This week the AI story stopped being about who has the smartest model. Washington quietly un-banned Anthropic's most restricted system, OpenAI floated giving the government a stake in the company, and a cheap Chinese model closed the gap on agentic coding. Here's what actually moved, and what it means if you're the one making the buying decision.
The Super-Trend Board
Models, Labs & the Capability Frontier — Advanced. What moved: Washington lifted restrictions on Anthropic's Fable and Mythos, a low-cost Chinese model closed in on agentic coding, and Nvidia's TwoTower started generating text in parallel blocks instead of one token at a time. So what (Capability-to-Consequence): the model that wins isn't the smartest one, it's the one you can still reach next quarter. Now what: build for multi-model portability, not single-vendor dependence.
Safety, Trust & Regulation — Advanced. What moved: OpenAI proposed a 5 percent government stake, the EU upheld a 4.1 billion euro fine against Google, and youth online safety legislation passed the U.S. House. So what (Systems-Over-Models): governments are becoming regulator, customer, and shareholder at once, and that changes how neutral any ruling can really be. Now what: watch how AI companies get treated relative to how platforms got treated. The playbook is transferring. I’m honestly not sure the Labs understand the transfer here, Google and Gemini are certainly the most immediately likely to notice.
Workflows, Not Models — Advanced. What moved: Argentina (remember that proposal to let AI run companies incorporate there?) proposed legal AI-run companies (now with a human on the hook for outcomes), Base44 launched its own model, and Meta explored selling spare compute. So what (Systems-Over-Models): the moat is moving from "which model" to "who owns the system around it." Now what: audit which parts of your stack you actually control versus rent. The scaffolding and harness and the things that make things work the way that you want is what every company will be (at least should be) looking to own.
Innovation & Health/Science — Advanced. What moved: Claude Science launched as a sustained-workflow research platform (look, another harness), COMPASS started predicting immunotherapy response across cancer types, and new research found medical AI privacy can look fine on average while failing rare cases. So what (Translation Layer, for a health system): a privacy audit that only reports the average is reporting the wrong number for your most vulnerable patients. Now what: ask your AI vendors for subgroup-level testing, not aggregate scores. This is a long known challenge in Healthcare analytics, it didn’t go away just “because AI”.
The Human Layer — Advanced. What moved: entry-level hiring kept softening even without mass layoffs, while a first-time founder used AI to build a 16-person company serving foster children. So what (Primary vs Secondary Effects): the first visible damage from AI in the labor market may be a missing rung, not a pink slip. Now what: rethink how junior talent gets trained if the entry-level job that used to teach them is quietly disappearing.
Compute, Capital & Power and Embodiment & the Physical World held steady: Meta explored monetizing spare AI compute, and China's Unitree filed for a $619 million Shanghai IPO. Worth noting, not a shift.
A few highlights
The most Chuck-brained headline of the week: OpenAI wants the government to co-own its horse in the AI race. The whole story, including why Anthropic says it hasn't had that conversation, is in this week's episode.
Argentina's answer to "can AI run a company" is basically "sure, but someone still has to take the blame." (The longstanding role of designated scapegoat remains) The full legal wrinkle is in this week's episode.
A founder with no tech background used AI to build a business plan, raise funding, and hire sixteen people to help kids in foster care. It's the most optimistic story in the pile this week, and it's in the episode.
One question
Would your organization switch AI vendors today if cost and access mattered more than raw capability? Does it already? What about the harness? Reply and tell me.
Watch the full episode: on YouTube
What would you like to see more of? Less of? Drop a comment or send a reply.
Chuck
Chief Technology Troublemaker chuckdevries.ai