In "Situational Awareness: The Decade Ahead", Leopold Aschenbrenner claims:
In 2027, a leading AI lab will be able to train a GPT-4-level model in a minute.
Resolves YES if by the end of 2027, a credible claim is made that:
Someone trained a GPT-4-level model in a minute.
or 1440 GPT-4-level models in a day, to account for parallelism
GPT-4-level will be evaluated using the same benchmarks as in the GPT-4 Technical Report Table 2.
I will use my best judgement if mostly the same benchmarks are reported.
"A minute" and "a day" refer to wall-clock times.
This is one of a series of markets on claims made in Leopold Aschenbrenner's Situational Awareness report(s):
Other markets about Leopold's predictions:
Results for round 4 of ML Perf are in. GPT-3 was trained on 11k H100s to a certain perplexity threshold in 4 minutes. Not clear how the threshold compares to "true" GPT-3 training, but it's getting fast!