Nathan Ng

1/ Today, we are thrilled to announce Generative Simulators, a new class of adaptive, auto-scaling environments for AGI training and evaluation 🤖🧵 Static datasets, hand-authored environments, and human-curated demonstrations do not automatically scale with the learning patterns of the trained model. We propose Generative Simulators as a principled alternative: environments that evolve, evaluate, and adapt to agent behavior over time. Technical Report: patronus.ai/generative-sim… Blog: patronus.ai/blog/introduci…

Thrilled to announce our work on Generative Simulation that will enable a new class of adaptive auto-scaling RL environments. If you’re excited by auto-scaling environments, I’d love to chat!

@wgrathwohl Sushi Sho and it's not close

Best ny omakase ?

Very proud to share this work from the Advanced Cell Engineering team at Genentech! We developed what we believe to be the best inducible CRISPR system yet that has nearly undetectable levels of leakiness and activation levels that match constitutive expression 🤩

@cleaming_ Donuts - J Dilla

Kind of specific but can anyone name some albums where the final track leads back into the opening track? Like the ending notes of the closing track connect right back into the first notes of the opening track, so the album can seemlessly be listened to on loop

@stanislavfort Can't seem to DM you but would love to chat! I'll try email

A decade ago, AlphaGo inspired me to leap from black holes to AI. Bittersweet to close my time at DeepMind, but thrilled to start a new chapter focusing directly on AI and security. Find me at NeurIPS this week or DM me here if you'd like to chat!

I'll be presenting this poster at #ICML2024 tomorrow morning! Stop by and chat about our work! I'll be around until Thursday, DM me if you'd like to chat about influence functions, uncertainty quantification, or robustness!

None of this work would have been possible without guidance from @MarzyehGhassemi and @RogerGrosse to whom I'm extremely grateful. This project took over a year of hard work from its initial conception and capped off the final year of my PhD!

Overall this is the paper I'm most proud of during my PhD, and I think there's a lot more interesting insights hidden in the MDL literature. If you want to read more, check out the arXiv: arxiv.org/pdf/2406.02745 blog post: nng555.github.io/research/confe… (should be up soon)

The minimum description length principle is an attractive Bayesian alternative for quantifying uncertainty, but how can we get it to work efficiently and accurately at scale? Excited to share our ICML work on measuring stochastic complexity with Boltzmann influence functions!