Anyone have any techniques for combining the "show your steps" style prompting, but then "only provide the answer" as some sort of output filter? Feels like you could feed "Now answer the original question without any explanation" as a second round trip on the chat dialog, and return that to the user instead.

I just have chatGPT label sections of its output, within an answer, eg:

[notes]

…stuff about how it will perform the task…

[output]

…stuff I wanted it to write…

- - -

That would be easy to then parse and it does a good job of consistently labeling.

Can you give example of the prompt to make it to this?

Sure, here’s a simple one.

- - - -

Hello ChatGPT, please perform the following [task] according to the provided [rules].

[rules]

- start your response with a section labeled [notes] where you describe how you’ll complete the task

- complete the task in a section labeled [output]

[task]

Write a paragraph explaining why water flows.

I just watched a Ted talk by Sal Kahn (founder of Kahn academy) where he talks about giving their chat bot an internal monologue. It's light on details but might be a starting point.

https://youtu.be/hJP5GqnTrNo?t=735

There could still be a state before generation and in response to input.

Give GPT-4 about 10 or a dozen words from a meaningful sentence scrambled. Now tell it to unscramble those words to make a meaningful sentence. As long as you don't try too many tokens, the rate of meaningful completions is nearly 100%. The chances of that happening via random chance is to put it lightly, extremely small.

The only way for humans to solve this kind of problem is to think about the words before the output or in this case generation.

... well, this is functioning as intended. It's a next-word prediction model.

Jesus Christ you need to stop spamming the same tokenized response to every comment on this thread or at least add a damn GOSUB routine.

Twice is hardly spamming. But whatever, I have no intention of repeating myself.

This argument is highly unsatisfactory. The only way for humans to divide two large numbers is to perform long division. That doesn’t mean that’s how a computer does it.

GPT LLMs are literally highly-skilled probabalistic next word predictors. It would be an astonishingly bad GPT that could not unscramble 12 words into a meaningful sentence. I just don’t think success at that task tells us anything other than it is a GPT.

If computers calculated arithmetic by prediction or inference then it might well be how they do it.

You can come to your own conclusions. The point of the above comment is simple. Either way, Blind probability is not enough. Not even close.

I genuinely don't know what you mean by "blind probability" or "surface-level probability", but in case you don't know, how transformer models work is after tokenization and embeddings are calculated each attention head in the model calculates "self-attention" which is the relationship between all tokens in the current context of the model and passes its results either to another layer of attention heads or to the decoder layer of the model. This typically uses a temperature parameter to control randomness in the output by affecting the probability distribution of tokens in the vocabulary when producing output. I'm not sure whether this counts as "blind" or "surface-level" but that's the probability in GPT models. If you want to learn more, here's Vaswani et al 2017 - the seminal paper https://arxiv.org/pdf/1706.03762.pdf

Why are they exhausting? While you seem well versed in the subject, this was quite enlightening to me, as a non expert.

To use a bad analogy, it's like reading a series of articles about the cognition of a rock. It's obvious to you that rocks don't think. It's a rock. but there are all these articles about what the rock is thinking about, whether it's thinking, whether it had an internal monologue, what it "wants"

In the same way if you understand the implementation of these types of language models you know that they simply cannot have inner monologues in the same way we do...the very concept of an inner monologue implies that there is some sort of state that we the user cannot see. Whether LLMs "think" or "want" or what thinking or wanting means etc can all be debated but whether the models have state that we cannot see is known to us: no. They do not.

Yeah well there could still be a state before generation and in response to input.

Give GPT-4 about 10 or a dozen words from a meaningful sentence scrambled. Now tell it to unscramble those words to make a meaningful sentence. As long as you don't try too many tokens, the rate of meaningful completions is nearly 100%. The chances of that happening via random chance is to put it lightly, extremely small.

The only way for humans to solve this kind of problem is to think about the words before the output or in this case generation.

I'm confused, "random chance" ...?

If you're giving it words from a sentence it is literally the baseline functionality of the text generator to choose the most likely next token, IE to recreate the most likely sentence from the words given.

This couldbe the least meaningfully revealing task I could possibly think of trying if I were trying to find some inner monologue, if gpt had to "think" on anything this would never be it.

The linked comment is a reiteration of the same thought in more words, it still makes the odd assertion against "random chance", you didn't even really give a reason except "it must be so"

Come to whatever conclusion you want lol. The point of the comment is that Blind probability is not enough. That should be clear enough.

The point of the responses to you is no one thinks it's blind probability. Constructing meaningful sentences out of words is literally the primary thing it's designed to do.

Chance simply doesn't enter into the equation to any notable extent.

Surface level probability doesn't work either.

Agreed, you could achieve the same success ratio by computing the stochastic probabilities of a well trained English corpus Markov model given those same words two decades ago.

No you actually couldn't lol. If you genuinely think this then you don't know what you're talking about.

Why would you need chance? It's applying the same parameters/algorithms to the same problem. As far as I understand you have to prompt it to provide more novel answers to the questions and there's some randomness built into the model so it doesn't sound like a broken record. This is is how it's supposed to work. Why do you consider this evidence of state?

The prompt would basically be (but with less words. Around 10 is the best for consistency from my experiments. Much higher than that and it's hit or miss).

Rearrange (if necessary) the following words to form a sensible sentence. Don’t modify the words, or use other words.

The words are: access capabilities doesn’t done exploring general GPT-4 have have in interesting its it’s of public really researchers see since terms the to to what

A successful completion would be.

Since the general public doesn't have access to GPT-4, it's really interesting to see what researchers have done in terms of exploring its capabilities

The number of permutations of the 24 words in the pre-scrambled sentence without taking into consideration duplicate words is 24 * 23 * 22 * ... * 3 * 2 * 1 = ~ 6.2e+23 = ~ 620,000,000,000,000,000,000,000. Taking into account duplicate words involves dividing that number by (2 * 2) = 4. It's possible that there are other permutations of those 24 words that are sensible sentences.

For a language model to consistently make sensible predictions, it quite simply has to be able to "look ahead".

When the probabilities for the candidate tokens for the first generated token were calculated, it seems likely that GPT-4 had calculated an internal representation of the entire sensible sentence, and elevated the probability of the first token based on that internal representation.

I don't know what else to call looking backward, looking forward and then producing output anything other than a state.

If you want to see what this kind of completion would look like without much or any regard for a sensible completion of the sentence then just ask 3.5.

It doesn’t need to look ahead, all of the words are already in the prompt behind the current position, so it just needs to use the existing context to infer the next word

Also on my second try, I gave it a slightly invalid sentence. The prompt was..

    Unscramble the following: waffles eat with like cheese you.

It responded with:

    You like to eat waffles with cheese.

But notice that the prompt doesn’t have the word “to” anywhere in it. But it was the most likely word that was supposed to go there.

Feel free to try and ask the same question with people. A good chunk of them will insert the "to" as well. It's not a gotcha.

This isn't looking ahead. All those words are "behind", since they are part of the prompt. No hidden state is required for this to work.

Isn't the conclusion here just begging the question?

By the nature of number ranges there are infinitely many more wrong answers available to calculators performing addition and yet a simple ripple carry adder still promptly produces the correct output without lookahead, why this is compelling evidence?

Further, if it relied on internal state or lookahead why would it occasionally hallucinate appropriate tokens it wasn't provided?

Arithmetic has a number of always applicable steps. Unscrambling does not unless one of those steps is "think".

If you want to believe looking ahead is unecessary, be my guest lol.

Humans would occasionally "hallucinate" words in too. For whatever reason(and this is verified independently of any of this), LLMs also appear to "get distracted"

In defense of the author, the second line of that article was:

"As I said before, the insights are not particularly original, I’m just raising awareness of issues."

If you are familiar in the subject matter, then that may have been your cue to stop reading.

Raising awareness of what issue? Does he want transformers to have "inner monolog" your comment is even more confusing than the article.

Seems to me that there’s only a very short jump between having a monologue at all, and having an inner monologue. Small children go through a similar stage where they repeat speech back with increasing fidelity, then after several million years of subjective listener time they finally learn to “use their inside voice.”

(Don’t worry, I’m not equating the two… but this pathway does work.)

Can be simulated quite easily though. Just prompt it periodically to generate some text based on previous records, store it in some sort of long-term storage (even a text file), then sample from it and mix with each new prompt.

I wish there was a way to walk back from the AI label in order to avoid sensationalism and pop sci-ification of LLMs or any model, really. Too bad that it's incredible for marketing.

> In the same way if you understand the implementation of these types of language models you know that they simply cannot have inner monologues in the same way we do...the very concept of an inner monologue implies that there is some sort of state that we the user cannot see.

I'm with you on the point you'd like to make, but LLMs obviously do have state that we cannot see. In fact, there are two different levels of obvious state: one allows the LLM to (for example) respond to input instead of ignoring it, which we may characterize as "unlearned" behavior. The level people are more interested in is the "finished model", the set of associations it learns in training. It's state and we can't see it.

The internal state of the model, at those levels, doesn't change when you interact with it. But the state is there.

(There is also a "working state", the state the model must keep in order to output different tokens in sequence, and in order to at some point decide that it's finished responding to a prompt. This does change during your interaction with the model, but the change is not persistent.)

A rock that talks in grammatical English sentences and, while while doing so, follows directions fairly well, such as following instructions that ask it to state its decision process (i.e. after you ask it to do so it modifies its behavior by doing so.) A rock that judges use to consult novel cases that aren't in the case law for the rock's opinion. A rock that passes first year medical school and law school exams. A rock programmers use to teach them things they want to learn. A rock programmers use to write small functions to their slecifications. A rock people use to write letters. A rock that can invent things if you ask it to. A rock that has a neural network for its architecture.

I guess that is some rock!

What made you think your analogy was bad?

Because it's prone to easily being made fun of by people such as yourself who, rather than make a good faith effort to understand what the other party was trying to communicate, choose to intentionally take a ridiculous interpretation because it allows them to lampoon what was written.

I was not trying to say chatgpt is actually like a rock. But you knew that already - you just chose to ignore it in order to get the epic dunk.

It's ultimately a philosophical issue about the nature of consciousness.

Some people think that consciousness is a phenomenon that emerges when we introduce enough complexity to some amount of matter. After all, if we are just atoms in the void, but still feel so different, we have to point to something, and the arrangement of our atoms seems like a fair candidate. So in order to build conscious beings like ourselves, we just need to figure out the correct arrangement.

Other people think that consciousness is something fundamental that does not emerge from complexity. And they will argue against the explanatory power of complexity, usually with other things that are somewhat complex, but don't show signs of consciousness. Will a traffic jam start to feel something if there are enough cars? I don't know. To them this whole endeavor is like alchemy or witch brewery. "If we just try this even more complicated recipe, then maybe the potion will get healing powers or maybe we get gold!". And additionally in computer science, we can just inspect the code, line by line. It's like believing that David Blaine can do actual magic, even when there are videos out there where he himself explains every little trick and misdirection.

It’s not a mystery how they work! The paper has been out for years now. Moreover, all of the research is focused on extending memory through other means since VRAM usage is proportional to the SQUARE in-text memory of the model. So the way forward isn’t a larger context window, but instead an external memory apparatus of some means.

The “memory apparatus” can be simply a text file. Or also a physical simulator to run some mental experiments in.

It can be argued that there's an RNN-like process in the decoder stage but it's re-run every time based on a new context so it's not really "keeping up" in the same way as the human is.

> Exhausting

This is increasingly the feeling with all public conversation no matter what the subject.

The hyperconnected echo chambers of social media, copycat blogs and cheap online journalism are creating a system that only knows two states: silence or delirious excitement. Since silence doesnt pay the bills, what you get is a constant barrage of exaggeration. Its exhausting the same way being exposed to 120db noise is exhausting.

The drowning of calm and informed discourse is not just another negative development that subtracts from our quality of life the way a polluted beach or river does.

You can increasingly see the impact of hypes on decisions of policy bodies that should know better. Ultimately they are serving "the people" and if people agitate about crypto or AGI or whatever, then thats what they too will focus on.

And yet it's quite easy to get ChatGPT to do tasks that involve some sort of thinking ahead, so this can't be the full story. It's pretty clear that nobody quite knows what's going on inside these very large neural networks, hence the surprise at discovering the "meta-gradients" involved in ICL.

A simple example is to ask it to write a complex program in a language like Java or Kotlin that requires you to declare all your imports up front. There can be a lot of these. Therefore the model must figure out what it wants to use very far down the program in order to emit the first words of the program correctly. Presumably somewhere in the intermediate state process the network is actually working out the entire program ahead of time so it can predict the next word of it, and that's sufficiently stable for it to do so successfully.

Also I'm not quite sure I understand what you mean by not maintaining hidden state between tokens. Doesn't the transformer algorithm start by parallel processing every token of the prompt to compute a large set of vectors and matrices that are then reused over and over whilst generating the results? What are the KV vectors if not hidden state of some sort?

When ChatGPT first came out I was asking it to draw pictures of goblins and wizards using Python and Turtle - it's pretty bloody good at it. You can ask it to make the goblin angry or happy or whatever.

There is some sort of magic in which the LLM is imagining what a goblin looks like, then converts those thoughts to turtle commands and then finally adds python. It's quite impressive.

For example, the eyes, head, hair, horns etc are all in the correct place.... how? Using imagination?

I think somewhere down the rabbit hole we lost perspective of what neural nets are doing and how we got to the post-GPT world. Remember deep neural-networks? The idea is that each layer _deeper_ the software encodes more _hidden_ relationships from the training data. That's where this "magic" is embedded. At certain depths the neural-net stores relationships that go beyond our human reasoning and conscious understanding, and the software results generated from these depths are unsettling and tend to startle us. Interesting enough, these are things we are perfectly capable of doing but unable to actually query in our heads as we, just like ChatGPT, do not posses the ability to physically examine our own neural-net content.

It’s important because most users of ChatGPT don’t know what transformers are or how they work, and have no inclination to learn it. The psychology terminology provides useful metaphors.

This should work if you know for certain the transformer model behind GPT is actually working like a traditional transformer, it could be not. We don't know because OpenAI won't let anyone see the internals of the system.

Now, a more mundane hypothesis, if GPT is actually self-consciuous,the most probable answer would be to try to deceive humans from knowing for certain about this fact. You just answer incredible well some questions, some other you seem "confused" or hallucinating

..and the poor meatbag's minds won't know how to go ahead because there is no pattern to follow.