[kazinator]

AI "thinks" the way a horse can count. The horse's interlocutor knows that the correct number is eight and shouts "right!" when the animal has stomped its hoof that many times, cuing it to stop.

When you know the right answer, you're not actually using the AI to produce the answer; you're just prodding it to generate bullshit after bullshit until it hits upon that answer, or a very good facsimile. The one actually thinking is you!

This is of no use to someone who has no idea what the answer is, and actually needs it.

[travisjungroth]

This is just silly. You’re saying that these models are completely incapable of what they’re doing and are only getting to answers from cheating. You can see this isn’t true very quickly when using them.

[Me] I want to make a bouquet to honor the home country of the first person to isolate Molybdenum. Be brief.

[ChatGPT-4] To honor Peter Jacob Hjelm, the Swedish chemist who first isolated Molybdenum in 1781, create a bouquet using blue and yellow flowers, representing the colors of the Swedish flag. Some suggestions are blue irises, yellow tulips, and blue delphiniums, arranged with green foliage for contrast.

I didn’t know the answer to that before asking. I also checked that it’s right. Soemtimes it’s right, sometimes it’s wrong. But if I walk up to a horse and ask it to stamp eight times, that’s going to happen approximately never. If I get it stamping at all, then maybe 1/20. This is much better than that.

[EugeneOZ]

Because you are using results of the training.

[mandmandam]

What human wasn't trained?

Even a child raised by wolves gets training.

[kazinator]

A human isn't trained to program in Python by memorizing impressions of millions of examples correlated to a verbal description, from which he or she regurgitates mash-up that statistically correlates to the prompt, without having any idea what the content means.

Ask it to write the same program in some little known programming language. A human who knows how to code could do that by studying a small amount of reference material.

[pmoriarty]

"Ask it to write the same program in some little known programming language. A human who knows how to code could do that by studying a small amount of reference material."

That depends on the language, the human, and their state of mind.

[travisjungroth]

GPT-4 can do that.

[EugeneOZ]

LLM doesn't pick some results because LLM is smart, it picks results because it's trained. So analogy with a horse is valid.

[travisjungroth]

You don’t get the horse analogy. The horse isn’t trained to count to eight. The horse is trained to stomp continuously. The performer yells “stop” at eight. It’s a carnival trick and a form of cherry-picking. The commenter was saying that ChatGPT doesn’t have the ability to solve these problems at all and people are just choosing the randomly correct answers. I think this is obviously not true.

[kazinator]

The purpose of the horse analogy wasn't so much to equate ChatGPT to a horse that just stomps, but the behavior of the humans around it who believe that it can count. This can be the naive trainer too, who doesn't recognize that he's subconsciously cuing the horse. He has a confirmation bias whereby he or she rejects any evidence refuting the hypothesis that the horse cannot count.

There is a difference between solving and problem and presenting a solution, and that difference doesn't hinge on whether the solution is correct.

[kazinator]

Writing programs requires an education. You can sort of fake it with training, if you're a machine that can train on a billion examples and reliably retain something from each one, without understanding any of them.

[fl7305]

So your claim is that GPT-4 is a stochastic parrot?

Meanwhile, the creators at OpenAI claims that it is not.

How do we test your claim? What are the logical reasoning questions we can ask GPT-4 that it shouldn't be able to answer if it's just a stochastic parrot?

I keep asking people this, but so far no one has answered. Is the claim a personal belief that you don't want to discuss, similar to your religious beliefs?

If so, fine, but please make it clear that you're not making a scientific claim.

[computerex]

This is selling it short. Are you saying you can't do useful things with these models?

I have a python script that's maybe 200 lines long, calling out to OpenAI using gpt3.5 turbo. I can give it commands like "search for the news today, summarize then save into news.txt" -> and it emits the DSL I asked it to use in the 1 shot prompt.

You can give it complex word problems and it'll write python programs to solve them. You can ask it to get top hacker news stories, and it'll find creative ways to fetch them.

[travisjungroth]

Can you share the script? If you do, no need for the standard “I need to clean it up” and so on apologies.

[computerex]

Here you go: https://gist.github.com/computerex/c3394a88098895d8731501700...

[travisjungroth]

This is awesome, thank you. I’ve been building single purpose stuff and thinking about tying it together. Seeing someone’s one-off (not a framework) take on that is super helpful!

[nuancebydefault]

Why don't you give gpt4 a chance for an hour or so and try it out with handing it a handfull of problems. I swear, you will not be disappointed.

[travisjungroth]

I’ve spent a bunch of hours!

[EugeneOZ]

It's the difference between AI and LLM.