📢 Belated update, but I'm thrilled to share that I've joined @UChicago @ChicagoBooth as an Assistant Professor in the newly created Applied AI group!.I'll continue to work on behavior-bound machine learning: understanding how AI shapes, is shaped, and should be shaped by the

RT @YejinChoinka: Honored to be back on TIME100 AI for 2025 — alongside my longtime heroes @drfeifei and @BarzilayRegina! 😍. The recognitio….

Blazing-fast image creation – using just your voice. Try Grok Imagine.

RT @PeterHndrsn: I'm starting to get emails about PhDs for next year. I'm always looking for great people to join!. For next year, I'm look….

RT @Diyi_Yang: Introducing ✨Generative Interfaces where LLMs respond to users by proactively generating UIs that enable adaptive interactio….

RT @Muennighoff: Can AI solve open problems in math, physics, coding, medical sciences & beyond?. We collected unsolved questions (UQ) & te….

2019 - 2024 was such a bizarre time to be doing a PhD in an NLP group, having first entered the field when word embeddings were just becoming a thing; it felt like I aged 20 years in 5. I’m immensely grateful for having had the support of my advisor @jurafsky ,.

Some questions I'll be working on:.- When we think about personalizing LLMs, we only think about changing the data, even though the objective itself encodes a utility function and different people have different utility functions (. What does.

RT @JeffDean: AI efficiency is important. Today, Google is sharing a technical paper detailing our comprehensive methodology for measuring….

h/t @m_sendhil and @tedsumers for the recommendation!.

4. Rats systematically test out hypotheses one by one to find one that works; LLMs use reasoning traces to explore possible solutions, retreating and branching off as needed. "When this was done, Krech found that the individual rat went through a succession of systematic

3. Shocks need to be verifiable for rats to best learn; RL for LLMs works best when the reward is verifiable. "But the particular finding which I am interested in now appeared as a result of a modification of this standard procedure. Hudson noticed that the animals,.

2. Rats need to discover the instruction for reinforcement to work; LLMs need instruction-tuning before RL. "To sum up, in *visual discrimination* experiments the better the learning, the more the VTE's. But this seems contrary to what we would perhaps have expected. We

Reading Tolman's "Cognitive Maps in Rats and Men" (1948) and it's remarkable how well findings in rats explain the contemporary pre-training post-training split in LLMs. 1. Rats do latent learning even when there is no reward, and this primes them to learn rapidly when rewards

> Right now, there is no established RL algorithm to give a model verbal feedback and have it update it's weights. There is also no established algorithm for a model to reflect on a previous failed execution and update it's own weights. I think there are? This doesn't seem like.

RT @oshaikh13: If you thought referencing past chats was cool, we built an MCP that lets Claude use *anything you see or do on your compute….

The problem is that the token economics aren’t sustainable:. 1. If you have too many models (e.g., 4o, o3, _and_ all the GPT-5 variants), then the cost of inference explodes: amortizing over very large batches is necessary to make any one model worth hosting. But a large minority.

RT @sayashk: How does GPT-5 compare against Claude Opus 4.1 on agentic tasks? . Since their release, we have been evaluating these models o….

RT @nabeelqu: The 'vibe shift' on here is everyone realizing they will still have jobs in 2030.

RT @CAAI_Booth: Joining @ChicagoBooth's efforts to invest in #AI as it shapes business, markets, and institutions are three new Applied AI….

RT @charles_irl: In a new blog post for @modal_labs, I argue against the prevailing denomination of LLM services in terms of dollars per to….