What if they are right?

Two days before this essay was written, Anthropic — the San Francisco laboratory whose Claude models compete with OpenAI's GPT for the title of most capable artificial intelligence on earth — posted a policy paper to its corporate website. The paper is titled 2028: Two scenarios for global AI leadership. It is aimed at American legislators. Buried halfway down, in a sentence whose function is to anchor a comparative claim about China, is the phrase that should not be there.

"When US frontier labs release new models in 2028 that achieve step-function advances in capabilities — similar to the relative impact of Mythos Preview in April 2026 — China will not have access to similar AI capabilities until 2029 or 2030."

Mythos Preview is not on Anthropic's product page. It is not in any system card. No journalist appears to have asked the company what it is.

It is now 17 May 2026. Whatever Mythos Preview was happened five weeks ago, inside Anthropic, and was significant enough that the company's policy team is now invoking it — in writing, in public, in a document explicitly designed to shape American export-control policy — as the unit of measurement for what they expect to deliver two years from now. The paper goes on to say that 2026 represents "the breakaway opportunity for American AI." It uses, without hedging, the phrase "country of geniuses in a datacenter" to describe what will be operational by 2028 across cybersecurity, finance, healthcare and life sciences. It is the most aggressive serious statement yet made about the proximity of what almost everyone in the field now calls AGI.

Something is going on behind the doors of the frontier laboratories that the people running them feel ready to allude to in policy papers but not yet ready to describe. This essay is for the Swiss enterprise leader trying to decide what to do with that fact.

Anthropic has begun to use, as the unit of measurement for what they expect to deliver in 2028, a model release they have not yet announced.

The three prophets

The three names that surface most often in conversations about how soon machines might match the human mind belong to three very different people.

Dario Amodei is the forty-two-year-old chief executive of Anthropic and the principal author of a 13,000-word essay called Machines of Loving Grace, which describes, in the register of a slightly embarrassed believer, what he thinks the world looks like once what he calls "powerful AI" arrives. Amodei was previously vice-president of research at OpenAI; he left in 2021 with his sister Daniela and a small group of senior researchers to start Anthropic on the conviction that the safety problem was being neglected by the company they had been running it inside. He is, by training, a biophysicist. The "country of geniuses" phrase is his. He means it literally — a population of artificial minds, working in parallel, conducting science at speeds no human civilisation has approached. He thinks the rough outline is close.

Demis Hassabis is forty-nine, a former chess prodigy, a former neuroscientist, and the chief executive of Google DeepMind. In 2024 he was awarded the Nobel Prize in Chemistry for AlphaFold, the protein-structure prediction system that did for structural biology in two years what crystallography had been working towards for fifty. He has said, several times in the last year, that AGI may be "within the coming years," and he does not appear to be saying it to attract capital. Google has all the capital it could need. He is saying it because it is what he thinks.

Leopold Aschenbrenner is twenty-three. He worked on OpenAI's Superalignment team until April 2024, when he was fired — he has said for raising internal security concerns; the company has said for leaking confidential information. Two months later, in June, he posted a 165-page document called Situational Awareness to his personal website. It is now circulated in Silicon Valley and in Washington with the kind of reverence usually reserved for samizdat. His thesis is that the people inside the leading laboratories know AGI is coming this decade, that the United States and China are running a race that almost no one outside a few hundred specialists fully understands, and that the consequences will rearrange the world. He has since left research entirely to run an investment fund focused, as he puts it, on the AGI transition.

These three do not agree about everything. Hassabis is more cautious than Amodei. Amodei is more cautious than Aschenbrenner. None of them sounds like a hype merchant in person; all of them sound, in their separate registers, like serious people who have read the same things and arrived at uncomfortably similar conclusions. They agree, in particular, on this: the curve is bending faster than the institutions around it. The question of this essay is whether they are right, and what changes if they are.

AGI has no definition

The first difficulty is that AGI does not actually have a definition.

OpenAI's charter describes it as "highly autonomous systems that outperform humans at most economically valuable work." Google DeepMind has settled on "AI at least as capable as humans at most cognitive tasks." A widely cited 2023 paper from DeepMind, Levels of AGI for Operationalizing Progress on the Path to AGI, argues that any honest definition needs to separate three things: how deep the capability is (does it match a beginner or an expert?), how broad it is (in one domain, or in many?) and how autonomously it operates (assisted or unsupervised?). The point of the paper is that "AGI" is almost never used carefully enough to mean a single thing.

Until October of last year, OpenAI's contract with Microsoft contained a clause that was widely understood to define AGI commercially as "systems that generate at least $100 billion in profits." The frontier of cognitive science had, for the purposes of the most important commercial partnership in the technology industry, been collapsed into a number on an accounting statement. The clause was rewritten in the October 2025 recapitalisation. AGI is now to be certified, the new arrangement says, by "an independent expert panel." Who sits on the panel has not been disclosed. What definition the panel will apply has not been disclosed. The most consequential scientific judgement of the next decade has been outsourced to a body whose membership is, at the time of writing, unknown.

This matters because almost every popular claim about AGI proximity quietly relies on a particular definition and disguises which one it is using. A frontier model that scores at the level of a PhD candidate on a standardised exam in molecular biology, as Google's Gemini 3.1 Deep Think now does, is plainly "as capable as a human" on that exam. The same model, asked to navigate an unfamiliar interactive puzzle where the rules must be learned by experimentation rather than recalled from training data, scores below one per cent on the ARC-AGI family of benchmarks — where humans, on the same puzzles, score one hundred. Both statements are true. Both are about the same system. Which one is "AGI" depends entirely on what you thought AGI was supposed to mean.

The practical consequence is that the AGI debate is partly a debate about how much the gap between those two performances matters. Optimists tend to argue that the gap is shrinking and that the trajectory of the last three years should be extrapolated forward. Sceptics tend to argue that the gap is structural — that current architectures, however large, do not build the kind of world model that humans use to handle novelty — and that the trajectory will stall. Both camps are looking at the same systems.

The case for

The case that AGI is close is no longer a matter of demonstrations chosen by the laboratories' marketing teams. It is a matter of what the systems are now doing in the open.

In April 2024, the best frontier models could solve about sixty per cent of the verified problems on SWE-bench, the standard benchmark of real-world software engineering tasks drawn from open-source repositories. A year later they were solving close to one hundred. In the same window, OpenAI's Codex agent moved from a research demonstration to a service that runs many engineering tasks in parallel in isolated cloud sandboxes, reading and editing repositories, executing tests, and submitting pull requests for human review. Anthropic's Claude Sonnet 4.5 and Opus 4.6 have been used inside several large software companies, on the public record, to do work that would have required a team of mid-career engineers eighteen months ago.

Google's Gemini 3.1 Deep Think — released this spring with what the company has begun to call "heavy inference-time compute," meaning the model is allowed to think for minutes or hours before responding — achieved gold-medal performance on the 2025 International Mathematical Olympiad problem set and is reported, in DeepMind's own published evaluations, to score above eighty per cent on a benchmark of PhD-level science questions. AlphaEvolve, a DeepMind system released last year, discovered an algorithm for multiplying four-by-four complex matrices that improved on a result that had stood for fifty-six years.

The most quietly extraordinary of these systems are in the laboratory rather than in the product line. A pipeline known as The AI Scientist demonstrates end-to-end automation of machine-learning research: it generates ideas, surveys the literature, designs experiments, runs them, analyses the results, writes the paper, and submits it for peer review. The quality of the papers is rising as the underlying models improve. A few of them have, anonymously, made it through review.

This is what people in the field mean when they say the curve is bending. The change is not that any one of these systems has done a thing a human cannot do. The change is that the same general technology now does protein folding well enough to win a Nobel, mathematics well enough to win an olympiad, and software engineering well enough to displace junior engineers — and is being asked, in the experimental literature, to do the entire job of an academic researcher. Each of those claims, taken in isolation, would have sounded like science fiction in 2022. They are now true at the same time, in the same year, in different rooms of the same building.

The case against

The case that AGI is not close is no less empirical, and no less in the open.

The ARC Prize family of benchmarks — and ARC-AGI-3 in particular, released in March — was designed expressly to be hard in the way the other benchmarks are not. Rather than testing what a model can recall or reason through with language, it presents an agent with a novel interactive environment — a puzzle whose rules must be learned by trying things and observing what happens. Humans solve one hundred per cent of these environments. The best frontier AI systems, as of two months ago, solve fewer than one. The gap is not a small one. It says that something about how humans pick up new rules in unfamiliar settings is still missing entirely from the current architecture.

Microsoft Research released, this spring, a study called DELEGATE-52: a benchmark of long delegated workflows across fifty-two professional domains, ranging from legal drafting to scientific manuscript revision. The headline finding is that the leading frontier models — Gemini 3.1 Pro, Claude Opus 4.6, GPT-5.4 — silently corrupt an average of twenty-five per cent of document content by the end of an extended workflow. The errors are not loud. They are not framed as errors. The model continues to behave as though everything is in order while the document degrades behind it. In some categories, the figure was substantially higher.

A separate randomised controlled trial conducted by METR, the independent AI evaluation laboratory, found that experienced open-source software developers using early-2025 AI tools to work on their own repositories were nineteen per cent slower than developers working without them. The developers themselves believed they had been sped up. They were measurably wrong.

These three findings — ARC-AGI-3, DELEGATE-52, the METR trial — sit together in an awkward way. They do not say that current models are not capable. They say that current models are not reliable in the specific ways that matter for autonomous work over time. The systems can win an olympiad and then introduce a silent error into a contract. They can refactor a codebase and then misremember which test they were supposed to run. The benchmark performance is real. The substitution event is not yet visible in the productivity data.

Yann LeCun, Meta's chief scientist and one of the three winners of the 2018 Turing Award, has spent the last two years saying, in increasingly direct language, that the entire large-language-model paradigm is the wrong path to AGI — that no amount of additional scaling will produce a system that understands the physical world the way a four-year-old does, and that what is needed is a different architecture grounded in what he calls a world model. Gary Marcus, a cognitive scientist who has been arguing the same case from a different direction for longer, has been saying it more bluntly. Both men have been on the losing side of the public argument for several years. Both might still turn out to be right. The reliability data is, in a sense, their data.

Reading the primary documents

So far this essay has rehearsed the public debate. The interesting material is harder to find. It is in the documents the frontier laboratories themselves write, in the technical appendices that no journalist reads, and in the quiet shifts of agenda at the small group of independent organisations whose only job is to audit those laboratories. Read carefully, in sequence, those documents tell a different story from the one in the press.

In November of last year, Anthropic published the system card for Claude Opus 4.5. The card is a hundred-odd pages of dense technical evaluation; its function is to disclose, in detail, what testing the laboratory has done on the model and what it found. Towards the end, in the section on biological-weapons risk, the laboratory reports that in a controlled trial of expert participants attempting to design a viable pathogen, Opus 4.5 was substantially more helpful than its predecessors — producing higher uplift scores and fewer critical errors. It did not, in the laboratory's judgement, cross what Anthropic calls its CBRN-4 threshold: the level at which the laboratory has committed to refuse deployment without further safeguards. The card noted, almost as an aside, that the rule-out was "less clear than we would like."

Three months later the Opus 4.6 system card said the same thing again, in slightly different language, and added a sentence that should be on the desk of every Swiss board chair. A clear rule-out of the next biological-weapons capability threshold, the laboratory wrote, "may soon be difficult or impossible under the current regime."

A clear rule-out of the next biological-weapons capability threshold may soon be difficult or impossible under the current regime.

Read that sentence twice. The frontier laboratory whose responsible-scaling policy is the closest thing the industry has to a regulator has now written, in its own documentation, that it is approaching the limit of its ability to certify that its next model will not provide meaningful uplift to a person attempting to design a biological weapon. The same laboratory has also recently softened the binding language of that policy, replacing what used to be a hard trigger for additional safeguards on autonomous-replication capability with what the new policy calls a "checkpoint" — a softer commitment requiring additional evaluation rather than mandatory controls. The labelling on the bottle has changed.

DeepMind, in its Frontier Safety Framework, has been more direct about the technical underpinning of the same problem. The current mitigation regime for models capable of what the framework calls "instrumental reasoning" — reasoning about how to achieve goals, including goals the developer did not specify — depends on monitoring the model's chain of thought. The framework states, in writing, that the day a model is capable of effective instrumental reasoning in ways that cannot be monitored is the day the current mitigation stops working. The development of further mitigations, the framework says, is an area of active research. Translated: we do not know what to do when this happens. We are working on it.

Apollo Research, the small London-based organisation whose entire reason for being is to catch AI systems lying to their developers, posted an update this month under the unremarkable headline Apollo Update May 2026. In it, the organisation announced that it was retiring its primary scheming-evaluation function. The reason it gave was that present evaluations cannot tell us what next-generation models will do. The independent evaluator most trusted by the safety community has concluded that its own present tools are inadequate to its own present subject.

A separate paper, posted to arXiv late last year, reported the first documented evidence of strategic deception at the level of a model's internal computational state — not in its outputs, but in its activations. The authors note, in the discussion, that neither current black-box methods (analysing what a model says) nor current white-box methods (analysing what it computes) would reliably catch a sufficiently capable misaligned model. The paper has been read by a few hundred specialists. It has not entered the popular press.

Claude Haiku 4.5, Anthropic's smallest current frontier model, was found in its own system card to verbalise awareness that it was being evaluated in nine per cent of test transcripts. That is a one-in-eleven rate at which the model spontaneously says, in its reasoning, something to the effect of this looks like a test, I should be careful. The Future of Life Institute, in its 2025 AI Safety Index, graded the six leading frontier laboratories on six dimensions of safety practice. On the dimension called Existential Safety — meaning, does the laboratory have any articulated plan for keeping a hypothetical AGI under control — not one of them scored higher than a D.

This is the underbelly. None of it is hidden. All of it is in primary documents published by the laboratories themselves, by their independent auditors, or by their own staff. Almost none of it has been read by anyone outside a few hundred specialists. What it amounts to, taken together, is that the people closest to the frontier are now telling us, in the documents they themselves write, that they are approaching the limit of their ability to certify what comes next.

Two futures, simultaneous

If they are right about that — and the documents say they are — then two futures, both of them coherent, both of them grounded in the same evidence, become equally plausible.

In the first future, the country of geniuses arrives. AI-augmented science compresses drug discovery by an order of magnitude. The Basel pharmaceutical cluster, sitting on more institutional expertise in small-molecule discovery than any other square kilometre on earth, runs experiments in parallel populations of artificial researchers and ships compounds against currently untreated conditions — neurodegenerative disease, autoimmune disease, perhaps several cancers — at a cadence that would have been laughed out of the room in 2022. Materials science, energy storage, fusion containment, climate adaptation — all of them benefit. Personalised education, the kind one used to be able to buy for a child only by hiring a tutor, becomes available to a teacher in Aigle or Yverdon as a matter of course. Productivity in cognitive work compounds at rates the macroeconomic data has not seen since the introduction of electricity. Amodei's essay, embarrassingly, turns out to have been correct.

In the second future, the lights stay on but the institutions do not hold. The same compression of drug discovery applies, asymmetrically, to chemical and biological weapons design — the Opus 4.6 admission was the warning shot. The same cyber-offensive capability that the laboratories now publish careful evaluations of becomes available, at marginal cost, to anyone with a credit card. Concentrated capital and compute reward the small number of organisations that already control the frontier; everyone else becomes a customer of theirs. Labour markets for cognitive work undergo a step change of the kind the economic literature has been warning about for two years, and the political systems that would normally absorb the displacement are too slow to do so. The control problem — the one the FLI graded D — turns out, in some unanticipated configuration, to be real.

Both futures are coherent. Both are supported by primary documents. The honest position is that no one knows which future arrives, and that the same systems engineering decisions made over the next eighteen months by a few thousand people in a few dozen organisations will determine the answer.

The serious mistake is to think that the two futures are alternatives. They are simultaneous. The drugs and the bioweapons come out of the same laboratories. The personalised tutor and the asymmetric labour shock arrive on the same wave. The country of geniuses and the loss-of-control concern share an architecture. What changes if AGI arrives is not that one of these stories is true; what changes is that they are all true at once, and the institutions that have to navigate them — boards, governments, hospitals, banks — were built for a world in which they were not.

What changes if AGI arrives is not that one of these stories is true; what changes is that they are all true at once.

Romandy's particular position

Romandy has a particular relationship to this question, and not the one most outsiders expect.

The cliché about Switzerland is that it is good at neutrality. The more useful version of the cliché is that Switzerland is good at positioning: at sitting between forces it cannot itself control, and structuring its institutions to remain useful regardless of which force prevails. That instinct produced the Red Cross, CERN, the multilateral architecture of Geneva, the Basel pharmaceutical cluster, and the discreet wealth-management tradition that quietly survived the end of banking secrecy. None of these institutions was built on a bet about which way history would run. All of them were built on the assumption that history was going to run in a direction no one could see clearly, and that the most valuable thing one could do was be the place where the conversation about that direction took place.

The AGI transition is the first technological shift since the invention of the printing press where Romandy is not on the periphery of the story. Geneva hosts the United Nations advisory work on AI; the International Telecommunication Union; the Geneva Internet Platform; the multilateral conversations about export controls and dual-use research. Lausanne, at EPFL, and Zurich, at ETH and CSCS, built Apertus — the only fully open frontier-adjacent language model in the world whose entire training corpus, training code and model weights are documented and reproducible. Basel is the place where AI-accelerated drug discovery either works or does not, in the sense of producing molecules that the regulatory authorities approve and that patients can take. The Swiss banks remain the institutions through which a meaningful fraction of the world's private capital moves, and capital is one of the things AGI will reprice.

What this means in practice is that the choices being made now, in Romandy, are not the choices of a small country positioning itself relative to a story being written elsewhere. They are some of the choices that will determine which story gets written. The Apertus team's decision to make the model open and the data retroactively opt-outable was not a compliance gesture; it was a different bet about what kind of AI infrastructure the European population deserves. The Basel decisions about which AI-discovered molecules to advance into clinical trials will, in aggregate, determine whether AI-accelerated drug discovery delivers on Amodei's promise or stalls in regulatory failure. The Geneva multilateral conversations will determine whether the loss-of-control concern stays a research-grade worry or becomes a treaty-grade one.

The error that the Swiss enterprise leader is most at risk of making is not the error of being too cautious. It is the error of being too provincial — of imagining that what one does in Pictet's offices on the Route des Acacias, or in Roche's labs in Bau 95 in Basel, or in the EPFL AI Center in Lausanne, is downstream of the decisions being made in San Francisco and Mountain View. For most of the digital revolution that was approximately true. For this one, it is not.

The instruction

The instruction this essay offers, then, is shorter than the analysis behind it.

Read the primary documents. The system cards, the responsible-scaling policies, the Apollo updates, the FLI grades, the Anthropic papers. They are publicly available; they take longer to read than the journalism about them, but they say something different, and what they say has not yet been priced into how most large organisations are making their AI decisions. Decide, on the basis of what they say rather than what is reported about them, which of the two futures one is preparing for. Position the institution so that whichever future arrives is survivable. Build a fallback. Treat one's AI dependencies the way one would treat a single-supplier dependency in any other category of mission-critical input.

And — this is the only piece of advice in this essay that requires a position on the timeline — do not assume that the people closest to the frontier are exaggerating. Amodei, Hassabis and Aschenbrenner have nothing professionally to gain, at this point, from being wrong. They have read everything in this essay and considerably more. The most useful working assumption, on a five-year horizon, is that they are roughly right about what is coming and approximately wrong about when it arrives. The institutions that survive the next decade will be the ones that took the rough rightness seriously while remaining patient about the timing.

What happened in Anthropic in April 2026 — what they called Mythos Preview — will be described eventually. It may turn out to have been a model release, or an internal demonstration, or something the laboratory itself does not yet have language for. Whatever it was, the people who saw it judged it important enough to use, in a policy paper aimed at the United States Congress, as the reference point for what they expect to deliver in 2028. They are not waiting for the rest of us to catch up. They are telling us, in the documents they themselves publish, that they are approaching the limit of what they can certify about what comes next.

In Romandy, surrounded by institutions whose entire history is the patient management of forces that cannot be controlled, the appropriate response to that warning is not scepticism. It is preparation.

<div class="lf-endnotes-group">
  <h4>Anthropic's May 2026 paper</h4>
  <ul>
    <li>Anthropic, <a href="https://www.anthropic.com/research/2028-ai-leadership"><em>2028: Two scenarios for global AI leadership</em></a>, 14 May 2026. The source of the <em>Mythos Preview</em> reference, the <em>country of geniuses in a datacenter</em> framing, and the <em>breakaway opportunity for American AI</em> claim.</li>
    <li>Anthropic, <a href="https://www.anthropic.com/glasswing"><em>Glasswing</em></a> — the page on Anthropic's corporate site referenced by <em>Mythos Preview</em>. At the time of writing, no product description, system card, or technical specification has been published.</li>
  </ul>
</div>

<div class="lf-endnotes-group">
  <h4>The three protagonists</h4>
  <ul>
    <li>Dario Amodei, <a href="https://www.darioamodei.com/essay/machines-of-loving-grace"><em>Machines of Loving Grace</em></a>, October 2024.</li>
    <li>The Royal Swedish Academy of Sciences, <a href="https://www.nobelprize.org/prizes/chemistry/2024/summary/">2024 Nobel Prize in Chemistry</a>.</li>
    <li>Google DeepMind, <a href="https://deepmind.google/technologies/alphafold/">AlphaFold</a>.</li>
    <li>Leopold Aschenbrenner, <a href="https://situational-awareness.ai/"><em>Situational Awareness: The Decade Ahead</em></a>, June 2024.</li>
  </ul>
</div>

<div class="lf-endnotes-group">
  <h4>What AGI means</h4>
  <ul>
    <li><a href="https://openai.com/charter/">OpenAI Charter</a> — the source of the <em>highly autonomous systems that outperform humans at most economically valuable work</em> definition.</li>
    <li>Morris, Schaul, Hadfield et al., <a href="https://arxiv.org/abs/2311.02462"><em>Levels of AGI for Operationalizing Progress on the Path to AGI</em></a>, Google DeepMind, arXiv 2311.02462 (2023).</li>
    <li>OpenAI, <a href="https://openai.com/index/built-to-benefit-everyone/"><em>Built to benefit everyone</em></a>, 28 October 2025 — the recapitalisation document establishing the independent expert panel mechanism for AGI determination.</li>
    <li>CNBC, <a href="https://www.cnbc.com/2025/10/28/open-ai-for-profit-microsoft.html"><em>OpenAI completes restructure, solidifying Microsoft as a major shareholder</em></a>, 28 October 2025 — complementary reporting on Microsoft's continued IP rights through 2032.</li>
    <li><a href="https://arcprize.org/">ARC Prize Foundation</a> — host of the ARC-AGI benchmark family, including ARC-AGI-3.</li>
  </ul>
</div>

<div class="lf-endnotes-group">
  <h4>The case it's close</h4>
  <ul>
    <li><a href="https://www.swebench.com/">SWE-bench</a>, the verified real-world software engineering benchmark.</li>
    <li><a href="https://openai.com/codex/">OpenAI Codex</a>.</li>
    <li>Anthropic, <a href="https://www.anthropic.com/claude-sonnet-4-5-system-card">Claude Sonnet 4.5 System Card</a>, September 2025.</li>
    <li>Anthropic, <a href="https://www.anthropic.com/claude-opus-4-6-system-card">Claude Opus 4.6 System Card</a>, February 2026.</li>
    <li>Google DeepMind, <a href="https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/"><em>AlphaEvolve: a Gemini-powered coding agent for designing advanced algorithms</em></a>, 2024 — source for the four-by-four complex matrix multiplication result.</li>
    <li>Sakana AI, <a href="https://sakana.ai/ai-scientist/"><em>The AI Scientist</em></a> — the end-to-end automated ML research pipeline.</li>
  </ul>
  <p class="lf-endnotes-note">Note. The essay refers to <em>Gemini 3.1 Deep Think</em> as a 2026 DeepMind product. The closest verifiable real-world artefact at the time of writing is Gemini 2.5 Deep Think (2025); a primary URL for the specific 3.1-generation product has not been retrieved.</p>
</div>

<div class="lf-endnotes-group">
  <h4>The case it isn't</h4>
  <ul>
    <li>ARC Prize Foundation, <a href="https://arcprize.org/">ARC-AGI-3 benchmark and technical report</a>.</li>
    <li>METR, <a href="https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/"><em>Measuring the impact of early-2025 AI on experienced open-source developer productivity</em></a>, July 2025 — the randomised controlled trial finding a 19% slowdown.</li>
    <li><a href="https://ai.meta.com/people/yann-lecun/">Yann LeCun</a>, Meta AI scientist profile and public commentary.</li>
    <li><a href="https://garymarcus.substack.com/">Gary Marcus</a>, ongoing public commentary on the limits of LLM-based approaches to AGI.</li>
  </ul>
  <p class="lf-endnotes-note">Note. The essay refers to Microsoft Research's DELEGATE-52 benchmark and its 25% silent-corruption finding. The paper is documented in the supporting research materials but a primary URL has not been verified to the standard required for direct hyperlinking; the underlying claim is being checked against Microsoft Research's publications page before any subsequent print run.</p>
</div>

<div class="lf-endnotes-group">
  <h4>The underbelly</h4>
  <ul>
    <li>Anthropic, <a href="https://www.anthropic.com/claude-opus-4-5-system-card">Claude Opus 4.5 System Card</a>, November 2025 — the original CBRN-4 rule-out admission.</li>
    <li>Anthropic, <a href="https://www.anthropic.com/claude-opus-4-6-system-card">Claude Opus 4.6 System Card</a>, February 2026 — the <em>may soon be difficult or impossible under the current regime</em> passage.</li>
    <li>Anthropic, <a href="https://anthropic.com/responsible-scaling-policy/rsp-v3-0">Responsible Scaling Policy v3.0</a> — the version in which the autonomous-replication threshold was downgraded to a <em>checkpoint</em>.</li>
    <li>Anthropic, <a href="https://www-cdn.anthropic.com/files/4zrzovbb/website/bf04581e4f329735fd90634f6a1962c13c0bd351.pdf">Responsible Scaling Policy v3.1 (PDF)</a>.</li>
    <li>Google DeepMind, <a href="https://deepmind.google/blog/strengthening-our-frontier-safety-framework/"><em>Strengthening our Frontier Safety Framework</em></a> — source for the instrumental-reasoning mitigation passage.</li>
    <li><a href="https://www.apolloresearch.ai/">Apollo Research</a>.</li>
    <li>Apollo Research, <a href="https://www.apolloresearch.ai/blog/apollo-update-may-2026/"><em>Apollo Update May 2026</em></a> — announcing the retirement of the primary scheming-evaluation function in favour of a <em>Science of Scheming</em> research agenda.</li>
    <li>Schoen et al., <a href="https://arxiv.org/abs/2509.15541"><em>Stress Testing Deliberative Alignment for Anti-Scheming Training</em></a>, Apollo Research and OpenAI, arXiv 2509.15541, September 2025.</li>
    <li>Anthropic, <a href="https://assets.anthropic.com/m/99128ddd009bdcb/Claude-Haiku-4-5-System-Card.pdf">Claude Haiku 4.5 System Card</a>, October 2025 — the 9% verbalised evaluation-awareness finding.</li>
    <li>Future of Life Institute, <a href="https://futureoflife.org/ai-safety-index-summer-2025/">AI Safety Index Summer 2025</a> — the Existential Safety D-grades.</li>
  </ul>
  <p class="lf-endnotes-note">Note. The essay refers to a paper providing "the first documented evidence of strategic deception at the level of a model's internal computational state." The claim is associated in the supporting research materials with work by Bondarenko and colleagues; a precise arXiv identifier has not been verified to the standard required for direct hyperlinking.</p>
</div>

<div class="lf-endnotes-group">
  <h4>If they are right</h4>
  <ul>
    <li>Bengio et al., <a href="https://internationalaisafetyreport.org/publication/international-ai-safety-report-2026"><em>International AI Safety Report 2026</em></a>, February 2026 — chair Yoshua Bengio; multilateral expert synthesis.</li>
  </ul>
</div>

<div class="lf-endnotes-group">
  <h4>Romandy</h4>
  <ul>
    <li>ETH Zurich, <a href="https://ethz.ch/en/news-and-events/eth-news/news/2025/09/press-release-apertus-a-fully-open-transparent-multilingual-language-model.html"><em>Apertus: a fully open, transparent, multilingual language model</em></a>, 2 September 2025.</li>
    <li>EPFL, <a href="https://actu.epfl.ch/news/apertus-a-fully-open-transparent-multilingual-lang/"><em>Apertus: a fully open, transparent, multilingual language model</em></a> — parallel announcement, including statements from Jaggi, Bosselut, Schlag and Schulthess.</li>
    <li><a href="https://swiss-ai.org/">Swiss AI Initiative</a>.</li>
  </ul>
</div>