Jack Clark, co-founder of Anthropic, predicts that AI capable of self-development may emerge by 2028, raising concerns about societal readiness.
Jack Clark, co-founder of Anthropic, predicts that by 2028, artificial intelligence systems capable of autonomously developing and training their successors may be on the market. He emphasizes the significance of this potential development, expressing concern about society's readiness for the changes it could bring.
“This is very significant. I don’t know how to comprehend it,” Clark stated. “The implications are so vast that I feel overwhelmed.”
Clark envisions a future where AI can autonomously manage the entire research cycle, which includes:
- defining research tasks;
- designing and conducting experiments;
- writing and testing code;
- optimizing training processes;
- enhancing the architecture of subsequent AI versions.
He describes this scenario as a “Rubicon into an almost unpredictable future,” estimating a 60% likelihood of its occurrence within the next two years.
Foundation of the Prediction
Clark's forecast is based on advancements in several benchmarks:
- SWE-Bench, which tests the ability to solve real engineering tasks, shows that leading models managed only 2% of cases in 2023 but are expected to reach 94% by 2026;
- CORE-Bench, which assesses the reproduction of scientific results, is nearing a 95.5% accuracy rate with current AI agents;
- MLE-Bench, which evaluates machine learning tasks, shows that top systems are achieving around 64-65% success rates.
According to Clark, these benchmarks indicate a rapid shift in AI capabilities from basic coding to executing complex engineering and research tasks.
Increasing Autonomy
Clark also pointed out that AI systems are now capable of performing tasks for longer periods without human oversight. In 2022, AI managed tasks that typically took humans seconds, but by 2025, this duration is expected to extend to six hours, with leading models currently able to work continuously for about 12 hours.
This trend is linked to the development of agent-based programming tools, allowing AI to maintain goals, evaluate interim results, and rectify errors throughout the research process.
Current AI Development Patterns
Clark noted that the AI development cycle involves studying materials, reproducing results, conducting experiments, retraining models, and identifying bottlenecks. The improvements seen in benchmarks suggest that AI is already taking over significant portions of this cycle.
He highlighted that AI is beginning to play a role in specialized tasks, such as designing GPU cores and improving the training of small open-source language models. As of spring 2026, the best neural networks are projected to achieve 25-28% of the target growth in model retraining.
AI's Role in Managing AI
Clark observed that modern AI systems are starting to coordinate the efforts of other AI agents, a development already seen in tools like Claude Code, which allows one assistant to assign tasks to several sub-assistants.
This capability is crucial as AI tasks often require managing multiple parallel processes, potentially reducing the need for human involvement significantly.
Creativity in AI Development
A key question raised by Clark is whether AI development is akin to discovering new scientific theories or merely assembling components. He acknowledged that while current large language models (LLMs) may not generate groundbreaking ideas, this may not hinder the automation of much of AI research and development.
“AI primarily advances through methodical execution of a cycle: taking a well-functioning system, scaling it, and correcting errors,” Clark remarked.
Potential Impact of Automation
Clark noted that major AI labs are already pursuing automation in research, with OpenAI planning to create an AI intern for independent scientific tasks and Anthropic working on automatic tuning to align AI with human values. If this trend continues, the industry may enter a phase of complete automation in AI development.
Should this occur by the end of 2028, society will face not just technological advancements, but also critical questions about safety, economic distribution, and the evolving role of human labor.
“If you forced me to name a probability for 2027, I would say 30%. If we don’t see this by the end of 2028, I think we will discover a flaw in the current technological paradigm,” Clark concluded.
Earlier this year, Anthropic CEO Dario Amodei predicted the imminent emergence of artificial general intelligence and potential job reductions.
