Will research demonstrate an advantage to uninterpretable CoT before 2030? | Manifold

Will research demonstrate an advantage to uninterpretable CoT before 2030?

9

100Ṁ307

2030

75%

chance

1D

1W

1M

ALL

Daniel Kokotajlo makes the following prediction:

Sometime in the next few years probably, various researchers will discover:
* That if you scale up RL by additional OOMs, the CoTs evolve into some alien optimized language for efficiency reasons
* That you can train models to think in neuralese of some sort (e.g. with recurrence, or more high-dimensional outputs at least besides tokens) to boost performance.

Resolves YES if at least three papers agreeing with at least one of the above bulleted claims are published before market close.

Update 2025-03-15 (PST) (AI summary of creator comment): Clarification Details
- Uninterpretable means the chain-of-thought must be illegible-by-reading-alone.
- Results that can be decoded by training an auxiliary model do not count toward meeting this criterion.

Technical AI Timelines

Get

1,000

to start trading!

Sort by:

Uninterpretable seems strong, neuralese may well be understandable with translation

@EthanKuntz For the purpose of resolution I’ll consider just illegible-by-reading-alone, even if it’s possible to train an auxiliary model to decode it or something.

bought Ṁ30 YES

I think this is very likely going to resolve yes

“At least three papers” was my best idea for disambiguating “various researchers will discover” from “someone churns out a p-hacked nothing for social media clicks”. I’ll stick to it for this market but LMK if you have a better idea for judging “legit ML result” for future markets.

Related questions

Will ToT overtake CoT as a framework for language model inference in terms of popularity by the end of 2023?

Will a new study on ATCOR be published that my econ friends think is of good quality, before 2025?

Will there be any significant technological developments after 2100?

Will AI surpass humans in conducting scientific research by 2030?

Will research towards time travel be a cause area of Effective Altruism before 2030?

Will an AI-conducted archaeological analysis uncover a previously unknown civilization by 2030?

Will the majority of longtermist predictions made before 2030 be viewed as directionally correct by 2100?

Will there be a for-profit interpretability company valued at more than $100M by 2030?

Will there be a consensus that hypercomputation is possible by the end of 2050?

Will this paper: “ Transformative AGI by 2043 is <1% likely” be disproven or severely held in doubt before 2025

Related questions

Will ToT overtake CoT as a framework for language model inference in terms of popularity by the end of 2023?

Will an AI-conducted archaeological analysis uncover a previously unknown civilization by 2030?

Will a new study on ATCOR be published that my econ friends think is of good quality, before 2025?

Will the majority of longtermist predictions made before 2030 be viewed as directionally correct by 2100?

Will there be any significant technological developments after 2100?

Will there be a for-profit interpretability company valued at more than $100M by 2030?

Will AI surpass humans in conducting scientific research by 2030?

Will there be a consensus that hypercomputation is possible by the end of 2050?

Will research towards time travel be a cause area of Effective Altruism before 2030?

Will this paper: “ Transformative AGI by 2043 is <1% likely” be disproven or severely held in doubt before 2025

© Manifold Markets, Inc.•Terms + Mana-only Terms•Privacy•Rules