I kid you not, early on (mid 2023) some guy mentioned using ChatGPT for his work and not even checking the output (he was in some sort of non-techie field that was still in the wheelhouse of text generation). I expresssed that LLMs can include some glaring mistakes and he said he fixed it by always including in his prompt “Do not hallucinate content and verify all data is actually correct.”.
Ah, well then, if he tells the bot to not hallucinate and validate output there’s no reason to not trust the output. After all, you told the bot not to, and we all know that self regulation works without issue all of the time.
The people who make them don’t really understand how they work either. They know how to train them and how the software works, but they don’t really know how it comes up with the answers it comes up with. They just do a ron of trial and error. Correlation is all they really have. Which of course is how a lot of medical science works too. So they have good company.
They can even get math wrong. Which surprised me. Had to tell it the answer is wrong for them to recalculate and then get the correct answer. It was simple percentages of a list of numbers I had asked.
Language models are unsuitable for math problems broadly speaking. We already have good technology solutions for that category of problems. Luckily, you can combine the two - prompt the model to write a program that solves your math problem, then execute it. You’re likely to see a lot more success using this approach.
Also, generally the best interfaces for LLM will combine non-LLM facilities transparently. The LLM might be able to translate the prose to the format the math engine desires and then an intermediate layer recognizes a tag to submit an excerpt to a math engine and substitute the chunk with output from the math engine.
Even for servicing a request to generate an image, the text generation model runs independent of the image generation, and the intermediate layer combines them. Which can cause fun disconnects like the guy asking for a full glass of wine. The text generation half is completely oblivious to the image generation half. So it responds playing the role of a graphic artist dutifully doing the work without ever ‘seeing’ the image, but it assumes the image is good because that’s consistent with training output, but then the user corrects it and it goes about admitting that the picture (that it never ‘looked’ at) was wrong and retrying the image generator with the additional context, to produce a similarly botched picture.
I once gave some kind of math problem (how to break down a certain amount of money into bills) and the llm wrote a python script for it, ran it and thus gave me the correct answer. Kind of clever really.
It’s easy, just ask the AI “are you sure”? Until it stops changing it’s answer.
But seriously, LLMs are just advanced autocomplete.
I kid you not, early on (mid 2023) some guy mentioned using ChatGPT for his work and not even checking the output (he was in some sort of non-techie field that was still in the wheelhouse of text generation). I expresssed that LLMs can include some glaring mistakes and he said he fixed it by always including in his prompt “Do not hallucinate content and verify all data is actually correct.”.
Ah, well then, if he tells the bot to not hallucinate and validate output there’s no reason to not trust the output. After all, you told the bot not to, and we all know that self regulation works without issue all of the time.
It gave me flashbacks when the Replit guy complained that the LLM deleted his data despite being told in all caps not to multiple times.
People really really don’t understand how these things work…
The people who make them don’t really understand how they work either. They know how to train them and how the software works, but they don’t really know how it comes up with the answers it comes up with. They just do a ron of trial and error. Correlation is all they really have. Which of course is how a lot of medical science works too. So they have good company.
Ah, the monte-carlo approach to truth.
They can even get math wrong. Which surprised me. Had to tell it the answer is wrong for them to recalculate and then get the correct answer. It was simple percentages of a list of numbers I had asked.
Fun thing, when it gets the answer right, tell it is was wrong and then see it apologize and “correct” itself to give the wrong answer.
In my experience it can, but it has been pretty uncommon. But I also don’t usually ask questions with only one answer.
Language models are unsuitable for math problems broadly speaking. We already have good technology solutions for that category of problems. Luckily, you can combine the two - prompt the model to write a program that solves your math problem, then execute it. You’re likely to see a lot more success using this approach.
Also, generally the best interfaces for LLM will combine non-LLM facilities transparently. The LLM might be able to translate the prose to the format the math engine desires and then an intermediate layer recognizes a tag to submit an excerpt to a math engine and substitute the chunk with output from the math engine.
Even for servicing a request to generate an image, the text generation model runs independent of the image generation, and the intermediate layer combines them. Which can cause fun disconnects like the guy asking for a full glass of wine. The text generation half is completely oblivious to the image generation half. So it responds playing the role of a graphic artist dutifully doing the work without ever ‘seeing’ the image, but it assumes the image is good because that’s consistent with training output, but then the user corrects it and it goes about admitting that the picture (that it never ‘looked’ at) was wrong and retrying the image generator with the additional context, to produce a similarly botched picture.
I once gave some kind of math problem (how to break down a certain amount of money into bills) and the llm wrote a python script for it, ran it and thus gave me the correct answer. Kind of clever really.