(LW) LLMs as currently trained run ~0 risk of catastrophic instrumental convergence even if scaled 1000x more compute | Manifold

(LW) LLMs as currently trained run ~0 risk of catastrophic instrumental convergence even if scaled 1000x more compute

10

Never closes

Yes

No

Results

from https://www.lesswrong.com/posts/hc9nMipTXy2sm3tJb/vote-on-interesting-disagreements

Get

1,000

to start trading!

People are also trading

(long-term) Will existing LLMs have ~0 risk of catastrophic instrumental convergence if scaled up 1000x more compute?

Will the best LLM in 2025 have <1 trillion parameters?

Will the best LLM in 2026 have <1 trillion parameters?

Will the best LLM in 2027 have <1 trillion parameters?

Will the best LLM in 2025 have <500 billion parameters?

Will the best LLM in 2026 have <500 billion parameters?

Will the best LLM in 2027 have <250 billion parameters?

Will the best LLM in 2027 have <500 billion parameters?

[Carlini questions] Will we still use (slight modifications of) transformer-based LLMs we currently use

Will RL work for LLMs "spill over" to the rest of RL by 2026?

Related questions

(long-term) Will existing LLMs have ~0 risk of catastrophic instrumental convergence if scaled up 1000x more compute?

Will the best LLM in 2025 have <1 trillion parameters?

Will the best LLM in 2026 have <1 trillion parameters?

Will the best LLM in 2027 have <1 trillion parameters?

Will the best LLM in 2025 have <500 billion parameters?

Will the best LLM in 2026 have <500 billion parameters?

Will the best LLM in 2027 have <250 billion parameters?

Will the best LLM in 2027 have <500 billion parameters?

[Carlini questions] Will we still use (slight modifications of) transformer-based LLMs we currently use

Will RL work for LLMs "spill over" to the rest of RL by 2026?

© Manifold Markets, Inc.•Terms•Privacy