if one figures out a way condense and get the "right" context for most physical labour in less than a million tokens.
they are going to make alot of money
do we need flow action heads on our VLAs?
introducing NAT a streaming action tokenizer
closes the inference latency gap with flow action heads
trained on 49k hours of data it can zero shot tokenize unseen robotic embodiements
GPT 6.2 Astra has scored a whopping 41.2%(!) on Ember-0, one of our robotics benchmarks.
This is a HUGE jump from early Sept, when GPT 6 Astra scored 19.6%.
We expect Ember-0 to be saturated very soon.
we are reconstructing the grid from first principles.
text your friends via radio, track tagged animals, and gather a ton of interesting atmospheric data via stratospheric balloon!
launch a stratospheric payload of your own https://t.co/OcJ2nhbYBA
Robot-use agents can now control real robots. But they are slow.
We made the GPT-6 Astra + Codex agent run much closer to real time:
- cup task: 11.1 min → 3.4 min
- towel folding: ~18 min → ~6 min
- robot moving 84% of the time, up from 24%
Blog: https://t.co/BIaiZ2P5ch
Code: https://t.co/MvWxzfMXiK
openai is totally mogging anthropic on transparency around rsi. nice reports today following that corposlop 6.2 launch video. the arc of time shifts closer to good