Gary Marcus prediction: physical reasoning of AI systems will improve before psychological reasoning

Ṁ1kṀ5.2k

resolved Oct 18

Resolved

ALL

https://garymarcus.substack.com/p/agi-will-not-happen-in-your-lifetime

Resolves very subjectivly; don't bet if you don't trust my judgement.

Physical reasoning includes things like driving cars, cleaning houses, and other use-cases that involve understanding directions and distances in the real world and manipulating physical objects.

Psychological reasoning includes things like understanding human thoughts and ideas, making art and mathematical discoveries, writing computer programs, etc.

This market resolves once it seems that AI's mastery of one of those domains has radically improved over the status quo in January of 2023. For example if before we reach AGI we get self-driving cars that are radically safer than humans in all situations, this will probably resolve YES. If AI comes up with some deep insight to solve a mathematical problem humans have failed to solve for many years, this will probably resolve NO.

(If they're both solved at the same time by the same system that has reached AGI, this resolves to 50% or N/A, whichever traders think makes more sense.)

Gary Marcus GPT-4 predictions

Get

1,000

to start trading!

🏅 Top traders

#	Trader	Total profit
1		Ṁ275
2		Ṁ232
3		Ṁ88
4		Ṁ70
5		Ṁ46

People are also trading

Gary Marcus 2026 AI Predictions

Gary Marcus 2029 AI predications

Will ARC's Heuristic Arguments research substantially advance AI alignment before 2027?

26% chance

AI honesty #2: by 2027 will we have a reasonable outer alignment procedure for training honest AI?

24% chance

Will AI continue to improve?

97% chance

Will Gary Marcus be accurate on at least 50% of his predictions on AI in 2029?

52% chance

Once AI can predict the future better than any human it will be AGI

49% chance

By what date will at least one state-of-the-art general-purpose AI system not be a reasoning model?

Will AI be capable of fast self improvement before physical AI robots are massively used for improving AI capabilities?

85% chance

By what date will all state-of-the-art general-purpose AI systems not be reasoning models?

Sort by:

@mods website is broken, please resolve this NO for me

bought Ṁ450 NO

From description: "If AI comes up with some deep insight to solve a mathematical problem humans have failed to solve for many years, this will probably resolve NO."

An AI has now solved a millenium prize problem while the same system is barely usable for robotics, and dedicated self driving car tech is only incrementally ahead of where it was at the time this prediction was made. It seems fair to resolve no at this point?

@2b3o4o Agreed, TY

I feel as though your examples should be reversed.

Much of what (in my opinion) makes self-driving a difficult problem is understanding the human psychology of other drivers, but the physical challenges of it I would consider to be more well-defined for computation, and I would expect them be overcome earlier.

How is mathematics "psychological reasoning"? Mathematics is not inherently psychological, is it?

@cloudprism I would absolutely bet YES if I understood the resolution criteria

Why is this market SOOOOOO HIGH?

The most impressive models to date are LLMs which understand the world through words. World's were designed for, and are best at, expressing thoughts and feelings. Of course that's where these models will excel.