Hacker News Reader

Comment by jstummbillig | Hacker News Reader

Disagree, because the very few very big AI players are (in contrast to the 90s) very solid. Yes, there is a breadth of absolute bullshit built on top of current AI, but if it were only ChatGPT-alikes and LLMs for coding from here on out, that in itself is enough real value, that requires very little imagination, a lot of implementation and there's more demand than can easily be satisfied right now for both.

View on HN

I’m an AI enthusiast, and it’s not clear to me that selling inference to a prop model is a winning business model, which is what Anthropic and OpenAI are doing. The open models are good enough today for many things, and are likely to only get better. Feels like inference is a commodity, and not clear how much money there is in it.

I would love to know how much of the inference I pay for is being paid for by VC cash: I suspect a lot of it.

> I would love to know how much of the inference I pay for is being paid for by VC cash: I suspect a lot of it.

Definitely a lot of VC subsidies for OpenAI and Anthropic, none for Google.

Sure, that doesn’t mean we’re not getting subsidised tokens from Google though

To be fair though though, the "open" models are fairly sketchy as well, from a business standpoint, because somebody is paying for a lot of GPUs and expensive talent. It's probably the least obviously sustainable open source product of all time, and it's not at all clear to me why that would change going forward.

Right now there seem to be roughly two paths, when it comes to frontier-level LLMs: Meta just not giving a fuck, spending instagram money and pretending it's a business, and whatever Deepseek is doing that might make it both good and also super cost-effective (and there it's even less clear, how much of it is real and what the actual costs are).

yeah, most of it

You can’t be serious, surely?

Most of the LLM applications are either entirely useless or trivially reproduced with much simpler free models (or even entirely non-“AI” methods).

Reply on news.ycombinator.com