The regulators arrive, and the error rate climbs to 34%

The read

The regulators moved, and on four fronts at once. The EU AI Office has sent its first formal requests for information to model makers — the first enforcement step under the AI Act, confirmed by Henna Virkkunen. An RFI is not a ban, and worth reading as the opening of a process rather than its conclusion. ChatGPT, Reddit and Roblox will be classified as very large online platforms under the Digital Services Act, each having passed 45 million monthly EU users — which puts a general-purpose assistant under a regime written for social platforms. The Bank of England's Andrew Bailey warned frontier AI "may have the ability materially to alter" cyber risk to the financial system. And the UK Greens want data centres stripped of the blanket critical-national-infrastructure status granted in 2024, on water grounds — a planning privilege, not a ban, but a meaningful one.

The politics is now electoral on two continents. The US data-centre backlash has become a midterm problem for both parties — with Republicans warning their own side in an NRSC memo, and the President criticising communities that oppose the build. Two polls disagree on the magnitude and we have carried both. In Australia, the Herald reports Labor calling AI the country's biggest sovereignty moment since the Second World War, attributed to Richard Marles — paywalled, so we have the framing as reported and have not characterised the policy behind it.

The money kept finding hardware and attention. OpenAI says its advertising business has reached a US$1bn annualised run rate, roughly 200 days after it began testing ads in ChatGPT — the company's own figure, presented as proof of the model. Nvidia is committing US$3.5bn to convertible bonds issued by MediaTek, pulling it into NVLink Fusion: debt instruments rather than equity, and the fabric is the load-bearing part. And CXMT says it beat the DRAM incumbents to LPDDR6 mass production — a Weibo self-claim with a September teardown as the falsifier, but if it holds it matters more than most of this week's silicon news.

The most useful number came from someone running the tools rather than selling them. Three months of operating an agentic harness took the error rate from 0% to 34% — zero for six weeks, then 15%, then 34%, and not because the model changed. That is the shape of the problem the harness argument keeps circling. Alongside it, OpenClaw 2.0 adds shared cloud sessions that keep an agent's memory across team members, with The Register questioning the security of exactly that, and a proposal to make agent memory a file format rather than a pipeline — a zip of markdown with an optional vector index, and that is the whole specification.

On the security side, Anthropic has moved against hijacked accounts being used for unpaid inference, harvested by infostealer malware that defeats MFA via session cookies. The report establishes unauthorised use, not resale — but if stolen inference ever becomes a traded good, it would price the underlying capability better than any vendor pricing page.

Three arguments to close, none of them settled. John and Mark disagree over whether Apple should be selling inference racks — recorded as an argument rather than a finding. A commenter's claim that if everything stopped at GLM-5.3 it would take the median programmer five years to catch up, which is one person's estimate and a useful way to frame the diffusion gap. And John returns to his own line: hire more designers, because engineering is not the constraint — a position he put six weeks ago and a week of conference conversation has not moved him off.

Read all 16 signals →