Researchers found that ChatGPT’s performance varied significantly over time, showing “wild fluctuations” in its ability to solve math problems, answer questions, generate code, and do visual reasoning between March and June 2022. In particular, ChatGPT’s accuracy in solving math problems dropped drastically from over 97% in March to just 2.4% in June for one test. ChatGPT also stopped explaining its reasoning for answers and responses over time, making it less transparent. While ChatGPT became “safer” by avoiding engaging with sensitive questions, researchers note that providing less rationale limits understanding of how the AI works. The study highlights the need to continuously monitor large language models to catch performance drifts over time.
My understanding is this claim is basically entirely false. The tests done by these researchers had some glaring errors that when corrected, show gpt-4 is getting slightly better at math, if anything. See this video that describes some of the issues: https://youtu.be/YSokS2ivf7U
TL;DR The researchers gave new GPT questions from two different pools. It’s no surprise they got worse answers.
You shouldn’t need to be a prompt engineer just to get answers to math questions that are not blatantly wrong. I believe the prompts are included in the paper so that you don’t have to guess if they were badly formatted.
The problem is they aren’t comparing apples to apples. They asked each version of GPT a different pool of questions. (Edited my post to make this clear).
Once you ask them the same questions, it becomes clear that ChatGPT isn’t getting worse at math, because it has been terrible all along.
I see. Thanks for clarifying
For me it’s like using a coffee machine as a stopwatch, and then complaining that it doesn’t always give the exact time lapsed.
This is the best comparison I have ever read my eyes just peaked reading that thank you very much!
If it’s a coffee machine that’s so advanced it was uninaginable a decade ago, you’d expect it not to perform worse over time.
My point was that a coffee machine is designed to make coffee, not to keep track of time. Maybe it always takes roughly the same amount of time to make a coffee, and so someone uses it as a proxy stopwatch. But it can very well suddenly take more or less time, without anything being wrong about it – maybe different coffee brands, cleaned pipes, or whatnot.
ChatGPT is an algorithm designed to parrot language, not to perform mathematical reasoning based on logic rules.
You are not wrong, but I think public perception is different. It doesn’t help, that OpenAI is pushing their models as problem solvers:
GPT-4 can solve difficult problems with greater accuracy, thanks to its broader general knowledge and problem solving abilities. (https://openai.com/gpt-4)
I didn’t know they made such claims. They’re borderline dangerous claims…
ChatGPT is an algorithm designed to parrot language, not to perform mathematical reasoning based on logic rules.
Mathematical language is a language, ChatGPT has been shown to come up with relationship between very distant elements of language that were not present in the training data… so there is nothing stopping it from, given enough mathematical training data, coming up with whatever relationships we call “logical rules”.
Mathematical language is a language, but mathematics is not just a language. It is a structure with internal rules that are not determined by pure convention (as natural languages are). We could internationally agree from tomorrow to call “blue” whatever it’s now called “red” and vice versa, but we couldn’t agree to say that “2 + 2 = 5”, because that would lead to internal inconsistencies (we could agree to use the symbol “5” for 4, but that’s a different matter).
This is also related to a staple of science: that scientific and mathematical truth is not determined by a majority vote, but by internal consistency. Indeed modern science started with this very paradigm shift. Quoting Galilei:
But in the natural sciences, whose conclusions are true and necessary and have nothing to do with human will, one must take care not to place oneself in the defense of error; for here a thousand Demostheneses and a thousand Aristotles would be left in the lurch by every mediocre wit who happened to hit upon the truth for himself.
If we want to train an algorithm to infer rules from language, we need to give samples of language where the rules are obeyed strictly (and yet this may not be enough). Otherwise the algorithm will wrongly generalize that the rules aren’t strict (in fact it’ll just see a bunch of mutually inconsistent examples). Which is what happens with ChatGPT.
Edit: On top of this, Gödel’s theorem and other related theorems have shown that mathematical reasoning cannot be reduced to pure symbol manipulation, Hilbert’s unfulfilled dream. So one can’t infer mathematical reasoning from language patterns. Children learn reasoning not only through language training, but also through behaviour training (this was pointed out by Turing). This is why large language models have intrinsic limitations in what they can achieve and be used for.
I apologize for my naivity.
but could openAI just introduce a flag into the decoder to highlight math questions and ports/transforms those math questions into a simple bash script to calculate the result instead of letting the LLM nodes “calculate” the formula?
I mean this would like straightforward give correct results. ChatGPT has a similar issue with counting as its nodes do not get the numerics. however a pc is capable of that. it would just rely on the encoder for parsing the question, and not going further the GPT route.
Yes. Apparently they are working on integrating wolfram alpha into ChatGPT. Then it would be able to solve a lot of math problems posed.
that sounds incredibly powerful.
I’m ok with this.
I’ve found it making up “facts” when I query it. I thought I was doing something wrong, but apparently, it’s just changing the way it works for some reason.
Same. Now I’m only using search engines that don’t have it.
It’s not changing the way it works. It’s making up shit when it doesn’t know.
And that’s how AI works, it’s all probability. It’s not answering 2+2, there’s a probability that the answer is 4 and it chooses that. If something convinces it that it should be 5 it’ll start answering 5
That’s how language models work. It’s grouped into AI as is so many things, but it’s not AGI. It could open the doors to AGI as a component, but isn’t actually thinking about its answers. And those probabilities are driven by training reinforcement which includes the bias of giving an answer the human will receive well. Of course it’s going to “lie” or make up things if that improves the acceptance of the answer given.
The best description I’ve heard to give to most people is that llms knows what the right answer looks like, not what it is.
If I wanted that I could just ask my daughter. She makes up shit all the time when she doesn’t actually know.
Would probably be more fun that way too.
Perplexity.ai has been a solid addition to my internet searches.
According to the Japanese zodiac, people born in May 1994 would have the zodiac sign of the Snake.
Expect it’s Dog, not Snake. Bing thinks it’s Ox. How did the entire field of AI go from surprisingly accurate to utterly useless in the span of under a year? I have no idea what benefits you personally see in this site.
How have you used Perplexity.ai?
Oh boy. I do research on it for various things. Florida released some laws changing alimony and I researched it via Perplexity to understand what the problem was. It worked. I understood the issue.
In any case, I do look directly at the sources. Perplexity.ai is useful for framing a topic, getting the gist of it, but for being sure I know wtf is going on, I personally need to look at the sources.
Thanks for this reply. That’s probably the best way to use LLMs - general definitions or framing / summarizing of issues. And then always check the sources to make sure it was accurate. I’ve played around with ChatGPT and Bard and I think my mistake has been to be a little too granular or specific in my prompts. In most cases it produced results that were inaccurate (ETA: or flat out demonstrably wrong) or only fulfilled a part of the prompt.
the best way to use LLMs - general definitions or framing / summarizing of issues. And then always check the sources to make sure it was accurate.
I agree. The criticism that they’re not accurate kinda misses the point of LLMs being tools. It’d be like complaining that a hammer doesn’t jam the nail in all the way after the first stroke. Hit it again…and maybe try hitting it straight this time instead of at an angle. It’s an iterative process that can be self-correcting when done thoughtfully.
To be honest I noticed a drop in quality of code generation via prompt by ChatGPT.
Still useful. Especially for boilerplate nonsense getting projects started. But it’s ability to understand complexities in code dropped drastically.
deleted by creator
