The Trillion Parameter Consortium, and a Vibe Shift

Trip 28 to 31 July 2025 · Written up 31 July 2025 · Last expanded 29 September 2026

Nine days, three towns, and three first-of-their-kind conferences, starting at TPC25 in Santa Clara.

The Trillion Parameters Consortium gathered researchers from the major national labs, all working the same question from different angles: how AI could integrate into large-scale physical science experiments, and eventually conduct them (conduct, as opposed to assist). It reminded me of work I had done at Georgia Tech two years ago, and it was a strange feeling to watch a room full of people arrive independently at something I had circled without quite naming.

It was a hoot, and I believe something important was started there. (I said at the time I would write it up shortly and then did not, because the networking ran long and there were people to see at Stanford and Cupertino in the evenings. This is that writeup, several months late.)

Energy in, insights out

What struck me most was a shift in framing, which I heard on both sides of a divide that usually does not share vocabulary.

Many of the national-lab discussions at TPC25 treated the whole thing as an energy-to-insights production chain (energy in, insight out, with everything between as plumbing) as opposed to the usual compute-to-results framing. Then I went straight from there to the Agentic AI Summit at UC Berkeley and heard the private sector describing the same chain in almost the same terms.

As far as I can tell these realisations are contemporaneous and not causal, i.e. neither side got it from the other. That pattern has shown up before at technology punctuations, through Web 2.0 and again at the start of GenAI, and it usually means the underlying constraint has become visible to everyone at once because it started binding.

In a way it puts us closer to what innovation looked like in America during the PC era, with public and private both contributing to the same expansion instead of one buying outputs from the other. I am not certain that holds, but if the binding constraint really is energy and not algorithms, the labs and the hyperscalers are going to keep converging whether they intend to or not.

Who I met

• Ed H. Chi, Google DeepMind. Two days later at the Berkeley summit, a small group of us got an extended and unexpectedly candid sidebar with him. He co-authored the 2022 Chain-of-Thought paper, and his group had just taken an IMO gold medal alongside OpenAI’s entry. The point that stuck with me: we train models heavily on what to do and almost never on what not to do, and that omission is what licenses them to hallucinate, or to “dream,” as he put it, because they were built to fill in blanks. He recast hallucination as a position on a human-to-machine imagination spectrum rather than a defect to be engineered out. I asked him about formal reasoning that goes past Olympiad-style problem solving toward real-world deductive rigour, and specifically whether such systems need explicit a-priori modules for identifiable thought-routines, a neuro-symbolic design, or whether those routines can be learned from data alone. That one is still open. I am still chewing on it.
• Stanford and Cupertino, the evenings. To be filled in.

Photographs

Links

The TPC25 programme, the Chain-of-Thought paper, and the Berkeley RDI Agentic AI Summit recordings. 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.