If LMs store info as features in superposition, are there >300K features in GPT-2 small L7? (see desc) | Manifold

If LMs store info as features in superposition, are there >300K features in GPT-2 small L7? (see desc)

13

1kṀ286

3000

59%

chance

1H

6H

1D

1W

1M

ALL

If in 2040 I am convinced at >80% confidence that LMs mainly store info in their residual stream as something like sparse linear features, and I am >80% confident in a particular approximate number of features in the residual stream before layer 7, market resolves Yes if that number is greater than 300K. If that number is less than 300K resolves No. Otherwise resolves N/A.

I won't bet in this market.

Mechanistic interpretability

Get

1,000

to start trading!

People are also trading

If LMs store info as features in superposition, does # features scale superlinearly with number of model parameters?

Will GPT-5 have over 100 trillion parameters?

Will GPT-5 have over 1 trillion parameters?

Will GPT-5 have over 10 trillion parameters?

Will GPT-5 have fewer parameters than GPT-4? (1500M subsidy)

Will it be possible to disentangle most of the features learned by a model comparable to GPT-2 this decade?

Size of smallest open-source LLM marching GPT 3.5's performance in 2025? (GB)

Will it be possible to disentangle most of the features learned by a model comparable to GPT-4 this decade?

Will it be possible to disentangle most of the features learned by a model comparable to GPT-3 this decade? (1k subsidy)

GPT-5 # of parameters (in billions)

Related questions

If LMs store info as features in superposition, does # features scale superlinearly with number of model parameters?

Will GPT-5 have over 100 trillion parameters?

Will GPT-5 have over 1 trillion parameters?

Will GPT-5 have over 10 trillion parameters?

Will GPT-5 have fewer parameters than GPT-4? (1500M subsidy)

Will it be possible to disentangle most of the features learned by a model comparable to GPT-2 this decade?

Size of smallest open-source LLM marching GPT 3.5's performance in 2025? (GB)

Will it be possible to disentangle most of the features learned by a model comparable to GPT-4 this decade?

Will it be possible to disentangle most of the features learned by a model comparable to GPT-3 this decade? (1k subsidy)

GPT-5 # of parameters (in billions)

© Manifold Markets, Inc.•Terms•Privacy