Building & thinking out loud
Joined April 2026
- Tweets 128
- Following 31
- Followers 8,287
- Likes 47
2 Photos and videos
friendly reminder that things get way less scary when you understand how to use them
and even less so when you understand how they work
and it serves the dual purpose of making you more resilient to people manipulating that fear to progress their own agenda
3
5
52
2,334
AGI doesn’t have a set definition, let alone be anywhere on the horizon. Treating it as an inevitable trillion dollar superweapon is science fiction and only adds fear and confusion.
agi is the most economically valuable asset of all time, there will be trillions in free market capital put into it
this is extremely unlike the manhattan project. this time, governments can only cooperate. we can't just pick a winner, or that winner will lose
2
806
ControlWiz retweeted
Jun 12
Being a PhD student is hard for many reasons. On paper, the flexibility can seem like an advantage: you often do not have a fixed schedule or many strict deadlines. In practice, however, that same freedom can become a trap. Since your progress depends so much on your own initiative, it is easy to convince yourself that every extra hour of work is an investment in your future. For students who are already inclined to overwork, this can blur the line between dedication and self-exploitation, eventually leading to burnout.
8
26
371
26,311
VLA based models can only interpolate over the support of its training/pre-training distribution. To become a generalist, it would need training coverage dense enough to span an enormous (possibly unbounded) task space, which is not scalable. A generalist needs to learn on the job and plan accordingly.
4
3
29
2,451
I do not want to do AI research that is reactive to what these companies are doing, or even what they're saying.
The entire field keeps chasing after product releases. Some spend more time reading marketing copy than their colleagues papers and I just... do not want to do that?
5
16
126
7,489
Jun 12
I have to stop arguing on here.
What work? This wouldn't even warrant a publication. The implementation would just be things like "don't ever follow closely enough that them braking could cause a collision", "don't ever enter their turning radius", etc. all the worst case bounds for any more-informed simulation
2
6
1,589
ControlWiz retweeted
Jun 10
Scientific research is fundamental to advancing civilization and helping people globally to solve the most critical problems, from medicine to materials, from brain science to physics, and much beyond. This is only possible when scientists have access to the best tools of the time to conduct scientific research, including having access to AI-based tools.
119
467
3,076
189,568
Jun 12
In multi-agent planning you react and replan continuously using your belief states at 30–50 Hz. If the behavior models of the other agents are known, many methods exist. When they are unknown, you need some predictive model, and collision avoidance still cannot be guaranteed without strong assumptions. I’m not aware of a general method that guarantees safe behavior without these, please share if you do.
No, you don't need some assumption about how other agents will act. It is possible to drive such that you always have an escape route. To never follow so closely that you can't stop or evade if they brake. To never cross ahead of someone that might speed up.
1
5
1,726
Jun 11
World models are not data sources.
Jun 11
What is the best data for training humanoid & robotics foundation models?
Pete Florence @peteflorence (CEO @Generalist, ex-Google DeepMind) dropped his live data tier list in this 7-minute clip on @tbpn:
- S-tier: Real-world robot experience (especially glove/sensor high-dexterity data)
- A/B-tier: Internet/YouTube videos. Surprisingly powerful for transfer learning (the “web data” moment for physical AI)
- B-tier: Text/common crawl (Reddit, books, etc.). Useful priors, but not enough alone
- C-tier: Motion Capture. Great for whole-body motion, weak on finger dexterity
- C or lower: Simulation / synthetic / world models. High potential, still waiting for strong real-world proof
Generalist has collected 270,000 hours of real-world manipulation data (scaling ~10k hours/week). And Pete stressed one key point:
“The quality of data is incredibly important.”
It’s not just about volume. It’s developing intuition for what actually drives performance through hands-on work.
As Physical AI scales, curated real-world and high-quality internet video looks like a winning combo.
h/t @yuji_fujima
12
2,407
Jun 11
You can only optimize if you have some model of the environment. Otherwise, you need a gazillion examples to learn it implicitly.
Jun 11
I think one of the cool things about world models is just how much more efficient they seem to be in both data and compute.
1
7
1,580
ControlWiz retweeted
Jun 11
Jokes aside, using frontier models to change the direction of human thought secretly sets an incredibly dangerous precedent - the fact that they were okay with setting this precedent speaks volumes about how extremist they actually are
62
110
1,709
43,587
Jun 11
Idk if human behavior is solvable. No car will likely reach level 5 autonomy, as long as there are human drivers.
Jun 11
just a lil speed bump.
6
11
3,400
Jun 11
We already have robots that automate many jobs. However, they are limited to structured environments and specialized tasks. I’m not sure we’ll have truly “general” autonomous humanoids without HITL in the next couple of decades. Companies are incentivized to be optimistic, they’re selling a product.
By end of this year, Figure and Optimus will have working demos of serious autonomous capability. Next year will be the launch pad moment.
2
10
1,473
Jun 11
You need at least three contact points for dexterous manipulation. This was proven around thirty years ago. However, three finger grippers have many more DoFs and are often more than 10× expensive. Dexterous manipulation is also a very challenging control problem, at least classically.
What are your thoughts on 2-fingered grippers? So more dofs than parallel jaw grippers but still lacking the third finger.
1
19
2,591
ControlWiz retweeted
Jun 10
Concentration of power, capabilities and economic wealth is the biggest risk in AI. We need open science and open-source more than ever!
111
479
3,087
159,862
Jun 11
It's well known that you need three fingers for stable manipulation. With two fingers you are limited to pick and place tasks. The missing part has been tactile sensing. There are decent ones but they are very expensive. The gap is robust multi finger manipulation with cheap force sensors.
Jun 11
What is your take on the current state of robotic grippers
5
26
4,406
Jun 11
It's not a point of being good enough. Without classical methods you can't guarantee stability/safety. Any autonomous sys that uses learning has classical methods underneath like CBFs or formal methods.
Jun 11
I agree that rn end-to-end models arent good enough. Hybrid models r the only way to make robots work . Dont think thats gonna be the case in the future tho. Iv taken lots of classical control classes and ml has gotten me better results than i would with finding the root locus
1
9
3,177
Jun 11
Autonomous driving is solved for roads without human drivers. Human behavior is what's left.
Jun 11
The Waymo driver and Tesla and zoox are astonishing embodied autonomous systems that will permanently alter the future of civilization for the better
And the data is only getting better
2
18
2,308