Hacker News Reader

Comment by vineyardmike | Hacker News Reader

> Not seeing Mistral Large as an open source model now has a bitter taste to it.

A company needs a product to sell. If they give away everything, they have nothing to sell. This was surely always the plan.

(1) They can give away the model but sell an API - but they can’t serve a model as cheap as Goog/Msft/Amzn who have better unit economics on their cloud and better pricing on GPUs (plus custom inference chips).

(2) they can sell the model. In which case they can’t give it away for free. Unlike open source code, there probably isn’t a market for support and similar “upsells” yet.

View on HN

> (1) They can give away the model but sell an API - but they can’t serve a model as cheap as Goog/Msft/Amzn who have better unit economics on their cloud and better pricing on GPUs (plus custom inference chips).

Which has a simple solution, release the model weights with a license which doesn't let anyone to commercially host them (like AGPL-ish) without your permission. That is what Stability.ai does it.

But who cares? Like not morally-open-access-sounds-nice but actual impact. Open-access to these models is nice but only commercial enterprises can afford the compute requirements for a massive LLM. So an AGPL type model would just sit on the hard drives of like 30 people who have a hobby to heat their home with GPUs.

It allows for research to continue, which might eventually benefit everyone. The primary advantage in my mind is giving academy a chance to learn from it and community to build cool stuff on top of it.

It's still not clear if model weights are under copyright protection at all.

> Unlike open source code, there probably isn’t a market for support and similar “upsells” yet.

_Like_ most open source code, there isn’t a market for support and upsells..

See The Linux Foundation, they don’t seem to have this problem.

I don't think a single sample makes a trend. Genuinely asking: what other companies are fully open source and still profitable? Redhat comes to mind, but nothing else, at least no thing on the scale necessary to run AI.

Let's imagine OpenAI open sourced what they were doing and provided training on the technology they've developed, do you think they'd have trouble finding sponsorship ? Honestly?

Sure they can find a sponsor… but Microsoft is spending billions on GPUs for them. Could the sponsor(s) cover those costs?

For better or worse, OpenAI monetizing GPT-3+ has been good for everyone. Hobbyists can’t afford to run the model anyways, and it pushed Google et al to start caring and develop their own models.

We probably wouldn’t have Gemini/LLaMa/etc see the light of day if OpenAI didn’t make LLMs go viral. It’s just disappointing that Google literally had an LLM good enough that was tricking employees into thinking it was sentient, and it took a competitor before they released it publicly.

I used to work for an opensource organization full time. They used to rely on contracting for revenue. It wasnt a lot but it was sufficient to keep the org going for 10 years.

Nowadays the market is much more competitive. "Just find some contracts" is not helpful. Clients expect fast delivery and immediate reactions. I have several people spending their full time on just looking for contracts, and it's still not enough.

If the models Open AI produced were open source, they'd have zero problems raising funding to train those models. Its not about that though, it's about monopolization of the tech.

Reply on news.ycombinator.com