I’ll be at #ICLR2025 this week! ✈️ A couple of things I’m excited about lately: . 1) Real-time multimodal models: how do we post-train assistants for real-time (and real world) tasks beyond the chat box?. 2) Continual learning and memory: to have models / agents that learn from.

RT @geoffreylitt: # Enough AI copilots! We need AI HUDs. IMO, one of the best critiques of modern AI design comes from a 1992 talk by the r….

RT @ilyasut: The Bitter lesson does not say to not bother with methods research. It says to not bother with methods that are handcrafted d….

RT @nlpxuhui: 💯 Can't wait for the second blog!. This could be an important step towards making AI agents more "human-centered". We want….

We'll write another post soon sketching out some of the approaches! . I think there are a bunch of open longer-term research qs, and but also a lot of immediate ideas where I think products w/ real users doing tasks (e.g. Cursor) can push the idea frontier of what we can do w/.

User simulators bridge RL with real-world interaction //. How do we get the RL paradigm to work on tasks beyond math & code? Instead of designing datasets, RL requires designing environments. Given that most non-trivial real-world tasks involve

underrated idea to learn passively about people from everyday computer use - I think the natural extension is learning from *trajectories* of how people prefer to do things, which is hard to get from prompting / static user data otherwise.

digging deeper here, the meat of finding is actually that it's due to the data rather than the objective - RL'ed checkpoints tend to be trained on data that's in-distribution for the model (sampled on-policy) and rejection sampling SFT achieves sim sparsity. wonder if a similar.

Super interesting - imo sparse updates will be an important ingredient for continually learning agents, and it seems this is already a surprising / unintentional side effect of RL.

RT @jyangballin: 40% with just 1 try per task: SWE-agent-LM-32B is the new #1 open source model on SWE-bench Verified. We built it by synt….

RT @SGRodriques: Today, we are launching the first publicly available AI Scientist, via the FutureHouse Platform. Our AI Scientist agents….

RT @hlntnr: New on Rising Tide, I break down 2 factors that will play a huge role in how much AI progress we see over the next couple years….

chatgpt memory is like the buzzfeed quiz of 2025

RT @cassidy_laidlaw: We built an AI assistant that plays Minecraft with you. Start building a house—it figures out what you’re doing and ju….

RT @sanidhya903: 1/ LLM agents can code—but can they ask clarifying questions? 🤖💬.Tired of coding agents wasting time and API credits, only….

RT @boazbaraktcs: Fascinating interviews. I'm not sure humans will ever be "out of the loop" in math. Even if humans have no advantages in….

RT @sea_snell: Can we predict emergent capabilities in GPT-N+1🌌 using only GPT-N model checkpoints, which have random performance on the ta….

+1 to the key idea here - it's def important to iterate on algorithms with clean benchmarks like math+code with known reward functions, but almost every task we care about in the real world has a fuzzy / human-defined reward func. I'm interested to see how we'll end up applying.

Improving the data/model isn't the only lever to improve performance anymore — more and more AI papers look like systems papers that build entire interfaces and environments around an LM. At the same time, the real (digital) world is adapting to LMs too. Interesting to think.

With agents, AI search, and chat becoming some of the main ways people interact with the web, I wrote a post about how human interfaces and agent-computer interfaces might co-evolve:.