Blog
Written by the people who broke it.
Engineering notes on things that actually happened — including the ones that were our fault — and long reads on what is moving in AI and technology.
More articles
-
EU AI Act: high-risk rules delayed, transparency duties already live
Regulation (EU) 2026/1744 moved high-risk AI obligations to December 2027 and August 2028. The transparency rules for chatbots, synthetic content and deepfakes have applied since 2 August 2026, with only a short marking grace period for generative systems already on the market.
-
Frontier AI now ships in tiers: what gated access means for builders
Claude Fable 5.1, Claude Mythos 5.1 and GPT-6 Astra all launched in early September with capability split across access tiers. Here is how gated release works, what it costs and how teams building on these models should plan.
-
Hours, not days: the defensive lessons of AI-run cyber operations
Anthropic's threat report, published on 10 September, describes a suspected state-linked espionage group, extortion crews and a lone hacktivist handing reconnaissance, exploitation and data theft to AI agents. The techniques are familiar, but the staffing and speed are not, and defenders need to adjust.
-
Nvidia's $96bn quarter and the shift to custom inference chips
Nvidia reported $96.2 billion in quarterly revenue on 26 August and guided higher while assuming no China data-centre compute sales. Thirteen days later, a Qualcomm filing revealed an Amazon warrant tied to up to $60 billion of chip purchases.
-
AI's power squeeze: PJM's backstop auction and Google's nuclear deal
PJM has asked regulators for a one-off auction to fill a 6,831 MW capacity gap, and for new data centres lacking their own supply to be curtailed before households in emergencies. In Finland, Google has signed for up to half of a nuclear plant's output.
-
FCC robot curbs and Unitree's IPO split the humanoid market
On 28 July the FCC barred new foreign-produced robots from US equipment authorisation, and on 19 August Unitree closed 460% up on its Shanghai debut. Here is how the rule works, what remains unclear and what operators should change.
-
Robotaxi fleets scale up as a federal audit tests the Cybercab
Nevada approved permits for up to 8,000 robotaxis and Waymo reached 14 cities, while NHTSA opened an audit of how Tesla certified its pedal-free Cybercab. An open audit is not a finding, and this piece explains why that distinction matters.
-
FDA's generative-AI device paper: a risk grid, not yet a rulebook
On 18 August 2026 the FDA published a discussion paper on medical devices built with generative AI, with feedback due by 19 October. It is not guidance, but its risk grid and testing ideas offer a practical lens for anyone deploying health AI.
-
Distillation disputes put Chinese open-weight models under scrutiny
US security agencies and Anthropic accuse Chinese AI labs of harvesting American models' outputs at industrial scale, and Beijing calls the claims baseless. Open-weight releases keep arriving regardless, leaving model buyers and heavy API users with new provenance and operational risks.
-
AI training copyright: US backs fair use, Munich rules against Suno
A Munich court has ruled that Suno's music models memorised and reproduced protected compositions, even though training happened in the US. Weeks later, the US Justice Department told a New York court that training language models on copyrighted text is fair use.
-
Fewer junior hires, not mass layoffs: what AI job data shows so far
Stanford researchers report that employment of 22- to 25-year-olds in highly AI-exposed jobs is 19% below where it would be had it kept pace with less-exposed peers. Meanwhile, AI slipped to fourth among reasons employers gave for August job cuts.
-
The export that swapped month and day
A date is ambiguous exactly when the day is twelve or lower. Above that the format gives itself away — which is why the bug hid for so long.
-
The metadata you can never capture again
We hold a corpus of scans and only recently understood what we had failed to record at capture time. Capture the context first.
-
Projection onto a curved surface is a geodesic problem
Standard decal projection is a box projector. It is fast, it is well understood, and it is wrong on anything that curves in two directions at once.
-
Your UI test is lying to you
The tests called the handlers directly. A real trackpad click jitters more than three pixels, and the drag guard swallowed it. Verification has to dispatch real events.
-
The bug that only appeared when someone was fast
The key allocator read the maximum, incremented, and wrote — with no guard between the read and the write. Two parallel writes computed the same key.
-
What an adversarial audit of our own stack found
They found a cross-project boundary that shared a workspace identifier and therefore trusted too much. Publishing what they found is the point.
-
Agents need a chain of command
That boundary is enforced at the authorisation layer, not by prompt, because a prompt is a suggestion and a policy is not.
-
We removed the buttons that did nothing
They were demo scaffolding that had outlived the demo. We deleted them rather than wiring them, because a control that lies is worse than a missing feature.
