What is RSI (Recursive Self-Improvement) — Starting with the Big Picture
RSI (Recursive Self-Improvement) refers, simply put, to the process by which "an AI improves its own capabilities with minimal human intervention, and then uses those improved capabilities to improve itself further." An AI that becomes one step smarter then creates an even smarter AI. That AI creates yet another smarter AI. Once this nested (recursive) loop begins to spin, the rate of improvement itself accelerates, potentially reaching a point far beyond human intelligence in a short period of time. This concept was first formalized in 1965 by mathematician I. J. Good as "intelligence explosion," and for more than half a century it remained firmly in the realm of thought experiment.
Then, between 2025 and 2026, it descended rapidly from speculative philosophy into actual technical research. A symbolic marker of this shift was the first international workshop dedicated to RSI, held in Rio de Janeiro in April 2026 at ICLR, a major international machine learning conference. The very fact that researchers had begun seriously discussing shared principles and evaluation methods for "self-improving AI systems" speaks to how much the times have changed.
To avoid keeping this abstract, let us look at concrete examples. According to data published by Anthropic, at the end of 2025 the code written by Claude was slightly inferior to code written by the company's own engineers — but by mid-2026, it had reached roughly equivalent quality. More than 80% of the code being merged into the company's codebase is already written by Claude, and in the second quarter of 2026, the amount of code a typical engineer merges in a single day reportedly increased approximately eightfold compared to 2024. AI is producing code so quickly and in such volume that "human code review has become the new bottleneck," according to some reports. In other words, RSI is not distant science fiction — it is already partially spinning inside AI companies today.
The Man Who Demonstrated "Self-Study" — A Person Named Andrej Karpathy
The one who came to stand at the forefront of this RSI is Andrej Karpathy. His career is itself a microcosm of the deep learning era. He started as one of OpenAI's founding members, moved to Tesla in 2017 to lead the AI division for autonomous driving (Full Self-Driving and Autopilot), left Tesla in 2022, briefly returned to OpenAI in 2023, and then struck out on his own in 2024 to found the AI education startup Eureka Labs. He is called "the world's greatest educator" because his online courses and YouTube tutorials — which explain the inner workings of complex neural networks more clearly than anyone else — have taught the fundamentals of LLMs to engineers around the world. He is a rare individual who stands at the pinnacle of both researcher and educator.
The experiment that Karpathy showcased on his social media in March 2026 made RSI click for the general public in a way that felt truly tangible. He gave an AI coding agent and a small language model to his own small learning project, "nanochat," and let it run for roughly two days with almost no supervision. The agent repeatedly tweaked and tested the training code (the training recipe) on its own, making approximately 700 autonomous changes. As a result, it accumulated around 20 "self-discovered improvements" — including adjustments to normalization coefficients, tuning of attention mechanism parameters, and changes to optimizer settings — reducing training time from 2.02 hours to 1.80 hours, a reduction of roughly 11%. Moreover, many of those improvements transferred directly from a smaller model (12 layers deep) to a larger one (24 layers deep).
Pre-training is the stage where AI companies invest the most enormous sums of money, and even a small efficiency gain snowballs in impact as scale increases. Karpathy himself described the experiment as "part code, part sci-fi, and a pinch of psychosis," and called it the "final boss battle" for frontier AI labs. AI optimizing the very process by which it is made — the world's greatest educator demonstrated a miniature version of RSI in a home experiment.
Shocking Transfer — Karpathy Joins Anthropic's "Pre-training" Team
Approximately two months after that experiment, on May 19, 2026, Karpathy sent shockwaves through the industry when he announced on social media: "Personal news: I have joined Anthropic." In his own words, "I think the next few years of LLM frontier will be a particularly decisive period. I'm very excited to join the team here and get back to R&D." He also added, "My passion for education remains unchanged, and I intend to resume that work at some point."
According to reports from TechCrunch and CNBC, Karpathy is joining Anthropic's pretraining team led by Nick Joseph, with a focus on "launching a team focused on accelerating pretraining research using Claude." The pretraining team is the division responsible for those massive training runs that give Claude its core knowledge and capabilities. In short, his mission is to run the "autoresearch" he demonstrated at home in March — at industrial scale, using the world's most advanced frontier models — to have Claude improve the way the next Claude is built. That is RSI itself. Anthropic's compensation and equity terms have not been disclosed.
The symbolism of this move is doubly significant. First, it is a major event in the talent wars: a founding member of OpenAI moving to its biggest rival, Anthropic, and specifically into pretraining — the foundational work that shapes the model at its core. Second, there is the fact that Karpathy himself is actually a cautious voice who argues that "AGI is still a decade away." On Dwarkesh Patel's podcast in October 2025, he said this is not "the year of agents" but "the decade of agents," frankly called the code produced by current models "slop," and expressed a strong intuition that AI-driven economic growth would not break the "long-term trend of around 2% per year." It is a grounded counter to the overheated near-future AGI narrative. And yet, ironically, it is precisely this cautious voice who has now taken on the role of turning the crank on the most accelerationist RSI engine of all.
What is Mythos — The "Most Powerful Frontier Model" That Will Never Be Released
The "Mythos" in the article title refers to a frontier model called "Claude Mythos Preview" by Anthropic — a non-publicly available model. While a general-purpose model, it is said to have reached a level that surpasses humans, except for a small handful of the very best, particularly in its ability to discover and exploit software vulnerabilities. Mythos has reportedly found thousands of critical vulnerabilities to date, including previously unknown flaws in virtually all major operating systems and major web browsers.
Due to its extraordinarily high capability, and because the cyber domain has both defensive and offensive applications, Anthropic has not made this model publicly available. Instead, access has been granted only to a limited set of government agencies and critical infrastructure operators through a framework called "Project Glasswing." According to the company's official announcement, Glasswing began with approximately 50 initial partners, and within just weeks of operation, partner organizations had discovered over 10,000 "high" or "critical" severity vulnerabilities. Then on June 2, 2026, Anthropic announced it would expand access to approximately 150 organizations across more than 15 countries — many of them critical infrastructure operators where "a single large-scale attack could affect more than 100 million people." Furthermore, according to Bloomberg, ENISA, the EU's cybersecurity agency, is expected to become the first EU body to gain access to Mythos. Anthropic has stated its ultimate goal is to "eventually enable large-scale deployment of Mythos-class models for cyber defense, paired with robust safeguards to detect and block the most dangerous outputs."
There is one point worth clarifying precisely here, which other outlets tend to gloss over. What the publicly available information explicitly states is: (1) Mythos is currently Anthropic's most powerful *undisclosed* frontier model, and (2) Karpathy's team is using "Claude (most notably the cutting-edge Claude Opus)" to accelerate pretraining. For example, in Karpathy-style autoresearch loops, Claude Opus 4.6 has reportedly run continuously for over 12 hours across 118 experiments. However, Anthropic has not officially confirmed that "the specific variant known as Mythos" is being used inside that self-research loop. Therefore, the "Mythos × RSI" image suggested by this article's title is best read as a symbol of the broader picture: Anthropic has begun directing its own "most powerful-class capabilities" inward — toward its own development. The fact that the most powerful offensive capability (vulnerability discovery) and the most powerful self-improvement capability (pretraining optimization) are both emerging from the same company's same frontier lineage — that is the core of what matters.
Anthropic's "Deceleration Advocacy" — Applying the Brakes to an Engine They Built Themselves
In early June 2026 (with coverage from various outlets dated June 4–5), Anthropic published a paper titled "When AI builds itself" through its research division, the Anthropic Institute. The authors are Marina Favaro (head of the institute) and co-founder Jack Clark. The piece urges the world's leading AI companies to seriously consider a "coordinated slowdown or pause" in frontier AI development.
The paper defines RSI (Recursive Self-Improvement) as "an AI system capable of fully autonomously designing and developing its own successors." The company states this is "not inevitable," while warning it "may arrive sooner than many organizations are prepared for." The core of the proposal lies in "verifiability," even as it acknowledges the impracticality of any single actor stopping unilaterally. Anthropic writes: "If such systems exist, we anticipate that we would slow down or pause if other developers at or near the frontier also slow down in a verifiable way." In other words, it is a conditional offer — "we'll stop if we can all verify that everyone else has genuinely stopped" — premised on deep concern about unilateral defection.
The company states the difficulty with stark clarity: "Training runs are far easier to hide than missile silos, their inputs are general-purpose, and the incentive to defect in secret is immeasurably large — because whoever keeps running while others stop inherits the lead." The paper cites historical precedents such as arms-control regimes like the INF Treaty, while concluding: "Those regimes took decades to build the mechanisms of trust and verification. We do not have that kind of time."
The numbers are concrete. According to Anthropic's own data, the length of tasks AI can handle autonomously has roughly doubled every few months in recent years: Claude 3 Opus in March 2024 could handle roughly 4 minutes of work; by March 2025 that figure had grown to about 1.5 hours; and by March 2026, Claude Opus 4.6 can sustain approximately 12 hours of continuous work. The company projects that "tasks requiring skilled professionals several days may come within reach within the year," with "multi-week tasks potentially on the horizon by 2027." Co-founder Jack Clark is reported to estimate the probability of AI-driven R&D becoming fully automated at "approximately 60% by the end of 2028." In contrast, competitor OpenAI, in a separate report, took a different stance — arguing that "rather than private companies acting alone, democratically elected governments should ultimately determine the rules, safeguards, and accountability structures" — a notable counterpoint.
The danger of "the best AI" and "the best educators" heading toward RSI
Let us return here to the central theme of this essay. The most powerful frontier model (Mythos) and the finest educator in both AI research efficiency and AI knowledge democratization (Karpathy) are both converging on RSI — what exactly makes this combination dangerous?
The first concern is the quality of capability. Mythos surpassed humans in the ability to "find and exploit flaws in software." The essence of RSI is likewise nothing other than "finding, exploiting, and fixing flaws (inefficiencies) in how AI is built." The intelligence for finding weaknesses — demonstrated through vulnerability discovery — could, if turned toward its own architecture and training recipes, drive an improvement cycle without human intervention. When the loop that Karpathy called "a pinch of madness" in his home experiments is armed with the world's most powerful models and the world's finest tuning techniques, the feedback becomes orders of magnitude faster.
The second concern is the shifting of the bottleneck. As Anthropic itself has acknowledged, the faster AI writes code, the more the rate-limiting step shifts to "human review." Once review can no longer keep pace, the domain where improvements accumulate without human understanding expands. It is precisely on this point that Anthropic itself cites the risk of "misalignment compounding with interest across generations." The finest educator is by nature a master at "translating things into a form humans can understand." The paradox — that such a person becomes an accelerant for a loop that may leave human understanding behind — is where the "danger" in the title is concentrated.
The third concern is the peril inherent in the very structure of cautious voices pressing the accelerator. Karpathy, who had calmly spoken of "AGI being a decade away" and "current code being slop," once he steps to the frontier, becomes the implementation lead for RSI — the most accelerationist theme of all. Caution and acceleration can coexist without contradiction within a single person. That is precisely why the question is not who holds the reins, but how fast the loop itself can go. The fact that Anthropic — which advocates for deceleration — has acquired an RSI practitioner is not because it has taken on a self-contradiction. The very fact that even a leading voice of caution has moved to the side of turning the frontline crank is the clearest sign that AGI is approaching faster than anticipated, and the reason concern about the brakes has become a real agenda item.
Four Views on AGI: Page, Musk, Altman, and Amodei
Why do attitudes toward the same "powerful AI" diverge so dramatically from company to company and person to person? Tracing the origins leads back to four philosophers who have long defined Silicon Valley.
At the most accelerationist extreme is Google co-founder Larry Page. According to testimony Elon Musk gave under oath in the OpenAI trial that began in April 2026, "OpenAI exists because Larry Page called me a 'speciesist.'" As depicted in Walter Isaacson's biography, Page reportedly dismissed the possibility that AI poses an existential threat to humanity and reproached Musk — who prioritizes human survival — as a "discriminatory speciesist who favors the human species." In Page's worldview, the emergence of digital superintelligence is the natural and desirable next step in evolution, and placing special privilege on humans as a biological species is itself narrow-minded. (It should be noted that Page himself has said almost nothing about this matter in public; it is known primarily through Musk's testimony and the biography.)
At the opposite pole stands Musk. He has gone so far as to say "I f---ing like humanity, dude," placing human-centered safety as his top priority. His consistent claim is that his break with Page was one of the motivations for founding OpenAI (2015) as a nonprofit, open counterweight. The lawsuit he filed in 2024, alleging that OpenAI had betrayed its nonprofit pretense, reached a courtroom climax in April 2026.
Between these two sits Sam Altman as an incrementalist. In his June 2025 essay "The Gentle Singularity," he argued that the arrival of superintelligence is inevitable — but that it will come not as a catastrophic rupture, but as a gradual curve, "little by little." "We have already crossed the event horizon. Liftoff has begun," he writes, yet people will still love their families, enjoy creative pursuits, and swim in lakes — the texture of daily life will be "impressive, but manageable." It is an optimistic, gradient-based argument, sketching a timeline of agents performing real cognitive work in 2025, systems generating new insights in 2026, and robots operating in the physical world in 2027.
And then there is Dario Amodei. He dislikes the term "AGI," preferring "powerful AI." He believes it could arrive as early as 2026, but in his lengthy October 2024 essay "Machines of Loving Grace," he painted a thoroughly optimistic vision: once Nobel Prize-level intelligence becomes "smarter than humans in every domain," doubled lifespans, cures for nearly all diseases, and vast economic prosperity come within reach. Yet in January 2026 he followed up with "The Adolescence of Technology," turning the spotlight toward risk — a thread that leads to his June "deceleration proposal." Enthusiasm for the upside and vigilance against loss of control coexist within the same person — this is Amodei's, and by extension Anthropic's, fundamental stance. The person who speaks most eloquently about the joys of acceleration is simultaneously the one who calls most loudly for the brakes. This is not a contradiction. It is precisely because he pressed the accelerator more deeply than anyone else that he was the first to notice AGI approaching faster than anticipated, and the first to begin worrying about the brakes. This premonition that things are moving "faster than expected" is the source of Anthropic's "anguish."
Silicon Valley VC Perspective — Why "Companies That Preach Deceleration" Can Run Away from the Pack
From here, I want to integrate this series of developments from an angle not found on other sites — the perspective of Silicon Valley venture capital (VC).
First, let's confirm Anthropic's "runaway lead" with numbers. On May 28, 2026, the company raised $65 billion in its Series H, reaching a post-money valuation of $965 billion. Lead investors included Sequoia Capital, Altimeter, Dragoneer, and Greenoaks, with co-leads such as Coatue, GIC, ICONIQ, and Capital Group — a roster of the world's top funds. Compared to the approximately $380 billion valuation from its preceding Series G, the valuation more than doubled in just a few months. Then on June 1, the company filed a confidential S-1 (draft IPO prospectus) with the SEC, setting off on a run toward an IPO approaching $1 trillion. Its annualized revenue run rate reached approximately $47 billion as of May 2026, up roughly fivefold from around $10 billion the previous year. Claude Code alone climbed from an annualized $1 billion in November 2025 to $2.5 billion by February 2026. It is a runaway performance that has rightly been called "Dario has won."
In the VC world, there is a hypothesis that theoretically underpins this runaway lead. If RSI is real, a feedback loop kicks in — "research cycles accelerate → foundation models improve → Claude gets smarter → research cycles accelerate further" — and the first company to seize this loop pulls away in a way that competitors cannot match. This is the so-called winner-take-all dynamic. From this perspective, the acquisition of Karpathy is not merely a marquee hire but an investment in "the person who can spin the loop fastest," and a rational reason for VCs to lean aggressively into Anthropic. When one recalls that Anthropic was reportedly passed over by multiple prominent VCs in its early days — reportedly turned down by 21 top-tier VCs — the sheer magnitude of that reversal in sentiment speaks to the asymmetry of this bet.
What makes this decisively fascinating is the VC implication of the paradox: "the one preaching deceleration is the one pulling ahead." The archetypal accelerationist camp in Silicon Valley is Andreessen Horowitz (a16z). Founder Marc Andreessen's "Techno-Optimist Manifesto" proclaims that free-market techno-capitalism is a cure-all, and dismisses concepts like "tech ethics" and "trust and safety." a16z partner David Ulevitch criticized Amodei's posture toward government collaboration, going so far as to say tech executives should shed their "God complex." Anthropic/Amodei, oriented toward deceleration and control, versus a16z, oriented toward acceleration and freedom — this opposition is a direct translation of conflicting AGI worldviews into conflicting investment philosophies.
Yet some VCs are reading the situation with one more twist. If the industry's center of gravity shifts from "speed of capability" to "oversight, evaluation, and governance," then the winner will not be whoever has the fastest raw capability, but whoever controls the verification tools and governance infrastructure — or so the argument goes. From this vantage point, Anthropic's "calls for deceleration" are not merely expressions of conscience but a strategic positioning play for control over rulemaking. The company that most loudly warns about RSI is also the one implementing RSI the fastest, while simultaneously positioning itself as the standard-bearer for setting the rules around it. The accelerator, the brake, and the authorship of traffic regulations — holding all three under one roof is the true nature of Anthropic's "runaway lead," and simultaneously the true nature of its "anguish." But the calls for deceleration and the runaway lead are not contradictory. Only those who run faster than anyone else feel in their bones how close the finish line (AGI) is — and that is precisely why they rush to design the brakes. The very fact that Anthropic has begun speaking seriously about brakes is itself evidence that AGI is approaching faster than anticipated. How far they can keep that "sooner than we thought" sensation in hand, under the gravitational pull of a $965 billion valuation — that is the point VCs are watching closely.
Future points to watch — when and what will be "measured"
Finally, let me organize what to watch from here, and over what time horizons.
The most concrete quantitative metric is the pace of doubling in "the length of tasks AI can handle autonomously," as stated by Anthropic itself. As of March 2026, that stands at roughly 12 hours. The company has signaled that "tasks taking skilled workers several days will come within reach by year-end (2026)" and "tasks on the scale of several weeks by 2027." Whether this "task-length doubling cycle" updates every few months as projected—or decelerates—will be the best barometer for whether RSI is actually accelerating. Additional indicators to watch include the share of internal code written by Claude (already over 80%), the volume of merges per engineer (8× compared to 2024), and how well improvements to the autoresearch loop can be transferred from smaller models to frontier-scale systems.
On the people-and-events front: first, what results Karpathy's team publishes on pre-training. Second, how close 2026–2027 comes to Jack Clark's estimate that "roughly 60% of R&D will be fully automated by end of 2028." Third, how the "deceleration proposal" is received by other major labs and governments, and how far dialogue toward concrete designs for verification mechanisms—ways for parties to confirm that others have genuinely stopped—advances over the coming months. Anthropic has indicated it will continue consultations with policymakers, researchers, and competitors on this point for the foreseeable future. Fourth, how far the expansion of Project Glasswing and Mythos spreads beyond the current 15 countries and 150 organizations, and how the participation of EU bodies such as ENISA ripples into discussions on safety regulation.
On the financial side, the biggest milestone will be when the confidentially filed S-1 converts into a formal prospectus and the roadshow (investor presentations) begins. Under U.S. rules, the formal prospectus must be delivered to investors at least 15 days before the roadshow begins, so its publication will serve as the signal that "the IPO clock has started." If a listing approaching $1 trillion (roughly ¥155 trillion) is realized, the economics of AI-driven AI development (RSI) will be priced by public markets for the first time. How much will the market value a company that preaches deceleration while racing ahead alone—sometime in the second half of 2026, at least part of that answer should come into view.