AI researchers seek tools for automated AI development slowdown

A group of AI workers is pushing for an automated AI development slowdown framework, asking the U.S. government to support international work on technical and governance tools that could pace AI-assisted model research. The open letter was signed by 1,134 employees at companies building frontier AI models, including Anthropic, OpenAI, Google and Meta.

The signatories also include workers at venture-backed AI labs such as Thinking Machines and London-based Inherent Laboratories. High-profile names on the letter include Anthropic Chief Executive Dario Amodei, several Anthropic co-founders and the chief scientists of OpenAI, Meta’s Meta AI unit and Thinking Machines.

The request targets a specific part of the AI race: using AI systems to automate the work of building, optimizing and testing newer AI models. The signatories argue that if model development itself becomes heavily automated, capability gains could speed up faster than companies or regulators can understand the resulting systems.

Why do AI researchers want to slow automated model development?

The concern is that AI tools are beginning to improve the infrastructure and research workflows that produce stronger AI systems. In the letter, the signatories said it is difficult to predict how much this will accelerate AI progress, but warned of a “real risk” that capability development could move beyond the industry’s ability to understand or control the systems it creates.

The letter asks the U.S. government to back an international effort to develop tools needed to “deliberately pace” the frontier of automated AI development. The signatories said those tools do not yet exist, which makes the request more about creating policy and technical capacity than imposing an immediate brake on any named company.

Recent disclosures from leading labs show why the issue is becoming less theoretical. In April, OpenAI said its GPT-5.5 model helped optimize the Nvidia GPU cluster on which it runs. One optimization produced by the model increased token generation speeds by more than 20%, according to OpenAI. The company has since released the GPT-5.6 series, which it said is much stronger at coding tasks.

Anthropic has also described using AI to accelerate AI research, though in a safety context. In April, the company said it had built a Claude-powered agent to study weak-to-strong supervision, a machine learning technique its researchers believe could help supervise future AI models more effectively.

The letter lands during a run of public AI policy interventions from researchers and industry executives. Nvidia and other large technology companies recently urged U.S. policymakers not to ban open-source AI models. Separately, a group of economists called for tougher AI regulation.

Google DeepMind Chief Executive Demis Hassabis also recently proposed a new standards body for AI safety. His proposal called for risk benchmarks that could be used to evaluate new large language models before broad release. Like the new letter, that proposal asks government to help organize the next layer of AI oversight rather than leaving evaluation solely to individual labs.

For AI companies, the debate cuts into a competitive advantage they are actively pursuing: using models to make the next generation of models cheaper, faster and more capable. The open letter does not provide a finished mechanism for slowing that cycle, but it shows that some senior people inside frontier labs want tooling in place before automation compresses development timelines further.

This story draws on original reporting from SiliconANGLE.