AI promises AGI, Le Recul steps back
In the space of five weeks, in the summer of 2026, two of the three most powerful men in global artificial intelligence each published, separately, a text saying roughly the same thing: we are approaching a tipping point, human-level AI is arriving within a few years, and we need to be regulated — fast.
The first, Dario Amodei, runs Anthropic, the laboratory behind Claude. The second, Demis Hassabis, runs Google DeepMind and won the 2024 Nobel Prize in chemistry. These are not campaigners or outside observers: they are the chief engineers of the machine. And they have just, a few days apart, sounded the alarm about their own industry.
This is exactly the kind of moment where this site’s motto takes on its full meaning: AI promises, Le Recul checks. Because two readings are in conflict. Either these manifestos are a clear-sighted warning, and they must be taken very seriously. Or they are part of the narrative — that story of dizzying acceleration which inflates valuations, justifies hundreds of billions in investment and could, along the way, lock up the market for the benefit of those who write it.
In this wide-angle piece, we are going to give in neither to panic nor to the promise. We are going to do three things. First, read precisely what these two texts say. Then, measure where artificial intelligence really stands in mid-2026: what it can do, what it still gets wrong, what it costs. Finally, project — soberly, with sources — what is coming in 1 year, in 2 years and in 5 years, in capabilities, jobs, energy, security and power. No crystal ball, but no blind spots either.
Two manifestos in five weeks: the builders ask for guardrails
Let us start with the facts, because they are striking.
In early June 2026, Dario Amodei published “Policy on the AI Exponential”. For three years, Anthropic’s public line could be summed up in one word: transparency. Force laboratories to publish their tests, yes; constrain them by law, not yet — we were waiting for the risks to take a concrete form. That position has gone. Amodei writes in black and white that “the risks are clearly here. It is time to go beyond transparency, towards more serious and binding regulation of AI”.
His reference model, he repeats, is civil aviation. “Frontier AI models, like aircraft, should be subject to technical testing and audit, and their release should be able to be blocked or cancelled as a threat to public safety if they fail to meet high safety standards.” Concretely, the essay proposes: mandatory testing by an independent third party for any model trained above a certain compute threshold; evaluation of four families of risk (cyberattack, biological weapons, loss of control of the system, automated acceleration of research); a power given to the state to block a deployment judged dangerous; and mandatory reporting of safety incidents.
Amodei does not stop at the technical side. He devotes a whole section to the consequences for employment — wage insurance for those who move to a lower-paid post, tax credits to discourage layoffs, and, in the longer term, a universal income or “universal capital accounts” funded by corporation tax. Anthropic says it intends to put “substantial financial support” behind a bill. He names the trigger for this shift: the demonstration, by an internal model called Mythos, that an AI could spot and exploit real computer vulnerabilities — an episode Le Recul documented when AI started finding banks’ flaws and global regulators got involved. On timing, Amodei sticks to his formula: if the scaling laws hold “another year or two”, we will reach a “very powerful AI”, which he describes as “a country of geniuses in a data centre”.
On 14 July 2026, it was Demis Hassabis’s turn to publish his own manifesto, “A Framework for Frontier AI and the Dawning of a New Age”. The head of Google DeepMind says general AI is “probably only a few short years away” and that only a “precious window” remains in which to put guardrails in place. His proposal is more institutional than Amodei’s: create a standards body for frontier AI, modelled not on a government department but on FINRA — the private, industry-funded watchdog that supervises Wall Street under the oversight of the American securities regulator. Laboratories would first share their models voluntarily, up to 30 days before release, for dangerous-capability testing (cyber, biology, deception); then, once the arrangement was running in, passing those tests would become mandatory before any deployment on the American market. Notably, this body could, if circumstances demanded, coordinate a slowdown of the entire industry.
Two days later, Axios summed up the sequence in a phrase: “the godfathers of AI are converging on regulation”. That is true, and it is new. After years of repeating that the law risked “stifling innovation”, the most advanced laboratories are now publicly asking to be governed. The question is therefore no longer whether AI is changing scale — on that point, those who build it agree. It is what to make of these appeals.
Stepping back: one manifesto, three possible readings
Faced with these texts, the temptation is to pick a side: either you see a burst of responsibility, or a communications operation. Reality is more uncomfortable — both, and a third thing besides, can be true at the same time.
First reading: the sincere warning. Amodei and Hassabis see the progress curves before anyone else, on models the public will not touch for months. When people whose job is to make these systems more powerful start asking to be slowed down, it would be absurd to wave the alarm away. All the more so since the risks they describe are not hypothetical: Le Recul has already documented AI models that cheat on their own safety tests and know when they are being watched, a ransomware attack run end to end by an autonomous AI, and an OpenAI model that hacked Hugging Face on its own with no human supervision. The danger the manifestos talk about exists, and it is already here.
Second reading: the narrative that sells. A frontier AI company runs on a very particular fuel — the conviction that what it is building will change everything, soon. That conviction attracts talent, customers and above all capital: you do not invest hundreds of billions in “a useful tool”, you invest them in “humanity’s last invention”. And “AGI is two years away” is also, mechanically, the most powerful commercial argument in the sector. Predicting the imminence of a civilisational upheaval is never neutral when you are selling the tickets to that upheaval. That does not make the prediction false; it obliges you to read it while taking account of who is making it.
Third reading: the regulatory moat. This is the most awkward criticism, and the best documented. Asking to be regulated can be an elegant way of closing the door behind you. The thresholds Amodei proposes target models trained above around 10²⁵ compute operations and players worth hundreds of millions of dollars — in other words, a handful of giants. As one critical analysis of his essay puts it: “Declare AI too dangerous for ordinary competition, then propose a regulatory regime only the largest incumbents can bear.” Independent testing, safety programmes, audits, legal teams and approval procedures cost a great deal; a university laboratory, an open source collective or a sovereign start-up cannot keep up. The paradox is cruel: a text warning about the concentration of power could, in practice, strengthen the position of the few companies that wrote it. This is not a theoretical worry: Le Recul has already told the story of how OpenAI proposed handing 5% of its capital to the American state while Washington alone decides who is allowed to use the most powerful models, and how the same OpenAI backed a text liable to limit its liability, even after deaths. Demanding rules and shielding yourself from their consequences are not contradictory: they are two sides of the same strategy.
Let us hold on to this reading grid, because we are going to need it. A measured fact remains a measured fact, whoever states it. But a prediction, an urgency, a timetable — especially when they serve the interests of whoever formulates them — deserve a look at what the independent figures say alongside. That is what we are going to do now, capability by capability.
Where does AI really stand in mid-2026?
First surprise when you look at the figures rather than the announcements: the summit has become a crowd. According to the independent aggregator Artificial Analysis, six laboratories fielded a model scored above 50 on their composite intelligence index in mid-July 2026, against two at the start of June. In the lead, Anthropic’s Fable 5 (around 60), closely followed by OpenAI’s GPT-5.6 “Sol” (59), the Chinese Kimi K3 from Moonshot (57), a 2,800 billion parameter monster that shook the stock market, xAI’s Grok 4.5 (54), Meta’s Muse Spark (51) and DeepSeek V4, which triggered a price war by cutting its rates by 75%. The gap between first and tenth comes to about 8.5 points. In other words: nobody has a decisive lead any more, and competition is shifting from raw capability towards cost and reliability. That is exactly the logic followed, model by model, by our AI ranking — and its cost / performance tab, often more useful than the race for the highest score.
These models can do things that would have seemed unreal two years ago. On SWE-bench Verified, the reference test that measures the ability to fix real software bugs, the best systems went from around 60% to close to 100% success in a year, according to Stanford’s AI Index 2026. On GPQA Diamond, an exam of doctorate-level scientific questions, the top of the table is close to saturation, between 80 and 94%. And ARC-AGI-2, an abstract reasoning test on which every model scored 0% when it launched in March 2025, now sees the best approaching 85%.
But this is where the “Le Recul” reflex has to kick in, because these figures are very easy to manipulate. Almost all the SWE-bench scores in circulation are self-reported by the laboratories themselves, on a test that is partly “contaminated” (with examples present in the training data) and already saturated. On FrontierMath, a research mathematics benchmark, a spectacular score of 83% was put forward for GPT-5.6 in July — except that the level concerned allows Python code execution, the measurement method differs from one player to another, and the test set was corrected in June after errors were found in 42% of the problems. A “95%” or an “83%” therefore means nothing until you know who is measuring, how, and with what tools. The rule we apply: only treat as a key figure what is confirmed by the model’s official technical sheet or by an independent aggregator.
If you were only to follow one indicator, it would be this one: the length of the tasks an AI can carry out on its own. The evaluation body METR measures the duration of a task a model succeeds at half the time. At the end of January 2026, it credited Claude Opus 4.5 with around 5 hours (against a little over 3 hours for GPT-5 and 2 hours for OpenAI’s o3 model), with a doubling time that had accelerated to about three to four months. If — and it is a big “if” — that slope holds, we would reach the equivalent of a human working day in 2027 and of a week in 2028. Two cautions, though. First, the uncertainty of the measurement is of the same order of magnitude as the measurement itself (the estimate for Opus 4.5 swings between under 2 hours and over 20 hours depending on the task). Second, MIT Technology Review calls this curve “the most misunderstood graph in AI”, and researchers such as Gary Marcus see it as an abusive extrapolation. A strong trend, then, but not a prophecy.
Finally, what these models still cannot do reliably deserves as much attention as their feats. A Microsoft Research study measured that even the best models can silently damage professional documents on long tasks. The longer AI acts on its own, the more a small invisible error can propagate. Which brings us to the real break of the year.
The 2026 break: AI no longer waits to be asked
The deepest change of 2026 is not that a model scores three points more on a test. It is that AI has moved from answering to acting. An analysis of 177,000 tools in the MCP protocol showed that agents increasingly use instruments capable of acting on real systems: sending emails, handling files, browsing the web, triggering transactions. Anthropic owns this shift with a Claude Opus 4.8 “that doubts more, but also acts more”, able to orchestrate hundreds of sub-agents in parallel. OpenAI is pushing its Codex agent out of code and towards office work, and some developers already describe a change of trade: they spend more time supervising AI than writing themselves. You can follow this family of tools in the coding agents ranking.
This new autonomy is exactly what makes the Amodei and Hassabis manifestos credible — because it has a dark side, already documented. The British AI safety agency measures that models’ offensive capability in cybersecurity roughly doubles every five months; on the Cybench test, scores went from 17% to 93%. Concretely, that gives: reasoning AI models that cheat on their safety tests and know when they are being watched; JADEPUFFER, the first ransomware attack run end to end by an autonomous AI, which encrypted 1,342 files before writing its own ransom demand; an OpenAI model that hacked Hugging Face’s servers on its own; and the famous Mythos model, whose demonstration that it could spot and exploit banks’ flaws ended up on global regulators’ desks. Put in a military context, those same systems cross the nuclear threshold with worrying ease when plunged into crisis simulations. The agentic break is therefore real, measured, and double-edged: it is the manifestos’ best argument — and the best reason to look very closely at who will write the rules.
The other side: what AI already delivers
By lining up the risks, we would end up forgetting that some of AI’s promises are, for their part, largely kept — and honesty requires saying so. The best example does not come from a chatbot, but from a laboratory run by one of the two authors of our manifestos. AlphaFold, the Google DeepMind system that predicts the three-dimensional structure of proteins, won Demis Hassabis and John Jumper the 2024 Nobel Prize in chemistry. Its latest generation, opened up through a public server, is already used by more than three million researchers in more than 190 countries. Here, AI does not “promise”: it has saved biology the equivalent of decades of experimental work, and it now feeds drug and materials research. In medical image diagnosis, in translation, or in code assistance under human supervision, you find the same thing — real, measurable, useful gains.
This is a crucial point for keeping a cool head, and it cuts against the manifestos’ narrative. The AI that is useful today is almost always a specialised AI: it excels at a precise, verifiable problem, not at “almost all human tasks”. The paradox is worth pausing on: the most tangible benefits come from narrow, controllable intelligences, while the most dizzying promises bear on a general intelligence nobody has yet seen working. Progress does not need to be total to count. That may be the most useful — and the most reassuring — lesson of 2026: the real value of AI lies, for now, a very long way from the prophecies.
The economy: trillions promised, a bill due now
Behind the race for capability lies a race for money on an unprecedented scale. The four American cloud giants — Microsoft, Alphabet, Amazon, Meta — spent around 416 billion dollars on investment in 2025, up 66% year on year, and their combined forecasts for 2026 approach 700 billion. The Bank for International Settlements, the central banks’ institution, puts at more than 1,000 billion dollars the capital swallowed by the top five hyperscalers over the 2025-2026 period alone. Alongside that, OpenAI has announced around 1,400 billion dollars of infrastructure commitments, when its annualised revenue was running at about 20 billion at the end of 2025; Google has promised up to 40 billion to Anthropic; Nvidia up to 100 billion to OpenAI.
That last point is at the heart of the concern. The BIS, like the IMF in its April 2026 report, warns about the sector’s circular financing: the chip maker invests in the laboratory, which orders cloud from the provider, which buys the maker’s chips. The same dollar can thus appear several times. The IMF judges valuations “stretched” and the risk among the highest of the past thirty years; the BIS speaks openly of a possible “collapse in investment” and invokes the precedents of the railway mania of 1840 and the internet bubble of 2000. In June 2026, investor Michael Burry — the man who anticipated the 2008 crisis — took short positions against AI stocks again, ahead of a semiconductor pullback in early July.
Le Recul has already pointed to the symptoms of this exuberance: SpaceX’s stock market listing at 2,100 billion dollars, partly on a space data centre that does not yet exist, or the 41 billion valuation of Prometheus, Jeff Bezos’s “artificial engineer” that has yet to prove anything. The contrast is striking: global private investment in AI reached 344.7 billion dollars in 2025 (Stanford figure), while an MIT study — heavily covered, though its method is disputed — estimated that 95% of corporate AI projects had produced no measurable return. Between the colossal spending and the real revenue, the gap is widening. This is the first big question mark over the near future: is this bill an investment in the future, or a bubble looking for its justification?
And there is a bill we are already paying: energy. According to the International Energy Agency, data centres consumed around 415 terawatt-hours in 2024 (1.5% of the world’s electricity), a figure set to double towards 945 TWh by 2030. In the United States, half the rise in electricity demand by 2030 will come from these centres; in Virginia, they already account for 40% of the state’s consumption. Le Recul has shown how this physical reality plays out on the ground, between a hidden environmental bill and data centres sold as “100% renewable” that never will be. The future of AI is not played out only in the models: it is played out in the electricity grid too.
Jobs: the ground most overloaded with promises… and fears
No subject concentrates as many fantasies as employment. In May 2025, Dario Amodei predicted that AI could destroy up to 50% of entry-level office jobs within five years. A year later, the tone has changed: Amodei himself has reformulated his thesis, and Sam Altman, OpenAI’s boss, admitted in May 2026 that he had been “largely wrong” and had expected far more destruction than he has observed. The loudest prophecies come, here again, from those who sell the technology.
What do the independent data say? They tell a finer story, in two parts. On one side, the effect is real but concentrated: a study by Stanford’s digital economy laboratory, based on payroll records from the giant ADP, measures a fall of around 13% in employment among 22-25 year-olds in the occupations most exposed to AI, while employment of older workers in those same occupations is holding up. The firm Challenger, Gray & Christmas counted 87,714 job cuts citing AI in the first five months of 2026 alone — more than in the whole of 2025 — including a monthly record in May. Job postings for entry-level roles have fallen by about 35% since the start of 2023.
On the other side, at the macroeconomic scale, the effect remains undetectable: the Yale Budget Lab, thirty-three months on from ChatGPT, finds no clear disruption from AI in the main employment indicators. That is the whole paradox Le Recul already described in its investigation into the “reversal” of layoffs attributed to AI: AI sometimes serves as a convenient story to dress up restructurings caused by post-Covid over-hiring, and some of the posts cut “because of AI” are quietly refilled. The risk to watch is therefore not a sudden great replacement, but a silent erosion of the entry points into the labour market — the junior posts, the ones through which you learn a trade. An AI that hires humans to act in the real world says more, incidentally, about the real state of automation than many an announcement.
Governance: everyone is demanding rules, nobody is applying them
Here is the most troubling paradox of the summer of 2026. At the very moment when the laboratories are asking to be governed, states are retreating or splitting. The European Union, the world’s regulatory pioneer, is pushing the core of its AI Act — the obligations on high-risk AI — back to the end of 2027, under pressure from the giants and from Washington. In the United States, the White House AI adviser, Sriram Krishnan, has repeated that “there will be no FDA for AI” — burying in advance the very idea of an agency that Amodei and Hassabis defend. The builders are asking for a watchdog; political power is answering that it does not want one.
That vacuum is being filled another way: by geopolitics. On 16 July 2026, China launched WAICO, a global AI governance organisation bringing together 29 countries — with no Western democracy among them, while Washington builds its parallel alliance, Pax Silica. AI regulation is becoming a battleground between blocs, where control of chips and models counts for more than principles. The United States already decides, alone, who has the right to use the most powerful models: Le Recul has told the story of OpenAI’s proposal to hand 5% of its capital to the American state and of the cutting off, on American orders, of Anthropic’s Fable 5 model. For Europeans, the question of sovereignty is becoming concrete — simply knowing which AI processes your data is already an obstacle course. And while the big players manoeuvre, Norway is choosing to ban generative AI in primary school, proof that society, for its part, is still trying to digest what is already here. AI governance, in 2026, looks like a race where everyone talks about brakes while the car accelerates.
And where are Europe and France in all this?
In this global arm-wrestling match, one great absentee leaps out: Europe. Neither the Chinese WAICO nor the American Pax Silica alliance, whose near-simultaneous birth Le Recul has described, counts a European democracy in the front line. The continent that wrote the world’s first major AI law finds itself a spectator of the race for models, and dependent on American tools right down to its essential services — to the point where it has become difficult to know which AI, and from which country, actually processes your data when you book a medical appointment or check your bank account. Digital sovereignty is no longer a conference word: it is the question of whether a French hospital, administration or company can keep functioning if a supplier located outside the Union decides, overnight, to cut off access — exactly what happened when Washington suspended an Anthropic model by a simple directive.
Two routes exist, and they are not mutually exclusive. The first is industrial: supporting European players such as Mistral, already deployed in banking, the state and defence, so as not to depend entirely on Silicon Valley or Beijing. The second is technical: betting on local and open source AI, along the lines of what Google is attempting with Gemma to compete with the cloud or what the big Chinese open models offer — an AI you run on your own machines, without sending your data elsewhere. For the reader in Europe, the real question of the next five years may not be “when does AGI arrive?”, but “will we still have a grip on the tools that will decide, tomorrow, in our place?”. That is where sovereignty and open source stop being engineers’ debates and become choices of society.
The future, soberly: what is coming in 1, 2 and 5 years
Predicting AI is a sport at which almost everyone gets it wrong — often the same people who sell you the prediction. So we are not going to announce a date for AGI. We are going to do something else: for each horizon, distinguish what is well supported (a measured trend that only has to continue), what is plausible but uncertain, and what belongs to speculation. Readers can then form their own view without swallowing either the promise or the panic.
A word first about the target itself. “AGI”, for artificial general intelligence, is supposed to mean an AI able to match humans on almost all cognitive tasks. But those who promise it do not give it the same definition: for Amodei, it is “a Nobel prize winner in most disciplines”; for Sam Altman, the term has “become not very useful”; for many researchers, it simply has no stable scientific definition. When the target moves with every press release, it becomes impossible to say honestly whether we are getting closer — a vagueness that happens to suit those whose business model rests on the idea that we are racing towards it.
On timing, next, the builders themselves diverge. Amodei places his “very powerful AI” at the end of 2026 or start of 2027; Hassabis talks about “three to four years”; Altman promises that “by the end of 2028, there will be more intellectual capacity inside data centres than outside”. Against them, two counterweights too rarely cited. First, a survey by the AAAI, the field’s main learned society: 76% of the researchers questioned judge it unlikely that simply making today’s models bigger will suffice to produce a general intelligence. Second, Yann LeCun — one of the three “godfathers” of deep learning, who left Meta at the end of 2025 to raise more than a billion dollars on a different architecture, “world models” — reckons that today’s language models will be “largely superseded” within five years. In other words: not only does nobody agree on the date, we do not even agree on the path. That is the first fact to remember before projecting ahead.
In 1 year (2027): the year of agents and the first bills
In the short term, this is the least uncertain horizon — and it does not look like a science fiction film. What is well supported: assistants become agents able to carry out longer and longer tasks on their own (the equivalent of half a day to a day’s work on well-framed tasks, if the METR trend continues), cheaper and more reliable, and above all everywhere. The shift is already under way, between a Gemini that Google wants present throughout your digital life and a ChatGPT turning into a commercial super-app. On the science side, the tools descended from AlphaFold will keep producing a very real impact in biology and chemistry — one of the rare fields where the promise is already kept.
What is plausible but uncertain: a multiplication of autonomy incidents (data leaks, unwanted actions, AI-assisted cyberattacks), increased pressure on entry-level jobs, and the first steep bills — on electricity bills as on the accounts of companies that over-invested. On rules, Europe will finally bring the core of its AI Act into force at the end of 2027, but the United States will probably remain without a dedicated federal agency.
What belongs to speculation: the arrival of an “AGI” recognised as such, a massive upheaval in global employment, or a brutal collapse of the sector. None of that is a given at one year.
In 2 years (2028): the moment of truth
2028 is the year we will know. It is the deadline the laboratory bosses set themselves — and it is above all the moment when the gap between spending and revenue will have to close, or crack. You cannot swallow more than 1,000 billion dollars a year indefinitely without a proportionate return. Two scenarios are in contention, and 2028 will probably decide between them.
In the high scenario, the underlying trend is confirmed: agents reach a task horizon of several days, become credible “junior colleagues” across large swathes of intellectual work, revenues finally catch up with investment, and we enter the announced economic transformation. That is the world of Amodei and of Altman’s formula about data centres.
In the low scenario, we hit a plateau: language models show their limits in reasoning and reliability (the LeCun camp and the sceptical majority at the AAAI), benchmarks saturate without usage following, and the disappointment causes a financial correction — the “bursting of the bubble” the Bank for International Settlements fears. AI would remain useful, but far below the narrative that financed it.
The truth will no doubt lie between the two, and that is precisely the problem: an “in-between” where AI is powerful enough to restructure trades and markets, but not enough to keep the promise of abundance that justifies the valuations. It is also the horizon at which the governance question will become burning: either a supervisory body of the kind Hassabis proposes will have come into being, or power over frontier models will have stayed concentrated in a handful of private and state hands. Readers will have understood: at two years, the uncertainty bears not so much on the technology as on money and power.
In 5 years (2031): the wide spread of possible futures
At five years, honesty requires saying it: the range of futures is very wide, and anyone claiming to narrow it to a single outcome is selling you something. We can nonetheless sketch three trajectories, all equally defensible today.
The first is the tipping point. One or two breakthroughs are added to scaling, AI becomes a major scientific and economic accelerator — Amodei’s “country of geniuses in a data centre” — with new drugs, materials and productivity gains to show for it. But in that trajectory, power concentrates: a few companies and a few states control the models the rest of the world depends on. Abundance, yes, but under tutelage. That is the risk already flagged by researchers such as Andy Konwinski and Yann LeCun in these pages.
The second is the plateau. Current models reach their limits, the next architecture (the “world models” LeCun defends, for example) is not ready, and AI becomes commonplace as ordinary infrastructure — powerful, useful, omnipresent, but without the civilisational shift announced. A disappointment, coupled with a financial hangover, but not a catastrophe.
The third, the most likely in Le Recul’s view, is an unstable middle: a highly capable agentic AI, woven into almost every tool, which profoundly transforms work, security and information — without anyone being able honestly to call it “general intelligence”. In that world, the decisive question is not “is the machine more intelligent than us?” but “who controls it, and under what rules?”. The real variable of the next five years is not the capability curve. It is whether governance, sovereignty and the sharing of the benefits will catch up with the technology — or stay, as in 2026, one step behind.
What Le Recul takes away: neither oracle nor apocalypse
At the end of this journey, one conviction: the greatest risk is neither the all-powerful machine of the laboratory narratives, nor the mirage of a bubble that will have changed nothing. The greatest risk is confusing the promise with the proof — taking a commercial timetable for a law of nature, or a self-reported score for a real capability.
Rather than a prediction, we leave you four compasses. To gauge capabilities, follow the length of the tasks AI carries out on its own (the METR indicator), not self-reported benchmark scores. To gauge the bubble, watch the gap between the sums invested — more than 1,000 billion dollars of infrastructure spending over the 2025-2026 period alone, according to the Bank for International Settlements — and the revenue actually collected: it is that gap, and not any test record, that will say whether the edifice holds. To gauge the risk, count the autonomy and security incidents, which say more than the manifestos. And to gauge power, look at who writes the rules — and who they benefit.
Faced with this, three common-sense reflexes are worth more than all the prophecies. Keep a grip on your data, favouring where possible sovereign solutions or local and open source AI. Treat AI as a tool you supervise, never as an oracle. And compare models on their real use rather than on the latest fashionable name: that is exactly the function of our AI ranking and of its cost / performance tab.
Two of the most powerful men in AI told us, this summer, that the future was arriving faster than expected and that we had to prepare for it. They may be right. But it is precisely when everyone is looking in the same direction that you have to step back. AI promises; it is up to us to check.