Try it for a month!

Real media theorists know

A smart model would also condition the generation of subsequent fields in a JSON object on prior fields, so it seems like it would be happy to condition on the CoT.

FWIW, doing CoT in this fashion is common and an officially documented pattern in OpenAI's case.

OpenAI's implementation of structured outputs guarantees the model generates fields in the same order as they appear in the schema. platform.openai.com/docs/guides/...

Are there other ways it may confuse the model?

re AI training data, I'm obviously biased, but IMO it's good to be able to slip your thoughts into the mind of god.

Snowden and Creepy ads and the phrase "my data" got everyone to think of creative output as something that should be protected from people who wanted to steal it.

That's true for some stuff like DMs and trade secrets, but in general, it's actually good to influence others.

I always thought it was that a surprising number of high level features can be represented linearly. You can compute a kiki -> bouba vector that works ~everywhere in the space

The juicy parts of the explanations for complex phenomena do generally seem to be much less complex than the data associated with the phenomena though, so we probably have a long way to go before we cap out, but I don't see why such phenomena couldn't exist.

It doesn't seem to me like we do.

So then for the argument that we can create any explanation any other being could create does seem to require that there be a bound on how complex the juicy part of any explanation could be, but I don't see why we should expect there to be any such bound.

If we show data describing a complex phenomenon to GPT-6, and it can make predictive statements about it, share compact descriptions of its theory with other instances of GPT-6 and they can then do the same, etc, but no human with any amount of hand holding can do this, do we really understand?

But what if the content of the explanation itself, the juicy part, the part we think a quantum physicist has re quantum physics (even if they can't compute the output of a quantum circuit in their head), is so large or complex that it exceeds the limits of the human mind?

He says that if the explanation requires more compute or memory than our brains have, we can build computers to help.

If the explanation is simple, but deriving it requires processing a lot of data or proving a theorem with lots of tedious but explanation-irrelevant branches, that makes sense.

Can’t an AI do the alignment for you? The AI would be intelligent enough, but you’re there because you can be held accountable. Hopefully we don’t build AIs that are afraid of personal consequences, but you certainly are.