Jaguar🔞

Jaguar🔞

@jaguar.barebottom.games

スケベが大好き! Lover of nudity/exhibitionism/ENF Founder of: @barebottom.games Discord: http://discord.gg/aJ6nysmWBY Currently learning Japanese, so I will occasionally write in Japanese. 今、日本語を学んでいる、では時々日本語に書きました。

1.9K followers161 following6.6K posts

LLMs will write complex Bash one-liners to test things, but I'm versed enough in Bash to be able to read it. If it wrote Batch or PowerShell, that would be a different story. I really don't like PowerShell.

The thing that has most impressed me with recent AI developments, is its ability to write deterministic scripts to prove correctness. The strawberry test is solved because a good AI/harness will write a temp Python script to deterministically count the number of r's in it.

LLMs will have some built-in knowledge, but they're not the most accurate or up-to-date; I never trust when they draw from it. With the ability to read codebases and search the web, I'm more willing to trust their output.

Hot take: I think LLMs are better than the average programmer. They don't compare to enthusiasts because they care deeply about their code, but they are better than the average programmer you'd find in a company. Which leads to using tools beyond their abilities and spewing slop everywhere.

It really is something to see some of the code I'm most proud of writing at work being copied by another team and utterly bastardized and declared sacrilegious. They had a really good implementation and utterly ruined it. And this was without AI too. It was ruined by people.

One the (many) concerns I have with AI is the fracturing of software as people make individual edits of software instead of collaborating together

If I had a nickel for every time this happened, I'd have two nickels. Which isn't a lot, but it's weird it's happened twice.

Jaguar🔞@jaguar.barebottom.games

I got three different packages ordered from three different sources at three different times delivered by three different shipping companies all delivered on the same day

I forgot that power draw was a major factor when I bought my Strix Halo PC. Buying a R9700 (or even two of them!) does make for very fast inference, but sucks up power.

There is a skill to using AI effectively, but you need to already have a strong foundation to best make use of it. It won't work if AI is your foundation.

One of the things I've done this weekend is setup specialized AI inference engine for my Strix Halo. It works and is significantly faster than before, but I've run into a problem: What do I do with it?

I do find it absolutely hilarious that it took not even a month between a team at work being allowed to use AI and absolutely fucking up their codebase.

I wish I could play Ocarina of Time for the first time again. But after watching an OoT speedrunner consistently for ~15 years, I've acquired so much secondhand game knowledge about it.

I keep trying to find a way you could incorporate AI into a game in a meaningful way, but it's a lot like NFTs: it's a solution looking for a problem. Every usage fails in comparison to an already existing solution. It's too slow to use and too unpredictable to control.

But I wonder if that comes from having a better design in mind beforehand. If their AI generated code would be spaghetti because their design is spaghetti to begin with.

I hate having to call places where they are just verbally repeating whatever they see on their screen and having to verbally input things. Just let me fill it out myself.