Completely agree with this point. How can you really take AI safety seriously without grappling directly and explicitly with the fact the world is open-ended? Rainbow Teaming is a great example!

The unpredictability of open-ended systems allow us to prepare agents for an unpredictable reality. AI development is itself a black-swan event, and if we cannot make AI systems to robust to the dramatic distribution shift they will cause, they will induce their own failures.

This got accepted to #ICML2025 as a *spotlight paper* (top 2.6%!) 🚀 --- work that Yi Xu did as an Msc student! Surely this will mark the start of an exceptional academic journey

Our UCL DARK MSc student Yi Xu managed to get his work accepted as a spotlight paper at ICML Conference 2025 (top 2.6% submissions) 🚀 What an amazing success testament to the outstanding supervision by Robert Kirk and Laura Ruis.

I always took the bitter lesson to mean "only work on whatever scales", not to never be in the loop

Huge congratulations to my academic sister Laura on getting a postdoc position at MIT! 🧠✨ So proud of everything she’s achieved — can’t wait to see all the amazing things she’ll do there. 🚀

Google's AI just made math discoveries NO human has! —Solved optimal packing of 11 and 12 hexagons in hexagons. —Reduced 4x4 matrix multiplication from 49 operations to 48 (first advance in 56 years!) and many more. AlphaEvolve is the AlphaGo 'move 37' moment for math. Insane.

The bar AI has yet to reach.

"open-endedness is all we'll need"...this is the study of a system’s ability to continuously generate artifacts that are both novel and learnable to an observer as a route to agi. excited to have Edward Hughes from Google DeepMind's open-endedness team join us at RAAIS 2025!

2025 is the year of open-endedness. Delighted to be giving a talk at RAAIS in a couple of weeks’ time!

Proud to announce that Dr akbir. defended his PhD thesis titled "Safe Automated Research" last week 🥳. Massive thanks to Murray Shanahan and Pontus Stenetorp for examining! As is customary, Akbir received a personal mortarboard from UCL DARK. Details 👇

Schmidhuber's Gödel Machine: AI "rewriting its code" if provably useful showed the dream of recursive self-improvement 🔄 Thrilled to share our practical realization, inspired by Darwinian evolution! Done with the amazing Jenny Zhang, Shengran Hu, Robert Lange Jeff Clune 😍

One promising direction is combining ideas from AlphaEvolve and the Darwin Gödel Machine. Imagine a self-referential system improving itself even at the lowest algorithmic levels at *scale* AlphaEvolve: deepmind.google/discover/blog/… Darwin Gödel Machine: arxiv.org/abs/2505.22954

What an enormous privilege to give the opening lecture at the OxML summer school this morning. Never have I had such a thought-provoking set of audience questions! Here's to the automation of innovation towards human flourishing alongside the next generation of researchers.

🚀Introducing “StochasTok: Improving Fine-Grained Subword Understanding in LLMs”!🚀 LLMs are incredible but still struggle disproportionately with subword tasks, e.g., for character counts, wordplay, multi-digit numbers, fixing typos… Enter StochasTok, led by Anya Sims! [1/]

Excited to introduce LLM-First Search (LFS) - a new paradigm where the language model takes the lead in reasoning and search! LFS is a self-directed search method that empowers LLMs to guide the exploration process themselves, without relying on predefined heuristics or fixed

Check out Alex’s amazing internship project using Quality-Diversity algorithms to create synthetic reasoning problems! 👇 💡Key takeaway: better data quality improves in-distribution results, while more diversity enhances out-of-distribution generalization.

Happy "The NetHack Learning Environment is still completely unsolved" day for those of you who are celebrating it. We released The NetHack Learning Environment (arxiv.org/abs/2006.13760) on this day five years ago. Current frontier models achieve only ~1.7% progression (see balrogai.com).

LLMs can be programmed by backprop 🔎 In our new preprint, we show they can act as fuzzy program interpreters and databases. After being ‘programmed’ with next-token prediction, they can retrieve, evaluate, and even *compose* programs at test time, without seeing I/O examples.

Much-needed multi-agent benchmark for LLMs 👥 Theory of Mind is key as LLMs act in agentic, interactive settings — yet remains underexplored and hard to measure. 💽 Decrypto offers an ToM-based evaluation of reasoning for agents operating in complex social settings. Great work!