ChatGPT et al

[Spearthrower]

It's like a genius 2 year old.

[THWOTH]

[Spearthrower]

This is a fantastic break-down of how transformer neural networks actually work step-by-step. It's taken me several reads to begin to grasp the basics, but others may have less trouble than me.

https://builtin.com/artificial-intellig ... al-network

One of the relatively minor points made therein that I never really even thought too much about - computers can't actually understand words, so even the words you're using undergo transformation and pre-modelled mapping into vectors.

[Animavore]

I see some guy got busted for creating AI child porn.

https://edition.cnn.com/2023/09/27/asia ... index.html

[Spearthrower]

It's going to generate all manner of unpredictable situations and challenges that we've never even conceived of, let alone dealt with before. Interesting times.

[THWOTH]

[Spearthrower]

Magic 8 billion Ball

[THWOTH]

[Spearthrower]

Very cool!

[THWOTH]

[electricwhiteboy]

Deleted posted in wrong thread. Oops. Sorry.

[Alan C]

Google's AI search is reportedly offering up valuable nuggets such as recommending you eat small rocks, glue;

https://www.howtogeek.com/google-search ... responses/

In a search asking about cheese not sticking to pizza, Google recommended adding “about 1/8 cup of non-toxic glue to the sauce,” possibly because it indexed a joke Reddit comment from 2013 about adding Elmer’s glue when cooking pizza. It told another person that “you should eat one small rock per day,” based on a parody article reposted from The Onion to the blog for a subsurface engineering company.

And a study found 52% of ChatGPT answers to programming questions are wrong [energy budgets well spent here?];

https://futurism.com/the-byte/study-cha ... wers-wrong

For the study, the researchers looked over 517 questions in Stack Overflow and analyzed ChatGPT's attempt to answer them.
"We found that 52 percent of ChatGPT answers contain misinformation, 77 percent of the answers are more verbose than human answers, and 78 percent of the answers suffer from different degrees of inconsistency to human answers," they wrote.

I would imagine people then need to come in and fix such borked code, possibly paid less because they're only correcting it rather than having wrote it all themselves to begin with.

[TopCat]

And a study found 52% of ChatGPT answers to programming questions are wrong [energy budgets well spent here?];

https://futurism.com/the-byte/study-cha ... wers-wrong

For the study, the researchers looked over 517 questions in Stack Overflow and analyzed ChatGPT's attempt to answer them.
"We found that 52 percent of ChatGPT answers contain misinformation, 77 percent of the answers are more verbose than human answers, and 78 percent of the answers suffer from different degrees of inconsistency to human answers," they wrote.

As a professional developer, and user of both Stack Overflow, and ChatGPT-4, I think the situation is a lot more nuanced than the study appears to show.

Plenty of SO answers are less than helpful, and it can sometimes be quite difficult to articulate a question precisely enough to be acceptable to the SO community. If your question isn't good enough, it gets closed, and then you get no answers at all. This isn't entirely a bad thing, as it does curate the quality of SO content. But it can make SO a difficult tool to get the best from, and of course you're reliant on someone knowing the answer. If your question is arcane enough (and mine often are, as I already know the basic stuff), you might get nothing back.

I use ChatGPT-4 quite a bit as a coding tool. Subjectively, I find the quality and relevance of its answers to be a lot better than with ChatGPT-3.5, much more so than the comments about the two versions in the paper suggest.

For sure, occasionally the answers are rubbish, but these cases are extremely obvious. Even when they're only a little bit wrong, they're often still helpful.

I also use GIT Copilot, which is an AI-based coding tool built into Visual Studio. It suggests code fragments in real time as you code, and again, sometimes it's dreadful, but quite often it's spookily good and overall saves me quite a bit of time.