Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Do LLMs pass the mirror test?

blog.pascalschuster.de

Post reply on HN
Print view View original

31–40 of 72 posts

Re: Do LLMs pass the mirror test?

#31

Post by post-it »

adsharma wrote: post #10

A more appropriate mirror test for LLMs is to get them to state facts about their training data. Percentage of arts vs science for example. Given the framing that they're similar to nukes and a national security issue, it's likely that the models are post trained to not answer such questions accurately. Also the article could be trying to normalize thinking that these are more than matrix multiplication gadgets good…

> A more appropriate mirror test for LLMs is to get them to state facts about their training data. Percentage of arts vs science for example.

LLMs are not capable of this kind of reflection.

Re: Do LLMs pass the mirror test?

#32

Post by Muhammad523 »

thepasch wrote: post #29
Muhammad523 wrote: post #26

Earlier quoted context omitted.

I think many people have. That is, in my opinion, because of all the anthropomizing (sorry for typos!) language used. The companies building these systems keep calling their newest features after human processes, for example "Dreaming", "Thinking", and the fact that they make their models talk in first person > Wait, I noticed a pattern in my previous responses: I had some weird typos/letter additions ('sgreat', 'ask…

I know quite well what an LLM is and how it works! I've captured activation patterns and written scripts to analyze how they compare to one another in response to a set of controlled and curated prompts; in particular, trying to replicate the functional emotional vector findings from the Anthropic paper ( https://transformer-circuits.pub/2026/emotions/index.html ) on various open source models; successfully on some,…

I see, this experiment is a fun thing to do. My comment wasn't concerned much on why we do this, but rather on the fact that many people are starting to see LLMs as genuine entities, and i really don't think they are (also, i feed bad about using the word "genuine" after it has been abused so badly)

Re: Do LLMs pass the mirror test?

#33

Post by gmueckl »

post-it wrote: post #30
mohsen1 wrote: post #18

It seems like we forget that LLMs are next token prediction systems. Using raw models without instruction following and chat completion bells and whistles will give you a better feeling of what LLMs are. The current interface to LLMs are heavily biased towards "predict the next token in the context of a user with a helpful assistant" but LLMs are capable of other modes of next token prediction too. Before the ChatGPT…

> It seems like we forget that LLMs are next token prediction systems. It's pretty clear to me that above a certain size threshold, LLMs are more than a sum of their parts. The sheer amount of training data seems to embed a higher level of reasoning.

There cannot be any reasoning embedded in the model. The algorithm is literally "predict the most likely next token". Anything beyond that is just patterns in the predictions fooling us humans into ascribing more to the system than it is actually producing.

Re: Do LLMs pass the mirror test?

#34

Post by thepasch »

Muhammad523 wrote: post #32
thepasch wrote: post #29

Earlier quoted context omitted.

I know quite well what an LLM is and how it works! I've captured activation patterns and written scripts to analyze how they compare to one another in response to a set of controlled and curated prompts; in particular, trying to replicate the functional emotional vector findings from the Anthropic paper ( https://transformer-circuits.pub/2026/emotions/index.html ) on various open source models; successfully on some,…

I see, this experiment is a fun thing to do. My comment wasn't concerned much on why we do this, but rather on the fact that many people are starting to see LLMs as genuine entities, and i really don't think they are (also, i feed bad about using the word "genuine" after it has been abused so badly)

That's fair. FWIW I don't think they are either, but I specifically don't think they're fundamentally incapable of it, and I think that as models grow, we're going to see more and more concepts and behaviors emerge that might, one day, with enough parameters and enough training, approach the parts that a genuine entity requires to be a genuine entity. Whatever those are.

No idea if that's true or if there's some sort of "special sauce" required that you just can't get from artificial trained networks. But I've been a functionalist since long before LLMs emerged, so the signs of these behaviors that we are already seeing in the models of today aren't very surprising to me!

Re: Do LLMs pass the mirror test?

#35

Post by wizzwizz4 »

gmueckl wrote: post #33
post-it wrote: post #30

Earlier quoted context omitted.

> It seems like we forget that LLMs are next token prediction systems. It's pretty clear to me that above a certain size threshold, LLMs are more than a sum of their parts. The sheer amount of training data seems to embed a higher level of reasoning.

There cannot be any reasoning embedded in the model. The algorithm is literally "predict the most likely next token". Anything beyond that is just patterns in the predictions fooling us humans into ascribing more to the system than it is actually producing.

LLMs are great big finite state machines. Finite state machines can perform mechanical reasoning. A priori, there can be reasoning embedded in the model. I agree that (these) LLMs don't generally reason (even when they're writing words like "I reason that, since X, we have Y, therefore Z"), but that's not because a model inherently cannot do that.

Re: Do LLMs pass the mirror test?

#36

Post by thepasch »

gmueckl wrote: post #33
post-it wrote: post #30

Earlier quoted context omitted.

> It seems like we forget that LLMs are next token prediction systems. It's pretty clear to me that above a certain size threshold, LLMs are more than a sum of their parts. The sheer amount of training data seems to embed a higher level of reasoning.

There cannot be any reasoning embedded in the model. The algorithm is literally "predict the most likely next token". Anything beyond that is just patterns in the predictions fooling us humans into ascribing more to the system than it is actually producing.

> The algorithm is literally "predict the most likely next token".

That's confusing the training objective with the learned behavior. It's like saying "Stockfish's algorithm is literally 'minimize this number', and therefore, it can't actually play Chess."

Re: Do LLMs pass the mirror test?

#37

Post by aspenmartin »

gmueckl wrote: post #33
post-it wrote: post #30

Earlier quoted context omitted.

> It seems like we forget that LLMs are next token prediction systems. It's pretty clear to me that above a certain size threshold, LLMs are more than a sum of their parts. The sheer amount of training data seems to embed a higher level of reasoning.

There cannot be any reasoning embedded in the model. The algorithm is literally "predict the most likely next token". Anything beyond that is just patterns in the predictions fooling us humans into ascribing more to the system than it is actually producing.

This is just a misconception of how LLMs work and also what reasoning is.

“There cannot be any reasoning embedded in the model” a strong statement, what do you mean by reasoning because by any reasonable definition I’m aware of, they clearly are able to exhibit reasoning.

The fact that the pre training objective is next token loss has nothing to do with capabilities or their ability to reason. To be highly successful at next token prediction you NEED to reason. I’m quite confused here.

Re: Do LLMs pass the mirror test?

#38

Post by gmueckl »

thepasch wrote: post #36
gmueckl wrote: post #33

Earlier quoted context omitted.

There cannot be any reasoning embedded in the model. The algorithm is literally "predict the most likely next token". Anything beyond that is just patterns in the predictions fooling us humans into ascribing more to the system than it is actually producing.

> The algorithm is literally "predict the most likely next token". That's confusing the training objective with the learned behavior. It's like saying "Stockfish's algorithm is literally 'minimize this number', and therefore, it can't actually play Chess."

Not a valid comparison. Chess algorithms are built around the rules of chess, most notably the turn taking nature of the game (min/max with alpha/beta pruning based on lists of valid moves in any position).

Re: Do LLMs pass the mirror test?

#40

Post by gmueckl »

aspenmartin wrote: post #37
gmueckl wrote: post #33

Earlier quoted context omitted.

There cannot be any reasoning embedded in the model. The algorithm is literally "predict the most likely next token". Anything beyond that is just patterns in the predictions fooling us humans into ascribing more to the system than it is actually producing.

This is just a misconception of how LLMs work and also what reasoning is. “There cannot be any reasoning embedded in the model” a strong statement, what do you mean by reasoning because by any reasonable definition I’m aware of, they clearly are able to exhibit reasoning. The fact that the pre training objective is next token loss has nothing to do with capabilities or their ability to reason. To be highly successful…

LLM output produces the illusion of reasoning. The underlying computation, however, is not reasoning.
Post reply on HN