←
Daphne Cornelisse796 posts
Daphne Cornelisse avatar

Daphne Cornelisse

@daphnesolves

(Multi-agent) simulation and RL • Ph.D. student NYU • Prev @Waymo @NVIDIAAI.

◷ Joined September 2017▧ 77 Photos and videos
Daphne Cornelisse avatar
Daphne Cornelisse@daphnesolves·Jun 19

New Paper: Human-like Autonomy Emerges from Self-Play and a Pinch of Human Data. We trained self-play RL on 60 years of simulation on 1 GPU in ~15 hours. Regularizing with 30 minutes of demonstration data produces much more human-like driving policies!

Omar Khattab avatar
⟳ Daphne Cornelisse retweeted
Omar Khattab@lateinteraction·Nov 16

hot take: nothing wrong with papers per se, but… here’s the recipe github.com/okhat/blog/blob/m…

Post attachment

i’m increasingly of the opinion that almost everyone agrees with this, but a big fraction of people aren’t comfortable saying it in public, because of what it implies for the bean counting mechanism the whole machine is built on (for awards/promo/hiring/etc).

Satpreet (Sat) Singh avatar
⟳ Daphne Cornelisse retweeted
Satpreet (Sat) Singh@tweetsatpreet·Nov 16

1/ Excited to launch my lab as a new assistant professor at UNC–Chapel Hill (@uncsdis)! 🧠🤖🐟 We'll study Natural and Engineered Intelligence through simulation, data science and theory, working closely with experimentalists. singhlab.cc

Brandon Amos avatar
⟳ Daphne Cornelisse retweeted
Brandon Amos@brandondamos·Nov 16

⚠️ Numerical issues can go on silently for decades like in the softmax/CE! Similar issues are present in other widely-used divergences like the KL and JSD I just posted this paper overviewing what they are and how to fix them: 📚 arxiv.org/abs/2611.18511

Post attachment

Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:

0:07
Lars Blackmore avatar
⟳ Daphne Cornelisse retweeted
Lars Blackmore@larsblackmore·Nov 15

Still in disbelief that this happened. Huge thanks to everyone that contributed. 7+ years in the making, and it would never have happened without thousands of absolute winners that I have the privilege of working with!

SpaceX@SpaceXNov 14

Mechazilla catches Starship! Congratulations to the entire SpaceX team on this incredible achievement!

1:33
Maximilian Igl avatar
⟳ Daphne Cornelisse retweeted
Maximilian Igl@MaxiIgl·Nov 12

Also, intern applications for next year are open, on lots of different topics. But especially if you're interested in RL, there's gonna be SO MUCH to explore here!

Super excited for this model release which contains multiple technical innovations, including, for the first time, using our new AlpaGym framework for closed-loop reinforcement learning in simulation! Ps.: We're also working on a paper and technical report specifically on AlpaGym!

Daphne Cornelisse avatar
Daphne Cornelisse@daphnesolves·Nov 6

I will be at CoRL next week! Come find me at poster session 3 on Wednesday, 11/11 (#60). Decided to riff on the cooking and spice analogy for the poster; here is a little sneak peek.

Post attachment
Kevin Miao avatar
⟳ Daphne Cornelisse retweeted
Kevin Miao@KJHMiao·Oct 29

We need to fundamentally rethink how we teach AI & CS. Had a great conversation with @BerkeleyCDSS about my journey from Berkeley to building AI for Vision Pro and Siri at Apple, to starting my own AI research lab at YC. We discussed why AI education shouldn’t chase the latest hype, but instead teach students the frameworks to think critically, experiment, and build at the frontier. Check the podcast out here: open.spotify.com/episode/5KY…

David Pfau avatar
⟳ Daphne Cornelisse retweeted
David Pfau@pfau·Oct 21

Making it possible to do things that were previously impossible is infinitely more exciting than speeding up things that were already possible.

Daphne Cornelisse avatar
Daphne Cornelisse@daphnesolves·Oct 20

Working on biology is humbling. It gives me a similar feeling to looking up at the stars at night and imagining how tiny the Earth is in relation to the rest of the universe.

Itai Yanai avatar
⟳ Daphne Cornelisse retweeted
Itai Yanai@ItaiYanai·Oct 17

Using AI to write for you - start to finish - is really just you robbing yourself of one of life’s greatest joys: to think through a problem and figure out what you think about it, and in the process becoming passionate enough to take on the new challenge of solving it.

Kyle🤖🚀🦭 avatar
⟳ Daphne Cornelisse retweeted
Kyle🤖🚀🦭@KyleMorgenstein·Oct 13

Octi/Nvidia + Abhishek's lab on a generational run. dex manip will be solved with sim.

(1/6) From @GoogleDeepmind’s AlphaGo to AlphaZero, we saw machines find moves that reached beyond what we’d imagined for them. No one taught them those strategies. They emerged from reinforcement learning: playing, failing, and trying again under a single binary task reward for success or failure. Inspired by them, we couldn’t help but wonder: what might a robot discover if we gave it that same freedom to learn? That question led us to AlphaNIST. Starting tabula rasa, with that same binary reward plus only a simple control cost, a single policy solves all 20 @NIST assembly tasks at a 98.2% success rate in simulation. More surprising still, AlphaNIST can complete 10 different tasks in a row, receiving that reward only once, after all 10 are done. Just as AlphaGo’s strategies surprised its creators, AlphaNIST discovered dexterous strategies we couldn’t have come up with ourselves. Paper: arxiv.org/abs/2610.13354 Website: alpha-nist.github.io

1:33