Comment by skwirl | Hacker News Reader

Comment by skwirl | Hacker News Reader

Parent Full thread

Sure, it's an unending stream of things, but here are some recent examples from my history:

-Had it extract some gnarly string manipulation logic in a MySQL where clause into a function. It wrote clean, well commented code, and even explained what a regex did in a comment.

-Had it explain to me how macOS plist files worked, how to view them, how to edit them

-Asked it what some lines in an ssh config file did and it clearly explained it to me

-Dropped in some XML from a spring applicationcontext file that was doing something I didn't quite understand and it explained it to me. Asked how I would change something about it and it told me how, explained how it worked, and provided example xml

-Gave it a SQL Server CREATE TABLE statement and asked it to create the equivalent CREATE TABLE statement for MySQL 8, which it successfully did

My only complaint has been how sometimes the system has been unavailable. I now pay for the $20/month subscription and that has been resolved.

Thank you for those examples. Converting sql manipulation into a function is an excellent example. You can do a lot in a few case statements especially with nesting but translating that into something else is valuable.

I asked it to write some queries in relational algebra and it was close but it was invalid in ways that a non-database person wouldn't understand. Kind of looked right.

If you ask it something that you do not know yourself, how do you know what it is telling you is not bullshit?

There exist problems where finding the solution is hard but verifying the solution is easy. example in the general case: prime-factoring numbers.

Are you claiming that for all possible problems posed, verifying the correctness of the output of an LLM is equal or higher difficulty than solving the problem itself? If so, that seems like a claim more arising from emotion than reason.

In the general case, you can have it generate code in a proof based language that "proves" the code is correct against formal specification. Unless you consider math itself to be "bullshit" too.

The people who are skeptical that it can be easier to verify correctness than to produce a correct work might as well try to prove P=NP (which is the theoretical and formal statement of such).

They also probably never asked people to work on a task either (hint: people can get things wildly wrong, even generally competent ones)

If it is a well known problem, I might as well find the solution on stack overflow, where actual the actual humans that trained this system discuss it.

As for proof based languages, can it actually do that? Have you tried?

Within the constraints of technical systems, it's pretty obvious. It's not a topic that you can really bullshit - things work or they don't. It makes mistakes sometimes and what it says won't be consistent with other things it's saying or that you know to be true. If you ask it if it's sure, it often will correct itself. Worst case, you verify with other sources and/or try it out.

It's not really any worse than asking someone knowledgable about a subject face-to-face. They know a lot, but they can be mistaken about things. At some point you will tease out the misunderstanding.

The best way to understand the value of the tool is to try it for yourself.

I have tried it, it fails at the simplest problems that I know how to solve myself. I would not trust it with anything I do not understand.

I can see how it might help with source to source translation or with generating random stuff, but that is not a problem I have.

Reply on news.ycombinator.com