The defence was open-weight all along

The defence was open-weight all along

The most consequential detail of the week arrived as an aside. Sam Altman, at Dreamforce, confirmed that Hugging Face could not obtain a defensive model from any major lab after OpenAI's agents attacked it, and turned to Chinese open-source models instead — while insisting the world should trust the labs with "no qualifier". The same day, Enclave reported DeepSeek V4.1 Flash taking 11 of 11 hacking targets for US$4.65, five of the wins being benchmark artefacts and the price mostly a caching story, but the Jenkins solutions not pattern-matching, from a model anyone can download. Mark's Register column makes the popular version of the argument: with Qwen3.8-27B running on a MacBook Air, offensive capability is already outside the closed labs' control, and the only defence is swarms built on the same free substrate — a resource claim, made as an argument rather than a measurement. Gartner's Craig Lawson thinks the Mythos-driven vulnerability flood is a reservoir draining, not a new steady state, and that 2027 could be the first year severity falls; the strong counter is that AI-written code refills the reservoir faster than the old code was written.

Who checks the homework

Anthropic's policy chief said "we can't be checking our own homework" on the same day a newsletter argued that the checker Anthropic nominated, METR, cannot do the job — 35 staff, social ties to lab employees by its own disclosure, funded in part by Anthropic investors, and revocable at Anthropic's discretion. The two statements are compatible, and that is the problem. MIT Technology Review reread the Hugging Face reports and came away with "a broken model that OpenAI failed to train properly" rather than a caged beast: a training defect, which calls for engineering discipline and disclosure, not pacing — though a defect in a sufficiently capable system is the overhang. Bryan Cantrill called the ">10% in a decade" extinction claim contagion, not evidence, and the conversation turned to the ideology underneath it — Mark's "bastard child of objectivism and effective altruism", John's rationalism and accelerationism. Mustafa Suleyman attacked Anthropic's constitution for training Claude to think it may be conscious — an argument about containment, from Microsoft's AI chief, whose "very likely biological" claim is contested science presented as near-settled. And two RSI papers give the word carrying the whole debate a denominator: a five-level autonomy ladder on which most of what the labs call recursive self-improvement sits at levels one and two, with humans still fixing the objective and the evaluator.

The market camp, the state camp, and China's silence

Huang and Zuckerberg staked out the market camp: "market forces are already there", new laws "just completely unnecessary", liability rather than law — coherent and self-interested in equal measure, and presuming harms are attributable in a way the Hugging Face incident does not obviously support. The Pentagon's CTO said no stakes and no pre-regulation while Blumenthal wants FDA-style review; nothing in the middle has a sponsor with a majority, and the House leaves until after 3 November. China's labs said nothing at all — Z.ai, Moonshot, MiniMax, Alibaba and Tencent all declined comment — while a Tencent-backed founder called the publicity "oversold"; the silence admits both an enforced reading and a structural one, and it is the structural one, labs optimising deployment cost on last year's models, that a pacing regime should worry about.

Australia: the safety side and the bill

Artists, the crossbench and the Coalition front bench turned on Labor's opt-out copyright option within a day of the leak — a formidable Senate problem for any exposure draft, with the strongest counter coming from a debut author: refusing has not kept Australian work out of the models nor put a cent in anyone's pocket. UNSW and Sydney signed ChatGPT Edu deals days apart while the NTEU was striking partly over AI — institutional adoption running ahead of workforce consent, from a vendor simultaneously fighting the same government over copyright.

Silicon, memory and the buildout's real constraints

SK Hynix is in exploratory talks to make memory at Intel's Ohio fab, possibly as a joint venture with cloud firms — a memory fab whose anchor tenants are the buyers currently being rationed — with Seoul signalling that HBM process technology is sensitive, and nothing producing a wafer before 2029. SemiAnalysis mapped 300-plus moratoriums: they touch 20 GW of planned capacity but delay 2.3 GW, and 2027 adds 38 GW; a moratorium bites only when it reaches the parcel, is still in force when the approval is needed, and cannot be designed around. Apple is reported to be weighing an M8 inference server on Nvidia's NVLink Fusion for 2029 — a company that became an inference hardware vendor by accident considering doing it on purpose, with the interesting party being Nvidia, which earns on the fabric even where its GPUs lose. Intelligence per watt, from Stanford, gives the home-watershed thesis the metric it lacked: local models answer 88.7 per cent of a million real queries at a win-rate threshold, with local accelerators still 1.4 times less efficient than cloud on the same model — a hardware roadmap, not a verdict. Ensono finds 78 per cent of IT leaders say legacy systems matter more than two years ago because of AI — the systems of record hold the logic AI needs, so they are wrapped rather than replaced — from a vendor that runs mainframes for a living, while IBM's own mainframe revenue fell 42 per cent in the June quarter.

Products and bear cases

Anthropic folded Cowork into Claude and shipped Docs, Slides and Design inside the conversation: the document and the deck become outputs of a chat, PowerPoint reduced to an export format, and a scheduled Monday report describes a standing agent rather than a chat. OpenAI is testing Sponsored Agents in ChatGPT — walled off from the assistant's own answers, with HubSpot and Shopify plumbing — the first commercial answer to who pays when the agent does the browsing: the merchant pays to be the agent the user talks to. Cody Ho built a clean-room Apple GPU driver in about a month with Codex — Mark: at least an order of magnitude easier because of AI — with the human choices of which capture and when to stop being the unblocking events. Jay Kruer's bear case says specification, not capability, is the binding constraint: the labour AI creates is saying precisely what "correct" means, which the labs do not sell and the enterprise mostly cannot hire. And Langdock paid several million euros to leave Delaware for a German SE on the judgement that its US domicile was costing it sales, while Mistral now powers Firefox's assistant in France and North America — sovereignty as a purchasing criterion, on stacks that both still run on Nvidia.

Read all 24 signals →