Anthropic's Jack Clark: 60% Chance AI Self-Creates by 2028
Jack Clark, co-founder of Anthropic, recently dropped a bombshell on the AI community: there is a 60% probability that AI will be capable of independent self-creation by the end of 2028.
This isn’t just another speculative prediction. It’s based on a rigorous analysis of public capability curves across programming, scientific research reproduction, and model training optimization—all of which are accelerating at an unprecedented rate.
Key Takeaways
- The 2028 Horizon: Clark estimates a 60% chance of recursive self-improvement (RSI) within 30 months.
- Engineering vs. Creativity: 99% of AI research is engineering "perspiration," which AI is already automating effectively.
- The Capability Surge: Benchmarks like SWE-Bench and CORE-Bench have seen near-total breakthroughs in under 15 months.
- Governance Gap: Anthropic itself admits society may not be ready for the "unpredictable future" of automated AI research.
The “Rubicon” of Automated Research
In his recent Import AI analysis, Clark argues that humanity is approaching a metaphorical “Rubicon.” Once an AI can train its own successor end-to-end, we enter an era of recursive self-improvement that makes the current pace of development look slow.
The basis for this conclusion isn’t internal secrets, but public data. Clark points to METR (Model Evaluation and Threat Research), which tracks the time it takes an AI to complete tasks previously requiring skilled human labor.
In 2022, GPT-3.5 could handle tasks taking a human 30 seconds. By 2026, Claude Opus 4.6 is handling tasks that take a human 12 hours. That is a 1,440x increase in just four years.
99% Perspiration: AI as its Own Project Manager
A common critique of the “singularity” is that AI lacks the spark of genius required for breakthroughs. Clark counters this by quoting Edison: “Genius is 1% inspiration and 99% perspiration.”
Most AI research is engineering work: data cleaning, experiment management, and hyperparameter tuning. These are precisely the tasks where agentic AI excels. We are already seeing:
- CORE-Bench Breakout: AI can now reproduce scientific papers with 95%+ accuracy, a task that was a major bottleneck just 15 months ago.
- Training Optimization: Internal Anthropic tests show AI optimizing training code to be 52 times faster than unoptimized versions.
- Automated Alignment: Anthropic researchers found that AI agents can tackle safety problems more effectively than human baselines.
When the “perspiration” of engineering is solved, the 1% of creative “inspiration” becomes the only remaining hurdle—and systems like Google’s AlphaEvolve are already showing signs of bridging that gap by inventing novel mathematical algorithms.
The Counter-Argument: Increasing or Decreasing Returns?
Not everyone is convinced. Pedro Domingos, author of The Master Algorithm, notes that while AI can build itself, the real question is whether this cycle brings increasing or decreasing returns.
If each generation of AI only provides marginal gains, the “intelligence explosion” may fizzle out into a plateau. Clark acknowledges this, giving only a 30% probability for 2027. The jump to 60% in 2028 assumes a “discontinuous capability event”—a breakthrough in creative intuition that could bridge the final gap.
Final Thoughts: The Business Implications
For enterprise leaders, the message is clear: AI is no longer just a productivity tool; it is becoming a research partner. If your competitor leverages a system capable of even 20% self-optimization, their technological lead will compound exponentially while yours remains linear.
We are standing on the edge of an unpredictable future. The question isn’t just if AI will create itself, but how we will govern a world where intelligence is no longer a human-limited resource.
Want to stay ahead of the curve? Explore our deep dive on AI governance frameworks for the enterprise to prepare for the coming shift.
