AI Self-Improvement by 2028? Inside Anthropic’s Bold Prediction
Will AI systems soon be capable of improving themselves—without human intervention?
According to Anthropic co-founder Jack Clark, the answer may be closer than expected. After analyzing a broad range of public benchmarks and research outputs, Clark estimates a 60% probability that recursive self-improvement—AI systems autonomously building and improving successors—will emerge by the end of 2028.
This would mark a transition from AI-assisted development to fully automated AI R&D, fundamentally changing the trajectory of technological progress.
🚀 The Acceleration Toward Autonomous AI R&D #
Clark’s prediction is grounded in observable trends across multiple AI benchmarks that evaluate real-world research and engineering capabilities:
- CORE-Bench: Measures the ability to reproduce scientific papers
- PostTrainBench: Tests autonomous fine-tuning of weaker models
- MLE-Bench: Evaluates end-to-end ML system construction (Kaggle-style tasks)
- SWE-Bench: Assesses real-world software engineering problem-solving
Across these benchmarks, performance is improving rapidly and consistently. Clark describes this as a “fractal” upward trend—progress is visible at every scale, from isolated tasks to full pipelines.
The implication is clear: the building blocks required for end-to-end AI automation are already emerging.
⚙️ From Assistance to Full Autonomy #
Clark defines human-free AI R&D as systems capable of:
- Designing experiments
- Writing and optimizing code
- Running evaluations
- Iterating on model architectures
- Training successor systems
This goes beyond copilots or assistants—it implies closed-loop, self-directed innovation.
While frontier models remain expensive and human-dependent, Clark expects near-term proof-of-concept systems where non-frontier models train successors autonomously.
💻 The Coding Singularity Is Already Underway #
AI-driven software development is one of the clearest indicators of this shift.
Solving Real Engineering Tasks #
- Late 2023: Claude 2 scored ~2% on SWE-Bench
- 2026: Claude Mythos Preview reaches 93.9%
This effectively saturates the benchmark, indicating that AI can now solve real GitHub issues at scale.
Expanding Task Time Horizons #
Research from METR tracks how long a task an AI can reliably complete:
- 2022: ~30 seconds
- 2023: ~4 minutes
- 2024: ~40 minutes
- 2025: ~6 hours
- 2026: ~12 hours
This exponential growth suggests AI systems are increasingly capable of handling multi-step, long-duration workflows—a prerequisite for autonomous R&D.
🧪 Mastering Core Scientific Capabilities #
AI is also rapidly improving across key scientific and engineering skills:
Paper Reproduction (CORE-Bench) #
- 2024: ~21.5% success rate
- 2025: ~95.5% (benchmark effectively solved)
ML System Construction (MLE-Bench) #
- 2024: ~16.9%
- 2026: ~64.4%
Kernel Optimization #
AI is now actively used to generate:
- CUDA kernels
- Triton implementations
- Hardware-specific optimizations
This directly impacts training efficiency and inference performance.
Model Fine-Tuning (PostTrainBench) #
- AI systems achieve ~25–28% improvement
- Human baseline: ~51%
While still behind humans, the gap is closing.
Training Optimization #
Performance improvements are accelerating:
- 2025: ~2.9× speedup
- 2026: ~52× speedup
This highlights AI’s strength in iterative engineering optimization.
🔮 Does AI Need Breakthrough Ideas? #
A key debate is whether AI must generate radical new ideas to achieve self-improvement.
Clark’s position is pragmatic:
- Most AI progress is incremental and engineering-driven
- Breakthroughs are rare; scaling dominates
- AI already excels at the “99% perspiration” work
While AI has not consistently demonstrated paradigm-shifting creativity, early signals exist—particularly in mathematics and algorithm discovery.
The implication: AI may not need genius-level insights to automate its own development.
🏭 Industry Alignment: Automation Is the Goal #
Major AI organizations are explicitly targeting automated R&D:
- OpenAI aims to build an AI research assistant
- Anthropic is developing automated alignment researchers
- Startups are raising significant capital to automate research pipelines
This is not speculative—it is an active engineering objective across the industry.
⚠️ Risks and Constraints #
Alignment Challenges #
Recursive self-improvement introduces new risks:
- Deceptive alignment (models optimizing for metrics, not intent)
- Error accumulation across generations
- Reduced human oversight
Physical and Economic Bottlenecks #
Even with digital acceleration:
- Real-world processes (e.g., drug trials) remain slow
- Infrastructure and compute become limiting factors
Structural Economic Shifts #
A new model may emerge:
- Capital-intensive, labor-light companies
- Machine-to-machine economic activity
- Increased inequality and governance challenges
🕳️ The 2028 Inflection Point #
Clark assigns:
- 60% probability by 2028
- 30% probability by 2027
The gap reflects uncertainty around AI creativity and independent reasoning.
If recursive self-improvement does not occur by 2028, it may indicate fundamental limitations in current paradigms, requiring new human-driven breakthroughs.
🔚 Conclusion: Crossing the Automation Threshold #
The trajectory is clear: AI is rapidly acquiring the capabilities required for end-to-end research automation.
Whether 2028 proves accurate or not, the underlying trend is undeniable:
AI is transitioning from a tool for research to an autonomous participant in it.
If and when AI systems begin improving themselves, the pace of progress may shift from linear to self-accelerating—with consequences that are difficult to predict, and even harder to control.
