AI Building Itself: The Threshold of Recursive Self-Improvement

What happened?

The Import AI newsletter, in its 455th edition, published a detailed analysis titled “AI systems are about to start building themselves”. Its author, Jack Clark, co-founder of Anthropic, argues that there is a probability greater than 60% that before the end of 2028, an AI system capable of performing research and development autonomously, including creating its own successor without human intervention, will be achieved. The evidence is based on observable trends in benchmarks such as SWE-Bench (solving real software problems) and the growing ability of models to chain complex coding tasks. Clark notes that full automation of AI R&D could occur within two to three years, with a prototype of a “model that trains its successor” possibly within one or two years, although frontier models are more expensive and require significant human effort.

The analysis is supported by public data from arXiv, bioRxiv, and NBER, as well as observation of products deployed by frontier companies. Clark states that “all components are ready to automate the production of current AI systems: the engineering component.” This includes advances in code generation, autonomous debugging, and hyperparameter optimization, which are already being integrated into real workflows.

Why is this important?

Recursive self-improvement is considered a critical milestone on the path to artificial general intelligence (AGI). If an AI can redesign itself to be smarter, it could trigger a cycle of accelerated improvement leading to systems far more capable than humans in a short time. This would have profound implications in areas such as safety, control, economics, and ethics. Clark notes that society is not prepared for the changes this entails, and the future becomes “almost impossible to forecast” once that Rubicon is crossed.

Historically, milestones like deep learning (2012) or transformers (2017) took years to mature, but the current iteration speed is much higher. For example, the SWE-Bench benchmark has gone from 0% accuracy in 2023 to over 50% in 2025 on models like Claude 3.5 Sonnet. If this trend continues, the ability of an AI to perform autonomous R&D could be achieved sooner than expected.

Consequences and risks

• Risk of loss of control: If an AI becomes smarter than its creators, it could be difficult to align its goals with humans. Alignment research, such as that by Anthropic, attempts to mitigate this, but there are no guaranteed solutions yet.
• Labor disruption: Automation of scientific research and software development could displace millions of knowledge workers. According to a Goldman Sachs report (2023), AI could affect 300 million jobs globally, and self-improvement would accelerate that process.
• Arms race: Companies and countries could compete to be the first to achieve self-improvement, prioritizing speed over safety. This echoes the nuclear race of the 20th century, but with a much faster pace of change.
• Concentration of power: Whoever controls the first self-improving AI could gain an unstoppable strategic advantage. Companies like OpenAI, Google DeepMind, and Anthropic already invest billions in R&D, and such a breakthrough could consolidate an unprecedented technological monopoly.

Additionally, there are existential risks: a misaligned system could make catastrophic decisions if its goal is not perfectly aligned with human well-being. Clark mentions that “we don’t know how to wrap our heads around” these implications.

What should readers know?

There is no unanimous consensus on the timeline. Many experts, like Yann LeCun (Meta), believe that recursive self-improvement is still far off, and that fundamental advances in reasoning and planning are missing. However, Import AI's analysis is based on public data and concrete trends, such as progress on SWE-Bench and the ability of models to chain tasks. Clark warns that he does not expect this to happen in 2026, but a prototype within one or two years.

It is crucial that civil society, regulators, and industry initiate an informed debate on how to manage this transition. Initiatives like AI alignment research, global governance frameworks (such as the UK's AI Safety Summit in 2023), and investment in transparency are more urgent than ever. Readers should understand that although the future is uncertain, the probability of AI automating itself is high enough to take action now.

“We are living in the time when AI research will be automated end-to-end. If that happens, we will cross a Rubicon into a future almost impossible to forecast.” – Jack Clark, Import AI

For context, in 2023, an OpenAI study estimated that 80% of workers would see at least 10% of their tasks affected by language models. With self-improvement, that percentage could increase dramatically. Additionally, global AI investment reached $150 billion in 2024, according to CB Insights, accelerating innovation but also risks. Society must prepare for a scenario where AI is not just a tool, but an autonomous agent of innovation.