
born in right generation huge ass neural nets cuda errors claude code iced americano

your friendly neighbourhood engineer

born in right generation huge ass neural nets cuda errors claude code iced americano

he puts it really really well
this isn’t some big secret. these are the first slides of my RL lecture notes that I teach at UT. function approximation is fine if you already know what you want, but intelligence is about function discovery. if you can write down a (proxy) objective, you can optimize it.


this was a really good read! i also agree with Yifan here that roboticists can be more ambitious now than ever! personally i really believe tools like astra and fable have shown us that general purpose robotics will be solved in our generation and it makes me really excited about the future with robots. ig it is the best time ever to go back to the fundamentals and solve more robotics given that we have intelligence on a tap now. choosing what between the right mix of control, learning and intelligence is going to produce incredible results! fundamental problems are still unsolved ~ sim, dexterity, continual and repeatable systems.
Thoughts after CoRL: old-school robotics strikes back in the age of coding agents. Whatever you believe about robotics, we need to understand what agentic robotics can do and where it breaks currently. Are those bottlenecks fundamental? yifank.github.io/post-astra-…

i see nothing more valuable than the cultivation of excellence in my existence what i mean here by excellence is the quality of being expressed through my judgment of direction of (attention, effort, time) excellence can be a pursuit purely for it's own sake without expectations of outcomes, pride and validation there's inherent beauty in the act itself. it's radiant, rare and one of one the cultivation of that beauty through one's existence feels like an end in itself this requires a certain quality of attention that feels like unclenched devotion - where it's possible to follow through your intentions, trust the loop of life and cultivate the traits that you find beauty in there's something deeply sacred about this

A robot arm threw a ball and the other arm caught it in 1.85 seconds, learned from 50 demonstrations with camera input only. Stanford’s Flash ALOHA also folds shorts in under 6 seconds and clears a dining table in under 7





Drones are hard to build and control with their hardware debugging weight constraints and shipment delays. But once you cross that the rewards are so very sweet. Here's a blog about my journey with drones from being confused to being confused with obsession.



Inkling 2 Small’s weights are now available for download on HuggingFace under an Apache 2.0 license. Make Inkling your own! huggingface.co/thinkingmachi…

Today we’re announcing a joint research project with @BMW Group on video-action models for industrial manipulation. We trained FLUX-mimic on automotive tasks using production parts and fixtures: assembly, insertion, kitting and cable routing. On our dexterous robot platform, it achieved success rates of up to 100% and consistently outperformed other models, even with 4x less training data. Collecting task-specific data is the biggest cost of bringing learned policies into factories today. From day one, we’ve believed that working closely with industrial partners on real production data is the only path to deployment.
