Dear reader,
Part 10 looks at AI designing the hardware that AI runs on — the loop that is supposed to close on itself. It is the post in this series where the headline result and the replication record diverge most sharply, and writing it required more care than any of the others.
My conclusion is not that the work is fake, because it is not. It is that what arrived was cheaper search rather than better judgement, and the two are very easy to confuse while iteration is nearly free. Search optimises exactly what you specified and nothing else, so an under-specified objective produces a design that is optimal against the metric you named and quietly wrong against the ones you did not. There is a second cost that I think is underrated: a searched design arrives without its rationale, and review, certification and failure investigation all assume a design can be explained. Norbert Wiener said the essential thing about this in 1960 and we keep rediscovering it.
This week's reflections are all, now that I line them up, about horizons — what compounds, what decays, and how badly we price both.
— Jan
This week's essay
Machines That Think, Part 10: The Design Loop
In 2021 a Nature paper described a reinforcement learning agent placing the macros on a chip in under six hours, with better power, timing and area than human engineers. The self-improving loop had apparently closed. Five years on it is the most disputed claim in the field: Cheng and Kahng at UC San Diego could not get the method to beat simulated annealing, an academic placer, a commercial placer or a human; ACM attached an expression of concern to one critique; Nature added a methodological addendum. The vendors report real gains that are not comparable to each other, and a Synopsys executive conceded that reinforcement learning in core placement and routing "hasn't really panned out". What actually arrived is cheaper search. When iteration costs compute rather than engineer-weeks, the specification becomes the ceiling. Next in the series: agents in the enterprise.
From the reflections
This week's five, in their one appearance in your inbox.
Slow is smooth, smooth is fast
I still think deadline pressure is useful, because it forces prioritisation. But pressure on the timeline is not the same thing as haste inside the work. Rushing the execution — skipping the test, pushing the change without thinking through what it touches — does not save time. It defers the cost, and the rework turns up later. Let the deadline decide what you work on, not how carefully you do it.
People are modest about luck in public, and in private the story almost always turns on one specific conversation with someone outside the usual circle. You cannot manufacture luck, but you can increase the number of surfaces on which it might land. A professional calendar crowds this out with familiar colleagues and familiar topics, so exposure to the unexpected has to be protected deliberately.
Nobody is watching that closely
The spotlight effect: we badly overestimate how much other people notice us. I can recall in detail a question I fumbled in a talk years ago — the pause, the inadequate answer, the particular quality of the silence — and I doubt one person in that room remembers it, because I cannot recall a single comparable moment from anyone else's talk. We are each the centre of our own view and a background figure in everyone else's.
The only thing that compounds for free
Skills decay, fitness disappears, knowledge expires. Reputation is the exception: once genuinely earned it accumulates while you do nothing, because professional life is a repeated game with a surprisingly small cast. The asymmetry is what makes it worth taking seriously. Building takes decades of unremarkable, consistent behaviour, and destruction takes an afternoon.
The half-life of what you know
A good deal of what I was taught as a young researcher has quietly expired, and none of it announced itself. Everyone's answer is that we must keep learning, which is the easy half. The hard half is discarding, because letting go of something you know means accepting that a piece of your expertise is now worthless — and the correction mechanism that operated on you at thirty has largely switched off by fifty.
If your organization is wrestling with any of this, I take on a small number of reviews and workshops each quarter — just reply to this email.