Automating AI Development Could Speed Up Progress 10x in 17 Months

We are used to discussing AI as a developer's assistant: it writes a function, finds a bug, assembles a prototype. A new study raises the question a level higher. What happens if models start automating not just ordinary programming but the very process of building the next models: experiments, analysis of results, preparation of training systems, and the search for new methods?

The authors call this scenario a software intelligence explosion. It does not require waiting for a new generation of accelerators. The acceleration can arise within computing infrastructure that is already available: AI helps researchers run more experiments, those experiments improve the AI, and the new version speeds up research even more.

How an acceleration loop could emerge

Imagine a lab of one hundred engineers. At first, AI only completes code and prepares tests. Then an agent receives a task like "check three architecture variants," changes configurations on its own, runs experiments, and consolidates the results. A human still approves decisions, but in a day reviews work that used to take a week.

The next step is automating a significant part of the research cycle. If a model can not only carry out an assigned experiment but also choose a promising direction, test hypotheses, and account for failed results, the lab's productivity begins to grow along with the model's capabilities. This is what creates positive feedback.

As an early signal, the authors cite Anthropic's internal figures. The share of accepted code written by AI rose from a few percent in January 2025 to more than 80% by May 2026. The share of autonomous R&D tasks performed under close supervision by specialists increased from 1% in March to 26% in August 2026. These numbers do not prove that an explosion is near, but they show how quickly the organization of development is changing.

The authors' model links such indicators to the economic returns on research effort. With central estimates of the returns parameter from 1.2 to 1.9 across different areas of AI, full automation of research could increase the pace of progress 10x in roughly 17 months. In that regime, the amount of work that used to take a year would be done in about five weeks.

Why this is not yet a forecast

The calculation has a wide zone of uncertainty. The lower bounds of the 90% confidence intervals fall below the level at which self-sustaining acceleration occurs. In other words, the available data is consistent both with sharp growth and with a far calmer scenario.

There is also a measurement problem. A growing share of AI-written code is not the same as growing scientific productivity. An agent can quickly write experimental scaffolding but fail to propose a good hypothesis. The metrics also conflate improvements in software methods with increases in computing resources. Finally, research runs up against hardware, data, energy, verification of results, and people's capacity to make decisions.

Therefore, the figure "10x in 17 months" should be read as the output of a model under given assumptions, not as a calendar promise. The work is useful above all because it turns a vague conversation about an intelligence explosion into a set of measurable quantities: the share of autonomous tasks, the quality of results, human oversight time, and the returns on additional research effort.

What teams should do now

Developers of AI systems should measure code generation and real autonomy separately. A useful metric is not the number of lines but the share of tasks that an agent carried to a verifiable result without human intervention. Alongside it, the cost of errors, review time, and the number of reverted actions should be tracked.

Product teams need agent action logs, limited permissions, isolated environments, and the ability to halt execution instantly. The longer an independent chain of actions, the riskier it is to rely only on the model's final answer.

For founders, the takeaway is practical too: development speed may change faster than familiar planning and hiring cycles. But it is too early to build a strategy around an inevitable "explosion." It is wiser to prepare processes for a rapid rise in autonomy while preserving control, reproducibility of experiments, and a human at the points where an error is especially costly.