Vrushank Desai

Meet Robo! i'm in the very early stages of building an accessible, general-purpose robotics platform, and looking for a few passionate people to help shape it from ground up. if you'd be excited about moving fast on a small, hands-on team I would love to chat -- my DMs are open! (demo is teleop, 1x speed)

@___Harald___ @marcospereeira this might be cope but i do think the loss landscape from 1990 to 2020 had a lot less 'fog of war' than todays world (US maintaining global geopolitical/financial stability + obvious tech tree to climb)

@marcospereeira Realizing how successful centralized power has been for China made me sad. It's like finding out Santa isn't real.

Now that China has the world's most dynamic market economy, isn' it embarrassing to have been one of those people that claims communism can't make iphones? The American "capitalists" are figuring out how to charge more for less, the Chinese "communists" are just building more.

@YuXiang_IRVL i've been wondering the same, i think we will soon run into hard limits of what the hardware can do bc robot hands/tactile sense sucks. the switch from software to hardware as bottleneck came faster than i expected

Is real-world RL becoming a cheat code for robot tasks? If we take a task A, run imitation learning, then fine-tune with real-world RL, the task will almost certainly work. So what am I missing? (We’re not talking about generalization to task B here.)

@chris_j_paxton i'm interested in seeing how this bet plays out but i am sorta bearish in the long-run. i think they rely heavily on expensive robots with harmonic drives to stay in-distribution with UMI data but ideally the policy can just handle cheaper arms that have backlash (ALOHA/ARX)

Generalist has probably the best manipulation videos around. The motions are so nice for very complex tasks.

@phethers ooh ive never tried flux on the wick, thanks!

@vrushankdes Use very thing copper wick and a tonnnnn of liquid flux on the part+wick, solder iron at 400 and just hold the wick on the area you're removing solder from

Regret choosing LQFP packages for some parts it has massively slowed down production, the legs provide a lot of surface area for solder which messes up some of the reflow process. Next iteration will have no legs on the larger fine pitch parts, you can swipe off any excess solder so easily

@lukebayes @phethers hmm i was having issues around 350C, i'll try a bit warmer and a wider tip, thanks!

Assuming you're not pushing super hard, it could be too much heat and melting the package. I've had issues with using too small an iron tip. This forced me to hold too long and damage parts. The fix (for me) was using a much larger wedge tip at ~380C, lots of flux and dragging (gently). Less time in contact and reflow happens before plastics melt. Search YT for, "Drag Soldering QFP"

@xiao_ted OK wow, just saw the paper. very impressive if the model generalizes without explicit alignment!

@xiao_ted is the output action space joint angles or end-effector position and is an overhead camera view part of the input? asking because UMI-style data collection can lead to 'motion generalization' but its more of a property of the IK-solver + standardized camera views/action space

It is very difficult to properly convey “zero-shot generalization”, it’s a very overloaded and overabused term these days! But let me try to add some color to why so many of my colleagues have been so shocked at what the Gemini Robotics 1.5 VLA has been doing… 🎨 The whole point of a VLA is to guarantee positive transfer by leveraging internet data to learn capabilities that you don’t need to cover with expensive real-world data collection on your robot. And this kind of works! We’ve now seen this positive transfer appear many times as semantic or visual policy generalization: understanding new visual-lingual concepts like “Taylor Swift” or “healthy snack” or being more robust to background and lighting changes. But *motion level generalization* has remained elusive. Back in 2023 with OXE/RT-X, we thought that naively pre-training a VLA on massively cross-embodied robot data mixtures might give us some of this right off the bat. We searched extensively for motion level improvements. For example, while our Google robot had never connected “pouring” task data, we knew that other robots had seen this data. But despite extensive evaluations, the main gains we found were semantic: understanding concepts like “next to” or “on top of”. In 2025, many *post-training* methods now aim to showcase motion-level transfer (ie. across different types of tabletop manipulators with different kinematics or from egocentric human to humanoid) by explicitly retargeting. But the dream, I would argue, has always been to discover a general motion-generalization approach which is extendable to *pre-training*. With the Gemini Robotics 1.5 VLA, we pushed hard to make a dent in this problem, but we were still very puzzled at how strong the initial evaluations were. We thought that there must be a bug, some data leakage. But after thorough debugging and validation, we began to believe the results. A *pre-trained checkpoint*, trained on disjoint data distributions on multiple robot platforms (Aloha, Bi-arm Franka, Apollo Humanoid) was showcasing motion level transfer! Despite the Aloha *never* seeing a tool hanging task (reaching at that height and horizontal depth with very specific asymmetric motion needed for a precise hang), it was able to repeatably succeed. Same story for Aloha to Apollo, for Bi-arm Franka to Aloha, all the bi-directional pairs. To be honest, the results still feel shocking. It feels like the very early days of RT-2 when we first started seeing promising signs of life with the first large-scale VLA training run. I have more questions than answers right now, but this is where some of the science and the real fun begins 🚀

it terrifies me how many talented people are contributing to a future they wouldn't want to live in. building something addictive. exploitative. destructive. something they'll ban their own kids from using. and for what? a paycheck? surely there's a better way to use your gifts?

@JimDMiller @Leo_Abstract @subygan @dwarkesh_sp @gwern i think bipedalism + hand is extremely underrated. an elephant's ability to channel more intelligence into evolutionary fitness is pretty bottlenecked by lack of appropriate end-effectors. whereas humans could write/dig and plant seeds/use spears/etc

@Leo_Abstract @subygan @dwarkesh_sp @gwern But what about getting smarter because of competition with other elephants?

A point I've heard both Carl and Gwern make is that with primates, evolution finally found both a scalable brain architecture AND a niche that rewarded marginal increases in intelligence. Some birds are really smart for the size of the brains. But they're in a niche that punishes bigger heavier brains - they'll fall out of the sky. The difference in neuron count between primate species' brains is proportionally to their brain mass, suggesting a scalable brain architecture, By contrast, for rodents and insectivores, neuron count scales sublinearly with mass. Another feedback loop like the cooking example: opposable thumbs allow us to make tools, which increases the value of having bigger brains to design said tools. Which incentivizes more dexterous hands...

@MechanizeWork dont take the bait guys dont do it

Open source is not a viable strategy for producing RL environments at scale. The future of AGI lies in closed source development.

humans are so cool, man bunch of apes walked out the jungle and went to the moon and made the sand think crazy fucking apes

americans wouldn't have such a strong distaste for silicon valley if any of you were willing to stand up for principles. but alas, you haven't vested yet have you

@faraz_r_khan have you tried V-blocks? amazon.com/Grizzly-H5608-…

Question for the mech engrs: how do I hold/clamp a bicycle head tube down super rigidly so that it doesn’t rotate at all? Trying to measure frame lateral / torsion flex but keeping the tube from rotating under a large torque is being difficult.

@jachiam0 mundane long-tails that accelerate day-to-day progress are still rough. for example, today gpt-5 web-search just hallucinated a handful of phone numbers for city food departments. when this is marketed as "Death Star Phd-intelligence" it can be frustrating...

This feels like an increasingly accurate description of the public reaction to new frontier models. In truth: progress is not slowing down. Each successive delta in model intelligence is just useful to fewer and fewer people.

@saranormous marshallbrain.com/manna worth a read!

3/ Machine-Verified Trades

1/ we're extending the application deadline for @Conviction Embed, our grant program for AI-native startups to 8/14 it's the best time in history to start a company

@willdepue @rationalaussie statement feels disingenuous considering how much effort is put into curation/expert dataset creation. consider a model pre-trained on everything from the internet OAI chooses to leave out of the corpus (SEO/auto-generated logs/etc). is that going to work well?

@rationalaussie dawg what do you think pretraining is. you just train on the firehose of the internet. actual opposite of verified. seems to work pretty well regardless?

everyone who mentions ‘continual learning’ as a problem is usually just talking about sample efficiency. clearly, you should ‘continually learn’ by continually training trajectories back into the model! there’s no mystery: this just doesn’t work with low sample efficiency.

@ErebiusWhite certainly a lot of scams would be exposed. like trying to trade powerpoint consulting services for a house and groceries... money is a dangerous abstraction

We do need nine to fivers, don't we? Or if everyone built something and traded it for something, wouldn't that work? Pure capitalism

@xlr8harder whAt aBoUT tiAnaMen SquAre?

It's a really sad state of affairs that if you want a minimally censored (at least on US politics) high performing open source AI model, you need to source it from China.