Hacker News Reader

Hacker News

Comment by Toorkit | Hacker News Reader

Full thread

Toorkit

Computers were supposed to be these amazing machines that are super precise. You tell it to do a thing, it does it.

Nowadays, it seems we're happy with computers apparently going RNG mode on everything.

2+2 can now be 5, depending on the AI model in question, the day, and the temperature...

View on HN

This, 100%, is the reason I feel like the sand's shifting under my feet.

We went from trusting computing output to having to second-guess everything. And it's tiring.

I kind of feel like if you're using a "Random text generator based on probability" for something that you need to trust, you're kind of holding this tool wrong.

I wouldn't complain a RNG doesn't return the numbers I want, so why complain you don't get 100% trusted output from a random text generator?

Because people provide that work without acknowledging it was created by a RNG, representing it as their own and implying some of level of assurance that it is actually true.

That's an interesting point of view. For some reason we put so much effort towards making computers think and behave like a human being, while one of the first reasons behind inventing a computer was to avoid human errors.

This is the most succinct summary of what's been gnawing at me ever since LLMs became the latest thing.

If Ilya Sutskever announced tomorrow that he'd achieved AGI, and here is its economic plan for the next 20 years, why would we have any reason to accept it over that of other human experts? It would literally be just another expert trying to tell us how to do things. And we're not short of experts, and an AGI expert has thrown away the credibility of computers as deterministically better calculators than we are.

These amazing machines weren't consistently able to tell if an image had a bird in it or not up until like 8 years ago. If you use AI as a calculator where you need it to be precise, that's on you.

I think the issue is that: I’m not going to be using as a calculator any time soon.

Unfortunately, there’s a lot of people out there, working on a lot of products, some of which I need to use, or will be exposed to, and some of them aren’t going to have the same qualms about “language model thinks 2+2=5”.

There’s a guy on Twitter scoring how well ChatGPT models can do multiplication.

A founder at a previous workplace wanted to wholesale dump data into ChatGPT and “make it do causal analysis!!!” (Only slightly paraphrased). These tools enable some frighteningly large-scale weaponised stupidity.

The problem is that’s what people do. And everyone else has to pay for it.

I think about it differently. Before computers had to be given extremely precise and completely unambiguous instructions, now they can handle some ambiguity as well. You still have the precise output if you want it, it didn’t go away.

Btw I’m also tired of AI, but this is one thing that’s not so bad

Edit: before someone mentions fuzzy logic, I’m not talking about the input of a function being fuzzy, I’m talking about the instructions themselves, the function is fuzzy.

> You still have the precise output if you want it, it didn’t go away.

For now. Given that most new devices seem to be fully hostile to the concept of general purpose computing (see phones, VR devices, TVs, etc), I wonder how long it will be before many of the computers that are sold are even more locked down than Chromebooks - just a few prompts for interacting with a preinstalled LLM.

Its a Large LANGUAGE Model and not a Large MATHEMATICS Model. People need to learn to use the right tools for the right jobs. Also LLMs can be made more deterministic by controlling it’s “temperature”.

Yep. ChatGPT will use the code interpreter for questions like is 2 + 2 = 5? as it should.

There's other forms of AI than LLM's and to be honest I thought the 2+2=5 was obviously an analogy.

Yet 2 comments have immediately jumped on it.

Hackernews comments and getting bogged down on minutiae and missing the overall point, is there a more iconic pairing?

But isn't this a great thing? I mean, this piece has been missing (no, I am not kidding). Computers have always had a hard time coping with situations that weren't 100% predefined.

Now, we have technology capable of handling cases that were not predefined. Yes, it makes mistakes, as do humans, but the range of problems we can solve with technology has been tremendously broadened.

The problem is how we apply AI. Currently, we throw LLMs on everything they might be able to handle without understanding how or if they have the capabilities to handle such a task. And that is not the LLM's fault but a human fault. Consequently, we see poor results, and then we blame the AI for not being able to solve a problem it wasn't designed to solve.

Sounds stupid, doesn't it?

Machines were not able to deal with non-formal problems.

There are areas it doesn't have to be as "precise", like image generation or editing which I believe better suited for AI tools

And by nowadays you mean since ChatGPT got released, that is less than 2 years ago (e.g. a consumer preview of a frontier research project). Interesting.

Had to laugh at this one. I think we prefer the statistical approach because it’s easier, for us …

This sounds to me like a straw man argument. Obviously 2+2 will always give you 4, in any modern LLM, and even just in the Chrome address bar.

Can you offer a real situation where we should expect the LLM to return a deterministic answer and should rightly be concerned that we're getting a stochastic one?

Y'all are hyper focusing on this example. How about something more vague like FOO obviously being BAR, except sometimes it's BAZ now?

The layman doesn't know the distinction, so they accept this as fact.

I'm not being facetious; I really can't think of a single good example where we need something to be deterministic and then have a reason to be disappointed about AI giving us a stochastic response.

Reply on news.ycombinator.com