Ed Chi on machine dreaming, and a vibe shift at Berkeley

Trip 2 August 2025 · Written up 2 August 2025 · Last expanded 29 September 2026

The first thing I put up from Berkeley was a photograph and four words, “Do you see Berkeley?”, and I left it at that, because the part of that visit worth writing down was a sidebar and not anything on the programme.

It was the middle stop of a nine day whirlwind, three towns and three firsts. The week started at TPC25 in Santa Clara, ran through the Agentic AI Summit at UC Berkeley in August 2025, and ended at a quantum AI conference in Bloomington, Indiana. The other two towns have their own write ups.

What we train models to do, and what we never train them not to do

A small group of us ended up in an extended and unexpectedly candid conversation with Ed H. Chi, Distinguished Research Scientist at Google DeepMind. Chi co-authored the landmark 2022 Chain-of-Thought Prompting paper, and his group’s newest foundation model had just earned an IMO gold medal alongside OpenAI’s entry, so I expected something of a victory lap and instead got a description of a blind spot.

He said we focus heavily on what to do when we train these models, and rarely on what not to do, and that this omission is what encourages an LLM to hallucinate, or to “dream,” as he put it, because the thing was built to fill in blanks and nobody ever taught it which blanks to leave alone.

From there he recast hallucination along a human-machine imagination spectrum, and another participant pushed that into what it might say about human imagination, which I have been chewing on since. If you call it a defect you have assumed the system was supposed to be a database, and if you call it dreaming you have assumed it was supposed to be a mind, whereas a spectrum lets you ask how much of it you want, which is a more useful question than how to get it to zero.

The inductive-deductive divide, and the question I asked

The longer thread was the inductive-deductive divide and when Chi thinks it could be breached. The notion (for anybody who has not run into the term) is that a model which surfaces its intermediate reasoning steps can be nudged from mere pattern completion toward principled logical inference, narrowing the gap between statistical learning and formal cognition. We also went round on whether genuine analytical reasoning can emerge from training data alone, or whether it still leans on retrieval and analogy.

My own question was narrower. Olympiad-style problem solving is a beautiful benchmark but it is not where real-world deductive rigor lives, so I asked him about formal-reasoning performance that goes past it. Underneath that sits the design question of whether you build these systems with explicit a-priori modules for certain identifiable thought-routines (a neuro-symbolic design) or expect those routines to be learned from data, and I was more interested in what he had observed than in what he thought.

A vibe shift, and something closer to the PC era

Across that week I picked up what I can only call a vibe shift, running the length of the production chain and through both the public and the private sector, though nobody said it from a podium. The private side in the Berkeley room had arrived at the same picture I had heard from the national labs a few days earlier, that what we are all building is an energy to “insights” production chain.

The interesting bit is that the two realizations look contemporaneous and not causal, i.e. neither side taught the other. That has been true of most of the tech punctuations of the past two decades, Web 2.0 and the opening of GenAI included, and it puts us closer to what innovation in America looked like during the PC era, with both sides contributing to the same expansion instead of one waiting on the other’s leftovers.

The question I left Berkeley with is a funding question and not a research one: if public and private are going to reach the same insight at roughly the same time anyway, I am not sure what the argument is for the sequencing our institutions are still built around.

One postscript, from a separate trip. Later in the autumn the Tianqiao and Chrissy Chen Institute and the UC Berkeley College of Computing, Data Science, and Society invited me out to the AI for the Accelerated Science symposium, one last run to Cali before the end of the fall and a nice capstone to a year of catching up with AI for science (I am working in it now and not just reading about it, which changes how you listen in a room like that!). The only thing I posted from it was a short clip captioned “Human-machine interface,” which tells you where my attention went.

Who I met

• Ed H. Chi, Distinguished Research Scientist at Google DeepMind, and co-author of the 2022 Chain-of-Thought Prompting paper. He said we train models heavily on what to do and rarely on what not to do, which is why they dream, and he put hallucination on a human-machine imagination spectrum instead of in the defect column. Also his thoughts on when the inductive-deductive divide gets breached, which is the reason I went.
• The Tianqiao and Chrissy Chen Institute and the UC Berkeley College of Computing, Data Science, and Society. They invited me to the AI for the Accelerated Science symposium later in the fall, which turned a year of reading about AI for science into a year of sitting in the rooms where it gets argued about.

Photographs

Links

The 2022 Chain-of-Thought Prompting paper, the Sequoia Capital podcast episode where the OpenAI team gave the down low on their own approach, and the programme for the AI for the Accelerated Science symposium. To be added.

Open entry
This page is not finished, and is not meant to be. I add photographs, people and links as I dig them out, so it keeps growing after publication. Why the whole site works this way is on the colophon.