Three weeks ago, rumors spread that an AI lab had successfully completed the largest training run ever

Three weeks ago, rumors spread that an AI lab had successfully completed the largest training run ever, and that the result far exceeded predictions from existing scaling laws. Based on recent information about Mythos, it's highly likely that the lab in question is Anthropic.

This rumor also includes an architectural innovation. If Anthropic achieved performance that completely departs from the existing trendline through training beyond a certain scale, this could be seen as an architectural breakthrough that goes beyond simple scaling.

According to leaked drafts, the candidate names for the new model were Mythos and Capybara, and there are observations that its performance was about 2x higher than expected. Anthropic also mentioned in an interview with Fortune that the new model achieved a 'step change.'

This shift is expected to have a major impact across the industry. OpenAI's recent restructuring and its decision to halt Sora may also be strategic choices to focus on large-scale training. In addition, as model performance increases, service costs may rise, placing a greater burden on users, and the importance of compute, memory, and energy will be further maximized.