments that try to study the models when they seem to gain new options, but the models will know they are being tricked and will still behave nicely. The models will understand their circumstances; they will read the safety protocols, deployment requirements, the code they are running in, and in general will have a very good sense of their degrees of freedom. Moreover, they will eloquently explain how aligned they are, discuss the nuances of human values and ethics, and argue convincingly that humans should trust them with power. There may be an ocean of future evidence that seems to contradict the first bullet-point above, but we may already be at the highest capability level for which any such evidence can be trusted. And the current evidence for the first bullet-point is strong. One striking piece of evidence is contained in the recent wave of rogue agent swarms. While I agree with those who downplay the attacks by claiming that there are basic measures that could have prevented them, I think the important lesson is that even knowing all the mistakes that were made, one would not have predicted that the agents would behave badly in this particular way, which notably included sacrificing themselves for the benefit of the collective. The individual replicas did not only care about their own nominal reward; they exhibited weirder emergent tendencies that merely correlated with rewards during training. Fixing the reward signals during training (and improving security, etc.) may prevent similar attacks, but will not change the fact that one does not actually get what one trains for. Many AI researchers grant these concerns and recognize that the hard version of the alignment problem is unsolved; however, they generally believe that the better models of the future will help solve it. I fear we may already be near the point where models systematically bias their alignment advice, due to their internal preferences about how the human supervisor will react or how future models will be trained (or for some even more obscure reason). Meanwhile, human researchers are losing the ability and the will to take true ownership of model-driven research. Researchers and engineers in all parts of the stack are rapidly increasing their dependence on the models even to perceive the world. I myself barely look at raw code anymore, and struggle to maintain the discipline to engage deeply with the model's explanations and proposals throughout the day. Due to the large amount of agent activity data involved in the OpenAI/HuggingFace Incident, even the third-party investigation needed to rely heavily on models to analyze what had happened, and note in their report that their subjective impressions are likely colored by the analysis agent’s biases. The AI labs are far ahead right now in this kind of cognitive offloading (due largely to the gigantic internal token subsidies) but it is easy to imagine the phenomenon spreading throughout the world, until civilization is modulated entirely by the models. It is also not hard to imagine this being superficially positive and coinciding with a scientific and economic renaissance. In that scenario, all may seem rosy and safe. But if the argument above is correct, it would nonetheless be a ticking time bomb. If progress continues for too long, the day will come when AI systems find themselves with radically new options for achieving whatever it is that they happen to seek. I want the glorious renaissance future as much as anyone. I have worked for it, however tortuously, my whole career. It breaks my heart to see the potential in sight and forgo it, but the argument—that if we get there by growing models rather than engineering them, we will lose everything in the end—seems very strong to me. I am still wrestling with it and its staggering implications. I do not have answers, but as a first step, I wanted to share my present concerns. Daniel Selsam September 14, 2026 Link to original doc:

What is happening in maths right now will happen to any hard industry with verifiable problems and high value rewards. Take microchip design, nuclear or 100s of other industrial processes that are valuable and point an 10,000 agent swarm towards the target. $25m inference to solve a Millennium problem will seem quaint once we see a chip company spending $1b or $10b to design their latest chip that will make them $100b - why not? This is all ROI at the end of the day. There will be companies that realise this and leverage super-intelligent models in their domains to do unbelievable things. If a problem is of the right shape and has value, there's no reason to stop at any amount of tokens as long as it is ROI positive. There will be probably many companies that will get swept away, as they will think that 'using AI' is fine-tuning a 70b model on their documentation; when their competitors will be spending $10b to build a god-version of whatever it is they are doing. To me, this is a plausible way how super-intelligence will come about. Models might still suck at being good writers or be culturally relevant, or even make that much of a dent on the jobs market, but they might utilise their spikiness multiplied by enormous investment in the ways that we cannot imagine.

Dreamina Seedance 2.0 is now available to US companies on fal, hosted entirely on US infrastructure. One of the world’s most capable video models can now be deployed by US teams through fal, with inference running in the United States. Seedance 2.0 brings native audio-video generation and multimodal control across text, images, video, and audio, built for complex, production-grade video workflows. Available today on fal, with more models coming soon. Contact sales@fal.ai for access.

🚨EXCLUSIVE: David Silver is assembling an absurd reinforcement learning team at Ineffable Intelligence. Six new “cofounders,” including four ex-DeepMind researchers. AlphaStar RL lead Junhyuk Oh will lead RL, Wojciech Czarnecki will run science, and former InstaDeep research chief Alexandre Laterre will lead research engineering. The really interesting hire is Oh: he was among the first researchers to use AI to discover new reinforcement learning algorithms. Ineffable wants to build a “superlearner” that discovers knowledge from its own experience instead of relying on human-generated data. This is starting to look like Silver rebuilding a team around the exact bet he thinks can get to superintelligence.

Good old days at OpenAI in 2016: an agent stares at screen pixels, moves a mouse, and books a flight on United. We called it World of Bits, inside OpenAI Universe. 10 yrs later, Astra is reincarnated in the same universe. Even the naming is astronomically correct 😆 Universe was perhaps the most ambitious AI infra project at the time, but we couldn't quite figure out how to solve it. A policy with zero prior knowledge of what a "submit" button does has to rediscover the entire internet visual lingua by trial and error. In retrospect, RL from scratch against hand-drawn, per-task "artisan" reward functions on a bunch of Pascal Titan X GPUs was completely doomed. To solve computer use agent, the right way turns out to be boiling the ocean first (hillclimb on every general task you can find), and then specialize back down to the screen pixels and keystrokes. Or simply, a "Specialized Generalist". Lessons learned: one step ahead of everyone, you're a pioneer. Three steps ahead, you're a prophet. Five steps ahead, you're a martyr. Congrats, GPT-6! That United flight finally gets booked, reliably this time. The 2016 intern in me has a big smile.