Poul-Henning Kamp6d ago

I think I can finally put words around what truly bothers me about the "AI changes everything" meme, as far as programming goes:

Programming is fundamentally hard for our brains, because it requires fractal perfection.

(By fractal perfection I mean that each code line must be perfect, but so must the overall architecture and all intermediate levels.)

They way your brain gets better at that is training by doing.

The oft quoted 10k hours is probably about right, also for programming.

1/2

Kristof Provost5d ago
Show threadPoul-Henning Kamp

But if you use a LLM to write your code, and only look at it, when it doesn't do what you want, you are not training to become a better programmer, you are training to become a better "incoming code quality inspector".

But there is no training-path from incoming code quality inspector to good programmer, because you are (deliberately) not investing the 10k hours.

At the core it comes down to passive vs. active vocabulary: You only grow your active vocabulary by expressing yourself.

2/2

May 19 at 5:02pmWeb
Show threadPoul-Henning Kamp6d ago

PS:

I really do wish there were more "incoming code quality inspectors" in the world.

I hope EU's CRA makes that a thing.

And then I hope they drive a stake through the m4-armor of the autocrap monster.

Show threadDr. Bruce Simpson6d ago
@bsdphk This dovetails with what someone posted here about "Interpassivity" yesterday: https://productpicnic.beehiiv.com/p/the-newest-term-in-the-ai-lexicon-is-interpassivity Also I was sciencing the shit out of the SoTA in software complexity over the weekend.
The newest term in the AI lexicon is "interpassivity"

The industry is reorienting around software that uses itself. But giving ourselves permission to lean back and relax merely indulges an old and toxic tendency.

The Product Picnic
Show threadElric6d ago
@bsdphk You mean I won't get bigger biceps from watching fitness videos? Damn.
Show threadGregor Schmidt6d ago
@elricofmelnibone @bsdphk you're not even watching fitness videos. You're just watching people who have big biceps, kind of. šŸ˜‚
Show threadMans R6d ago
@schm @elricofmelnibone @bsdphk You're watching fail videos shot at the gym.
Show threadPaco Hope6d ago
@elricofmelnibone I donā€™t watch fitness videos. Iā€™m vidmaxxing. I orchestrate a fleet of agents to watch the fitness videos and send me a bullet point summary. Iā€™m gonna get ripped so much faster than those folks who waste all that time lifting weights by hand.
@bsdphk
Show threadtrademark6d ago
@bsdphk The LLMs allows people who aren't programmers to create their own software, I think that's a good thing. I'm seeing lots of people creating small things that they find useful for themselves or familiy/friends. Here's one which is generally useful: https://randomwire.com/mapthread/
Mapthread: Tell a Story with a Map

In 2016, Craig Mod and Dan Rubin published Koya Bound, a beautiful photo book covering an eight day hike on the Kumano Kodō pilgrimage trail in Japan. Alongside this they made a companion website wā€¦

Randomwire
Show threadAttila Kinali6d ago

@trademark @bsdphk Unfortunately, LLMs allow people to create programs way more complex than they can handle. And the unintended consequences of that can be devastating.

IMHO we should put in the same legal safeguards into programming as we do with, e.g. cars or airplanes. Just because you can build one using modern manufacturing, doesn't mean you are allowed to drive/fly it.

Show threadtrademark6d ago
@attilakinali @bsdphk Pretty sure that if your idea gets any kind of traction, what the law will say in the end will be something like "No releasing any software before a Mythos-class AI says it's OK"...
Show threadPoul-Henning Kamp6d ago

@trademark

I'm not against LLMs.

But we need to recognize that if the pretend-competence they enable, leak into spaces where actual competence is required, bad things happen to innocent people.

We do not allow people with pretend-competence to perform brain-surgery, design high-rise buildings or fly airplanes, but we are OK with them staging Shakespeare.

The traditional solution to such line-drawing is licensing, responsibility & mandatory insurance.

Show threadtrademark6d ago
@bsdphk Yes, but that problem has always existed. E.g. the Tea Dating App data-leak had many people loudly claim "only an LLM could do something so stupid, no human ever would". Turns out the bug dates back to 2022, way too early to blame AI. In fact the LLMs are getting so good at security that I expect any kind of "professional rules" would mandate their use.
Show threadMikeStok6d ago
@bsdphk @algernon ā€œapparent incoming code quality inspectorā€
Show threadRafael PĆ©rez šŸŒ±6d ago

@bsdphk This reminds me of Norvig's "Teach yourself programming in ten years."

https://www.norvig.com/21-days.html

Teach Yourself Programming in Ten Years