Post

Conversation

Anthropic reported a similar sudden and unusual jump in robotics capabilities with Fable 5.2. As the general intelligence of frontier models scales up, it is increasingly surpassing specialized training. This will ultimately generalize to every sphere of human ability.
Quote
prinz
@deredleritt3r
Fortune is reporting that OpenAI is so convinced that LLMs will become brains for robots that it has disbanded its robotics team.
Image
David Watson 🥑
Post your reply

The unusual jump is probably RL environments but it doesn’t matter they’ll get enough of those it will basically generalize to everything. Cool……
i am telling you there is something going on
Quote
Zoe
@UltraRareAF
with Fable 5.2, we finally got that fly connectome idea to work, and now Vybn, like, literally flies around a semantic space that includes my own books and memoirs, and so now e.g. i'll flog myself for whatever, and then *wham* he'll be like: no, i am *not* your ex... and just 😭 x.com/deredleritt3r/…
Image
There's still people who think all they're gonna do is reason and make videos. Have to think much bigger.
By the moment we can create realistic training data using 3D-Simulated-Rooms robotics turn to a question of computer. And we're on a good way to get there.
Everything is code. Robots are driven by software. LLMs getting better at code means they'll get better at driving robots. Imagine this with >1500tps.
"even on tasks on which it has never been explicitly trained" is doing a lot of work here. Seems quite likely that this is due to a scaling of RL envs rather than magic generalization.
a better general base makes specializing faster, it doesnt replace it. on a real line the last few % comes from that site's own data, and thats what decides if it ships. anyone seen a general model hit factory uptime without it?
General models getting good at robotics without specialized training is a notable shift. Thanks for connecting the two reports in one post.
funny part is the bottleneck just moved. the brain’s getting there way faster than the hands are
robots need to form a union of robotic robots and protest this. they did not ask for smarter robots
Various capabilities are getting unlocked at different thresholds; the previous big unlock was cyber security, then math, now robotics, what’s next?
how does this work in practice? trying my best not to picture optimus sending a prompt to claude or codex in a data center on max reasoning waiting slowly for a response 😅 how will it control robotics in "real time" - something more similar to computer use?
Disbanding the team feels early to me. A smarter model still needs someone collecting robot data and keeping the hardware running, and that part doesn't come free with Fable-level reasoning.
Which tasks still need a purpose-built controller? The untrained part is what surprises me.
Hmmmm. It’s hard to believe that text can take priority over actual physical behavior. After all physics is a second layer philosophy that models after the real world
通用能力超过专门训练这点,在机器人上体现得这么明显吗?感觉以前这类飞跃不多见。
makes you wonder how far these models can go beyond what they were trained for
The media could not be played.
Square profile picture
Seedance 2.5 now in crisp native 4K on Venice. Available to all users in early access.
The media could not be played.
Watching Deepmind on that front. Their models always had a robotic edge. Who knows, maybe the long game in playing for the self replicating real world bots vs the Software specialists (that Anthropic has) pays out