You know you're the target market when the game has an entire help system that kicks in when it detects you haven't played the game in a while

Is that Olli Vänskä?

“I used the LLM to...”

...

“Oh sorry, I used ‘ChatGPT’ to...”

I find this other Scott Sumner quote explains why so many have trouble understanding this
www.themoneyillusion.com/the-autistic...

The starvation of the Rohingya people is absolutely tragic, however I'm skeptical of these claims that N people have “already” died. They're based on linear interpretations of number of lives saved, which is not an accurate way to measure deaths in response to funding cuts

That's what I do today with agents. The thing I don't know how to square is the ability to fix manually with the "real code is the spec". Perhaps some sort of metaphorical equivalent of __asm__ inline blocks

Re: code gen, the link is that when you find problems with the output, most of the time you can effect a fix by adding to the spec, but some of the time it won't get it or will break something else.

It's like answering quantum physics questions with English. Infinite prose can't replace the math

Agreed on AI having better design sense than most developers (if not today, then soon)

Probably nothing comes of it, but I find it hard to invest effort into any platform run by a mad king... just sitting there... at the top

Code gen is an old idea. Historically it didn't work out (beyond scaffolding) because inevitably you would need to tweak enough things in the generated output such that you could no longer regenerate. With LLMs you can go a lot farther, but the problem is still there for most real world use cases

I see, you're wondering when we get human coder replacement, while retaining human spec writer. I'm sure it will work in many areas, and steadily grow over time. Good luck in the domains I've seen though

John Titor tried to warn us

I'm sure people will try this paradigm and tackle the complexity problem by decomposing the app. Or store the prompts and regenerate sometimes but less frequently than every build. But I have a hard time seeing this paradigm becoming common

While making the LLM reasonably deterministic is likely solvable, the bigger problem is models can't one-shot solutions beyond a certain threshold of real world complexity. That threshold will rise over time, but color me skeptical it will rise above the complexity of most real world applications

You tried an Italian restaurant better than Olive Garden, interesting. New Battlefield, eh, yeah quality gaming. Figuring out how to hang displates at the new place, that'll look cool

It will be. People first have to figure out how it fits into their tribal worldview

That's a small one. Reminds me of telephone exchange buildings

I wonder if there's a way to get the audio to loop cleanly. Then someone could make a whole account out of this

If you love this era, do yourself a favor and give @retrotechdreams.bsky.social a follow