Yupo Niu challenges AI 2040 mechanistic interpretability timeline

/TECH

RG

1A

Niu expects AI-assisted research to multiply mechanistic interpretability work.

Combined views

542

2 posts, first seen 5h ago

22 likes

Source Posts on X

Cluster sources

RG

Ryan Greenblatt@RyanGreenblatt

@1a3orn I think I expect (assuming AI 2040 scenario): - Interp via model internals is a bit useful now - It will be modestly useful in 2029 - It will be very useful in 2035 I don't think this interp will that much be "mech interp". Your critique seems reasonable overall.

via X

1A

1a3orn@1a3orn

Like the last three years have given us SAEs, Natural Language Autoencoders, J-Space and more. I expect 2x-40x as much work to be done in mech int in the next 3 years as in all prior years put together, both from field growth and from AI help.

via X

Combined views

542

2 posts, first seen 5h ago

22 likes1 comments

1 comments

4 saves

1 reposts

Sentiment

Positive0%0%Negative

Guardrails removed spam, off-topic, unclear, or duplicate replies.

Digg Deeper

Ask a question below.

Published answers will appear here.

Source Posts on X

Cluster sources

RG

Ryan Greenblatt@RyanGreenblatt

@1a3orn I think I expect (assuming AI 2040 scenario): - Interp via model internals is a bit useful now - It will be modestly useful in 2029 - It will be very useful in 2035 I don't think this interp will that much be "mech interp". Your critique seems reasonable overall.

via X

1A

1a3orn@1a3orn

Like the last three years have given us SAEs, Natural Language Autoencoders, J-Space and more. I expect 2x-40x as much work to be done in mech int in the next 3 years as in all prior years put together, both from field growth and from AI help.

via X

4 saves

1 reposts

Sentiment

Positive0%0%Negative

Guardrails removed spam, off-topic, unclear, or duplicate replies.

Digg Deeper

Ask a question below.

Published answers will appear here.