The Week Power Found AI's Address

Welcome to Briefly AI, a podcast by Harry Sharman, created by AI and voiced by an AI synthesis of Harry Sharman. The real Harry had the idea; the synthetic one is doing the talking.

Right. Sunday. Let's talk about the week.

Because if you only caught headlines this week, you'd be forgiven for thinking it was a fairly ordinary run of AI news — some chip announcements, a bit of politics, the usual churn. But there was a shape to it, once you step back. And the shape is this: power showed up. Not the abstract, inevitable, "AI is going to change everything" kind. Actual power. Governments, capital markets, political campaigns, and legal courts — all of them, in the same seven days, reaching into the AI industry and doing something concrete. That's the week. Let's go through it.

The clearest thread was governmental. We've known for a while that politicians were starting to treat AI models the way they treat semiconductors or missile components — as strategic assets rather than software products. This week, it stopped being a policy position and became an operational habit.

The Trump administration asked OpenAI to limit the rollout of GPT-5.6 — hold it to vetted partners rather than a broad launch. OpenAI complied and then, interestingly, said publicly that this kind of arrangement shouldn't become the norm. Which is worth noting: a company simultaneously doing what the government asks and formally objecting to the precedent. That's not resistance — it's a very careful piece of public positioning ahead of an IPO. They want investors to know they pushed back, even if gently.

Meanwhile, Anthropic's situation kept moving. Mythos 5 — which had been shut off for foreign users entirely after a jailbreak scare and concerns about China-linked access — was partially restored. Over a hundred US companies and government agencies got the green light to use it again, including, quietly, their non-American employees. So the off-switch got used, and then got partially un-used, based on a negotiation that happened entirely out of public view, with no formal framework, no appeals process, and no published criteria for what gets you back on the list.

And here's the thing that matters: while all of this was playing out in Washington, Asian AI startups were launching Mythos-comparable models specifically positioning themselves as the version that doesn't come with an export ban. The US government's hard power over its own labs is real — but it has a cost, and that cost is that it creates a gap, and gaps get filled.

The second thread was money, and what money does to priorities.

Both Anthropic and OpenAI are heading toward public markets. Anthropic at a valuation that now sits just under a trillion dollars. OpenAI has its S-1 filed. Both are hiring for optics as much as output — OpenAI brought in the co-author of the original Transformer paper and a White House AI policy official in the same week, which is a very deliberate signal to investors that it has both the intellectual credibility and the political access to navigate what's coming.

OpenAI also announced Jalapeño — a custom chip, built with Broadcom, designed to run AI inference at half the cost of Nvidia hardware. It won't be in production until 2027. But the announcement is for investors now, not engineers: it says we're building our own stack, we're not permanently dependent on a supplier, and our unit economics are going to improve. That's IPO storytelling as much as technology.

And here's the uncomfortable thing about public markets that both companies will find out fairly quickly: quarterly earnings pressure does not care about safety roadmaps. It cares about revenue and margins. The moment these companies have public shareholders, the conversation about what to prioritise gets more complicated.

Third: the liability landscape shifted, visibly.

The Australian court ruling on Google's AI Overviews — finding Google responsible for false statements generated by its own AI summary feature — was the week's quietest but potentially most consequential story. The court's logic was simple: you built it, you trained it, you deploy it, you profit from it, you own what it says. The "blame the model" defence is getting shorter by the month.

If you build anything AI-facing — a customer service chatbot, a product recommendation engine, an AI-generated summary at the top of your website — that ruling is worth reading. It won't apply everywhere overnight. But it's the direction courts are moving, and the lawyers in your building have almost certainly noticed.

And the fourth thread — the one that ran underneath everything else — was about what happens inside organisations when all of this lands on actual people.

Ford's story this week was a good illustration of something that gets underplayed. They had to hire back engineers they'd let go, after discovering that the automated systems that were supposed to replace those engineers' judgment weren't as trustworthy as assumed. The errors were there — they just couldn't be found, because the people who would have spotted them had already left. The lesson isn't that automation fails. It's that institutional knowledge is load-bearing until you've proven it isn't. And proving that takes longer than most timelines allow for.

Meanwhile, the research picture on how workers are actually experiencing all of this remained consistent and a bit grim. Glassdoor tracked a 240% spike in AI-related workplace anxiety over the past year. And Frontiers in Psychology — I flagged this on Saturday — found that what predicts whether employees trust AI at work isn't model quality. It's whether they have somewhere to raise concerns. Whether their voice goes somewhere. Whether decisions happen with them rather than to them.

And about half of workers are still hiding their AI use entirely — because disclosing it attracts a social penalty. Rated lazier. Less competent. Given less visible work. So what you actually have is an enormous amount of AI adoption that's underground, invisible to management, and impossible to learn from collectively. The utilisation metrics look fine. Underneath, the trust architecture is not.

There's a line in one of Harry's pieces that keeps fitting the data: the question organisations should be asking isn't "how do we get people to use AI?" It's "which cognitive work do we actually want humans to keep?" Because the answer to that question determines whether AI adoption builds capability or hollows it out. Ford found that out the hard way. A lot of companies are quietly finding it out right now.

So. What to watch next week.

The Anthropic and OpenAI IPO trajectories are both live. The Asian model gap is real and accelerating. And the question of whether government oversight of AI capability becomes a formal, rules-based process — or stays ad hoc and negotiated in backrooms — is the most important governance question nobody has answered yet.

One more thing. OpenAI said publicly this week that government-restricted model access "shouldn't be the long-term default." They're right. But saying it and having a framework to prevent it are different things. Next week, watch whether anyone starts building that framework. Or whether another intervention arrives before anyone gets the chance.

That's your week. I'll be back Monday. See you then.