tech, ai & politics nerd Il got info you think i'd be interested in? let's talk! [email protected] || dono: https://teapawt.net/t.co/VTujrKRyey
Joined February 2024
- Tweets15
- Following232
- Followers4
- Likes0
🚨 Anthropic "do not currently have plans to release [the most powerful model they use internally] externally"
Oof. Might have to U-turn on this after Astra drops lmao
🚨 OpenAI have launched a waitlist for 5.6 Sol Ultrafast, a version of Sol that streams at up to 750tps - 14x faster than standard
xAI have made an incredible comeback
From the days of Grok 4 -> 4.3, where they were trailing the frontier by far, they're now arguably the 3rd best lab in the world - behind only Anthropic and OpenAI
Holy shit lmao. Grok 4.6 benchmarks are out, and it's near SoTA on AA's Intelligence Index (61, matching Sol, 1 point behind Fable 5 and 2 points behind Opus 5), SoTA on GDPVal-AA v2, SoTA on AA-Briefcase and Harvey LAB (Vals), and 2nd only to Fable 5 on CursorBench v3.2,Frontier
Terminal-Bench 3.0 has been released! Claude Opus 5 Max takes the crown from GPT-5.6 Sol, achieving a SoTA score of 43.5% to Sol's 34.6%, with Fable 5 in 3rd within Sol's margin of error (34.1%).
this is some great work, and i think the actual headline here is bullet 2: "a model that has seen zero robot trajectories gets better at robot tasks as you feed it more human video"
so the embodiment gap is more of an adaptation than a knowledge one (and thus easier to solve)
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
🧵
Can confirm - after concluding Astra meets the threshold for "Critical" on their Preparedness Framework's cyber category, the model's release has been indefinitely postponed for further safety work in cooperation with the US Govt
Seedance 2.5 is a lot of fun to play around with. Asking for a "very realistic iPhone video" seems to get you some especially cool outputs like the one below -
So... Seedance 2.5 is pretty good. This was a one-sentence prompt
Inspiration for the scenario goes to A24's Civil War - the Seedance classifiers are pretty aggressive and refused to take any prompts specifying an A24 style though
