Recorded Talks
Below are some talks that have been recorded, listed reverse-choronologically, which means, as you go down the list, it deteriorates in quality. My best talks are the ones that have not been recorded.
- What can(‘t) next-token prediction do? @ NeurIPS workshop, 2025 [30 min talk]
- Next-token prediction and creativity @ NeurIPS 2025 oral [20 min talk]
- Pitfalls of next-token prediction @ Simons, 2024 [1h talk]
- Out-of-distribution generalization @ CMU AI lunch, 2025 [1h talk]
- Failure of uniform convergence bounds @ Princeton, 2020 [10 min talk]
- Failure of uniform convergence bounds @ NeurIPS 2019 oral [20 min talk]
- Algorithm configuration @ Simons, 2016 [30 min talk]
Slides
Below are slides I have designed. I try not to put too much text (an insight that has blossomed only with age), but I also don't like to skip all text (a sedated audience member must be able to follow when they are unlucky enough to wake up in the middle); I rely significantly on visualizations.
- Creative limits of next-token prediction [Oral Slides][1h Talk Slides]
- Pitfalls of next-token prediction [Slides]
- Out-of-distribution generalization [Slides]
- Downscaling LLM slides[Slides]
- GAN optimization [1hr talk - slides] [NeurIPS Oral slides]
- Algorithm configuration @ Simons, 2016[Slides] </a>
Posters
Same philosophy as my slides. But if I'm being candid, my earlier posters (like uniform convergence and GAN optimization are too dense and cluttered in hindsight, a sentiment that applies to how my taste in research has aged too.
- Creativity ICML’25 poster
- Downscaling LLMs ICLR’24 poster (A poster done in collaboration with Tian)
- Pitfalls of NTP ICML’24 poster
- Interpretability theory ICLR’21 poster
- Out-of-distribution generalization ICLR’21 poster