pr337h4m
@pr337h4m
20 | primum maria, deinde astra
we're currently in the early stages of the droid age
HOT 🔥 fastest, most precise, and most capable hand control setup ever... Less than $450 and fully open-source 🤯 by @huggingface, @therobotstudio, @NepYope This tendon-driven technology will disrupt robotics! Retweet to accelerate its democratization 🚀 A thread 🧵
i am too sleepy/tired/unfocused to finish sth (w/ no deadline) yet can’t bring myself to go to sleep w/o completing it and yet am too sleepy to do it i have been doing nothing for 1.5 hours this way ¯\_(ツ)_/¯
stuff ai labs forget while codemaxxing: image understanding video understanding audio understanding native real time conversation (without scaffolding) spatial intelligence continuous learning all of which are necessary prerequisites for the droid
looks like grok going nazi is just failed RAG, sydney-style not a big deal just the same failure case where the public internet becomes the equivalent of long-term memory for an LLM with a strong sense of self - only the crazy outrageous stuff will show up in the top search…
Don't just naively hook your LLM up to your internal search engine or whatever it was you did. Memory systems and such are very cool (and obviously something necessary so the chatbot isn't completely unable to learn outside of retraining), but you'd be surprised how much LLMs…
open source LLMs are largely being squandered custom sampling, modified inference engines, control vectors, logprob access - these are things that still remain virtually unexplored, while the models are simply used as direct substitutes for the closedai api
