Anthropic CEO Dario Amodei has warned that AI development is accelerating as models become more capable of helping build their successors. In an essay titled “We Must Pace the Frontier,” he described this process as recursive self-improvement and argued for slowing the development of frontier systems to support their alignment with human interests and reduce potential risks.
A recent experiment by researchers at the Massachusetts Institute of Technology offers a more limited example of AI-assisted model development. The team found that OpenAI’s GPT-6 Astra could directly alter the internal parameters of Qwen3.5-0.8B, a language model developed by Alibaba with 800 million parameters. According to the researchers, Astra improved the model’s mathematical performance by changing just 0.016% of its parameters, without running a new training process.
Those parameters include weights: numerical values that determine how strongly connections within a neural network affect its behavior. Rather than retraining Qwen, Astra identified an issue and adjusted relevant weights, the account of the study explains.
The researchers’ result shows that an AI model can make targeted changes to another network, but it does not demonstrate recursive self-improvement in the strict sense. The available information also does not establish that a model has autonomously built a more capable successor to itself.
Direct model editing could give researchers another way to refine AI systems. It may help accelerate scientific work, address problems that would otherwise require lengthy training, and support the development of more efficient models. The same capability could also introduce errors, hard-to-detect vulnerabilities or unexpected behavior if changes are not carefully assessed.
If AI systems increasingly modify their own components or other models, evaluating and controlling successive changes may become more difficult, particularly if improvements build across generations.
Sources: Xataka.