Anthropic Co-Founder: 60% Chance AI Builds Itself by 2028

Anthropic co-founder Jack Clark has issued one of the most striking predictions in recent AI discourse: there is a 60% probability that AI systems will be capable of recursive self-improvement (RSI) before the end of 2028. If true, it means an AI could autonomously design, train, and deploy its own successor — without human intervention.

Clark, who also founded and writes the influential newsletter Import AI, published his analysis on May 4, calling it 'a view I hold reluctantly' and admitting he is unsure whether society is prepared for the consequences of fully automated AI research.

Key Takeaways

• 60% probability of recursive self-improvement by end of 2028, according to Clark
• RSI means an AI system could end-to-end train its own successor model
• Clark does not expect this to happen in 2026 but sees proof-of-concept demos on non-frontier models within 1-2 years
• He compares the moment to 'crossing the Rubicon' — a point of no return
• The prediction is based on publicly available research from arXiv, bioRxiv, and NBER
• Anthropic's own Claude models are among the systems pushing toward this frontier

What Is Recursive Self-Improvement and Why Does It Matter?

Recursive self-improvement refers to an AI system's ability to modify, retrain, or entirely rebuild itself to become more capable — and then repeat the cycle with each new, improved version. It is a concept that has long existed in theoretical AI safety literature but has recently moved closer to practical reality.

The idea is straightforward but profound. Imagine a model that can identify its own weaknesses, design architectural improvements, gather and curate training data, run the training process, evaluate results, and then deploy the improved version — all autonomously.

Unlike incremental improvements driven by human researchers, RSI could trigger an exponential capability curve. Each generation of the AI would be smarter, faster, and more capable of producing an even better successor. This is what some researchers call an intelligence explosion, a term coined by mathematician I.J. Good in 1965.

Clark's Reluctant Warning Signals Growing Industry Anxiety

What makes Clark's prediction especially notable is its source. He is not an outsider or alarmist commentator. As co-founder of Anthropic — the $60 billion AI safety company behind the Claude model family — he sits at the very center of frontier AI development.

His framing reveals deep unease. In his Import AI newsletter (Issue #455), Clark wrote that this is 'a view I hold reluctantly: its implications are so vast that they make me feel small, and I am not sure society is ready for the changes that automated AI R&D would bring.'

He explicitly compared the moment to crossing the Rubicon River — the ancient metaphor for passing a point of no return. Once AI can build better AI without human oversight, the pace of progress would no longer be constrained by human cognitive speed, funding cycles, or institutional decision-making.

Clark does not believe RSI will arrive suddenly in 2026. Instead, he envisions a more gradual trajectory:

• 2025-2026: AI systems increasingly automate portions of the research pipeline — literature review, hypothesis generation, experiment design, and code writing
• 2026-2027: Proof-of-concept demonstrations where a non-frontier model successfully trains a functional successor end-to-end
• 2027-2028: Frontier labs potentially achieve full recursive self-improvement with state-of-the-art models

The Evidence Trail: Public Research Points Toward Automation

Clark's 60% estimate is not based on insider knowledge alone. He explicitly grounds his analysis in publicly available research papers from arXiv, bioRxiv, and the National Bureau of Economic Research (NBER).

Several recent trends support his thesis:

• AI-driven coding agents like Anthropic's Claude Code, OpenAI's Codex, and Devin from Cognition are already capable of writing, debugging, and deploying complex software — including ML training scripts
• Automated ML (AutoML) research has advanced significantly, with systems now capable of neural architecture search, hyperparameter optimization, and data augmentation without human guidance
• Self-play and self-improvement techniques, popularized by DeepMind's AlphaGo and AlphaZero, have demonstrated that AI systems can improve through self-generated feedback loops
• Scaling laws research from Anthropic, OpenAI, and DeepMind has made the training process more predictable and therefore more automatable
• Agentic AI frameworks now allow models to chain together multi-step workflows, including setting up cloud compute, managing training runs, and evaluating results

The gap between 'AI assists human researchers' and 'AI replaces human researchers' is narrowing faster than many expected. Each of these capabilities represents a building block. Combined, they form a plausible pathway to full automation of the AI research loop.

How This Compares to Other Industry Predictions

Clark is not alone in sounding this alarm, though his timeline is among the more specific. OpenAI CEO Sam Altman has repeatedly suggested that superintelligent AI could arrive within 'a few thousand days' — a timeline that aligns roughly with Clark's 2028 window.

Dario Amodei, Clark's co-founder at Anthropic, published a lengthy essay titled 'Machines of Loving Grace' in late 2024, exploring how powerful AI could transform science, medicine, and governance — implicitly acknowledging that automated AI research is on the horizon.

Meanwhile, Google DeepMind CEO Demis Hassabis has taken a somewhat more cautious public stance, emphasizing the difficulty of achieving true general intelligence. Yet DeepMind's own research into AI-driven scientific discovery — including AlphaFold for protein structure prediction — demonstrates the potential for AI to autonomously advance entire fields.

The key difference with Clark's prediction is its specificity. A 60% probability by a named date from a named insider gives the AI safety community — and policymakers — something concrete to plan around.

What This Means for Developers, Businesses, and Society

If Clark is even partially right, the implications ripple across every sector of the tech economy and beyond.

For AI developers and researchers:
- The competitive moat of human expertise in ML research could erode rapidly
- Teams that build robust AI evaluation and safety frameworks will become exponentially more valuable
- Open-source AI communities may face a critical choice: embrace or restrict self-improving systems

For businesses:
- Companies relying on AI-powered products should prepare for a dramatic acceleration in capability — and unpredictability
- Investment in AI safety and alignment is no longer optional; it is a business risk issue
- The cost of frontier AI development could drop precipitously if AI automates its own R&D

For society and policymakers:
- Regulatory frameworks like the EU AI Act may need urgent updates to address self-improving systems
- International coordination on AI governance becomes critical — RSI does not respect national borders
- Public understanding of AI risk needs to move beyond chatbots and deepfakes to existential-scale questions

The Safety Paradox at Anthropic's Core

There is a deep irony in Clark's warning. Anthropic was founded in 2021 specifically to build safer AI systems. The company's Responsible Scaling Policy is designed to impose capability thresholds — known as AI Safety Levels (ASL) — that trigger additional safety measures as models become more powerful.

Yet Anthropic continues to push the frontier with Claude 3.5, Claude 4, and its agentic coding tools. The company recently raised $2 billion from Google and other investors, bringing its total funding to over $7 billion. It is simultaneously building the technology that could enable RSI while warning the world about its dangers.

This tension is not unique to Anthropic. It reflects the broader AI safety paradox: the organizations most capable of building dangerous systems are also the ones best positioned to understand and mitigate the risks. Whether this dual role is sustainable — especially in a race against competitors like OpenAI, Google, Meta, and xAI — remains one of the defining questions of the AI era.

Looking Ahead: Preparing for a Post-RSI World

Clark's 60% estimate leaves significant room for doubt. A 40% chance that RSI does not arrive by 2028 is far from trivial. Technical barriers remain substantial — including energy costs, data limitations, hardware constraints, and the fundamental difficulty of AI alignment.

But the directionality is clear. The question is shifting from 'Will AI automate AI research?' to 'When and how fast?' Clark's contribution is to put a number and a date on a prediction that many in the field hold privately but rarely articulate publicly.

If proof-of-concept demonstrations appear on non-frontier models within the next 1-2 years, as Clark anticipates, the policy window for establishing meaningful governance frameworks will shrink dramatically. The Rubicon metaphor is apt: once crossed, there may be no turning back.

For now, Clark's prediction serves as both a forecast and a call to action. The AI industry, governments, and civil society have roughly 3 years to prepare for a world where the most powerful technology ever created might start building itself.