By Thorsten Meyer — May 2026
On May 4, 2026, Jack Clark — Anthropic co-founder and head of policy — published Import AI #455 with the title “Automating AI Research” and a single sentence near the top that constitutes one of the most consequential public statements ever made by a frontier-lab leader on takeoff timelines: “I reluctantly come to the view that there’s a likely chance (60%+) that no-human-involved AI R&D — an AI system powerful enough that it could plausibly autonomously build its own successor — happens by the end of 2028.”
That sentence deserves to be read slowly. No-human-involved AI R&D. An AI system that can train its own successor. By the end of 2028. With 60%+ subjective probability assigned by the head of policy at one of the three frontier labs that would actually know.
The reaction in the Bay Area technical community over the weekend after publication followed a predictable pattern. The accelerationist contingent treated the essay as confirmation of what they had been saying for two years. The safety contingent treated it as the most honest public statement yet about timelines they had been forecasting privately for longer. The skeptical contingent treated it as a marketing exercise designed to maintain Anthropic’s positioning ahead of its IPO. All three readings miss what the essay actually is. It is the head of policy at a frontier lab publishing a probabilistic forecast that explicitly says — in its own concluding lines — “we may be about to witness a profound change in how the world works.” This is not a forecast in the abstract analytical sense. It is a policy statement.
This dispatch is the structural read on what the statement is and what it means. The evidence Clark assembles, the benchmark cascade, the compounding-error problem, and the machine-economy implication get their own pieces. What this piece is about is the fact of the statement itself — who said it, when they said it, under what institutional conditions, with what positioning, and what the act of saying it commits Anthropic and the broader frontier ecosystem to.
Sixty percent
by twenty-twenty-eight.
A frontier-lab co-founder publishes a probabilistic forecast on automated AI R&D arrival. The institutional weight exceeds the analytical weight.
May 4, 2026 · Import AI #455 contains a single sentence that constitutes one of the most consequential public statements ever made by a frontier-lab leader on takeoff timelines. The fact of the statement matters as much as its content. The AGI debate is now closed for the people who would know. The question is what we do during the window the forecast describes.
Clark fills the empty seat.
The takeoff-timeline forecasting discourse has been continuous since 2022 but conducted almost entirely by researchers, ex-employees, and outside commentators. No sitting frontier-lab co-founder had published a numerical probability on a specific takeoff threshold within a specific timeframe. Until May 4, 2026.
Perplexity AI, A Practical Guide to Powered Research: Smart Search, Workflow Automation, and Knowledge Discovery in 2026 (The Practical Guide to Modern AI Tools)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Public forecasts create commitments.
Senior executives publishing probabilistic forecasts create operational obligations even when presented as personal analysis. Anthropic must now act as if the forecast is approximately right — internally, regulatorily, and in coordination with peers.
Generative AI Design Patterns: Solutions to Common Challenges When Building GenAI Agents and Applications
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Five disagreements. Five different magnitudes.
Not every credible observer will share Clark’s 60%/2028. The honest disagreement isn’t about whether AI capability is improving — it’s about whether the curve continues, whether compute supply binds first, whether shocks intervene.
Building Agentic AI Systems: Create intelligent, autonomous AI agents that can reason, plan, and adapt
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Four stakeholders. Four obligations.
The Clark essay doesn’t change capability trajectory. What it changes is the public-domain epistemic situation. Anyone modeling AI deployment must now account for the institutional position.
The AGI debate is now closed for the people who would know. The question that remains is what we do during the window in which we still have time to act.
AI Training Infrastructure Architecture: Designing Scalable, Distributed, and Cost-Efficient Training Systems for Enterprise Machine Learning & Foundation … (AI Systems Architecture Series Book 4)
As an affiliate, we earn on qualifying purchases.
As an affiliate, we earn on qualifying purchases.
Why a co-founder saying this is different from a researcher saying it
The AI takeoff timeline discourse has been continuous since 2022, but it has been almost entirely conducted by researchers, forecasters, and outside commentators. Ajeya Cotra’s biological-anchors work. Daniel Kokotajlo’s AI-2027 scenario. Leopold Aschenbrenner’s Situational Awareness. The METR team’s time-horizons measurements. Various Anthropic and DeepMind researchers publishing in personal capacity. Even Sam Altman’s various tweets about AGI timelines — which technically come from a frontier-lab CEO but read as marketing copy more than serious forecast.
What none of this prior discourse contains is a senior frontier-lab executive publishing, in their official institutional voice, a numerical probability estimate on a specific takeoff trajectory within a specific timeframe. Clark’s 60%/2028 is, as best I can determine, the first such statement. It matters in a way that exceeds its content because of who is making it under what institutional constraints.
Jack Clark is Anthropic’s co-founder. He is the head of policy. He communicates regularly with the U.S. government, with foreign governments, with regulatory bodies, with congressional staff, and with the broader policy community on Anthropic’s behalf. His public statements on AI policy are read by people who have actual power to shape AI regulation. A 60%/2028 estimate published in his official capacity is, by definition, the position Anthropic is comfortable having its head of policy hold publicly. That has institutional weight that a personal-capacity blog post from a researcher does not.
It also creates institutional commitment. Clark cannot now, in 2027 or 2028, walk back the estimate without making Anthropic’s policy positioning look performative. If the trajectory turns out slower than 60%/2028 suggests, Clark has staked his credibility on a forecast that did not materialize. If the trajectory turns out faster than 60%/2028 suggests, Clark has demonstrated that the frontier-lab insider view in May 2026 was an understatement, which is the more dangerous failure mode from a societal perspective. Either way, the statement has weight that researcher commentary does not — because it was made by someone who will be held to it.
The closest historical analogue I can construct: when Geoffrey Hinton resigned from Google in May 2023 and publicly stated his concerns about AI risk, his statements carried institutional weight specifically because of his prior position at a frontier lab. Hinton resigned in order to speak freely. Clark is making a comparable statement without resigning, while still inside his institutional position. That is a different kind of move. It signals that Anthropic, as an institution, has decided this view can be expressed in public by its head of policy. The signal is meaningful even before you read the content.
What the essay actually argues, briefly
Clark’s argument is structurally simple. AI systems have improved dramatically on benchmarks that test the specific skills required for AI engineering: writing code, reproducing research papers, fine-tuning models, designing kernels, managing other AI systems. The improvement curves are public, monotonic, and currently accelerating. Frontier labs and well-funded neolabs are explicitly targeting automated AI R&D as a product goal. The capital being deployed against this goal is in the hundreds of billions of dollars. Given the trajectory and the alignment of incentives, the probability that we cross the threshold of “AI system trains its own successor without human involvement” by end of 2028 sits somewhere around 60%.
The essay’s main intellectual move is the distinction between AI engineering and AI research. Engineering is the methodical work of writing code, running experiments, debugging, optimizing, scaling. Research is the work of generating novel ideas about what to try next. Clark argues that current AI systems can probably already automate “vast swatches, perhaps the entirety, of AI engineering,” and that the question of whether they can automate AI research depends on whether AI research requires genuine inventiveness or whether — Edison-style — it is 99% perspiration that can be done by skilled imitation of existing patterns. Clark is non-committal on this question but notes that even if research requires genuine creativity, AI engineering automation alone produces a profound acceleration that materially changes the world.
The essay then lists three implications: alignment becomes critical because recursive self-improvement compounds errors, AI introduces massive productivity multipliers that raise allocation and inequality questions, and the economy bifurcates into a “machine economy” growing inside the “human economy.” Each implication is sketched rather than developed. The essay closes with the 60%/2028 estimate and a remarkable closing line: “I have written this essay in an attempt to coldly and analytically wrestle with something that for decades has seemed like a science fiction ghost story. Upon looking at the publicly available data, I’ve found myself persuaded that what can seem to many like a fanciful story may instead be a real trend.”
That closing sentence is the entire essay in miniature. A frontier-lab insider, examining publicly available data, concluding that science-fiction scenarios may be becoming science fact. The honest emotional register of the sentence — reluctantly, dwarfed by them, I’m not sure society is ready — is also worth noting. This is not triumphalist. This is not marketing. This is, by every textual indicator I can read, a person genuinely wrestling with something they did not expect to find.
The position-relative-to-historical-statements question
How does Clark’s 60%/2028 estimate compare to the existing landscape of public AI timeline forecasts? A non-exhaustive survey:
Sam Altman, OpenAI CEO, has made various public statements on AGI timelines, generally trending toward “soon” without specific probability assignments. His tweet about “automated AI research intern by September 2026” (referenced by Clark) is the closest he has come to a specific operational target. Altman’s general public framing has been promotional rather than analytical.
Demis Hassabis, DeepMind co-founder and CEO, has been more circumspect, generally citing 5-10 year horizons for AGI without pinning specific probabilities to specific years. DeepMind’s institutional voice has been the most measured of the big three frontier labs.
Dario Amodei, Anthropic CEO, has published longer-form essays (Machines of Loving Grace, October 2024) that explicitly model “powerful AI” arrival around 2026-2027 with various caveats. Amodei’s framing has been more concrete than Altman’s but has not, to my knowledge, attached specific numerical probabilities to specific takeoff thresholds.
Leopold Aschenbrenner, ex-OpenAI, published Situational Awareness in June 2024 with an explicit forecast of AGI by 2027 and superintelligence by 2030, with specific compute trajectories backing the forecast. Aschenbrenner’s work is the most concrete prior forecast in the public domain — but Aschenbrenner was speaking as an ex-employee, with no institutional commitment to defend.
Daniel Kokotajlo, ex-OpenAI, published the AI-2027 scenario (with co-authors) in April 2025, describing a specific trajectory by which superintelligence emerges by end of 2027 through recursive self-improvement starting from automated AI R&D. Kokotajlo’s scenario is structurally similar to what Clark describes but resolves earlier (2027 vs Clark’s 2028 with 30% probability). Like Aschenbrenner, Kokotajlo speaks as ex-employee without institutional constraint.
METR researchers, including Ajeya Cotra (cited in Clark’s essay), have published time-horizons measurements that imply specific extrapolation curves but typically resist attaching probabilistic forecasts to specific takeoff thresholds.
Clark’s 60%/2028 estimate is more specific than Altman, Hassabis, or Amodei, less aggressive than Aschenbrenner or Kokotajlo, and uniquely backed by current institutional authority. It is the first time a sitting frontier-lab co-founder has publicly committed to a specific probability on a specific takeoff threshold within a specific timeframe. The history of AI timeline forecasting in 2026 has a clear before-Clark and after-Clark.
The closest prior statement from inside Anthropic is Dario Amodei’s Machines of Loving Grace essay, which projected powerful AI around 2026-2027 but framed the trajectory in capability terms rather than specifically about automated AI R&D. Clark’s essay is the operationalization of what Amodei sketched — a specific threshold, a specific probability, a specific date. The two essays read in sequence as the increasingly precise public statement of Anthropic’s institutional view on what is happening.
What the statement commits Anthropic to operationally
Public probability forecasts from senior executives create operational obligations even when they are presented as personal analysis. Three obligations follow from Clark’s essay:
First, Anthropic must now act as if the forecast is approximately right. A frontier lab whose head of policy publishes a 60%/2028 estimate on automated AI R&D cannot operate as if that threshold is 10+ years away. The internal safety case, the alignment research portfolio, the compute allocation toward interpretability research, the timeline for various policy engagements, the structure of the Long-Term Benefit Trust governance, the IPO disclosure language — all of these must be calibrated to the forecast. Anthropic has, in effect, told its regulators and the public that it operates inside a 32-month window in which a specific civilizational threshold may be crossed. The institutional behavior must match.
Second, Anthropic must now share evidence of its operating assumptions. If 60%/2028 is the institutional view, the public, regulators, and Anthropic’s customers have legitimate questions about what Anthropic is doing about it. The Responsible Scaling Policy (RSP) framework, the deployment caution that delays product launches, the alignment-research-as-product-development integration, the work on automated alignment research (referenced in Clark’s essay) — these all become more legible as responses to a concrete near-term threshold rather than as abstract good-corporate-citizenship gestures. Anthropic will be asked to show its work in greater detail than it has historically been comfortable with.
Third, Anthropic must now coordinate. If 60%/2028 is the forecast, the question of what happens when (not if) automated AI R&D arrives becomes a coordination problem across labs, governments, and the broader public. A lab that publishes this forecast and then races to the threshold without coordination has effectively admitted to creating the danger it claims to be trying to manage. Anthropic’s stated position on coordination — that it would prefer to slow down if competitors would slow down too — gets tested in concrete ways over the next 24-36 months. Clark’s essay raises the stakes on every coordination question.
The implicit reading: Clark’s essay is partly a request for outside pressure to make coordination possible. By publishing the forecast in his official capacity, Clark gives regulators, journalists, and the broader public a documented institutional position they can hold Anthropic to. He is, in effect, asking to be held accountable to the implications of his own analysis. This is a sophisticated move. It is also a high-trust one — it requires believing that publishing this forecast will produce more useful response than not publishing it. Whether that bet pays off is one of the things we’ll find out over the next 24-36 months.
What the statement does to the broader frontier-lab discourse
Clark’s essay changes the public discourse in three concrete ways even if no other lab leader responds:
One. It raises the floor on what counts as a credible public statement about takeoff timelines. Future statements from lab leaders that decline to engage with specific probabilities and specific timeframes will read as evasive in a way they did not before May 4. Hassabis, Amodei (separately from Clark), Altman, and any future frontier-lab executive will be asked: do you agree with Clark’s 60%/2028 estimate? If not, what is your estimate? The question becomes legitimate to ask in a way it was not before.
Two. It makes the existing public forecasts (Aschenbrenner, Kokotajlo, AI-2027) more legible. Outside forecasters who had been making concrete predictions can now point to Clark’s essay as evidence that their framing was not eccentric — it was the same framing the frontier-lab insiders were using internally, now made public. The accelerationist and safety communities both get vocabulary they can use without sounding fringe.
Three. It puts implicit pressure on competing labs to either match Clark’s transparency or explain why they aren’t matching it. OpenAI’s communications around takeoff have been less specific than Clark’s; if 60%/2028 is the Anthropic view, what is the OpenAI view? Are they more confident or less confident? Are they planning for a faster trajectory or a slower one? The pressure to answer these questions becomes harder to deflect.
The structural reading: by going on record first, Clark has framed the conversation. Subsequent statements from other lab leaders will be read against his benchmark. If competitors come in higher (say 75%/2027), the public conversation becomes about how to coordinate during an extremely short window. If competitors come in lower (say 30%/2030), the question becomes why Anthropic’s internal view is more pessimistic — and Anthropic will be asked whether they are seeing something the others aren’t.
What honest disagreement with the estimate looks like
Not every credible observer will share Clark’s 60%/2028 estimate. The reasons to disagree, in order of intellectual seriousness:
Benchmark saturation does not equal capability transfer. The fact that AI systems are saturating SWE-Bench, CORE-Bench, MLE-Bench, and PostTrainBench does not mean they can do AI research. The benchmarks measure specific tasks under specific conditions; AI research in practice involves taste, scientific intuition, ability to identify productive research directions, and a host of meta-skills that benchmarks may not capture. A 95.5% score on CORE-Bench means an AI system can reproduce existing research papers; it does not mean the AI system can generate the next research paper. This is a legitimate concern and Clark addresses it explicitly but does not fully resolve it.
The METR time-horizons curve may not extrapolate. The curve has been roughly exponential for four years, with doubling times around 7 months. Extrapolating that curve to end-2028 implies AI systems capable of tasks taking tens of thousands of hours of human researcher labor — i.e., entire research projects. The extrapolation may break down at some point, either because we hit a fundamental capability ceiling, or because the curve was always more sigmoid than exponential, or because the metric stops meaning what it has been measuring. The track record of “this exponential will continue” forecasts in technology is mixed.
Compute supply may bind before capability ceiling. Even if the algorithmic capability exists to automate AI R&D, the actual deployment of automated research requires compute resources that are constrained by physical buildout (data centers, GPU supply, power, water cooling, transmission infrastructure). The “compute reckoning” thesis covered in the recent dispatch on Anthropic-SpaceX is real. The 60%/2028 forecast assumes that compute scaling continues at projected trajectory. If compute scaling slows due to physical constraints or capital reallocation, the timeline slips.
Geopolitical / regulatory shocks may reshape the trajectory. A major safety incident at a frontier lab, a serious policy intervention by a major government, a major escalation in AI export restrictions, a Chinese frontier capability breakthrough that changes the competitive dynamic — any of these can change the trajectory in ways the current forecast does not model. 32 months is a long time for geopolitical shocks.
The forecast may be self-defeating. If Clark’s essay produces sufficient policy response, sufficient public pressure, sufficient coordination among labs, or sufficient alignment-research investment, the trajectory may slow specifically because of the forecast itself. This is the most interesting failure mode — Clark’s prediction failing because Clark’s prediction caused enough response to bend the curve. From a societal-welfare perspective, this is the failure mode we should hope for.
I would put my own subjective probability somewhere between Clark’s 60%/2028 and a more conservative 40%/2028, mostly because I weight the compute-supply and regulatory-shock factors more heavily than Clark does in the essay. But I want to be precise about what I’m disagreeing with: Clark is making a forecast about what can happen if current trends continue and no major shocks intervene. The shocks may intervene. I think they will intervene. But the existence of shocks that bend the curve does not make the underlying capability trajectory wrong. It just means deployment will be slower than the capability would predict.
The Anthropic IPO context
Clark’s essay arrives in a specific corporate-finance context that deserves naming. Anthropic is in late-stage preparation for an IPO at a reported $900 billion valuation, with Q4 2026 the likely timing per multiple reports. The May 4 essay sits inside the IPO disclosure preparation window — the period in which lab leadership is making public statements that will become part of the IPO’s documented public-record context.
There are two ways to read the essay in this context. The cynical reading is that the essay is IPO marketing — Anthropic positioning itself as the frontier lab with the most honest public take on takeoff, in order to differentiate from OpenAI and Google in the eyes of public-market investors. The cynical reading is not unreasonable; Anthropic does benefit from being the lab with the most credible safety-and-public-interest framing, and Clark’s essay reinforces that framing.
The non-cynical reading is that Clark publishes the essay precisely because the IPO context makes it more difficult to walk back later. By publishing 60%/2028 in advance of IPO disclosure, Clark creates a public record that the IPO documents will be expected to acknowledge. This is the inverse of the cynical reading: Clark is constraining Anthropic’s future flexibility by publishing the forecast first.
Both readings are partially correct. The essay does serve Anthropic’s IPO positioning. The essay also constrains Anthropic’s future actions in ways that pure marketing would not. The two effects coexist. The honest synthesis: Anthropic has decided that the cost of publishing the forecast publicly (operational obligations, IPO-disclosure constraints, reputational risk if wrong in either direction) is lower than the cost of not publishing (coordination failure, public-interest argument loss, broader frontier-lab discourse remaining unhealthy).
This is a calculated bet by Anthropic. It is also, by my read, the right bet — the alternative of frontier-lab insiders holding 60%/2028 views privately while the public discourse remains uninformed produces worse outcomes than the forecast being published, even with all the marketing-positioning issues attached. Clark deserves intellectual credit for making the calculation visible.
What changes for everyone reading this
The Clark essay does not change capability trajectory. The capability trajectory is what it is; Clark is observing it, not creating it. What the essay changes is the public-domain epistemic situation:
For frontier-lab investors. Anthropic is now on record with a 60%/2028 institutional view. Apply this to valuation models, particularly to discount rates on terminal-value calculations. If the forecast is even directionally correct, the trajectory through 2028 may compress decades of value into a 24-36 month window — which has implications for IPO valuation, for compute capex deployment, and for the structural value of frontier-lab equity. The valuation models that assume gradual AGI emergence over 2030-2040 are now in tension with the public statement of one of the three labs whose work the models depend on.
For policy professionals. The Clark essay raises the institutional weight of policy positions on AI takeoff. Anything you have been working on that depends on a slower trajectory needs to be re-examined. The U.S. Executive Order framework, the EU AI Act implementation timeline, the UK AI Safety Institute’s evaluation cadence, the various federal agency efforts — all of these were calibrated to a trajectory that was implicit rather than explicit. Clark has made the trajectory explicit. Policy calibration follows.
For workers in cognitive-task labor markets. The 60%/2028 forecast is about AI R&D specifically, but the implications generalize. If AI can do AI research, it can do a substantial fraction of all knowledge work. The labor displacement piece already covered the early signal in junior software engineering cohorts. The Clark forecast implies the signal becomes the trend faster than most workforce planning currently assumes. Workforce-level response — reskilling programs, transition support, social-safety-net adjustments — needs to operate on a faster cadence than current institutional capacity supports.
For everyone else. The phrase “we may be about to witness a profound change in how the world works” was published on May 4, 2026, by a person institutionally positioned to know. This is not science fiction. This is not marketing. This is the head of policy at one of the three frontier labs explicitly saying that the current path may produce, within 32 months, a system that can train its own successor. Read it. Sit with it. Make whatever decisions you need to make about your own position, your own work, your own life, in light of the possibility that the analysis is correct.
The closing read
Jack Clark’s Import AI #455 is one of the most significant public statements ever made by a frontier-lab leader. The 60%/2028 estimate on automated AI R&D arrival is now part of the public record. The institutional weight of who said it under what conditions exceeds the analytical weight of the underlying argument. The argument itself is rigorous, well-sourced, and intellectually serious. The act of publishing it is a policy intervention as much as a forecast.
The four pieces that follow this one — the benchmark cascade, the compounding-error problem, the machine-economy implication, and the synthesis centerpiece — work through the substance of Clark’s argument in detail. This piece is about the fact of the statement itself. The fact of the statement is the news. The substance of the statement is the situation we now occupy.
Anthropic’s head of policy publishes 60%/2028 on automated AI R&D. Read it, again, slowly: the head of policy at Anthropic believes there is a 60%+ chance that, within 32 months, an AI system will be capable of training its own successor without human involvement. That is the situation as of May 4, 2026. The next 32 months will be spent learning whether the forecast is approximately right, approximately wrong, or some combination — and learning what we collectively are able to do during the window the forecast describes.
The honest assessment from where I sit: the AGI debate is now closed for the people who would know. The question that remains is what we do during the window in which we still have time to act.
About the Author
Thorsten Meyer is a Munich-based futurist, post-labor economist, and recipient of OpenAI’s 10 Billion Token Award. He spent two decades managing €1B+ portfolios in enterprise ICT before deciding that writing about the transition was more useful than managing quarterly slides through it. More at ThorstenMeyerAI.com.
Related Reading
- The Benchmark Saturation Cascade — Piece 2 in this series
- The Compounding Error Problem — Piece 3 in this series
- The Machine Economy — Piece 4 in this series
- The Co-Founder’s Black Hole · Synthesis — Piece 5 in this series
- The State of AI Replacing Jobs in 2026
- The Anthropic IPO Disclosure Document
- Post-Labor Economics
Sources
- Jack Clark · Import AI 455: Automating AI Research · May 4, 2026 · jack-clark.net
- Anthropic · automated alignment researchers research note · 2026
- METR · time horizons measurement curve · 2022-2026
- Sam Altman · “automated AI research intern by September 2026” · X / Twitter
- Recursive Superintelligence · $500M raise · automating AI research goal · Financial Times
- Mirendil · neolab · “building systems that excel at AI R&D”
- Geoffrey Hinton · public statements on AI risk · May 2023
- Daniel Kokotajlo et al. · AI-2027 scenario · April 2025
- Leopold Aschenbrenner · Situational Awareness · June 2024
- Dario Amodei · Machines of Loving Grace · October 2024
- Ajeya Cotra · METR · time horizons forecasting work
