Hacker News Reader

Comment by ninkendo | Hacker News Reader

When typing this comment to you right now, my brain is coming up with the next word for me to type, just in time, as I'm typing. You could be reductive and call my brain a simple next token prediction engine too. Mozart was just a next note prediction engine when making his compositions. Einstein was just a next word prediction engine when forming his theory of relativity, etc.

The point I'm making is not that GPT is as intelligent as a brain, but rather that the ability to produce the next token in a way that forms arbitrarily intelligent responses to arbitrary questions, may well require something resembling real, actual intelligence.

View on HN

You also decided that a response was needed and had planned an argument of sorts.

You decide which comments to do this for as well.

How does GPT decide such things?

If GPT4 were given this whole thread as an input, it might reasonably be expected to predict a message that is a reply to a particular comment in the thread and presents a relevant and coherent argument based on the priors and "personality" encoded somewhere in its huge parameter space.

Sure, LLMs are not inherently goal-oriented. But that doesn't mean they fundamentally cannot exhibit goal-oriented behavior if either prompted or trained to do so. And if we remove the "one-off" restriction and allow the retention of some sort of an efficiently encoded state between invocations, who knows what might happen?

----

EDIT: Tried this with ChatGPT (GPT-4) and a subset of this thread, and the results are pretty good, even though the tree structure was lost in translation (simple copypaste). Certainly its thoughts of the differences and similarities between token-prediction and reasoning are more informed than those of many human commenters. It's also able to seamlessly include references to other comments, for instance (it deduced that "dougwmne" was the nickname of the comment author by itself):

> Additionally, it's worth noting that GPT-4's ability to interact with novel systems, as mentioned by dougmwne, might indicate some form of generalization and adaptation capabilities. Although it may not possess the same kind of reasoning as humans, it demonstrates an ability to understand and apply new information in a useful manner. This could be considered a form of reasoning, albeit a different and less sophisticated one than humans possess.

My prompt was:

> Here is an excerpt from a Hacker News discussion about GPT-4 and whether it can do reasoning. Write a comment that contributes to the discussion, either by bringing up a new idea or deduction, or by replying to another comment and arguing for or against it. You can quote another comment by prefixing it with ">".

followed by the copy-pasted subthread.

Also, we're all here because there is some kind of human intelligence behind everything. We are writing here because there's a small chance that a single comment written will affect all of humanity. (just one pathway for example is that someone popular will read your comment and tweet it out later because it was insightful).

GPT has no "reason" behind its words. Its actions are not tied to the general betterment of humankind. When we all sit here with our coffees and donuts (and farting) we're generally trying to push others to see our worldview so that maybe, just maybe, the world can change a little bit.

If we eventually have a way to tie an AI to all these moods (hunger, pain, tiredness) and all the other things of the human experience, it would ground the thoughts and text it generates at some point such that it would want to DO things it wasn't doing before. It also can't right now anyway, unless we give it that ability (ReAct pattern could).

We can, of course, simulate all these human experiences, with "best practices" but then the reason won't be really tied to human reasoning. It will be just more text autocomplete with additional instructions.

I decided such things when given an input of the comment to which I replied. If GPT were given such a comment, it could have reasonably come up with a similar answer (it probably wouldn't look the same, as GPT isn't very intelligent, but calling it "just" a next token predictor is a gross mischaracterization.)

Which comments should GPT respond to on HN?

All of them? If not, how to determine which ones?

There is a difference here between you and the machine

Looking at the LLM algorithms, they do appear to be next token predictors. Why do you ascribe something more to them when the algorithm is known?

Given we know that causal patterns are in the training data, and that LLMs are capable of embedding algebra, it is understandable that they appear to "reason", while we as humans can also describe the algorithm and mechanics actually at work here.

If your only point is that “the modality of ChatGPT is that it responds to prompts, therefore it’s a next token predictor”, then sure. You’re not wrong, but it’s an extremely oversimplified explanation of what GPT is. It’s basically the most boring argument you could make about GPT.

Yes, GPT is code that only runs when you, erm… run it. It’s not an autonomous being that’s cut loose on the internet and given free reign to use as much compute resources as it wants, a la skynet, replying to random forum posts. But nobody’s really saying that it is, so it seems a strange argument to say “it’s just a next token predictor”, if by that you really mean “it only runs when you give it a prompt.”

The modality of GPT is that it responds to prompts, but the interesting question is “are the responses given a sign of intelligence?”. Saying “no, it’s just a next token predictor” is IMO a sign of a motte-and-bailey fallacy. Your motte is that it only works in response to prompts (trivially true and uninteresting), and your bailey is that it has no intelligence (very debatable and IMO false.)

> All of them? If not, how to determine which ones?

How does it determine how to respond to any prompt? Exactly. It might reasonably predict responses to those comments that, according to its black box, are somehow most likely to receive a reply. Different GPT instances prompted or trained to have different personalities would choose differently based on their priors and biases.

I'm pretty sure the vast majority of human beings are able to reason without really having the slightest idea as to how exactly they do it. Educated humans can use language to talk about reasoning in terms of logic, and even make logic engines out of doped silicon and whatnot, but as of now, none of us has any idea of how reasoning is implemented in the mass of fat and protein that's inside our heads.

I think you missed the point of the parent comment. My experience with ChatGPT is that it responds to each and every thing sent to it. It doesn't seem to have the ability to not respond.

The original question was, given that there were X comments which could be replied to on this page, how and why did the poster decide to respond to the one to which they responded, rather than either "all" (like ChatGPT would) or "none?"

If ChatGPT is given an entire message (sub)thread as input, it will decide which parts, if any, it responds to. Sure, that's a bit different from how humans tend to read and reply to these threads.

If you mean feeding it the messages one at a time, then, sure, it will respond to each of them because that's what its default "goal function" is. (I assume that we're talking about a single "conversation" here so the context is retained – feeding it each comment out-of-context would be very unfair.)

If you prompt it to only reply to messages it considers "interesting" in some way, and return "I'll skip this" for the others, it would certainly do that.

Another interesting aspect to consider is that GPT needs priming & prompting to get it to generate a relevant response. I doubt blindly providing comment text would create valuable responses.

Yet another case, many times we will respond (or not) to a questions with another question, to elicit deeper considerations. In my experience, GPTs will answer any question without thought, pun intended :]

ChatGPT in particular is trained to predict dialogue and helpful answers to questions, so it would naturally expect some instructions on what exactly it should do. But for example vanilla GPT-3 has no such bias and merely tries to answer the question "what is the most likely completion to this input based on my corpus?" so it might well simply start generating plausible comments that extend the thread.

I've had it add the question mark as the first token in the response when I forget it, or intentionally leave it off

If GPT were to comment, it would comment to everyone on this board, instantly and blandly and there would be a lot of duplication. It would seem like a farce and no one would comment on HN anymore. It would be stupid and "filled with bots".

Well, none of the GPT models are self-willed, so it would be a matter of how the GPT powered bot was written that would determine that. I mean, sure, if you pasted every comment here intonthe ChatGPT UI and post the response back as a comment, that would hapoen, but the UI is specifically designed for people who want a response every time.

> When typing this comment to you right now, my brain is coming up with the next word for me to type, just in time, as I'm typing.

So you saw the parent comment about hype and clicked into this comment box without a thought of what you might say other than "When?" Really?

Did you know whether you agreed or disagreed with the parent comment, or did it take until the word "reductive" for you to firm that up?

If you're saying that the part of your brain that controls your fingers is doing so one word at a time, then sure, obviously that's true. But if you're making the opposite claim, that LLMs like Chat-GPT are "deciding" what to say and only then working out the finer details of how to express that decision, that's a claim I'd like to see explicitly.

If you are suggesting that your approach to intelligence is parallel to that of an LLM, then either you're arguing for something far more radical than most as an emergent property of LLMs, or you're suggesting that you only form opinions one word at a time as you type, which I think would make you, well, I won't insult you, but I think you're probably doing a little more than that.

Reply on news.ycombinator.com