@yb_effect

General Intuition's model makes so much sense when you look at @PimDeWitte's career. > started in gaming with a RuneScape private server > worked on satellite data labeling with Google > happened to sit next to the 2016 DeepMind team > started a video game clipping product called Medal > collected the best video game data over the last decade > remembered that Demis used games & simulations to create general agents > now locked in on creating a frontier model for spatial and temporal reasoning with Medal data

https://t.co/tkEWsuNGKp

I watched @DrJimFan talks on robotics at Sequoia Ascent from last year and the one he gave a month ago. The key insight is that video and world models are getting much better at predicting the next world state. This is putting robotics on the same trajectory as LLM development (hence great parallel). Up until recently, robotics was bottlenecked by VLAs (vision-language-action) models and teleoperation. - VLAs are heavily dependent on telling robots what the scene is through words and images. They’re too focused on nouns and environment descriptions. A good starting point but robots need to understand cause and effect…verbs and physics! - Teleoperation is a ridiculously cumbersome process to collect data. It requires people to control robots (i.e. through VR headsets) from a distance as if they’re playing a very slow video game But! Now we’re seeing developments in World Action Models (WAMs) which allow robots to take action in thousands of simulated environments. If it moves this cup at this angle with this much force, what happens to the world state right after? Robots are learning so much from these models that they’re able to one-shot and generalize new tasks much faster. Additionally, there’s been a ton of new data collection through egocentric videos which are first person data collection from people using wearables. These videos, along with simulated world scans, are becoming the fuel for robotics pre-training. Effectively what the open web common crawl gave LLMs. To bring down teleoperation needs, there’s also been a shift to motion capture (MoCap) gloves to help with action fine tuning. This lets humans wear gloves and help the robots understand more dexterous parts of the task. Jim’s bigger point: more compute results in more simulated environments that robots can train in resulting in a huge increase of training data.

If someone asked me to create a framework on the tech space with what I know right now: 1. Agentic infrastructure: what does a web for the agents look like? Search, payments, privacy, hosting, context tooling, and identity. This includes companies such as Cloudflare, Nous Research, and OpenRouter but also companies such as Tempo, Stripe, Parallel, and Exa. Of course, Anthropic and OpenAI operate here also but will keep them in the frontier AI category for now. 2. Software 3.0: using frontier model capability to change the way people work and live. The new app store with LLMs as the platform change. Of course includes companies such as Cognition, Cursor, and Harvey but also Software 2.0 companies such as Salesforce, ServiceNow, Figma, and Databricks embracing the transition. 3. Frontier AI research: this one is obvious. Those working to improve model capabilities at top labs. I also include distributed and decentralized AI companies such as Prime Intellect and Pluralis in this bucket. Also companies such as Harmonic focusing on niche, advanced problem solving use cases. 4. Chip & Data center design: Think MatX, Etched, Nvidia, Groq, the CPU stocks, and so on. Companies actively trying to improve the energy-token efficiency by attacking a layer of the hardware stack. Additionally, I also include the more quieter research focused on non-transformer based AI in here also. 5. Onchain Finance: the institutional and real world asset side of crypto. Think Hyperliquid, Coinbase, Circle, Stripe, Robinhood, and the long tail of mid-stage stablecoin companies and neobanks. 6. Prediction markets & Hypertokenization: “the rebellious child of tech”. These are the companies that will be forever debated as gambling. Think trading trend coins, prediction market betting, sports markets, and so on. The hypertokenization movement and market for everything thesis. Kalshi, Polymarket, Pump Fun, and even traditional companies like DraftKings. 7. Vertical Robotics: in the traditional sense. Robots built for specific use cases and verticals through a laborious training data collection process. Lots of industrial and factory use cases. 8. Physical AI & World Models: the “new robotics” that many people are catching on to recently. General purpose training. Robotics post 2023 with AI as the main driver of teaching robots how to act even through times of uncertainty. Obvious companies here are Physical Intelligence and Skild. Also putting World Labs type companies here because it most applicable for robotics as of right now (gaming as well I guess). 9. Rare Earth Minerals & Energy: building world class mines, people focusing on copper, Nuclear (Antares / Aalo Atomics) mircoreactors, Base Power type companies, and so on. My knowledge here is limited but obviously extremely important and the biggest bottleneck to pretty much all the other categories. 10. Transportation & Defense: drone companies such as Fleet and Aloft. Also new plane companies such as AstroMechanica and Boom that are in the middle of commercialization of fast planes but working on defense as well. And then of course you have the Anduril and related companies locked in on true defense. 11. BCI, Healthspan, & Longevity: Obviously NeuraLink. But also NewLimit, Isomorphic, Altos, Acquired Coefficient, Retro, and the many startups that I have no idea even exist as of right now. 12. Deep sea mining & Space: on either extreme, important stuff happening. In space, obviously data centers and other production processes more efficient up there. All I know is Varda and SpaceX right now. And with deep sea mining, helping solve the energy crisis through more serious extraction capabilities. I literally know nothing about this other than the fact that venture investment is heating up here.

just read the Semianalysis Unitree post this morning and now I'm seeing G1s walking around MSG screaming "Brunson egg and cheese" and "fuck Trae Young" https://t.co/tjSOXxNrlp