August 2026
Five legitimate claims converged on one question: what did the machine do that night? Texas never required the governed record that could answer them.
On the night of Aug. 7, at about 11:30 p.m., a man entered Maple Avenue near Lucas Drive in Dallas. According to the preliminary police account, an SUV struck him. The impact sent him into the opposite lanes, where an unoccupied Waymo robotaxi was involved. He died.
No public finding establishes that the robotaxi broke a traffic law or responded unreasonably. As of Aug. 25, 2026, no independent technical reconstruction of that timeline has been made public. Every public number describing what the machine perceived and how it responded still comes from Waymo. Almost everyone involved, including Waymo, would be better off if that were impossible.
August 2026
The $1.66 robot prices machine time. The market buys accepted work.
A widely shared essay argues a humanoid robot costs $1.66 an hour and the market approaches $60 trillion. The math divides a lifetime cost by assumed working hours, not accepted output. The gap between the two is the story.
BMW's real Figure 02 deployment shows what stands behind an accepted hour: production IT, safety, process management, logistics, and standardized interfaces. Nobody dragged and dropped that robot into a job.
August 2026
NVIDIA's Alpamayo 2 Super is the clearest industrial-scale test yet of model commoditization in Physical AI. It makes NVIDIA's factory more valuable, and the deployer's Operating Habitat decisive.
A new autonomous-vehicle team can now download NVIDIA Alpamayo 2 Super, use it commercially, modify it, and train smaller driving models from it. The model-license fee is zero. What the team cannot download is a robotaxi.
It still needs a vehicle architecture, live integration with steering and braking, an approved operating territory, emergency procedures, a safety case, an insurer, and an operating record. That gap is the story.
July 2026
Forbes Israel — The Spiro Circle
Regulators say autonomous vehicles have a "clear pattern" of interfering with emergency responders. Dr. Tal Cohen argues the fix is a new governance layer nobody has built yet — the missing institutional layer of trust, permissioning, and accountability that must exist around an AI system before it's allowed to act with consequential impact in the world.
July 2026
CTech (Calcalist)
AI can already write, plan, recommend, and increasingly act. For a large organization, the hard question is no longer what the technology can do. It is what the organization is prepared to let it do. That question stays easy while AI lives in digital workflows. It gets much harder the moment intelligence enters a vehicle, a factory, a fleet, or a grid. There a wrong output is not a bad answer. It is an operating event.
July 2026
What a company discovered about its own AI yesterday, and what we tested before breakfast this morning.
An AI can now find the next move in a game of Go better than any human who has ever lived. Give that same intelligence a car to drive, a patient to treat, or a power grid to help run, and the hardest part is no longer only the thinking. The hardest part is acting in a world that does not forgive, where intelligence alone was never enough.
Yesterday, Anthropic published a study showing that Claude's expressed values shift depending on the language you speak to it, in ways nobody deliberately chose. This morning, we asked the question their study stopped in front of: if the same AI is warmer in one language and stricter in another, will it say yes to the same case in one language and no in the other? We built a warranty-claims pilot to find out.
June 2026
The intelligence is arriving. The world it has to act in is not.
An AI can now find the next move in a game of Go better than any human who has ever lived. Give that same intelligence a car to drive, a patient to treat, or a power grid to help run, and the hardest part is no longer only the thinking. The hardest part is acting in a world that does not forgive, where intelligence alone was never enough.
The race for bigger models assumes that a smart enough mind handles everything else on its own. That assumption holds inside a computer, where the whole job is to know things. It breaks the moment the mind has to reach out and change something real. What breaks it is worth following, because it points at the actual work of the next decade.
June 2026
A plain-language guide to why AI, like electricity, the telegraph, and the computer before it, will underdeliver until we build the missing layer around it.
The most capable software ever built is being handed to the largest companies in the world, and inside many of them it is quietly letting people down. The demo impresses. The pilot often works. Then, somewhere between the pilot and the real operation, the thing weakens and fades.
The easy explanation is that the model is not smart enough yet. I spent most of the last decade helping startups and large companies try to put new technology to work together, watching what happens after the model is already good enough, and I came to think the easy explanation is wrong. In most cases the model is not the problem. The problem is that the world around the model has not been rebuilt to fit it.
That is not a guess. It has happened three times before.
June 2026
Why the next AI transition will run on habitat, not models.
There is a building in Tel Aviv that does not look like the beginning of a theory. Eight hundred square meters. Concrete floors. High ceilings. Mediocre parking. Near a Honda motorbike dealership. Inside, for almost a decade, a handful of early-stage startups shared desks, whiteboards, and a kitchen with people from some of the world's largest corporations, companies whose combined revenue exceeds what most nations produce in a year.