What will be true of the SOTA AI on the FrontierMath benchmark, before 2028?

Plus

Ṁ215

2028

63%

Transformer-based architecture

55%

Narrow domain of knowledge. ie Does not know random facts such as when Google was founded, or who won the 1960 presidential election.

50%

Developed by Google Deepmind

50%

Developed by OpenAI

50%

Part of the AlphaProof family of models (AlphaProof N and variations)

50%

Part of the o1 family of models (o1, o2, etc. and variations)

50%

Based on Symbolic AI (https://en.wikipedia.org/wiki/Symbolic_artificial_intelligence)

50%

Energy-based Model (https://en.wikipedia.org/wiki/Energy-based_model)

42%

Over 1T parameters

31%

Part of the GPT-N family of models (GPT-5, GPT-6, and variations)

26%

Developed by a non-British and non-American company

An option resolves YES if it is true about the AI model, or program, known to be State of the Art in terms of the FrontierMath benchmark, at the end of the year 2027. It resolves NO otherwise.

You're welcome to add any interesting facts that might or might not be true about the state of the art in math problems, as defined by achieving the highest score on the FrontierMath benchmarks.

I reserve the right to cancel any option that is too vague, too improbable, etc.

/Bayesian/what-will-true-of-the-sota-ai-on-th-ROldIhZZgt
/Bayesian/what-will-true-of-the-sota-ai-on-th-RQptyR5uO8 (this market)

/Bayesian/will-an-ai-achieve-85-performance-o-hyPtIE98qZ

/MatthewBarnett/will-an-ai-achieve-85-performance-o

/Bayesian/will-an-ai-achieve-30-performance-o

This question is managed and resolved by Manifold.

#AI

#️ Technology

#Technical AI Timelines

#IMO Grand Challenge

#Math

Get