A week in the cursed town: SciFM25 in Ann Arbor

Trip 28 to 30 May 2025 · Written up 30 May 2025 · Last expanded 29 September 2026

I spent a week in the cursed town of Ann Arbor (O H!), a thing a Buckeye says out loud and then quietly regrets when the work turns out to be good. I had been tipped off a few months earlier that U-M is laying claim to being the premier center for scientific foundation models in the nation, and SciFM25 was where I went to check.

By the time I sat down for Bill Dally’s talk I had to admit I have hit the conference circuit too hard, because this was the third Nvidia talk of his in a year. Georgia Tech late last spring on the start of GenAI, then GTC25 interviewing Yann LeCun on the longevity of Transformer based LLMs, and now SciFM on how Blackwell chips apply to training science foundation models. (The fourth turned out to be the Berkeley RDI Agentic AI Summit, one short of Bill Dally bingo.)

Why still chase the node?

The exchange I keep coming back to happened in the Q&A, when a professor from Michigan asked Dally why NVIDIA still invests so much effort in node improvement at the semiconductor level, when so much is already possible at the algorithmic level with existing chips on current nodes, i.e. why chase diminishing returns down in the hardware.

Dally cited a 2.5× efficiency gain from moving to smaller nanometer nodes, and his point, from what I heard, was that the 2.5× is about energy efficiency as much as performance. When policymakers in D.C. ask how we stay ahead of China, this is one of the few quantifiable advantages NVIDIA can still push forward at the hardware layer, and he noted that China has made do with larger-nm chips that are less energy efficient.

What he did not say is that China’s comparative energy production costs are structurally lower, and increasingly so. Huawei’s CloudMatrix architecture shows how bundling technologies on less advanced silicon can still yield competitive or superior performance when the underlying energy and manufacturing inputs are cheaper and more abundant. If the American advantage sits only at the bleeding edge of chip research, and China has scale advantages in energy and integrated stack deployment, then NVIDIA has to run very hard just to stay in place (the Red Queen problem).

The silver lining for Nvidia is domestic: if they reach some kind of stability under that pressure they will profoundly dominate the competition inside the US ecosystem, and I would not be surprised if they go parabolic.

El Agente, and the incentive underneath it

Dally also walked through El Agente, a RAG-like (?) framework for QC and MatSci where a 1 billion parameter model with tool use beats a 1 trillion parameter model without, though I am not sure that always needs to be true. Nvidia has staked a claim on these frameworks as an early engineering workaround to minimize hallucinations, and there is an incentive underneath: it is of course great for Nvidia if you must hit the foundation model several times for one good output, and better still when a system has multiple agents all doing it.

Accountants, calculators and Polanyi’s Paradox

Thomas Mason, of Los Alamos, responded to a familiar worry. If AI now handles the mechanics of physics, the Feynman-like tricks that resolve a nasty triple integral, will the next generation ever learn any of it? He likened today’s students to yesterday’s accountants who reconciled ledgers by hand and who, once calculators arrived, upgraded their skill sets instead of eulogizing long division.

Polanyi’s Paradox is lurking in that analogy, since much of what we rely on day to day is tacit knowledge we cannot fully write down, and losing the practice costs you the intuition you never knew you were using. I hear the same tension on campus. An applied math professor teaching Complex Analysis told me over lunch at Georgia Tech’s “The Science of Intelligence” symposium that she was having real trouble thinking about testing in this early era of LLMs, and my suggestion was to ask indirect questions that surface the reasoning moves models still botch (and they botch a lot).

What we still need is a strategic audit of the non-negotiable human skills. Trigonometry looked obsolete the moment graphing calculators hit classrooms in the late ’80s, and yet fluency with the identities still powers creative leaps in analysis and physics. The gnarly part is that we may not know in advance what will turn out to be useful, which makes it hard to decide what we gladly offload and what students must master with their own neurons.

I left impressed by Michigan’s momentum, which is a hard sentence for a Buckeye to type. Ohio State friends, what is our next move in quantum information and AI? I would love to see an equally visible cross-college effort of our own.

Who I met

• Bill Dally, Nvidia. Talked about Blackwell chips applied to training science foundation models, defended node improvement with a 2.5× efficiency gain, and framed it as energy efficiency rather than performance alone.
• Thomas Mason, Los Alamos. Answered the worry about AI doing the mechanics of physics by comparing today’s students to accountants who upgraded their skills when calculators arrived. It stuck because Polanyi’s Paradox sits right underneath the analogy.

Photographs

Links

Dally’s SciFM25 talk, the El Agente framework, Huawei’s CloudMatrix architecture, Georgia Tech’s “The Science of Intelligence” symposium, and Polanyi’s Paradox. 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.