Ready for a challenge? Microsoft is partnering with Queen Mary University & CrowdAI to host a reinforcement learning competition using Project Malmo. Discover more about Learning to Play: The Multi-Agent Reinforcement Learning in MalmÖ (MARLÖ) Competition: aka.ms/AA1vi10

Great, easy-to-read recap of how TerpreT (BTW: It's open source now: github.com/51alg/TerpreT) showed that using SGD (+ variants) for searching programs doesn't work well: dselsam.github.io/the-terpret-pr…

It’s not every day that the work of nearly 15 years goes open source!! Very proud of this :)

Exciting news from MIT! news.mit.edu/2018/mit-resha…

Come do a PhD with me, Joel Fischer 😶 stuart reeves 😶 Sean Rintel, and Kenton O'Hara. Microsoft Research PhD Scholarship in Machine Learning and HCI for Games and Mixed Reality. Apply at: nottingham.ac.uk/jobs/currentva…

If you're working on reinforcement learning and natural language then consider submitting to our workshop on text-based games.

"Wordplay: Reinforcement and Language Learning in Text-based Games" with Adam Trischler and Katja Hofmann starts tomorrow. The workshop will highlight research that investigates existing or novel RL techniques for text-based settings: aka.ms/AA3fl0x #NeurIPS2018

Our paper 'Inverting Supervised Representations with Autoregressive Neural Density Models' (arxiv.org/abs/1806.00400) has been accepted to #AISTATS2019. Many thanks to Chris Williams and Nate Kushman. Can't wait for Japan!

Frontiers in Artificial Intelligence is a series of public lectures at Microsoft Research Cambridge. Join us on Wednesday 16 January as Hoda Heidari from ETH Zurich discusses "What Can Fair ML Learn from Economic Theories of Distributive Justice?" aka.ms/frontier-ai-ho…

Looking forward to presenting our work "Learning Robust representations via Multi-View Information Bottleneck" at #ICLR2020. Join the poster session this Wednesday at 8 am or 8 pm GMT. Anjan Dutta Nate Kushman Zeynep Akata poster link: iclr.cc/virtual/poster…

After many great years Microsoft Research I'm excited to have made the move to @DeepMind this week.

#Alphacode I'm really excited to finally release what I've been working on recently! The results are way beyond what I thought possible and I'm excited to see if we can push it even further!

We compare many ways to finetune chain-of-thought models and find a combination of techniques that leads to SOTA results on the grade school math dataset (GSM8K). If you’re interested, come to my talk tomorrow at 1:20 at the MATH-AI workshop NeurIPS Conference.

#AlphaCode—a new #AI system for developing computer code developed by @DeepMind—can achieve average human-level performance in solving programming contests. Learn more this week in Science: scim.ag/Nn

Our #AlphaCode Science paper is released!

Gemini 1.0 is out! It's a privilege to have been a part of this massive effort. Gemini Ultra beats all other models on 30 out of 32 benchmarks and is especially strong on reasoning and coding matching expert human performance on MMLU with a score above 90%.

Also excited to have been a small part of AlphaCode 2 which is built on top of Gemini. This new iteration of AlphaCode has moved from the middle of the pack for competitive programmers to near the top, beating ~85% of Codeforces participants.