Hacker News Reader

Comment by shzhdbi09gv8ioi | Hacker News Reader

> The danger of "AI" is that we actually believe the plausible fabrications it produces are "intelligent".

100% this!

I'm fed up with seemingly rational people who just can't comprehend that the AI hype is just sales talk.

I had to convince a customer the other day we cannot write whole apps with ChatGPT. As far as I know, not a single example exists of a full app written by ChatGPT. It just can't be done, because ChatGPT is not reasoning!!! It is not intelligent.

It can just spit out seemingly coherent, but often incorrect, snippets of code.

I think it's best described as a word-calculator, or autocomplete on crack, in the sense that it is great on guessing what comes next. But it has no reasoning behind it's predictions, only statistics.

EDIT: The AI apologetics are downvoting me LOL

View on HN

Are we using the same ChatGPT?

It cannot write a full app, but that is a matter of context size.

It absolutely can "reason" within bounds, attention is a much more powerful system than you realize.

Chain-of-Thought is a well-established technique for LLM prompting which can produce great results. And the snippets of code GPT-4 generates for me are usually on point, at least they were until OpenAI dumbed the model down in the last few months with GPT-4-Turbo.

I get it not only to write novel code, but explain existing code all the time and it reasons about my code very well.

Ahh.. I didn't pay for the premium version, that is why I am not convinced. /s

I don't know what to tell you if you think you're entitled to a completely free cutting-edge service that takes multiple dedicated GPUs to serve your requests.

>autocomplete on crack

I have had some really frustrating conversations about this very topic with people insisting on using it for complex tasks like designing business strategies. It doesn't know how to actually answer your question, it's just spitting out probable words associated with the topic in a plausible sequence! It's like Quora with spellcheck and an aura of legitimacy which makes it infinitely more seductive.

I’ve actually used a mates replit chained gpt-4 app to build short single function apps from a single command that function on first run, so it’s definitely possible, and will only get better - but building code is just following a logical set of instructions down a pathway

"Single function app".. how is that different from a snippet?

Show me a proper app generated by ChatGPT.

You can't, because ChatGPT can't.

It will only get better, based on what? A true believer just knows.

Reminds me of the latest excuse from OpenAI: https://news.yahoo.com/ai-needs-nuclear-power-breakthrough-1...

(its not the tech that are wrong, we just don't have enough energy in the world).

Sam Altman is also a non-trivial investor in various nuclear energy R&D companies, e.g. $375 million into Helion, so take that "need" with a pinch of salt. He made a bet, presumably because he already believes, but it's human nature to try to justify your decisions after the fact and not just before.

> I'm fed up with seemingly rational people who just can't comprehend that the AI hype is just sales talk.

The reason why you see rational people believe the AI hype is that AI is a lot better than you give it credit for. They are right.

> I had to convince a customer the other day we cannot write whole apps with ChatGPT. As far as I know, not a single example exists of a full app written by ChatGPT. It just can't be done

It can. When it first came out, I wanted a macOS menu bar app so that I could access ChatGPT conversations from the menu bar. I’d never written a macOS app before. I told it what I wanted, it wrote the whole thing in one go. There was one minor compile error (a type signature had changed from a previous version, if I remember correctly), which was a one-line fix. I iterated a couple of times, telling ChatGPT what improvements I wanted to make to the app. It did them.

Would I use it to build a complex app? No. But it is capable of building a whole app.

> The reason why you see rational people believe the AI hype is that AI is a lot better than you give it credit for. They are right.

I am sorry, but you are mistaken. There are no AI. The "AI" that is being mongered today is just a salespeople buzz word for Large Language Model technology, which is just that, a language model. It can be used to generate seemingly coherent text, but it can not reason.

> Would I use it to build a complex app? No. But it is capable of building a whole app.

It only works for you as you accept lowering the bar of that would pass as an "app".

I mean of course any kind of normal app. Say Microsoft Paint for example.

Show me ChatGPT able to create a full app like that based on business logic you prompted.

> It only works for you as you accept lowering the bar of that would pass as an "app".

I asked it to create an app, it created the app I wanted. That’s not “lowering the bar”. You’re just redefining words to prop up your incorrect claims.

Okay. Enjoy your AI infused career where you can knock out hello world apps in just a few minutes.

Prove it.

> it can not reason.

As someone with an A-level in philosophy: What is reason?

> I mean of course any kind of normal app. Say Microsoft Paint for example.

The original, or the current one? Ironically, I think either would be a case where ChatGPT works best. The features in a raster graphics tool are very well isolated from each other.

I'd expect ChatGPT to face-plant much more with any meta-level stuff, so a database that doesn't have a 1-1 correspondence with whatever appears in a UI for example.

Or physics sims where you've got several layers of indirection that all need to be correct. I've never been able to get a complete working Navier–Stokes simulator out of it, for example.

> As someone with an A-level in philosophy: What is reason?

Maybe you are better off answering that question then. But I would define ability to reason as the ability to make logic conclusions based on input from a feedback loop.

Let's use programming as an example.

You write some code you think will work. You run it. It outputs the wrong values. You reason about the code, realize your mistake. Fix the code. Run again. It outputs the correct values.

ChatGPT cannot do any of this. All it can do, is spit out text that it deems probable to belong to previous text it already generated. This is based on probability, but it completely lacks a reasoning part and a feedback loop.

> The original, or the current one? Ironically, I think either would be a case where ChatGPT works best.

Show me a prompt that produces either one.

> ChatGPT cannot do any of this. All it can do, is spit out text that it deems probable to belong to previous text it already generated. This is based on probability, but it completely lacks a reasoning part and a feedback loop.

50% of the time, if you give it source code and an error, it can fix the source code and tell you why. The other half I still get paid for.

> Show me a prompt that produces either one.

K, if I remember by the time I've done the higher priority things, I'll send you a repo with as much of it as I can be arsed to ask for to prove the point, along with the share link to the chat session.

> Show me a prompt that produces either one.

Note how you only need to ask for each feature and in it goes. Sometimes there's a bug or three, but you can generally fix them by describing them.

https://chat.openai.com/share/e2a6b513-c088-449d-8d8f-4f4c73...

I won't ask for the rest of the features, because it's boring.

What I bring to the market right now, might be merely knowing what to ask.

Oh yeah, I said repo as well. Note the commits are copy pasting blindly. I don't do JS professionally, I'm an iOS dev: https://github.com/BenWheatley/JSPaint

You must be kidding right?

Your chat demonstrates most of what I brought up:

* it only spits out snippets

* it cannot produce an app

* it is hallucinating

* it is creating more work than just doing the work yourself

And its a web app, but that's on you for requesting that for some reason.

I asked for a MS Paint app. That would mean something that produce a Windows compatible EXE, not a some html fest.

> I won't ask for the rest of the features, because it's boring.

Or, because it's infeasible and you know it.

What is reason? No matter.

and what is matter? Never mind.

Doesn't need answering. All we need to know is that reason requires logic. LLMs do not employ logic, thus they are not able to reason. They sometimes act as if reasonable by happenstance, but fundamentally they are not.

Transformer models are universal, which isn't very surprising given they're made of things which are also universal when configured right, so they absolutely can learn the rules of logic.

Some say that reason requires consciousness, which I'm… frustrated by given the 50 common meanings of the word "consciousness"; but to merely use logic as your standard here?

Why, of all the things people could object to in an LLM, why is logic what people want to pick on? It's the weakest possible objection IMO.

The set of models we're discussing haven't.

You were derailing with "reason" for no reason, so I pointed it out. That doesn't mean logic should be applied as some sort of universal standard.

> The set of models we're discussing haven't.

Haven't what?

> You were derailing with "reason" for no reason

I was quoting someone else who said LLMs can't reason, and I'm asking them what they meant by that because ChatGPT sure acts like it reasons no matter what's going on "inside". I assume the inside to be a Transformer model because otherwise the naming is weird, but whatever it is, it acts like it learned to reason.

And I'm saying that despite wanting this to be a repeat of Clever Hans so I can go back to feeling optimistic about my economic future.

Why didn't the customer write the whole app themselves with chatgpt?

They tried to show me what they meant.

ChatGPT proceeded to spit out a snippet of rust code, starting with a dependecy outdated by 4 years, proceeeded to mix API from several versions of the dependency and completely ignored most of the request.

Customer do not know rust so it looks amazing to them.

ChatGPT is a great con-man for the gullible.

> ChatGPT proceeded to spit out a snippet of rust code, starting with a dependecy outdated by 4 years, proceeeded to mix API from several versions of the dependency and completely ignored most of the request.

If that's what you meant, I think everyone here, including the people disagreeing with examples (e.g. me), would have agreed.

What I mean with my examples is: it can do all the bits. Is it perfect? Nope. Do you still need a human in the loop to understand the failure modes? Yup. That MS paint app you asked for 45 minutes ago? Features are decomposable, so it can do this step by step, so long as you're OK adding one feature at a time instead of trying to do it all in one go — 75% of the results are fully functional and do what you asked for (which isn't always what you wanted, "tool selector" can be the CSS or the widget itself).

Some people used to write crap code stitching together snippets from stack overflow, now the same people write crap code stitching together what ChatGPT spit out after trained on said stackoverflow clusterfuck of buggy or flat out incorrect solutions. You do you.

And those people are now your competitors and customers.

Worse (for all of us in this industry) sometimes there's a bug or three, but it can generally fix those bugs by you giving the source code (if the source didn't already come from the LLM) and simply describing the issue. Even when the bug is graphical in nature, which is pretty wild for a text based chat engine that can't see (like 3.5 is).

And that's without applying my experience of any language or frameworks, e.g. when asking it for solutions in languages I don't have meaningful professional experience in, like JavaScript.

When I can apply my experience, I cringe at the choices it comes up with, and find I want to use it as a rough guide rather than a complete answer. But you know who doesn't care about cringy code? Everyone who isn't a developer. Even QA only cares if it passes the tests.

> And those people are now your competitors and customers.

For now, yes.

> When I can apply my experience, I cringe at the choices it comes up with, and find I want to use it as a rough guide rather than a complete answer. But you know who doesn't care about cringy code? Everyone who isn't a developer. Even QA only cares if it passes the tests.

This is a race to the bottom, surely you must understand that?

I expect some of the biggest tech companies to start limiting developer use of LLM:s soon due to the monumental technical debt they are creating.

More context: https://visualstudiomagazine.com/articles/2024/01/25/copilot...

What matters is if the tool is useful or not.

You cannot claim that this tool is not useful to people and use cases you are entirely unfamiliar with.

Debates of what is real intelligence are akin to running around in circles.

"""That's one of those irregular verbs, isn't it? I give confidential security briefings. You leak. He has been charged under section 2a of the Official Secrets Act.""" - Yes Minister.

To rephrase for the subject: We're only human, mistakes are to be expected; They are a idiots, mistakes are to be expected; that thing is just a glorified calculator, mistakes are to be expected.

(There's a short story I found I couldn't bring myself to finish, “Zero for Conduct” by Greg Egan, where the lead character is bullied by someone who has a similar disregard for her intelligence; I know one cannot use fiction to learn about reality, so I will instead say that this disregard of human intelligence by other humans happens a lot in real life too, the racism and xenophobia of βαρβαρίζω can still be found today in all the people who insist that ancient structures like the pyramids couldn't possibly have been built by the locals and therefore it must have been aliens).

> AI hype is just sales talk

But where does the hype end and the reality begin?

> I had to convince a customer the other day we cannot write whole apps with ChatGPT. As far as I know, not a single example exists of a full app written by ChatGPT. It just can't be done, because ChatGPT is not reasoning!!! It is not intelligent.

I will agree ChatGPT does indeed make incoherent solutions — one test project was making a game in JS, it (eventually) gave me a vector class with methods for multiply(scalar) etc., but then tried to use mul(scalar).

But ironically, I've also made a functioning (basic, but functioning) ChatGPT API interface… by bolting together the output of ChatGPT. I won't claim it's amazing or anything, because I'm an iOS developer and the thing "I" made is a web app, but it works well enough for my needs (just don't paste HTML into the query section, because I stopped adding to it when it was good enough for my needs and therefore only have a very basic solution to code being shown in the chat list, there's a lot of stuff that would be improved by using simple libraries but I didn't want to).

> I think it's best described as a word-calculator, or autocomplete on crack.

And that's the same error in the opposite direction.

If I understand right, GPT-3 is literally the complexity of a mid-sized rodent. Thus the metaphor I use is:

Imagine a ferret was generically modified to be immortal, had every sensory input removed except smell, which was wired up to a computer. The ferret and computer then spend 50,000 years going through 10% of all the text on the internet, where every sequence is tokenised and those tokens are turned into a pattern of olfactory nerves to stimulate, and the ferret is rewarded or punished based on how well it imagined the next token.

You're annoyed that this specific ferret's jokes are derivative, their code doesn't always compile, that they make mistakes when trying to solve algebraic problems, that their pecan pie recipe needs work, and that they make mistakes when translating Latin into Hindi.

I'm amazed the ferret can do any of these things.

> If I understand right, GPT-3 is literally the complexity of a mid-sized rodent.

You are not understanding things right, then.

Intelligence is comprised of multiple complex systems. ChatGPT only ever claimed to focus on the language part. It does not contain reasoning.

Even a rodent can reason.

Much less complex organisms can reason, too.

> You are not understanding things right, then.

GPT-3 has about the same number of free parameters as the number of synapses in a mid-sized rodent brain.

> Intelligence is comprised of multiple complex systems. ChatGPT only ever claimed to focus on the language part. It does not contain reasoning.

It (ChatGPT) is doing abstract symbol manipulation, unless you want to expand the idea of language to include chess positions (even if mediocrely), algebra, and the application of the rules of formal logic to the symbols used to represent those rules when it is so prompted.

> Even a rodent can reason.

> Much less complex organisms can reason, too.

For which values of the meaning of the word "reason"? From your other comment "But I would define ability to reason as the ability to make logic conclusions based on input from a feedback loop."

1. This is a terrible waste of computer power, in the order of multiple-trillions-to-one ratio, given logic is the underlying thing used to represent the numbers and floating point operations that approximate linear algebra that in turn is used to build LLMs.

(A similar ratio is also found in human "logical" cognition, and is why we don't use logic all the time, hence the classic example of "a baseball bat & ball that cost $1.10 together, the bat is $1 more expensive than the ball, how much does a bat cost?" which so many get wrong).

2. LLMs can do that anyway. Even literal flow charts are still doing that anyway, and LLMs can build flow charts.

It still makes mistakes, sure, yup, but then the question is "how many compared to a human?" and that's terrifyingly close to humans and I'm just hoping it will remain on the same side, slightly worse — I mean, your own example is humans being fooled by it, so clearly you also know humans can be wildly wrong.

Reply on news.ycombinator.com