I Have Blocked OpenAI | Hacker News Reader

I Have Blocked OpenAI

gavinhoward.com

1 pointgavinhoward9 comments

Genuinely and wildly stupid logic betraying a fundamental misunderstanding of how LLMs function.

How so? How is it not statistics?

A) if your software is "unique" such that an LLM can't figure it out, it is a terrible UX, and B) the LLM will have learned about responding to your customers specifically; it will know when to pull from elsewhere and when to pull from your docs.

A) if all the other software has bad UX, then why does being different mean my software will definitely have bad UX?

B) No, because that's not how statistics work. All it takes is some others talking about my software and getting details wrong, and GPTx ingesting their stuff too. It will inevitably mix my stuff and the stuff people write about my stuff.

By definition "all the other software" can't have bad UX, as a large part of UX is predictability.

And no, talking about your software "wrong" won't matter meaningfully. Also people would have to talk about your software at all for that to even matter, which is not happening.

By what definition? Also, I'm making a build system, which are pretty universally hated. And that's not to say I don't follow patterns that other build systems do; I do. In fact, a lot of software has bad UX by not acting as users assume they do. My software take those assumptions and works with them.

But that doesn't explain how the LLM can't get it wrong.

People talking about my software wrong will matter because LLM's predict the next word. If my documentation does not win the prediction war in a particular run, the LLM will get it wrong.

And people are already talking about my software, so claiming they aren't is wrong.

If what you say is true then an LLM will do just fine, exactly how it does with literally every other piece of software that has hey do work well with. Like I said, this reveals a wildly wrong and terrible understanding of how LLMs work.

Your software is not special, unique, or difficult for an LLM to understand, and people talking about using it wrong would not screw with an LLMs ability to provide help.

Additionally, no. Nobody is talking about your software, not to any level relevant to this conversation.

Uh no. You're wrong on all counts.

The worst is that LLM's do not "understand" software. If you think they do, you are the one with a bad understanding of them and how they work. So I can stop talking to you.

And any talking about my software is a level relevant to this because we don't know the weights an LLM has.

In the context of LLMs, it’s common to say “understood” as an anthropomorphic shortcut to describe the weights on their neural networks. You don’t get this, I understand, but the phrase is perfectly acceptable here.

And I absolutely do not need to know the LLMs weights to understand generally how they work. Further, I’ve successfully used them for exactly what you’re claiming they’d fail at, as have literally millions of people.

They simply wouldn’t be useful if they worked as you described here.

Reply on news.ycombinator.com