LLMs effectively compress the current state of human knowledge into a model. But can such models be used to go beyond the knowledge they were trained on, and discover new results on open problems? That’s what we study in our paper appearing in @Nature.

Excited to be participating in this event next week!.

Fantastic to see the endless possibilities of AlphaEvolve, the new generation of FunSearch. Congrats @matejbalog, @SashaVNovikov, @pushmeet and team for the amazing work!.

AlphaTensor not only discovers faster algorithms for matrix multiplication - it can make quantum computing more efficient.

RT @GoogleDeepMind: In @NatMachIntell, we show how our AI system AlphaTensor-Quantum can make quantum computing more efficient. 🖥️⚡. By opt….

RT @GoogleDeepMind: For the first time ever, we show how a large language model can help discover novel solutions to long-standing problems….

RT @erikbryn: This is undoubtedly just the beginning of a new age of scientific discovery.

سعيد جداً بنشري بحثاً في مجلة @Nature. لأول مرة في مجال الذكاء الاصطناعي، تحقيق اكتشافات علمية جديدة بواسطة نماذج لغوية كبيرة (Large Language Models) مثل الChatGPT أو Gemini. لقراءة المزيد:.

Also see this great thread!.

Thanks to my amazing collaborators @ber24, @BarekatainAmin, @SashaVNovikov, @matejbalog, @Mpawankumar123, @emidup, @franciscuto, @JSEllenberg, @PengmingWang, Omar Fawzi, @pushmeet.

Beyond maths, we also applied FunSearch to find new heuristics for the bin packing problem, which is central to many applications - eg, scheduling jobs in data centers. FunSearch evolved a simple heuristic towards a better one - which reduces waste - for the data at hand.

The program is more insightful, and is easier to understand. It’s also crucially more concise, and this is one reason why FunSearch works so well in problems with structure. In fact, FunSearch favors solutions represented as short programs - or low Kolmogorov Complexity.

We showcase results of FunSearch on a problem in combinatorics (the cap set problem), where FunSearch constructs cap sets that are larger than known. Crucially, FunSearch describes how to construct the solution with a program (left), rather than what the construction is (right).

In practice, we evolve functions (in code) so we can run and evaluate these automatically. Fortunately, in mathematical sciences, many problems admit an efficient evaluator: these problems are typically hard to solve, but easy to verify/evaluate - think NP-complete problems!.

Our approach, FunSearch, combines the creativity of an LLM with an automated evaluator that guards it against incorrect ideas. The process is an iterative back-and-forth between a frozen LLM and evaluator, to evolve initial (trivial) ideas into more complex (but correct) ones.

Naively asking the LLM to generate a new mathematical result / find better algorithms than the known ones will not work: at best, you will get a known result - and by pressing it for new results, you will end up with a hallucinated answer which doesn’t solve your problem.

RT @arkitus: We're hiring a research scientist to join our Quantum Chemistry and Materials team ⚛️🚨. The team is working on using machine l….

Honoured to be featured in the list alongside amazing innovators!.

If you have ideas on further initiatives or would like to be involved, please get in touch with me or any of the #nassma members.

Enjoyed a lot my time in Rabat. Thanks to all speakers, attendees and organisers who made this workshop a reality. This workshop is only the beginning of a series of initiatives which will bring cutting edge AI research to the MENA region!.