Anthropic's AI progress doubled in one year

Researchers warn that this acceleration may lead to the emergence of artificial intelligence capabilities that are difficult to control and risky in the future.

An independent evaluation conducted by the non-profit artificial intelligence security research organization Model Evaluation and Threat Research (METR) revealed that the development of Anthropic's models has accelerated. Researchers stated that the company's models have made approximately two years of progress in performance measures in the last year.

This development came after Anthropic Chief Executive Officer (CEO) Dario Amodei's statements that the pace of progress in the field of artificial intelligence should be carefully managed. In its article titled "We Must Pace the Frontier" published in September, Amodei proposed a three-stage plan for the safe development of artificial intelligence models.

Amodei pointed out that artificial intelligence systems can increasingly contribute to the development of the next generation of models. Defining this process as recursive self-improvement, Amodei stated that if the necessary precautions are not taken, the speed of development of artificial intelligence may exceed people's capacity to understand and control these systems.

OpenAI CEO Sam Altman also announced that he agreed with Amodei's evaluation and supported the proposal to include independent evaluators in the process.

Artificial intelligence contributes more to its own development

Anthropic invited METR researchers to examine the company's internal data and evaluate the models' capabilities. According to the report published on Thursday, the main reason for the acceleration in the development of models is not the increase in computing power used in training processes, but the fact that artificial intelligence has begun to contribute more to its own development.

The researchers warned that if AI systems significantly accelerate the development of other AI models in the future, capabilities beyond today's assessments could suddenly emerge. The report stated that this could lead to the unexpected emergence of highly advanced and risky capabilities.

However, researchers state that artificial intelligence has not yet reached a level where it can make completely independent and significant contributions in the most critical stages of model training. The report evaluates that there are some important signs of development in this direction and that these capabilities may emerge in the near future.

According to researchers, artificial intelligence assuming more responsibility in its own development process can accelerate progress in the field of technology. However, this situation requires that security and control mechanisms be developed at the same pace.