特徴量の気持ち retweeted
Two years ago, we released Genesis, a universal physics engine for robotics.
Today, I’m excited to announce 🍎 Eden — a robot learning framework built from the ground up for agents, unleashing full potential of Genesis.
One task definition. From simulation to real robots. Sim-to-sim and sim-to-real just got easier.
Coding agents build tasks 1.6–4× faster than in Isaac Lab.
🧵 1/11
特徴量の気持ち retweeted
MotoGP championship standings after GTA V's release vs. right before GTA VI. #MotoGP
LISTEN: Air traffic control audio captures the moment a nationwide ground stop is issued, halting departures across the U.S.
Transportation Secretary Sean Duffy says a “communication issue impacting all Air Route Traffic Control Centers” prompted the FAA’s order. Technicians are investigating.
Departing aircraft are instructed to remain grounded until further notice. International arrivals will be handled on a case-by-case basis.
Download the Fox News app or visit FoxNews.com for the latest updates.
特徴量の気持ち retweeted
~1 year ago: AI coding didn't work
~0.5 year ago: agentic tasks failed
~now: nobody checks code + full automation
What's missing?
My first (real) attempt at AI coding was mid last year, mostly playing around with Copilot. It was still very much a vibe-coding experience and not ready for serious work. People oversold it -- lots of bugs and way too many iterations to be productive. In fairness, I gave up on it quite quickly.
The real game changer was Opus 4.6 -- it needed precise instructions, but its execution was on another level. If you iterated on the plan until it was well-specified, it would turn each task into pretty decent code. However, it still struggled with longer-term reasoning, and with every iteration the chance of breaking something else went up -- agentic tasks in particular mostly failed.
This changed over the past few months: agentic tasks now work quite well, and I'm trusting agents with much more complicated prompts -- if the underlying methods are well known, this works remarkably well. I'm particularly stunned by how well the VLMs inspect results and propose fixes on their own; 3D modeling is a prime example here but there are many other examples.
Despite that progress, there are still missing pieces:
- the main bottleneck seems to be prompting and the imperfect nature of language itself, i.e., I find it quite difficult to tell agents what I actually want them to do.
- ideation / brainstorming still struggles with specific solutions that aren't mainstream. Especially for research projects, agents rarely come up with good ideas and auto-research loops are mostly limited to hyperparameter optimization.
Given the current insane speed of progress, I'm extremely excited about what the next year brings. I've heard so many times now that "AI can't do x" -- and now we all take this for granted. I also really don't see things slowing down, since everything keeps getting easier to try out. Really looking forward to it!
特徴量の気持ち retweeted
arXiv submissions are skyrocketing
42K new papers in October. Up 81% since January and nearly double the monthly volume at the end of 2024
Pangram analysis coming soon 👀
Meet GR00T 2, our new generation of world-action models (WAMs): it imagines how the physical world will unfold and generates the actions that get it there. One model runs on parallel grippers, dexterous hands, and humanoids.
research.nvidia.com/labs/gea…
特徴量の気持ち retweeted
12/3(木)にマルチモーダルAI技術勉強会 #sumoai の第五回をIVRy株式会社にて開催します。
sumo-ai.connpass.com/event/4…
sakana.aiの久保陽太郎 @yotarokubo さんに音声に関する研究活動、東大生産技術研究所の篠田理沙 @dahlian0 さんに動物を対象としたCV研究についてお話し頂きます。













