3 ms·

Today I had a small amount of Google Sheets work to do, to extend somebody else's work, and I had both Claude and ChatGPT hallucinating mythical function names,

Today I had a small amount of Google Sheets work to do, to extend somebody else's work, and I had both Claude and ChatGPT hallucinating mythical function names, and not understanding array functions with utter confidence and certainty. StackOverflow results from a web search gave me the answer faster, in the end.

b112 2y ago
Not to mention, such sites often have a discussion, with pros and cons of certain tacts, why some methods may be pitfalls, and more. Such sites help you understand why, far more than a distilled down response of a LLM. Such sites help you learn. Knowing the why, is often far more important than the how.
- epistasis 2y ago
  Ironically enough, the LLM was better at explaining trade offs and new syntax, and I learned more about the topic from the LLMs, largely from observing the unintended effects of their incorrect instructions. It was just dead wrong on several key details and when it didn't know how to do something it answered with BS with just as much confidence as the insightful parts, which would be infuriating from a human, but merely time wasting from the LLM.
ncruces 2y ago
LLM are the reason I get bug reports like… Them: “why doesn't this thing work I your OSS lib, whereas it works in this other OSS lib?” Me: “does it? it's not supposed to.” Then: “oh idk, ChatGPT told me so.”
- epistasis 2y ago
  There should be reputation points on bug reports where somebody loses the ability to file after a certain amount of bad behavior like that...
arkh 2y ago
> hallucinating I hate the use of this word to describe when a sentence completer gives a wrong result. To hallucinate you have to have a concept of reality. Which those models don't have. They're just mathematical tools which some times manage to give a result humans think as human-like.
- epistasis 2y ago
  I don't work in LLMs, but I use a ton of similar tools to model biology, and I really like the term "hallucinate." I would argue that the LLM does have a concept of reality, in the form of a probabilistic distribution over sentence structure. And that probability distribution does not match the reality that I live in, even if it matches the LLM training data and training procedure fairly well. I would argue that even a simple random forest model has a concept of reality, or maybe is a concept of reality, though a faulty one.
- scarface_74 2y ago
  Do you also hate when prooor use the terms like debugging, serverless, non relational databases (all databases model relationships), etc?

b112 2y ago

Not to mention, such sites often have a discussion, with pros and cons of certain tacts, why some methods may be pitfalls, and more. Such sites help you understand why, far more than a distilled down response of a LLM. Such sites help you learn. Knowing the why, is often far more important than the how.

epistasis 2y ago
Ironically enough, the LLM was better at explaining trade offs and new syntax, and I learned more about the topic from the LLMs, largely from observing the unintended effects of their incorrect instructions. It was just dead wrong on several key details and when it didn't know how to do something it answered with BS with just as much confidence as the insightful parts, which would be infuriating from a human, but merely time wasting from the LLM.

ncruces 2y ago

LLM are the reason I get bug reports like… Them: “why doesn't this thing work I your OSS lib, whereas it works in this other OSS lib?” Me: “does it? it's not supposed to.” Then: “oh idk, ChatGPT told me so.”

epistasis 2y ago
There should be reputation points on bug reports where somebody loses the ability to file after a certain amount of bad behavior like that...

arkh 2y ago

> hallucinating I hate the use of this word to describe when a sentence completer gives a wrong result. To hallucinate you have to have a concept of reality. Which those models don't have. They're just mathematical tools which some times manage to give a result humans think as human-like.

epistasis 2y ago
I don't work in LLMs, but I use a ton of similar tools to model biology, and I really like the term "hallucinate." I would argue that the LLM does have a concept of reality, in the form of a probabilistic distribution over sentence structure. And that probability distribution does not match the reality that I live in, even if it matches the LLM training data and training procedure fairly well. I would argue that even a simple random forest model has a concept of reality, or maybe is a concept of reality, though a faulty one.
scarface_74 2y ago
Do you also hate when prooor use the terms like debugging, serverless, non relational databases (all databases model relationships), etc?