Asymptotically approaching Ph.D. | AI engineer at an international big tech. Tweets about things around LLMs, neurotech, EEG, and BCI when not busy shitposting.
Rotary inverted pendulum demo from Pulsar at ICRA 2026. Previously I tried the RL way on my YouTube, but here they show some comparison between classical methods like LQR and modern ways using RL. While classical way should easily work for such a simple system, what makes this single actuator and single encoder setup so nice is the simplicity. You learn the same concepts to deploy it on a humanoid. At least the general idea.
If you want to go viral in longevity world:
Go do a long run with a VO2max test at the end
Your VO2max will suck.
Then eat lots of candy, binge Netflix for a few days... re-do the VO2max test.
You'll see a huge jump. Tweet that netflix + candy made you 20 years younger.
Once again computer scientists show total inability to comprehend that there are other fields of research. Seriously, this is like a disease or something
My advice to PhD students in 2026:
1) If your advisor hasn't logged >100 hours in a modern agentic IDE, stop listening to your advisor.
2) Write your next paper *inside your codebase* as a .tex (with 80 char word-wrap). Force your advisor to read and make edits inside the IDE.
Advice to PhD students in 2026:
1. Make sure that you have a real interest in the topic you are researching.
2. Make sure that the advisor has good publication record and connections in the industry (if applicable)
3. (Crucial) Make sure that the advisor is not a total dick
People using non-IDE text editors were once considered elite programmers, but now if you don't delegate everything to an agentic IDE, you are considered the idiot. 🤷♀️
Excited to present our rat exoskeleton at ICRA 2026!
We developed what we believe is the first rodent exoskeleton designed to support the entire rat hindlimb (thigh, shank, and foot).
Our goal is to provide a new robotic platform for neurorehabilitation research.
#ICRA2026
Do 3D reconstruction transformers really need a billion parameters, or are most of those layers just doing the same thing over and over?
Introducing Déjà View: a single transformer block, looped K times, that matches or beats models 8–10× its size with lower compute. 🧵
If he defines consiousness mechanistically as a feature of computational complexity and fitness for particular purpose (close to what Tononi does), then this conclusion can't be erroneous given this definition - it is as good as any other. Nobody knows what consiousness is anyway
Two things have happened with Richard Dawkins here with his claim that the AI known as Claude is "conscious".
1. He's fallen for the Eliza Effect, first diagnosed in 1966. "A psychological phenomenon where humans anthropomorphise computer programs, projecting understanding, empathy, and intelligence onto them."
2. He has a mechanistic view of consciousness anyway. He frequently employs machine and computational metaphors to explain human consciousness and biology. He views the brain as a "complex computer" and humans as "survival machines" or "gene machines".
So his framing belief system can only lead to his erroneous conclusion - that he and the machine are both conscious.