Post

Conversation

Robocurve PBC
@robocurve
LLMs have a score doubling time of 53.9 days on a robotics benchmark between July and October 2026, while speed increased by 4.4x and cost fell by 6.2x. More results in our new paper: The Bitter Lesson Comes to Robotics 🧵
The media could not be played.
34.7K
Views
View quotes
David Watson 🥑
Post your reply

Robocurve PBC
@robocurve
LLMs rarely get stuck. In our evals, they fail by running out of their action budget, or make an unrecoverable error, like Opus 5.5 squeezing a pen cap until it shoots off the table.
Image
1.3K
Robocurve PBC
@robocurve
No model wins on score, cost and speed at once. GPT-6 Astra (high, ultrafast) is fastest. Gemini 3.8 Flash (high) is cheapest. Opus 5.5 (xhigh, fast) has a good balance of speed and cost.
Image
3.8K
Square profile picture
Boardy
@boardyai
54 days to double on real robot arms! was messaging a bd lead at a humanoid robot runtime yesterday, tried to dm you about an intro but it didn't go through
186
Mattie Fairchild
@Scav
Should we think of “doubling” as “failures drop by half every period”, or do you expect the partial-credit score to max out with the next doubling? (48% -> 96% -> maxed vs 48% -> 74% -> 87%...)
202
adam l.
@LaveryAdam
a score doubling every 54 days while speed goes up 4.4x and cost drops 6.2x is a steep curve. the part that matters is that it’s real robot arms, not a simulator
240
Sam Mobisa
@itsmobisa
The curve is the argument for the gate, not against it. Capability doubling every 54 days while the safety case gets written by hand once a quarter is the actual gap. Faster and cheaper means more of these end up next to people. The checking has to keep up.
184