Comment by chx | Hacker News Reader

Comment by chx | Hacker News Reader

Looking at https://stackoverflow.co/labs/

> The future of collective knowledge sharing.

The future is feeding all answers into a meat grinder and then forming word like nuggets from the paste and hope we like this? I am sorry but you lost your mind. There is a very fundamental disconnect between LLM hype and reality.

Here is my latest experiment, I needed to search for a string in dangling commits and asked Bing: https://i.imgur.com/pQtpkH5.png You will notice my question was a link because I copied the title from https://stackoverflow.com/q/21903972/308851 and apparently that made it into a link (I had the question open because every time I need to do this I just jump to this question, I outsourced storing my knowledge to SO ages ago). Every single answer given by Bing is wrong, of course, because git log -g is not appropriate for this as the answer states. It's right there: "Commits that are (still) referenced by the reflog are considered reachable and not dangling". And here's the best part: the Learn More 1. Stackoverflow.com links to this very question. The language model have seen "git log -g" in the vicinity of the string "Search for string in dangling commits in Git" and thought it's appropriate. The irony.

The answer I got with the prompt "You are a git expert. How do I search for a string in dangling commits?"

Was: Dangling commits in Git are often referred to as commits that are no longer reachable from any branch, tag, or the HEAD. These commits usually occur when you create a commit and then move the branch pointer away from it without creating a new branch or tag. They can also occur during rebases or other history rewriting operations.

Git has a garbage collector (`git gc`) that will periodically clean up these dangling commits, but until that happens, they're still in your repository and can be examined.

Unfortunately, Git doesn't provide a direct command to search for a string in dangling commits. However, you can find these commits using the `git fsck` command, which checks the filesystem for errors and dangling commits.

Here is a way to do it manually:

1. List all the dangling commits using `git fsck --lost-found`. This will print out all the dangling commits and store them in `.git/lost-found/commit/`.

2. For each dangling commit, you can search for a specific string using `git log -p <commit-id>` or `git show <commit-id>` and then pipe the output to `grep`.

Here's an example of what this might look like in a script:

  mkdir -p .git/lost-found
  git fsck --lost-found
  for commit in $(ls .git/lost-found/commit/); do
      git show $commit | grep -q "your string here" && echo "Found in commit $commit"
  done

This script will print the IDs of any dangling commits that contain the string "your string here".

Remember to replace "your string here" with the string you are searching for.

*NOTE*: This operation could be time-consuming if you have a lot of dangling commits or a large repository. Also, this approach might not work if the garbage collector has already cleaned up the commits. Make sure to be careful when manipulating Git's internals like this, as it could potentially corrupt your repository if done incorrectly.

I don't like the profit motives, but it's tough for me to argue that the experience of using stackoverflow is better than just jamming my question into chatgpt.

Except the information value coming out of a chatgpt answer is zero. You are literally best off if you ignore whatever it says.

It's really not though. I ask it questions about typescript all the time: "i want to do X with types, how do i express this in TS syntax?" and it answers correctly. No disagreement that i don't always agree with its subjective judgment, but that is on par with subjective opinions on stack overflow.

You're ignoring all the times the stackoverflow answers also provide zero information, so this is not a balanced assessment. Prompting LLMs correctly to get high quality answers is a skill.

Seeing value in LLM answers is apophenia. It's an understandable human condition but still.

There can be value in SO answers, there can not be in LLM as it's just automated plagiarism without understanding anything.

There's immense value in ChatGPT's answers, you just need to accept that it's fallible.

I think most of the critique is coming from the people that made their mind up already or used only GPT-3.5 in the past.

Then it becomes a self-fulfilling prophecy: a) "ChatGPT is stupid! Also, plagiarism! No value!" b) " I wrote a carelessly bad prompt and it responded with a seemingly bad answer!" c) " This confirms my biases!"

I find that when I provide it with enough context and a set of requirements (as I would with a non-AI engineer), it's very good at problem solving.

How does the issue of plagiarism bear on its utility? That’s like saying it’s impossible to get value out of listening to a torrented mp3.

I’m not sure understanding is an inherent issue either. When I Google ”weather lisbon,” Google spits out a number. The app doesn’t ”understand” anything, but it still helps me plan my day.

> automated plagiarism without understanding anything

I'm old enough to remember the memes campaigning to rename "software engineering" to "googling StackOverflow", so plagiarism isn't a point against the value of an LLM, and given the case of an LLM made by StackOverflow using StackOverflow data it's also not a legal issue.

> there can not be in LLM as it's just automated plagiarism without understanding anything.

Conjecture. We lack a mechanistic understanding of "understanding", so you can't make this claim. Human answers are also fallible, so how do you know that human cognition isn't also statistical pattern matching similar to what's used in LLMs? If it is, then they are understanding, their understanding just has flaws.

> Prompting LLMs correctly to get high quality answers is a skill.

That's like saying "reading the future from chicken entrails is a skill".

I don’t know how to square this with the fact that I routinely ask GPT-4 for code and get back answers that are in fact correct. You’re making it sound like it’s purely random, like I’m just getting back a string of zhrueksnbd8373&;! or something.

But that is what you get, it's just dressed up to sound like the answer. But it's not the answer , it can not be. If it accidentally is the answer, that's an accident, nothing more. See my example above about dangling commits...

> If it accidentally is the answer, that's an accident, nothing more.

Accuracy rates much, much higher than chance are not accidental. While you're pointing to a valid failure mode of LLMs, you're not using the correct words to describe it which is leading you to say nonsense.

yes, it is nonsense, all of it, we agree.

All of what you wrote is nonsense? Ok then.

No, all that LLMs produce is nonsense.

> That's like saying "reading the future from chicken entrails is a skill".

If I could read the future from chicken entrails with as high a reply accuracy as I get from ChatGPT, then I would fully endorse that too.

Or the times Stack Overflow is correct, but the answer is from 12 years ago so you don't trust it and move on.

This is so wrong it’s not even funny.

ChatGPT recently suggested a fix for a buggy service worker I have and it worked perfectly. I am eternally grateful.

Along those lines, how is that position not randomania?

You’re basically saying that the signal to noise ratio is zero so its utility is zero. I guess you must think all of the comments here where someone mentioned how it helped them solve a problem, also have a 0 (or at least dismissible) signal to noise ratio?

What signal to noise would be needed for you to say it’s not devoid of any information? What would the minimum accuracy need to be? 80%? 95? 99? 100?

The thing is there are many questions where it is a short path to a solution that can then be verified. The most useful way to use it productively is to induce it to give you pointers to things you had no idea existed. Then dereference those pointers via traditional means (ie google, man pages). Be your own null ptr (hallucinations) segfault handler. Critical thinking a hundred percent preserved.

It’s a little bemusing how people take issue with next token prediction trained with a log likelihood loss. In particular they think that getting a high accuracy extremely low perplexity LLM is unimpressive because of the low accuracy high perplexity autocorrect on their phone. No- the llm’s emergent generalization capabilities as nll drops really is a capital p Phenomenon.

But I feel that I don’t need to try to change anyone’s mind anymore. Everyone had a threshold logp where they’ll change their mind about this, whether it’s attained in the next week or long from now. Just state the nonzero perplexity that you would accept from a person thinking, and wait for the loss curve to meet it.

Don't bargain on that. For some the threshold is way too high given how far they've sunk.

As always, the ones most vocal about it haven't even bothered trying out guanaco 70b 2bit.

Probably don't know what ggml, gptq, cformers or any of the others are.

They still base their assumptions on the openai gold standard.

Reply on news.ycombinator.com