Analysis: "Reallocation of learning → inference resources" behind AI speed control theory

Major AI companies, including Anthropic and OpenAI, are shifting their focus from developing larger models to improving the performance of existing ones. OpenAI's recently released GPT-6 Astra model is a prime example, which achieved a 99.9% score in the ARC-AGI-3 benchmark by increasing computations in the inference stage. This approach allows for more accurate problem-solving but raises safety concerns due to the opaque nature of the model's decision-making process.

Experts, including Yoshua Bengio, are warning about the risks of uncontrolled AI development and calling for international safety measures.