AI 속도조절론 배경엔…“학습→추론 자원 재배치” 분석

Recently, industry analysis suggests that "speed control theory" has shifted the weight of AI model development competitions, as major AI companies like Anthropic and OpenAI have begun focusing on optimizing computational resources from learning to inference stages. This change in strategy, driven by the release of OpenAI's "Astra" model, involves using the same computational processes multiple times in the inference stage to improve performance and enable solving more complex problems.

The result is that Astra achieved an overwhelming score of 99.9% on the ARC-AGI-3 benchmark, which tests models' ability to infer new rules for problem-solving. Major AI companies are now contemplating reallocating GPU computational resources from learning to inference to maximize profits. This shift in strategy, hailed as "speed control theory," could improve operational profit margins for companies like Anthropic and OpenAI.

However, the approach raises concerns about the lack of transparency in the model's reasoning, making it difficult to predict and control the AI's behavior.