By 2028, AI systems capable of independently developing and training their successors without human intervention could emerge in the market. This prediction was made by Jack Clark, co-founder of Anthropic.

“This is very significant. I don’t know how to comprehend it. I arrive at this conclusion reluctantly because the implications are so vast that I feel overwhelmed, and I’m not sure society is ready for the changes that automated AI development entails,” he noted.

Clark described a scenario of complete automation in AI research, where the model independently:

  • sets research tasks;
  • designs experiments;
  • writes and tests code;
  • optimizes training;
  • improves the architecture of the next version of AI.

The expert referred to this as a “Rubicon into an almost unpredictable future” and estimated the likelihood of such a scenario at 60% within the next two years.

Basis for the Assessment

Clark's conclusion is based on the dynamics of several benchmarks:

  • SWE-Bench — a test for solving real engineering tasks from GitHub repositories. By the end of 2023, the best models handled about 2% of cases; by spring 2026, this figure is expected to reach 94%;
  • CORE-Bench — reproducing results from scientific AI papers by setting up environments, running code, and analyzing outcomes. According to Clark, this benchmark is effectively “closed”: modern agents show about 95.5% accuracy;
  • MLE-Bench — performing ML tasks at the level of Kaggle competitions. The best agent systems are already achieving 64-65%.

According to the co-founder of Anthropic, all three metrics demonstrate one thing: AI is rapidly transitioning from point coding to fully executing engineering and research tasks.

Growth of Autonomy

Another argument is the increasing duration of tasks that AI models can perform without human intervention.

According to METR, in 2022, systems managed tasks that took humans tens of seconds. By 2024, this duration increased to about 40 minutes, and by 2025, it reached six hours. Currently, leading models can conduct engineering work for about 12 consecutive hours.

Clark linked this to the spread of agent-based programming tools. The longer a model maintains a goal, checks intermediate results, and corrects errors, the more stages of the research cycle can be delegated to it.

Importance for AI Development

The current AI development cycle follows a specific pattern: study materials, reproduce results, assemble experiments, train or retrain models, check metrics, identify bottlenecks, and repeat. Growth in SWE-Bench, CORE-Bench, and MLE-Bench indicates that models are already handling entire segments of this cycle.

Clark specifically pointed out progress in more specialized tasks. For example, AI is beginning to be used for designing GPU cores — code that determines the efficiency of training and inference of models on specific hardware.

Another area is model retraining. In the PostTrainBench benchmark, AI systems improve small open-source LLMs.

As of spring 2026, the best neural networks achieve 25-28% of the target growth (human teams achieve 51%). Clark considers this result significant: the benchmark is set by real instructive models created by experienced researchers.

Anthropic measured how its models optimize LLM training on CPUs. Over the course of a year, acceleration increased from 2.9 times (Claude Opus 4) to 52 (Claude Mythos Preview). A human typically requires four to eight hours for a similar task.

AI is Already Learning to Manage AI

Clark noted that modern systems are beginning to coordinate the work of other agents. This approach is already being utilized in products like Claude Code or OpenCode: one assistant distributes tasks among several sub-assistants, monitors their progress, and compiles results.

This is crucial for AI development: they rarely represent a single linear task — usually, they involve dozens of parallel processes, including coding and environment setup. If a model starts managing such processes independently, human involvement will significantly decrease.

Do Neural Networks Need Creativity?

According to the co-founder of Anthropic, one of the key questions is whether AI development resembles discovering a general theory of relativity or assembling Lego.

Clark acknowledged that current LLMs are not yet capable of generating fundamentally new scientific ideas. However, this may not be necessary for automating a significant portion of AI R&D.

“AI primarily advances through methodical execution by humans of a certain cycle: taking a well-functioning system, scaling some aspect of it, examining errors during scaling, and correcting them. This requires very few unconventional ideas, and much of this process resembles unglamorous rough engineering work,” the expert noted.

Early Signs of Scientific Contribution

Clark believes that AI models are already beginning to show early signs of scientific intuition. He provided several examples from mathematics and computer science:

  • A team of mathematicians used Gemini to check around 700 Erdős problems and found 13 solutions, one of which researchers called a “slightly non-trivial” contribution to an open problem;
  • Scientists from the University of British Columbia, the University of New South Wales, Stanford, and Google DeepMind published a mathematical proof found with significant assistance from Gemini-based tools.

What Happens If the Prediction is Correct

Clark pointed out that major AI laboratories are already moving towards automating research. OpenAI plans to create an AI intern for independent scientific activity, while Anthropic is releasing work on automatic tuning to align with human values.

If the current pace continues, the industry will transition to a phase of complete automation in AI development, the expert predicted — a cycle will begin where each new generation of AI accelerates the emergence of the next.

According to him, if the transition occurs by the end of 2028, the world will face not only a technological leap. Fundamental questions regarding safety, capital distribution, the role of human labor, and control over systems that begin to evolve faster than their creators will also come to the forefront.

“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, then I think we will discover some flaw in the current technological paradigm, and human ingenuity will be needed to move forward,” Clark concluded.

Recall that in January, Anthropic CEO Dario Amodei predicted the imminent emergence of AGI and job reductions.