Dear reader,
Part 13 is on regulation, and I went in expecting to write about whether the rules are too strict or too loose. The more interesting question turned out to be what the rules can attach to at all, because law cannot regulate a capability. It can only impose obligations on something it can identify, count and audit.
Two things are worth flagging. The first is the deferral pattern in the Digital Omnibus: the prohibitions and the transparency duties landed on schedule, while conformity assessment — the part that actually verifies a high-risk system does what it claims — slipped by a year or more. Not because the risk changed, but because the standards, the notified bodies and the trained auditors do not yet exist. That quietly moves authority from legislatures to whoever establishes practice first, and auditors tend to ratify incumbent practice because they have nothing else to measure against. The second is where the attention sits: the regime regulates the frontier, which is roughly 25 companies, while hiring, credit scoring and benefits — the uses that actually touch citizens — were pushed out to late 2027.
This week's reflections share a shape I did not plan. In each of them, the presentable version of something turns out not to be the true one.
— Jan
This week's essay
Machines That Think, Part 13: Regulation Meets Silicon
The debate about AI regulation is almost always conducted as a question of quantity: too much or too little. But law cannot regulate a capability, only something it can identify, count and audit — which in practice means compute, documents and physical shipments. Article 51 of the AI Act presumes systemic risk above 10^25 floating-point operations of training compute, a line roughly 25 models had crossed by 2025 and one that is backward-looking: it ignores inference-time compute and small models deployed in critical infrastructure. Compute gets used anyway because a regulator can check a reported number against a cloud invoice. And weights cannot be embargoed, because they copy and cross borders as files, while shipments can be stopped at a border — which is why the most binding AI regulation today is trade policy on accelerators rather than the AI Act. Next in the series: the synthesis.
From the reflections
This week's five, in their one appearance in your inbox.
When weighing a commitment — an investment, a research student, a hire — we scrutinise the plan and under-weight the person. The idea will change several times before a good one emerges, so what matters is how someone responds when reality proves the plan wrong. Look at the record: how they have handled being wrong, whether they seek out people who disagree, whether they can name the weakest point in their own argument unprompted.
Work reaches a point where it functions — the code runs, the paper makes its case, the deck covers the material — and that is where most of us stop. It is not laziness: the effort required rises just as the sense of reward falls. But that last stretch separates work people respect from work they merely accept, and as agents make competent output cheap, care and coherence are most of what is left to distinguish anything.
Serious about the work, not about yourself
The people I most enjoy working with have very high standards and almost no self-importance, and will tell you about their own mistakes unprompted. Self-importance makes being wrong expensive, so errors get defended, explained or hidden rather than simply fixed. It also teaches everyone around you not to raise problems. Low ego is what makes high standards sustainable.
Ask someone why they did something and the first answer is almost always true, defensible, and not the actual reason. "I did not have time" is often "I did not want to start, because I suspected it would be harder than I was admitting." The practice is to ask twice: accept the first answer, then ask what else was going on. It arrives quickly, which suggests it was never hidden, only unexamined.
In almost every organisation I have worked with, the people closest to the work could have told you what was going to go wrong, and often did. Nobody hid anything; each summary was a reasonable compression of the one below it. The engineer's three specific problems, two of them without owners, become "a risk being managed" and then "on track with some risk". The specifics evaporate in the second summary.
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.