Post

Amias Channer
@amias.net

I expect this era of vibe coding to fizzle out. The cost of the tokens are getting non trivial and will only go up so usage will slow. Defensive rules to block giant PRs without prior discussion and requiring specific consent to use agents on a codebase can help.

0 reposts12 likes

cooperativescience.bsky.social
@cooperativescience.bsky.social

Additionally, we know that token costs are heavily subsidized, see: all of that debt. They will need to stop being subsidized at some point. So token costs have definite economic reasons to rise as well as plausible technological reasons for them to fall. (And this is neglecting externalities.)

cooperativescience.bsky.social
@cooperativescience.bsky.social

And remember that much of the cost of LLM agents is bound up in day-to-day inference costs. Even if we stopped all LLM research right now and stopped making new systems, just running what's currently operational to end of life could easily cost another trillion.

Amias Channer
@amias.net

The ability of a.i. will diminish as it encounters more and more of it own untrustable output as input. It's certanty is built on blind trust so its hallucinations will become exponential and it will not be able to detect them.

Amias Channer
@amias.net

no it isn't , the cost to do the same thing goes up every time the model gets more complex. there is also the way that you have to fix A.I problems with guardrails and defensive prompting that only cost you more tokens. i get it you like A.I , whoop de do , don't lie about it.

Kobold Commando
@kobold.bug.sex

They aren't though. They never have. People always say that. It's never been true. "They're only becoming more capable!" and then nothing happens. The same old scams and misdirection and AI finds a new field to move the grift to. It's pathetic.

🥶
@eggmansasshole.bsky.social

This is a complete misunderstanding of how models work - it doesn't "poison" itself with feedback like people intuitively think, and the top models for a few years now have actually been trained on more carefully curated corpuses that largely consist of AI generations because they're more useful