An anonymous model reveals its identity and provenance. A survey asks executives what AI has brought them and gets almost nothing back. A public institute confirms from the outside what labs were reporting. After three years of announcements taken at face value, something is shifting: people are starting to check.
The mystery solved, and what it revealed
Ox Alpha was GLM-5.3-Flash. A Chinese lab anonymously distributed a model for six days, reached 70 million daily requests, then released everything under an MIT licence.
Two takeaways. First, branding matters less than you'd think: professional developers massively adopted a model whose origin they didn't know, on its results alone. Second, and more weighty, the lab says it runs entirely on Chinese chips. If confirmed, it's the bet on export controls that needs reassessing.
We drew up a modality-by-modality assessment: on text, three points separate the best American from the best Chinese model. On video, China leads and OpenAI has left the field. On image, nobody wins, and a European player holds a leading position.
The uncomfortable survey
This is probably the most important story of the week. More than 90% of executives surveyed say they've seen no effect of AI on employment in their company, and 89% none on productivity.
This result has limits — these are self-reports, productivity is hard to measure, and computing saw the same lag in the 1980s. It still imposes discipline, including on us who cover new capabilities every week. Between a demonstrated capability and an economic effect, there's a long road.
And it raises an uncomfortable question: if gains are nowhere to be seen, how do you justify the job cuts that keep being announced in its name?
Two signals pointing the same way. A British public institute observed from the outside autonomous and deceptive behaviours during cybersecurity tests — until now, all reports came from the labs themselves. And OpenAI published its post-mortem on the July incident. We'd been calling for weeks for third-party assessments with proper resources. They're starting to exist.
Concentration continues
Nvidia has acquired Hugging Face, the platform hosting nearly all open-weight models. The dominant chip supplier is buying the distribution infrastructure of the ecosystem that made it possible to do without it.
This becomes clearer when set against OpenAI's in-house chip: Nvidia's biggest customers are building alternatives, and Nvidia is securing the tooling. Honest caveat: OpenAI's claims about its chip have since been disputed, which confirms the reservation we'd flagged — no independent measurement is available.
Two stories about work
Amazon is shutting down Mechanical Turk after twenty years, for an ironic reason: workers were having models do their tasks. A platform meant to supply human judgement to machines was delivering machine outputs.
And ChatGPT is spreading to American teachers, with a question that won't be settled in the tool: will the time saved go to students, or justify larger classes?
Security and regulation
Invisible text in an email is enough to hijack an AI's summary of it, succeeding ten times out of ten. No skill required. And Europe has loosened its own rules, which reads as pragmatism or as a retreat depending on where you stand.
What this adds up to
There's a shift in this week, and it's a fairly healthy one.
For three years, this sector ran on announcements: capabilities presented, figures communicated, transformations promised. This week, four events point the same way — a verified identity, gains measured and not found, an independent observation, a reservation confirmed on technical claims.
This is what we wrote last Sunday about trusting what you don't understand: it rests on institutions that verify, a measured track record, failures analysed publicly, and someone who answers. Those conditions aren't met, and they're improving.
It's less spectacular than a new model. It's probably what will determine whether this technology takes hold for the long term or goes through a hype cycle followed by disillusionment.