This is not a demo, it is a specification — and specifications decide what will be possible in eighteen months far more reliably than demos do. The move from text to action is happening here, quietly.

In your organisation, who would sign off on letting software operate a machine? IT, the executive team, a committee — or nobody, for now?

The word that matters is “gradual”. Nobody decides to take the human out of the loop: they drop out because validating takes time and the machine is right nine times out of ten. The tenth remains.

After how many consecutive clean validations do you actually stop checking? The honest answer is rarely “never”.

This matches what I see on assignment: each agent is tested alone, and behaves correctly alone. The chain is tested by nobody — often because it belongs to nobody on the org chart.

In your organisation, who would answer for an error born of two tools chained together, both of them working correctly?

The sample is broad enough to move past anecdote, which remains rare in this field. Bear in mind it is made up of people advanced enough to configure an agent: the average organisation is a long way behind.

Do you delegate the tasks you master least, or the ones that bore you most? The consequences are not remotely the same.

The terminal is where an agent holds the most power with the fewest guardrails: no confirmation, no undo, no trace a non-technical reader can follow.

Should a tool with no “undo” button be allowed to run without confirmation — or is that precisely where one is needed?

A threshold has been crossed: this is no longer AI used by an attacker, it is an agent producing the attack in the course of a legitimate task. How quickly the report was published is, by contrast, rather reassuring.

When the attack is born of a legitimate task, where does the fault sit: with the vendor, with whoever launched the task, or in the very design of autonomy?

Read alongside the investigation above, not instead of it. Public post-mortems are a healthy practice and still too rare here; this one is nonetheless written by the party concerned, and the gap between the two accounts is instructive.

Should such post-mortems be made mandatory and public, as in aviation? And who would be tasked with verifying them?

What makes this piece worth reading is the accumulation rather than any single case: taken together, the incidents stop being accidents and start describing a category.

At what point does a run of accidents become an insurable risk — and who should carry it, the vendor or the user?

When direct competitors sign together, it usually means the risk is real and regulation is coming. Both motives can perfectly well coexist. Read it against the refusal, noted a few days earlier, to explain how a model gone rogue would be contained.

Can a standard written by those it is meant to constrain protect anyone other than them? I have no settled answer.

The most useful piece in this section, and the least spectacular. It says something simple: a rule written into an instruction is a suggestion, a rule written into access rights applies.

Does your AI policy exist anywhere other than in a document? And if so, where exactly?

If it is confirmed, this is the structural event of the year. Hugging Face is the de facto crossing point of open source in AI; the chipmaker would own the square where you choose the models that run on its hardware.

Does open source still guarantee independence when its main host belongs to the hardware manufacturer?

The word used is “circular”, and it deserves a pause. It does not say the technology is overvalued — it produces measurable value. It says the measurement runs through a largely closed loop.

From the outside, how do you tell real demand from demand financed by the seller?

45 billion dollars is roughly the annual budget of French public research. I cite it to give a sense of scale, not to provoke: we have moved from a competition in engineering to a competition in access to capital.

Should Europe keep aiming for frontier models, or accept building on other people's? Both positions can be defended.

The notable figure is energy efficiency rather than speed: it is what decides the real cost of a large-scale deployment, and the footprint you will soon be asked to account for. These are, however, the manufacturer's numbers on its own workloads.

Would you require an independent test before committing infrastructure on these results — and does a third party capable of producing one even exist?

Free cooling, continuous solar power, no land or neighbourhood constraints: the physical logic holds up better than it first appears. Maintenance and applicable law, much less so. I file this under long-range watching.

Which law applies to data processed in orbit? The question looks remote; it will not stay that way.

The most important fact of the week for the general public. It is worth reading precisely what is written: “assisted”. The hand remains human, and so does the responsibility.

Should that division — augmenting the gesture without replacing it — serve as a model for other sectors, or is medicine too particular to generalise from?

Far less visible than the London operation, and probably more consequential. A model that is wrong remains manageable; a model that is wrong with confidence does not.

In which of your processes is AI currently allowed to answer “I don't know” without it being treated as a failure?

This is not a model, it is an annotation method: the invisible work everything else depends on. In most of the projects I support, the blocker is not the model — it is that nobody was willing to fund the structuring of the data.

Why is the model so readily funded, and the data that makes it possible so reluctantly?

This answers a real blind spot: extreme events are by definition rare, therefore poorly represented in the record — and a shifting climate makes that record less and less predictive.

Can you trust a forecast that rests on no precedent? Local authorities and insurers will have to decide before meteorologists do.

Read alongside “Redwood”, published the next day: an accelerator designed and deployed in two weeks by an AI. Two signals converging on systems that produce results nobody asked for.

How do you assess a result nobody commissioned, and nobody could reproduce by hand?

My favourite piece of the week. A four-year-old has heard a few tens of millions of words; a model has ingested a trillion. This is not to say AI is weak — it does things far beyond us.

If the gap does not close with scale, what exactly are we measuring when we talk about “progress” in artificial intelligence?

For the teachers reading this — and there are several of you — it is the most directly applicable piece here. What I observe in my own courses: banning moves the use out of sight, allowing without a frame removes the effort of learning.

Should we assess the output, or the path to it? And if it is the path, are we equipped to do that with thirty students?

Read this with a cool head: the subject lends itself to alarmism and the studies are young. The cognitive offloading effect is nonetheless well documented, going back to GPS.

Which skills do you accept losing, and which do you decide to keep exercising? It is a trade-off — and it makes itself if you do not make it deliberately.

Here is the business model of AI in higher education, and it is arriving faster than the debates about cheating. You sell access to the brand while cutting the cost of the teacher. The teaching quality may well be perfectly adequate.

What does a student buy for 699 dollars: democratised access to knowledge, or a brand with the human presence taken out?

The neuro-symbolic choice is justified here for a precise reason: when you flag a student as “at risk”, you must be able to say why, both to them and to the institution.

Would you accept an algorithm flagging your child as “at risk” without being able to explain what it based that on?

“It's complicated” is an honest answer here, not an evasion. The uncertainty will probably not be resolved by one unifying ruling, but by sedimentation, market by market.

If a model deployed in good faith were ruled unlawful tomorrow, who should bear the cost: the vendor, the company using it, or nobody?

The diagnosis deserves a hearing. The remedies seem more fragile to me: a robot tax presupposes we can define a robot, and reserved jobs would create a category hard to hold over time.

Does a call to slow down carry the same weight depending on whether it comes from the one who accelerated or the one living with the effects?

A third is considerable, and the technical implication is the interesting one: the next models will therefore train largely on text produced by the previous ones. This is called model collapse, and the critical threshold remains unknown.

Should labelling generated content be made mandatory? And above all, who would be in a position to enforce it?

Read this if you have ever promised someone their data could be removed from a model. The short answer is: far less than is claimed. It leaves the right to erasure in an uncomfortable position.

When erasure turns out to be technically out of reach, should the law adapt or the use be restricted?

Little commented on, potentially structural: one of the first cases where an AI vendor successfully challenges a risk designation imposed by a state.

Who should have the last word on whether an AI system is dangerous: the regulator, the court, or the vendor?

The same shift as in section 01, but this time in the hands of two billion people, with nobody installing anything. This is how agentic AI will become ordinary: by update.

Which of your everyday tools have gained the right to act on your behalf without you explicitly agreeing to it?

The most profitable use of AI for a freelancer or a small organisation remains transcription: little risk, immediate gain, no change management to run.

Is it better to start with what impresses, or with what saves time from Monday morning?

A long-awaited feature, and a double-edged one. An assistant that remembers becomes markedly more useful; it also becomes a place where information about your clients and your files piles up without inventory.

Do you know what your assistant has retained about your files? And if you had to show a client, would you be comfortable?

At 399 dollars, robotics fits the budget of a secondary school, a technical college or a fablab. Physical-AI skills will probably be formed there, rather than in multi-million-dollar labs.

Does the gap between handling and simulating justify the spend, in an already tight teaching budget?

I flag it mainly for the cadence: video generation is advancing faster than institutions can organise themselves around proof by image.

If your work rests, even partly, on the authenticity of a video — would you know how to establish it today?

The best technical piece of the week, and the principle carries well beyond the technical. A colleague handed the entire file works less well than one told what is a constraint, what is a preference and what is an assumption.

Do your instructions distinguish those three registers — for your agents as much as for your teams?

Almost every RAG project I have seen fail did so for reasons on this list, and never because of the model. The costliest point recurs everywhere: documents are chunked without regard for their structure.

Do your documents have a structure a machine can exploit, or only a layout meant for the eye?

The indispensable counterpoint to the paper in section 01. Human oversight does not disappear out of ideology, but because it costs time; any solution that fails to address that cost ends up being bypassed.

How many minutes a day are you prepared to spend on oversight? A figure is the only answer that commits you.

A modest and useful piece, because it touches a structural limit: a model has no notion of lived time, only descriptions of time. The estimate it produces imitates a discourse.

Should we ask a model for what it can, by construction, only mimic — or is it for us to learn not to ask?

Read it straight after the piece on complexity between agents: this one shows the mechanism, the other shows the bill. Delegation between agents solves a capacity problem and creates a traceability one.

When the chain has five links, who made the decision? And would you be able to reconstruct it afterwards?