Alex Lamb and I are looking forward to tutorialing on discovering agent-centric world models at 4pm today at ICML icml.cc/virtual/2023/t… . And in the meantime, a jetlag sunrise bonus.

An empirically very useful deep active learning algorithm.

And we're live! Join us now for the workshop on Interactive Learning with Implicit Human Feedback at #ICML2023! w/ Akanksha Saran Andi Peng Tengyang Xie Anca Dragan John Langford Pierre-Yves Oudeyer

Join us for the U of T Department of Computer Science Distinguished Lecture Series, featuring John Langford of Microsoft Research. 🗓️ Tuesday, August 8, 2023 🕚 11 am – 12 pm 📍 BA 1200 ℹ️ Open to the public. No registration is required but space is limited. uoft.me/9wJ

The Hamas attacks in Israel were horrific. The most analogous situation+outcome I see at the moment is the battle of Mosul en.wikipedia.org/wiki/Battle_of….

We also have an intern req posted now: jobs.careers.microsoft.com/global/en/job/… .

I joined AI Frontiers, a new multilocation org in MSR (microsoft.com/en-us/research… ) with the goal of advancing the frontiers of AI.

There is quite a bit of potential here, and we are actively looking at applicants. Take a look and please consider applying.

This paper: arxiv.org/pdf/2403.11940… is making an important point related to the tutorial (bit.ly/agentstate ) that Alex and I did at ICML. You really need forward dynamics on top of a multistep inverse model to fully capture an agent's endogenous dynamics.

This takes random exploration data over a continuous control space as input and learns a representation that can regularly accomplish ad-hoc tasks at run time with _zero_ additional samples once the task is known.

This: marktechpost.com/2024/07/11/mic… is the first time I've seen a single GPU model outperform GPT3.5 across a wide array of metrics.

It's rare to see a benchmark jump from 42% to 89% (see: allenai.org/data/musique), but apparently that's what happens with multihop question answering when you jointly train a set of agents.

I'm attending ICML and will be at the Microsoft booth 11:30am Wednesday for a general AMA around AI Frontiers. Drop by if you want to chat (or email me to chat separately).

As part of the Microsoft Research forum (x.com/MSFTResearch/s… ), I'm running a panel on multimodal models (x.com/msftresearch/s… ) which seems like a fun topic. Register/join if that's your jam :-)

We are doing a post-covid restart of the NYC ML Symposium (events.nyas.org/event/ml2024/s… ) on Oct 18th. It's a great chance to meet folks in the area, so drop by if interested. Abstracts for posters are due Sept 6th, in about 10 days. (These are super quick & easy.)

Looking forward to this.

A new self-driving OS benchmark that's fully parallelized, with a range of initial benchmarks, and an open source helper screen scraper.

An evaluation system from AI Frontiers folks covering semantic vision & language tests.