On May 4, 2026, Anthropic co-founder Jack Clark posted on social platform X: “I now believe there is a 60% probability that recursive self-improvement will occur before the end of 2028.”
Just minutes after the post was published, Eliezer Yudkowsky, a long-active researcher in the field of AI safety, replied with: “Then we will all perish.” He followed this by quoting an analogy pointing to the design flaws of the Chernobyl nuclear reactor’s RBMK, suggesting that the system being activated is one that no one truly knows how to shut down.
This conversation, completed in seconds, lit a match that ignited discussions long buried in technical papers and internal evaluations. Recursive Self-Improvement (RSI)—the concept that AI systems can not only optimize their outputs but also autonomously improve their own improvement processes, ultimately building successor systems stronger than themselves—has been placed by Anthropic’s co-founders into a countdown clock with a 60% probability of occurring by the end of 2028.
A month later, Anthropic officially published a lengthy article titled “When AI Builds Itself.” Written jointly by Marina Favaro and Jack Clark and released by the Anthropic Institute, which was established in March, the article presented an externally calibrated acceleration signal using a series of previously unpublished internal data and a meticulously crafted narrative structure. The signal conveyed both “We have not yet arrived there” and “But it may arrive faster than most institutions are prepared for.”
In the same month, DeepMind CEO Demis Hassabis used an unprecedented phrase on the Google I/O stage: humanity is standing at the “foot of the singularity.” In a subsequent interview, he adjusted the timeline for artificial general intelligence (AGI) from “not long after 2030” to “2029 is a real possibility,” and admitted that his use of dramatic language was “intentionally provocative,” aimed at creating urgency among governments, economists, and the public.
Two leading institutions, known for their emphasis on security and long-standing role as moderating forces in the AI industry, nearly simultaneously adjusted the volume and tone of their public communications. This timing itself warrants being treated as a standalone event.
A meticulously calibrated long-form article
Anthropic’s lengthy article, released on June 4, immediately laid out its narrative goal: to argue not just for a technological trend, but for a directional, accelerating process. To support this, it unveiled a set of internal data never before made public.
The first set of numbers points to a structural shift: as of May 2026, over 80% of merged code in Anthropic’s codebase was written by Claude. Two years ago, this figure was in the low single digits. The same data also shows that, in the second quarter of 2026, the typical Anthropic engineer merged eight times as much code per day as in 2024.
One can imagine the reaction of anyone unfamiliar with the AI industry upon first encountering these two numbers. However, Anthropic itself acknowledges several important caveats in its footnote: leadership has publicly estimated that, when including scripts and experimental code, Claude writes over 90% of the code, with 80% representing a more conservative measure of merged code; lines of code are an “imperfect metric” and may overstate actual productivity gains; and the code attribution pipeline itself has “gaps.”
The way these footnotes are written is itself worth analyzing. On the surface, their presence constitutes an honest concession, but in reality, they serve to make the numbers in the main text appear as if they have undergone careful self-filtering, thereby enhancing their credibility. This is a two-layer narrative structure: the main text conveys the signal, while the footnotes provide disclaimers.
The second set of numbers relates to speed. On code optimization tasks, Claude Opus 4 achieved approximately a 3x speedup in May 2025, a level that would take a skilled human researcher 4 to 8 hours to match. By April 2026, Claude Mythos Preview pushed this figure to about 52x. The maximum duration for which AI could independently complete tasks also doubled every four months, rising from 4 minutes in March 2024 to 12 hours by March 2026. The very pace of doubling every four months creates a highly memorable, geometrically imaginative point that is easily spread.
Another set of data comes from an internal survey of 130 Anthropic research team members conducted in March 2026. The median respondent estimated that output using Mythos Preview is about four times higher than without AI. The footnote again notes that METR’s prior independent research suggested that developers’ estimates of AI’s productivity gains may be generally overstated. The same two-tier structure appears again.
The third set of figures indicates that AI is approaching the boundary of human researchers' judgment. In November 2025, Claude Opus 4.5 outperformed human researchers in selecting research directions in 51% of cases. By April 2026, this figure rose to 64%. The sample size consists of 129 cases, and Anthropic notes in the footnote that these cases were deliberately selected by humans as moments where human choices had room for improvement.
Isolate any single number, and it can fit into different interpretive frameworks. But together, the direction is clear: speed is increasing, the gap is narrowing, and all of this is happening within Anthropic’s own codebase and labs—not as theoretical projections on an external benchmark.
After presenting these data points, the long-form article outlines three future scenarios.
The first is trend stagnation, entering an S-curve plateau. Anthropic's wording is: "We do not believe this is likely."
The second is composite efficiency gains, where AI continuously replaces humans across a broader range of R&D tasks, while humans still set the direction and define success criteria. Anthropic assesses this as "evidence suggesting we are likely moving toward this scenario."
The third is complete recursive self-improvement, where AI autonomously designs, trains, and deploys successor systems more powerful than itself, with humans no longer part of the loop. The phrasing is "possible."
The arrangement and tone allocation of these three scenarios form a complete narrative gradient. The first is lightly presented, serving to accommodate skeptics; the second is anchored in “evidence,” lending the article a rational veneer; the third, through phrases like “possible” and the conditional “if technological trends continue,” pushes the boldest assumption to the edge of the reader’s imagination without requiring it to bear the burden of proof.
At the very heart of the article, Anthropic’s stance is condensed into one sentence: “We haven’t gotten there yet, and recursive self-improvement is not inevitable—but it may arrive faster than most institutions are prepared for.”
From "willing to pause" to "unilateral suspension only allows the reckless to catch up"
If the June 4 long-form article was a carefully composed snapshot, placing that snapshot on a timeline reveals a longer trajectory.
In 2023, Anthropic released the Responsible Scaling Policy (RSP). The core commitment of this policy document is that the company will halt training more powerful models if their capabilities exceed the company’s safety controls. This is not merely a verbal statement, but an internal governance document with an evaluation framework and specific triggering conditions. At one point, this document was regarded by the AI safety community as a practical example of voluntary regulation.
In 2024, CEO Dario Amodei published a widely circulated article suggesting the possibility that "powerful AI" would arrive by 2027. At that time, Anthropic still presented itself as an independent voice focused on safety, maintaining a restrained stance toward scaling and acceleration narratives.
On January 26, 2026, Amodei published a 38-page essay titled “The Adolescence of Technology” on his personal website. In it, he made a statement that has since been widely cited: “Since AI is now writing most of the code inside Anthropic, it is already substantially accelerating our progress in building the next generation of AI systems. This feedback loop is gaining momentum month by month and may be just one to two years away from the current generation of AI autonomously building the next.” In the same article, he described the impending “powerful AI” as “a genius nation inside a data center.”
This was nearly the starting point for Anthropic’s systematic signaling that a self-improvement feedback loop is underway. The timing of this blog post coincides with the company’s transition from a $350 billion valuation to a higher valuation range.
Less than a month later, a turning point arrived.
On February 25, 2026, CNN reported that Anthropic revised its Responsible Scaling Policy, removing its core commitment to pause training of more powerful models if capabilities exceed safety controls, replacing it with a non-binding “Frontier Safety Roadmap.” That same week, U.S. Secretary of Defense Pete Hegseth issued an ultimatum to Dario Amodei: withdraw the safety red lines or lose a $200 million Department of Defense contract.
The report quoted Anthropic’s Chief Scientific Officer, Jared Kaplan, responding to Time magazine: “We believe stopping training models doesn’t help anyone… if competitors are going all out.” The wording in this response is highly noteworthy. “Doesn’t help anyone” is not a technical argument, but a statement rooted in stakeholder dynamics. “If competitors are going all out” mirrors the exact narrative structure of “Unilateral pauses only allow the least cautious participants to catch up”: it replaces the original logic of suspension based on one’s own safety capabilities with a speed-based logic defined by competitors’ actions.
Anthropic still emphasizes in CNN’s reporting that it has maintained two red lines: not using AI systems to control weapons and not using them for large-scale domestic surveillance. This is significant because it shows that Anthropic has not abandoned its safety stance entirely, but rather made selective compromises and坚守 on different dimensions of safety. Yet this selectivity itself is a key clue in narrative strategy analysis: the boundaries of where it conceded and where it held firm define the new calibration of safety.
On March 11, the Anthropic Institute was officially established, led by Jack Clark and positioned as a "public interest research organization." Less than two months later, on May 4, Clark posted the "60%" message.
Once aligned, the signal density and timing of this timeline are not random. From January’s personal article preview, to February’s policy changes, to March’s institutional establishment, to May’s founder probability forecast, and finally to June’s official long-form post, this is a clearly paced narrative pipeline with progressively escalating language. One cannot directly conclude that “all of this was pre-planned,” but the sequence itself presents a question analysts must confront: Does this sense of rhythm indicate that Anthropic has incorporated “narrative acceleration” into its public communications strategy?
Hassabis's deliberate provocation
If only Anthropic had adjusted its stance in the first half of 2026, analysts would have had sufficient reason to focus on the internal decision-making logic of the company. However, since DeepMind CEO Demis Hassabis made a nearly simultaneous adjustment in the same direction, the argument that this is a “single-company case” no longer holds.
January 20, Davos Forum. Hassabis maintained his long-standing assessment: a 50% probability of AGI by 2030. Three weeks later, on February 18, at the India AI Impact Summit, he softened his stance: “AGI could arrive within five years.”
May 20 to 22: Google I/O. At the keynote, Hassabis said humanity is standing at the “foot of the singularity.” Around the same time, OpenAI released GPT-5.3-Codex, claiming the model played a key role in its own creation—including assisting with debugging the training process, managing deployment, and analyzing evaluation results. The timing gap among the three leading labs has now been compressed to within weeks.
After Google I/O, Hassabis was interviewed by Axios. This interview was later widely cited, with its most crucial statement being his acknowledgment that using phrases like “the foothills of the singularity” was “intentionally provocative,” aimed at raising awareness among governments, economists, and the public about the urgency of AI’s accelerated development. He also adjusted the AGI timeline from the previous “not long after 2030” to “2029 is a real possibility,” though 2030, plus or minus one year, remains the widely expected timeframe.
Hassabis told the Seoul Economic Daily more directly: “Five to ten years from now, when we look back at 2026 and 2027, we’ll say, ‘That was the moment we entered the AGI era.’”
The phrase “intentional provocation” is worth careful consideration. It is a rare, firsthand admission of narrative intent. It acknowledges that at least some of the language he used was not a passive reflection of technical facts, but an active choice as a tool for communication. This admission does not negate the possibility that he genuinely perceived a technical turning point; rather, it explicitly separates “narrative” from the shadow of “fact,” allowing it to be examined as a distinct object.
Hassabis’s self-interpretation of his wording opens a side door to understanding this round of synchronized signals. His “intentional provocation” and the “disclaimer footnote” in Anthropic’s extensive data argument reveal the same amphibious stance: one hand pushes out signals potent enough to shake public opinion, while the other retains the safety net of retreating to “this is merely one possibility.”
The same data, completely different interpretations
When Anthropic and DeepMind jointly constructed a narrative framework claiming that "AI is accelerating its own evolution," independent external researchers offered alternative interpretations of the same data and phenomena. These interpretations matter not because any side holds the ultimate truth, but because they reveal the extent of interpretive flexibility within the official narrative.
The sharpest response came from Eliezer Yudkowsky, who not only replied to Jack Clark but continued to speak out on multiple subsequent occasions. MindStudio’s blog documented his full position: he compared current AI system safety designs to the Chernobyl RBMK reactor. The core argument of this analogy is that if the control rods and accelerator are linked within the same system, attempting to slow down causes the system to lose control even faster.
Nathan Lambert of the Allen Institute for AI introduced the concept of "Lossy Self-Improvement" (LSI). His argument directly challenges the "acceleration flywheel" model: as systems become increasingly complex, each generation's improvement process introduces friction and loss, much like signal degradation over long-distance transmission. According to this logic, the improvements that enable 80% or 90% of code to be written by AI cannot be infinitely replicated in subsequent generations, as later systems will face more complex problem spaces, and the noise and errors within AI-generated outputs will be amplified across generations.
Dean Ball, a senior fellow at the Foundation for American Innovation, offered a more straightforward linguistic framework to contextualize Anthropic’s data. He told IEEE Spectrum: “Perhaps eventually they will automate genius, but not next year. Next year, they’re automating grunt work.” This distinction cuts to the heart of the ambiguity behind the claim that “80% of code is written by AI.” If AI is automating repetitive patterns in codebases, bulk generation of parameters, or end-to-end pipeline configurations, then such tasks indeed correspond only to “grunt work” in the context of software engineering. The remaining 20% likely encompasses architectural design, strategic direction, and trade-offs made under incomplete information—these are the elements of genius.
David Scott Krueger of the University of Montreal, founder of the AI safety nonprofit Evitable, proposes the trigger红线 for a pause as “99% of the code written by AI.” He told IEEE Spectrum, “I think we may already be crossing that line.” The tension between his framework and Anthropic’s own loosened commitment to a pause is one of the most significant structural contradictions in this narrative.
UBC computer scientist Jeff Clune, in an interview with IEEE Spectrum, takes the opposite stance, saying, “We are at a tipping point for recursive self-improving systems.” If this statement is validated, it would mean that Yudkowsky’s warning has struck the right note.
Four groups, each with distinct perspectives—even within the same direction, there are internal tensions among radicals. Yet their commonality lies in their rejection of official narrative frameworks; instead, each derives independent judgments about the same set of phenomena from its own methodology. The very diversity and conflict among these judgments constitute the strongest rebuttal to the notion that any single narrative can fully capture the truth.
Coupling of valuation curves with narrative beats
In January 2026, Anthropic completed a funding round at a $350 billion valuation, with investors including Microsoft and NVIDIA. This figure had been previewed by some media outlets by the end of 2025, but the official announcement coincided precisely with Amodei’s release of “The Adolescence of Technology.”
In February, another round of financing totaling $30 billion was completed, with the valuation remaining in the range of approximately $350 billion. Also in February, the security policy was updated to remove the suspension commitment, and the threat of a $200 million Pentagon contract was lifted.
In May, Reuters, The New York Times, and TechCrunch nearly simultaneously reported that Anthropic completed a funding round of $65 billion, valuing the company at $965 billion. This figure surpassed not only its own valuation from two months prior but also OpenAI’s $852 billion valuation in March 2026. The New York Times additionally cited Dario Amodei at a developer conference, stating that the company’s annualized revenue reached $30 billion, with Amodei even joking that he hoped the 80-fold revenue growth this year would not continue, as it would be “crazy.”
On June 4, the Anthropic Institute published the long-form article "When AI Builds Itself."
Lining up these time points does not imply a precise arrow pointing to a chart. If someone claims there is a causal relationship between these elements, they must provide direct evidence. Without internal decision records, no analyst can or should make such an assertion.
On the other hand, completely ignoring and failing to document the correlations between these milestones is equally unreasonable. A company increased its valuation from $350 billion to $965 billion in just five months—nearly tripling in value—while simultaneously undergoing a major shift in its security policy, building a narrative pipeline led by independent research institutions, and having its co-founder issue a 60% probability forecast. When all these events are compressed into a six-month period, investors are at least entitled to ask: To what extent, if at all, did these signals serve to communicate to the market the message that “we are at the forefront of acceleration”?
The very act of asking this follow-up question holds analytical value. The answer may never be singular. But once a question is clearly posed, it cannot be easily retracted.
Global AI market funding reached $297 billion in the first quarter of 2026, with the top five deals accounting for a significant portion of this total. Under these conditions, all leading labs face the same pressure: you must convince investors that your technological curve will be steeper than your competitors’. Your risk warnings must also be loud enough that, when regulators eventually step in to establish rules, your voice is already embedded within the policy framework. Your narrative must simultaneously be compelling enough to attract top researchers to your lab and alarming enough to preserve your foothold in the safety community.
These demands are inherently contradictory. Anthropic’s narrative shift in the first half of 2026 can be seen as a realignment of the balancing point among these conflicting demands at the linguistic level. The weakening of safety commitments, the strengthening of acceleration signals, and the repeated use of the argument “we cannot unilaterally stop” together form a set of vectors pointing in the same direction.
The signal has been sent, then
We need to return to the most fundamental question: Do these signals more closely reflect a technical turning point, or are they a rhetorical upgrade aimed at capital and regulators?
Existing public evidence does not allow for a simple checkmark between the two options, because the evidence used in both interpretations is, in fact, the same set of data. The 80% code占比, 52x performance improvement, and task duration doubling every four months can be used to support either “a turning point is approaching” or “we are conveying to the market a trend awareness our own engineers have personally experienced”—the boundary between these two is blurred.
But there are some facts that are certain, and there is no need to take sides between the two interpretations.
First, Anthropic’s narrative shift completed in the first half of 2026 is not an isolated case. DeepMind’s Hassabis made a nearly identical adjustment—similar in direction, differing in degree but identical in essence—during the same quarter. OpenAI’s Sam Altman stated at the India summit that “the world is not yet ready,” and in February 2026, released GPT-5.3-Codex, claiming it played a “critical role in its own creation.” If this were merely Anthropic alone sending signals, one might analyze it from a corporate strategy perspective. But when the three leading labs simultaneously amplified their messaging within a concentrated period, it constitutes an industry-wide narrative shift.
Second, there is a precisely traceable temporal correlation between the timing of these signals and the rhythms of financing, policy adjustments, and institutional restructurings. This correlation itself does not need to prove anything—it simply needs to be presented honestly. After it is presented, everyone’s own methodology will determine how they interpret it next.
Third, Anthropic itself labeled the third scenario—“complete recursive self-improvement”—as “possible,” not “likely.” This indicates that, within the company’s own internal judgment framework, their narrative of acceleration has not yet been fully closed. The forces that have conditioned them to include qualifiers in academic papers and blog posts still hold back their public wording.
Fourth, Hassabis’s “intentionally provocative” confession confirms a mechanism that, while widely suspected, had rarely been openly stated by those involved: at least some leaders of cutting-edge labs deliberately choose their wording with clear communication objectives in mind. This means that any interpretation of their statements must analyze two layers simultaneously: the facts they claim, and the rhetorical strategies they employ in making those claims—as actions in themselves.
Those who carefully read Anthropic’s full data receive a completely different signal than those who only remember the two numbers: “80% of code written by AI” and “52x acceleration.” But in this case, “how it’s remembered” may be more important to analyze than “what was actually said.”
This long article itself is a precise example of the phenomenon it describes: it constructs a sense of imminent acceleration through data, while using footnotes and qualifiers to retain room for retreat; it calls for global coordination and verifiable slowdown, yet has already withdrawn its prior commitments to pause in earlier policy revisions. This is not hypocrisy, nor mere inconsistency between words and actions—it is an institutional balancing act between technological uncertainty, commercial pressures, and public responsibility. And Hassabis’s confession of “intentional provocation” inadvertently confirms from the side door that this balancing act has become a consciously employed method within leading laboratories.