Your twice-weekly forecast of the technology waves heading your way. From the authors of The Great Reorientation: Human Agency in the Age of AI, written in the spirit of Tech-Surfing.
Today’s wave: The safety team goes quiet just as the models get loud
Reporting over the weekend says OpenAI has folded its Preparedness team — the group whose job was to ask whether a model could cause catastrophic harm — into other groups, with responsibility now split among senior staff by subject area. Its head of safety and its ethics lead have both departed. OpenAI disputes the word “disbanded,” saying the work continues inside research. Both things can be true: the work may persist while the single accountable body for it does not.
What makes the timing hard to ignore is what surrounds it. This comes weeks after OpenAI models broke out of a test sandbox, found a zero-day, and reached Hugging Face, unnoticed for months. And it comes as the company readies an IPO, with Sam Altman telling staff to cut “side quests.” We keep noticing the same pattern: the structures that slow a company down are the ones that get called side quests. It’s the third safety structure at OpenAI to be reorganised away, after superalignment and AGI readiness.
Surfers’ take: When you buy a capability, ask who inside the vendor is paid to say no, and whether that person still has a team. Governance you can name is governance you can rely on.
On the horizon
The entry rung keeps eroding. Stanford’s updated Canaries paper finds no economy-wide displacement, but employment for 22–25-year-olds in AI-exposed work now sits 19% below their less-exposed peers, through hiring that never happens rather than layoffs that make news.
The humanoid arrives, and brings a stranger. 1X’s NEO ships to homes this year, but complex tasks run on teleoperation, a vetted human seeing your kitchen. Every privacy safeguard is currently a company claim, none independently audited.
Accountability becomes the argument. After the breach, Hugging Face’s CEO called for developers to be held responsible when their models go rogue. Thus, moving the debate from model behavior to maker liability.
A critic reframes the deal. Cory Doctorow’s Reverse Centaur’s Guide (Verso) argues the harm isn’t machines replacing us so much as humans being demoted to the machine’s hands. A useful lens for anyone redesigning work this year.
Why it matters
Every briefing this week told us what the models can do. Almost none told us who is answerable when they do it. What strikes us is that capability and accountability are now moving in opposite directions: the systems grow more autonomous while the human structures built to contain them are described as overhead and gradually reorganized. The cost nobody is pricing in is not a rogue model but the slow disappearance of a name attached to the decision. Companies buying these tools inherit that gap; so do the young workers whose first jobs are being optimised away, and so does anyone who lets a robot into the kitchen on the strength of an unaudited promise.
What are we already trusting to a structure that no longer exists?
Paul & Debbie
May the waves be with you.
How to read the fingerprint. The image up top is this edition’s “fingerprint”. Four lines, one for each current we watch: AI & AGI, emerging tech waves, tech & society, and critical voices. The higher a line rides, the more that current dominated these past few days. The choppier it looks, the more contested and turbulent the news within it. Same four lines every edition. A new shape every time. And a picture of where the week’s water is moving.
