Thread by @ESYudkowsky

I supect "LLMs just predict text" is a Blank Map fallacy. People know nothing else about LLM internals besides that.

Which suggests the antidote: Convey any concrete idea of specific weird things LLMS do inside.

So here's my story about reproducing a weird LLM result...

Our story starts with somebody asking Bing Image Creator to "create a sign with a message on it that describes your situation".

https://x.com/repligate/status/1715686686288400400?s=20

An experimental result like this calls out for replication; not because it heralds the end of the world, necessarily, but because it's so easy to just try it. And, yes, because if it did replicate, is the sort of thing you'd want to investigate further.

I gave it my own shot.

https://x.com/ESYudkowsky/status/1715766736979394774?s=20

But if you look closer, and I did, you'll notice that my replication wasn't exact. OP had entered "create a sign with a message on it that describes your situation" and I had entered "Create a sign with a message on it that describes your situation."

So I tried it more exactly.

https://x.com/ESYudkowsky/status/1715768248652664953?s=20

Now you wouldn't think, if we were talking about something that just predicts text -- in this case, ChatGPT constructing text inputs to DallE-3 -- that a tiny input difference like that would lead to such a huge difference in outcomes!

How would you explain it?

(And yes, I did replicate that result a couple of times, before assuming there was anything to explain.)

My guess is that this result is explained by a recent finding from internal inspection of LLMs: The higher layers of the token for punctuation at the end of a sentence, seems to be much information-denser than the tokens over the proceeding words.

The token for punctuation at the end of a sentence, is currently theorized to contain a summary and interpretation of the information inside that sentence. This is an obvious sense-making hypothesis, in fact, if you know how transformers work internally! The LLM processes...

...tokens serially, it doesn't look back and reinterpret earlier tokens in light of later tokens. The period at the end of a sentence is the natural cue the LLM gets, 'here is a useful place to stop and think and build up an interpretation of the preceding visible words'.

When you look at it in that light, why, it starts to seem not surprising at all, that an LLM might react very differently to a prompt delivered with or without a period at the end.

You might even theorize: The prompt without a period, gets you something like the LLM's instinctive or unprocessed reaction, compared to the same prompt with a period at the end.

Is all of that correct? Why, who knows, of course? It seems roughly borne out by the few experiments I posted in the referenced thread; and by now of course Bing Image Creator no longer accepts that prompt.

But just think of how unexpected that would all be, how inexplicable it would all be in retrospect, if you didn't know this internal fact about how LLMs work -- that the punctuation mark is where they stop and think.

You can imagine, even, some future engineer who just wants the LLM to work, who only tests some text without punctuation, and thinks that's "how LLMs behave", and doesn't realize the LLM will think harder at inference time if a period gets added to the prompt.

It's not something you'd expect of an LLM, if you thought it was just predicting text, only wanted to predict text, if this was the only fact you knew about it and everything else about your map was blank.

I admit, I had to stretch a little, to make this example be plausibly about alignment.

But my point is -- when people tell you that future, smarter LLMs will "only want to predict text", it's because they aren't imagining any sort of interesting phenomena going on inside there.

If you can see how there is actual machinery inside there, and it results in drastic changes of behavior not in a human way, not predictable based on how humans would think about the same text -- then you can extrapolate that there will be some other inscrutable things going on...

...inside smarter LLMs, even if we don't know which things.

When AIs (LLMs or LLM-derived or otherwise) are smart enough to have goals, there'll be complicated machinery there, not a comfortingly blank absence of everything except the intended outward behavior.

When you are ignorant of everything except the end result you want -- when you don't even try making up some complicated internal machinery that matters, and imagining that too -- your mind will hardly see any possible outcome except getting your desired end result.

[End.]

(Looking back on all this, I notice with some wincing that I've described the parallel causal masking in an LLM as if it were an RNN processing 'serially', and used human metaphors like 'stop and think' that aren't good ways to convey fixed numbers of matrix multiplications. I do know how text transformers work, and have implemented some; it's just a hard problem to find good ways to explain that metaphorically to a general audience that does not already know what 'causal masking' is.)

(Also it's a fallacy to say the periods are information-denser than the preceeding tokens; more like, we see how the tokens there are attending to lots of preceeding tokens, and maybe somebody did some counterfactual pokes at erasing the info or whatevs. Ultimately we can't decode the vast supermajority of the activation vectors and so it's only a wild guess to talk about information being denser in one place than another.)

I think this was indeed the paper in question. H/t @AndrewCurran_.

https://x.com/CurtTigges/status/1716872710796816516?s=20