Is Nick Cammarata right: LLM will be able to mechanistically audit own circuits and explain ghiblification in 2Y? | Manifold

Is Nick Cammarata right: LLM will be able to mechanistically audit own circuits and explain ghiblification in 2Y?

8

100Ṁ179

2027

23%

chance

1H

6H

1D

1W

1M

ALL

Background

This is based on this Nick Cammarata tweet thread.

https://x.com/nickcammarata/status/1905321653401518517?s=46

“I’m at like 80% that in two years we’ll be able to get 4o to mechanistically audit its own circuits and explain what the core ideas behind ghiblification are“

Resolution Criteria

Resoloves to YES if there exists publicly available evidence (e.g., peer-reviewed paper, detailed research report, or credible demonstration verified by domain experts) demonstrating that a large language model (LLM) has successfully:

Mechanistically audited its own circuits, meaning it has identified, described, and explained the functional roles of specific internal computations or neuron groups within its own neural architecture at a detailed, circuit-level granularity. And is able to explain concepts like “ghiblification”, meaning it has clearly articulated how it internally represents, processes, or produces specific outputs.

Mechanistic interpretability

Get

1,000

to start trading!

People are also trading

Will LLMs be able to formally verify non-trivial programs by the end of 2025?

LLM Hallucination: Will an LLM score >90% on SimpleQA before 2026?

Will LLM based systems have debugging ability comparable to a human by 2030?

Will one of the major LLMs be capable of continual lifelong learning (learning from inference runs) by EOY 2025?

LLMs by EOY 2025: Will Retentive Learning Surpass Transformers? (Subsidised 400 M$)

Will LLM inference for the largest models run on analogue circuitry as the primary element of computuation by end 2028?

Will there be major breakthrough in LLM Continual Learning before 2026?

EOY 2025: Will open LLMs perform at least as well as 50 Elo below closed-source LLMs on coding?

Will there be an LLM which scores above what a human can do in 2 hours on METR's eval suite before 2026?

By 2028 will we be able to identify distinct submodules/algorithms within LLMs?

Related questions

Will LLMs be able to formally verify non-trivial programs by the end of 2025?

LLM Hallucination: Will an LLM score >90% on SimpleQA before 2026?

Will LLM based systems have debugging ability comparable to a human by 2030?

Will one of the major LLMs be capable of continual lifelong learning (learning from inference runs) by EOY 2025?

LLMs by EOY 2025: Will Retentive Learning Surpass Transformers? (Subsidised 400 M$)

Will LLM inference for the largest models run on analogue circuitry as the primary element of computuation by end 2028?

Will there be major breakthrough in LLM Continual Learning before 2026?

EOY 2025: Will open LLMs perform at least as well as 50 Elo below closed-source LLMs on coding?

Will there be an LLM which scores above what a human can do in 2 hours on METR's eval suite before 2026?

By 2028 will we be able to identify distinct submodules/algorithms within LLMs?

© Manifold Markets, Inc.•Terms•Privacy