Will any transfer learning model, trained for any amount of time on one Atari environment, outperform the median human learning curve on most other Atari environments when transferred by 2026?

8

Ṁ560Ṁ267

resolved Jan 6

Resolved

NO

1H

6H

1D

1W

1M

ALL

- Median human learning curve will be based on whatever the best available data for the human learning curve at the time is. If there is no good data available the question resolves N/A

Market context

Technical AI Timelines

Ancient Markets

Get

1,000

to start trading!

🏅 Top traders

#	Trader	Total profit
1		Ṁ88
2		Ṁ81
3		Ṁ0
4		Ṁ0

Sort by:

This one is closer than I expected: JOWA has pretty impressive transfer learning but doesn't actually met the criteria: https://arxiv.org/pdf/2410.00564.

Any idea what the current performance is here?

@StephenMalina Essentially nonexistent. There are scattered attempts at transferring between a few games. AFAIK no one has made any notable progress and no one has tried transferring across all games.

People are also trading

In 2028, will an AI be able to play randomly selected computer games at human level without getting to practice?

-4% 1d49% chance

[ACX 2026] Will an AI model reach a 3 hour time horizon with 80% reliability during 2026?

Benchmark Gap #6: Once we have a transfer model that achieves human-level sample efficiency on many major RL environments, how many months will it be before we have a non-transfer model that achieves the same?

Will a publicly known AI model achieve an 80% time horizon that is an 1 hour and 30 minutes by September 2026?

Will models be able to do the work of an AI researcher/engineer before 2027?

Will an AI achieve >80% performance on the FrontierMath benchmark before 2027?

Will an AI achieve >85% performance on the FrontierMath benchmark before 2027?

Will a new lab create a top-performing AI frontier model before 2028?

In 2028, will an AI be able to play randomly-selected computer games at human level, given the chance to train via self-play?

Will it take <=12 months for open-source AI forecasting to go from consistently worse than humans to broadly superhuman?

Related questions

In 2028, will an AI be able to play randomly selected computer games at human level without getting to practice?

[ACX 2026] Will an AI model reach a 3 hour time horizon with 80% reliability during 2026?

Benchmark Gap #6: Once we have a transfer model that achieves human-level sample efficiency on many major RL environments, how many months will it be before we have a non-transfer model that achieves the same?

Will a publicly known AI model achieve an 80% time horizon that is an 1 hour and 30 minutes by September 2026?

Will models be able to do the work of an AI researcher/engineer before 2027?

Will an AI achieve >80% performance on the FrontierMath benchmark before 2027?

Will an AI achieve >85% performance on the FrontierMath benchmark before 2027?

Will a new lab create a top-performing AI frontier model before 2028?

In 2028, will an AI be able to play randomly-selected computer games at human level, given the chance to train via self-play?

Will it take <=12 months for open-source AI forecasting to go from consistently worse than humans to broadly superhuman?

© Manifold Markets, Inc.•Terms•Privacy