Robotics Pundit. Should be trusted as much as a regular pundit. Interested in the intersection of Robotics/Product/Research
- Tweets 1,518
- Following 372
- Followers 60
- Likes 1,378
If you can’t see this as an obvious burning need for more liberal arts education… you needed to take more liberal arts classes.
Hello computer.
Decompile and distill Grand Theft Auto.
Make the game set in my hometown -- use Google Earth to reconstruct it in pixel-perfect detail.
Make the character me and have my spawn point be in the mansion I always look at on Zillow.
Review my chat memory and make every mission about my hopes and dreams.
Make it playable online so my friends can join.
When you're done, sell all my video games on Ebay and use that money to fund tokens for my next game.
7
Strange that a person this educated would say something so silly.
Half of math is crazy debatably useless conjectures that no “thinking system” would ever posit.
Because AI is not allowed to be absurd.
1
28
This is great!
Really nice (brief) description on how gains were made.
To level set this:
1 failure in 50 gets a human fired by EOW.
Good work! (not a product yet)
Two months of ablating π0.5 finetunes on a real manufacturing task. The policy now succeeds 98% of the time.
We're publishing all results, all the data, and every run, including the ones that went nowhere.
What we learned:
1/ 5x more data took us from 63% to 76%. Pure scaling was the weakest lever we found.
2/ Diversity got more out of the same hours. 4h spread across five scenes beat 4h in the eval scene by 30pp.
3/ Quality did even better. One clean, slower hour on top of the 21h: 76% → 90%. That single hour added more than the 17 hours before it.
4/ So we tried quality on its own: 1h of clean data plus 240 rollouts with human interventions. 1.7h in total, 28% → 88%. That's 12 points above the 21h model with a twelfth of the data.
5/ Inference settings add more on top, with no retraining. They took the 21h model from 76% to 93%, and the 21h+1h model from 90% to 98%.
Just as useful: what didn't matter. Learning rate, batch size, relative vs. absolute joints, image augmentation. All less important than we expected.
Full write-up linked below.
1
25
Before you enabling comma_ai self driving, the buyer should have to watch 30 minutes of geohotz talking. Obvious that safety is not a priority to this person.
techcrunch.com/2026/09/23/co…
25
No amount of pacing changes the scale of AI’s impact. Current models can already replace a huge share of white-collar
These models will deliver productivity and demand. But the time constants for job displacement will be shorter than those for labor gains to materialize.
2
27
t.will retweeted
Our robots only learn when they're running -- the people who keep the fleet up are core to π.
We're hiring robotics technicians (2+ yrs experience) to own our fleet, apply here: pi.website/join-us?ashby_jid…
5
12
87
9,066
Because humans have flexible backs and high DOF spines.
These opinions on “humans will never run like this” are … not … intelligent.
This is crazy.. It's almost like AlphaGo's move 37.. Robots are showing how momentum can be used to propel forward. But honestly I don't think any human can run like this.
1
114
t.will retweeted
I'm giving three workshop talks tomorrow (Sunday) at ICLR!
- self-improvement w/o reward bottleneck, incl meta-harness
(10:30 am, RSI workshop)
- emergent physical generalization, incl π0.7
(11:40 am, multimodal workshop)
- RL for robustness
(1:45 pm, CAO workshop)
7
12
255
21,551
t.will retweeted
LLM post-training used to mean fine-tuning to a downstream task
Robotics has been stuck in this setting, needing task-specific fine-tuning for best performance
π07 changes this: It works out of the box & outperforms fine-tuned specialists
Details: pi.website/pi07
32
67
638
83,107
