Another incredible launch by the Dyna team! 🚢

Today we are introducing Dyna-2, a world-action model pre-trained on one million hours of human video. At this scale, for the first time, we discovered several new scaling laws: • world-action models exhibit scaling law on human data across four orders of magnitude, from 1000 to 1,000,000 hours, • this human data scaling law implied a scaling law on never seen robot data, • both data and objective matter; world modeling and scaling on video data are essential for cross-embodiment scaling transfer to emerge 🧵

Great work to the team across this effort @itsghvr @PranavKuppili @camrzadki @Kaiwen66lc Zac Gong @jasontheutopian

Evals are the backbone of understanding and improving model capabilities The WetLab Benchmark will be updated as capabilities progress and we unlock a scale up of scientific work in the real world Explore the deep dive and all runs here: mecka.ai/blog/wetlab

Astra exceeds Opus 5.5 on science lab tasks, but it burned ~4x the tokens It had the highest success rate and shortest median attempt time, but the highest mean estimated API cost: $8.64 per attempt Opus 5.5 nearly matched its success rate at $2.39 per attempt

Hallucination on task completion is still a challenge Models claimed 192 attempts as complete despite human evaluators finding 79 partial completions and 10 failures Astra hallucinated the least, whereas Opus 5.5 had to be corrected by humans on over half the completions

From successes, models were able to do a mix sequence, place a test tube, and pick a top plate from a stack of 24-well plates Human evaluators recorded 110 successes, 246 partials and 184 failures across 540 attempts Astra led with 25.6% and Opus 5.5 closely followed at 23.3%

Accidents like a knocked over beaker didn’t end the run Opus 5.5 stood it back up and kept going. In another run, it correctly pulled a stirring rod from a fallen rack. Neither attempt fully succeeded, but both showed signs of mistake recovery and emergent behaviors

Lab work demands delicate and precise movements but Fable 5.1 crushed one of the beakers despite instructions to be gentle Other failure modes showed missed grasps due to shadows, grasps from inside the beaker which lead to cross contamination, and bad planning for the tasks

Are robots ready to accelerate science labs? We created the WetLabs Benchmark to test 9 critical tasks, from easy to hard, and let 3 leading models take 20 attempts per task Astra leads but it comes close...

personal update: i've joined @MeckaAI to work on robotics, RLEs, data, and deployment for frontier research It’s bittersweet. Thank you to Nikita, Matt, Ruslan, Pratyush, and everyone who built BWL with us through the ups and downs. To Thomas, Mark, Latif, Clay, Ali, and the rest of our investors: thank you for backing us and being in the trenches with us. At Mecka, I’ll be working on robotics, RL environments, data, and robotic deployments for frontier research in physical AI. If life is just hillclimbing and verifiers, I can’t think of a more ambitious hill to climb than getting AI to do useful work in the physical world, from manufacturing to science. Super excited to work with @joshavata and @jasontheutopian , two young Canadian builders who are among the hardest-working, most humble founders I’ve met. We’re growing fast and hiring across the company in SF, New York, and Toronto. Come attack the bottlenecks with us.

we have a few extra seats for a robotics poker game in SF tonight! DM

hosting a private robotics dinner with @jasontheutopian of @MeckaAI in sf on weds 9/23 if you are curious about building at the frontier of data and deployment for physical ai, dm us! 🤖🧤

come get the alpha leak

If you are obsessive, technical and have maniacal urgency, then consider applying to join the team at Mecka

Spotted @MeckaAI on the @BusinessInsider Robotics list Thanks @ryajetha for writing about us!

Huge congrats to the @JasonMa2020 and the Dyna team on a monumental milestone!