On September 12, Anthropic CEO Dario Amodei published one of his most controversial essays. He titled it ‘We Must Pace the Frontier’, and in his text he defended its necessity Development is slow Frontier artificial intelligence (AI) models aim to ensure their proper alignment with human interests and minimize as much as possible the risks arising from the development of the most advanced AI.
This passage in their article is particularly insightful: “As of this summer, AI is advancing very rapidly, driven primarily by its increasing ability to develop next-generation models. This dynamic is called iterative self-improvement, and it is beginning to occur across industries, including anthropology.” We understand iterative self-improvement as the ability of a language model to autonomously improve new models on its own.
It is clear from Amodei that the most advanced models in anthropology and other laboratories are already participating in the development of new, more capable language models, although at the moment none of these companies has publicly demonstrated that their models have achieved iterative self-improvement in a strict sense. However, an unexpected twist has just occurred: a group of researchers at MIT (Massachusetts Institute of Technology) has shown that GPT-6 Astra is able to directly change the behavior of other AIs.
What does Astra really do and what does it mean?
Several researchers at the MIT Artificial Intelligence and Computer Science Laboratory have verified that GPT-6 Astra, OpenAI’s AI model, is able to directly change the internal parameters of another neural network to improve its behavior. Interestingly, these scientists conducted their experiments on Qwen3.5-0.8B, a language model with 800 million parameters developed by the Chinese company Alibaba. What Astra has managed to do is improve its performance in arithmetic operations by leveraging parameters of only 0.016%.
As the MIT researchers explain in their paper, it’s interesting that Astra was able to improve Qwen3.5-0.8B’s math performance so much by making such a modest change. However, it is also surprising that this improvement was possible Without adopting a new model training process. What Astra did was change the peso value of Qwen. To understand why this is important, we are interested in reviewing what an AI model weighs.
Instead of retraining Kuen, Astra identified a problem in its performance and directly manipulated the numerical values that triggered it.
Great language models are great at recognizing patterns. In fact, this capability is at the heart of its operation and is achieved through billions of numerical parameters, including weights, that adjust the model during training. An AI is a vast network of billions of tiny connections. During the training process, AI analyzes large amounts of data and adjusts those connections to learn to recognize patterns.
The numerical value of each of these connections determines how much each of them affects the behavior of the model. These values are weights. Instead of retraining Kuen, Astra identified a problem in its performance and directly manipulated the numerical values that triggered it. This is a hitherto unprecedented technique in the hands of an AI that opens a new path in accelerating the development of new models, although we are not facing the iterative self-improvement that we talked about a few lines above. Not in the strict sense.
The good news is that AI laboratory researchers have one more tool in their arsenal when refining their frontier models. An AI model directly intervening in other neural networks has the potential to speed up scientific research, correct problems that until now required very long training processes, and facilitate the development of more efficient models. However, this scenario also poses risks.
If more advanced language models begin to intervene more autonomously in their own components or in other models, AI labs will need to find a way to ensure that these changes do not trigger errors. Weaknesses are difficult to identify or abnormal behavior. Furthermore, if improvements accumulate over successive generations of AI models, we run the risk of making it more difficult to evaluate and control the evolution of frontier models.
Picture | Google DeepMind (edited)
More information Rikiya Taki (MIT) | anthropological
In Xataka | China has plans to win the AI war against the US. And Dipsik’s champion
