SOTA TV

SOTA TV ep 2 with Aaron Lou, leading OpenAI strategic explorations team.

I asked Aaron: if you were a dataset, which format would you be? He said a Lean proof. He’s a theorist at heart.

In 2021 Aaron started his Stanford PhD on geometric deep learning. Then ChatGPT launched and he asked himself: does it matter?

He switched to generative modeling, and his diffusion LM paper SEDD won the ICML 2024 Best Paper Award. But he dropped out to join OpenAI at the craze of o1.

Towards AGI: autoregressive models with speculative decoding, or diffusion LMs? He worked with us on a chalkboard on the design and trade-offs.

We also talk about latency vs throughput (a billion ChatGPT users vs robots that can't wait), Astra as the brain driving robot arms.

0:00 Intro

1:06 If you were a dataset, which format would you be?

1:41 From a Stanford PhD to OpenAI

7:38 Diffusion vs autoregressive: the chalkboard and the bet

10:16 Speed and cost: latency, robots, and serving

17:28 Self-improvement, and human + AI vs AI alone

22:10 Life with unlimited Astra computer use

Aaron Lou: https://x.com/aaron_lou

Lily Zhang: https://x.com/lily_gpupoor

B-roll: Astra paints the Golden Gate, via @cdngdev (https://x.com/cdngdev)

SOTA TV, a new podcast with leading minds in AI. Madison Kanna and I sat down with Graham Neubig, CMU professor and OpenHands co-founder. We talked about his new AI agents course at CMU, which AI agents will still be around in five years, why he believes human-AI teams will beat AI alone for a very long time, and how research and conferences should change as submissions explode.

He is one of the most agent-forward professors in AI, but he doesn't have a personal agent yet because of privacy concerns. Graham leads OpenHands, one of the most popular open-source coding agent platforms. It just hit 90,000 GitHub stars.

Guest: Graham Neubig, https://x.com/gneubig
Host: Lily Zhang, https://x.com/lily_gpupoor
Producer: Madison Kanna, https://x.com/Madisonkanna
OpenHands: https://github.com/All-Hands-AI/OpenHands

Chapters
0:00 Trailer
0:40 Intro
1:10 The agent tier list: which AI agents will be around in five years?
3:37 What makes an agent sticky?
5:01 Will personal and work agents merge?
6:19 Where OpenHands fits
7:56 Graham's new AI agents course at CMU
14:28 ICLR hits 62,000 submissions: reshaping the reward function for good researchers
16:34 How universities can keep up with AI
18:39 Which parts of research should we automate?
21:49 Can conferences fix the paper inflation?
25:31 Open vs closed weights

More conversations from COLM 2026 coming this week!
#AIagents #OpenHands #COLM2026 #LLM