soon the big labs will train HUGE MODELS that won't be served to the public

Discussions and reports from major labs indicate that the latest generation of models, such as Astra and Fable, are significantly accelerating internal R&D for future models. It is anticipated that these labs will soon train massive models, potentially 20T parameters, which will be too expensive to serve to the public, costing hundreds of dollars per million tokens. Instead, these powerful models will be used internally to further boost R&D efforts.

This report highlights a shift in AI model development, moving from public-facing deployment to internal R&D tools due to escalating costs, unlike previous generations aimed at broader access.