Epoch's recent report predicts that 2e29 training runs will likely be feasible by 2030.
Resolves YES if:
- A credible source publicly confirms that an AI model was trained using at least 2e29 FLOP
- A reasonable analysis estimates that a model was trained using at least 2e29 FLOP.
Resolves N/A if:
- Reasonable analyses disagree about whether there was ever an AI model that was trained using more or less than 2e29 FLOP.
-There are compelling but questionable rumors of a training run that used at least 2e29 FLOP. (Compelling enough to be covered in major news outlets, questionable enough that it's not widely accepted as true)
Resolves NO otherwise.
@vluzko Great question, I didn’t think of that. I think that with the way I’ve worded things so far, I’ve gotta say that that wouldn’t count