Hacker News Reader

Comment by Legend2440 | Hacker News Reader

There well may be an AI crash, and everyone who invested into datacenters may lose a lot of money.

But in ten years we'll be using LLMs even more than today. I'm sure of it.

This is a lot more like the dot-com boom than like crypto or IoT. The internet really was a trillion-dollar industry, and AI will be too.

View on HN

Exactly. People didn't go back to mail ordering things from the Sears catalog because pets.com was overhyped and went out of business.

When you’re paying the real, unsubsidized price? $1 for $1 of compute?

Maybe.

This is flawed thinking - the models will exist even if every company who trained them fails, and then someone else will rehost them and offer them at inference rates. Inference is _astonishingly_ cheap.

If anything, prices would go _down_ as the rehosters wouldn't be saddled with debt from training models.

How long does a model stay relevant? You’re assuming that a huge and lasting plateau will emerge that can still adapt to changing conditions?

Even if models stopped improving now, I would use them. They provide so much convenience, I'll never go back. There's no need to keep changing.

In fact I'm more or less at the point where I don't even want the model to get better, I care more that it outputs faster.

For coding in existing languages? Probably indefinitely. And huge swathes of human civilization run on decades-old codebases.

We're on said plateau. In terms of normal coding usage, models have not become wildly more useful in the last few months. The existing open models that are useful today will remain useful.

Qwen 3.8 runs fine on my laptop with no subsidy.

OpenRouter is full of inference providers for open weight models that are only months behind models like Opus, with a fraction of their price per token. And those inference providers don't do charity.

If optimizing inference was a priority, OpenAI and Anthropic could also focus on it and get the prices down. They are just focusing on their dumb AGI race instead.

Running open models from commodity hosts doesn't appear to be particularly subsidized today. It's not a massive expense, because the biggest cost that's causing the frontier labs to hemorrhage money at the moment is training. Inference isn't terribly expensive for what you get, and seems to be quite comfortably profitable.

I’m already paying the real unsubsidized price with my own GPU.

Maybe it can’t solve a Millennium Prize, but it’s pretty good at writing CSS.

IoT did happen, too. You just won't notice it unless you work in specific industries. Just couple weeks ago I chatted with a dairy farmer who had crazy amount of telemetry from each of their cows.

You're saying LLMs, but the future holds a full system using it as a tiny component we can't really imagine. Like transistors became unimportant or should I say trivial by themselves, the complex systems built on them are truly unimaginable. Except well if you're David Bowie: https://youtu.be/8tCC9yxUIdw

Reply on news.ycombinator.com