Today we’re introducing UniversalAGI. Airplanes take 20 years to build. Cars take 5. Drones take 1. None of that is required by physics. It’s required by how we test. Engineers can’t test every design, so they test almost none. Simulation was supposed to fix this. It didn’t. Complex setups, hours-long runs, and compute costs mean most of the design space never gets looked at. The best machine you could build is usually the one you never had time to try. We’re building the physics AI that ends that. Our first architecture is LIFT, the Latent Interaction Field Transformer. It predicts industry-grade, 4D high-fidelity surface and volume fields across physical domains. The conditions on an object and the flow around it are learned together in a shared latent space, not separately. It’s faster and more accurate. In a controlled comparison, LIFT delivered 6.27× the inference throughput of AB-UPT and 1.72× that of Transolver’s separate surface-and-volume models. In one aircraft design search, LIFT evaluated 5,800 candidates in 19 minutes. CAD upload to optimized design. Every candidate screened before a single high-fidelity run or physical test. Customers loved using LIFT in their design workflows, rockets, ships, drones, cars, oil & gas reservoirs, bioreactors, batteries and turbo-machinary. We’re building foundation models for the physical world. Aerospace, defense, energy, manufacturing, maritime, automotive. The time to build something new should be set by ambition, not compute. Architecture and results: universalagi.com/research/li…

(part 2) I asked people which neolabs under $10B have the highest talent density: Full list: frontierlist.com/neolabs 1 vote each: - Biography - Chai Discovery (@joshim5, @jackdent, @Mattmcpartlon1, @j_boitreaud) - CTRL-G (@GregoireGambatt, @pacovilletard) - Emulate (@jparkerholder, @MattMcGill_, @philipjohnball) - Engram (@dan_biderman, @EyubogluSabri, @RealJessyLin, @jxmnop, @scott_linderman, Chris Ré) - Factory (@matanSF, @EnoReyes) - Fulcrum - General Intuition (@PimDeWitte, @EloiAlonso1, @AdamJelley2, @micheli_vincent, Iggy Harmsen) - Isomorphic Labs (@demishassabis) - Olo Din (@mishayagudin) - Pantheon (@calixo888, @Jannikschg) - Sakana AI (@hardmaru, @YesThisIsLion, @renitomail) - Simile (@joon_s_pk, @msbernst, @percyliang) - Stealth - Stealth text-diffusion lab (ex-GDM) - TypeSafe (@CompleteSkeptic, @EGafni, @hackgoofer) - UniversalAGI (@aha_ml) - Wafer (@gpuemi, @gpusteve)

As a founder, your number one job is to eat shit. You're on the call with the upset customer. You need to fire people by telling them directly. You're listening to the engineer who wants to quit due to burnout. You have to read the 40-page vendor contract, which is a nightmare to go through. Every hard, uncomfortable thing lands on your desk. That's the whole job. You're here to absorb the bad stuff so everyone else can keep moving.

no one of us gives a fuck who's right and who's wrong just please continue responding to each other so that we have more memes material

Combining the benefits of RL and SFT with on-policy distillation, a promising approach for training small models for domain performance and continual learning.

Exciting progress on Middle East. Go @jaredkushner archive.ph/8Fxsm