If GPT-5 can do recursive self-improvement, will it first be via fine-tuning on its outputs?
AIOne-Off12d
Manifold MarketsGeen KYCGeverifieerde afwikkelingsgegevensRedelijk gekalibreerdHuidige gemeenschapsvoorspellingNee 63.4%
Voorspellers
7
Vraagtype
binary
Methodologie
Play-money forecasting platform
Brontype
Voorspelling
Marktdata
Bijgewerkt 9 minuten geleden
30 mei 23, 17:28 UTC1 nov 26, 06:59 UTC
Trends
Regels
If this market resolves Yes: https://manifold.markets/NathanHelmBurger/will-gpt5-be-capable-of-recursive-s and I agree with the resolution, will it be because GPT-5 is fine-tuned on its outputs? There can be an initial RLHF or similar training stage, but the core of the self-improvement shouldn't include humans or RL, just supervised learning.
- That is, for each training step: - There is some set of input datapoints. - With minimal prompting/scaffolding, GPT-5 produces some output given each input datapoint. - GPT-5 is fine-tuned (normal supervised learning) on some simple function of the input data and GPT-5's output.
- If Nathan's market resolves No or N/A, this market will resolve N/A. If I disagree with the resolution, this market will resolve N/A. Note this part of Nathan's market description: > As clarified in the comments, if the [Noa: steps of the?] recursive self-improvement can't be clearly demonstrated using less than 3% of the FLOPs used in training GPT-5, then it doesn't count.
