Hacker News Reader

Comment by LeoPanthera | Hacker News Reader

This will be for real, when I can run this stuff at home.

Stable Diffusion is there. But GPT-3 still has to be run on my behalf by a company that could pull the rug at any moment, or add more stuff that they don't approve of to their TOS. (You already can't use it for "sexy" content. Want help writing a love letter? Oh no, you're banned.)

The "lighter" versions these text models that can be run at home aren't good enough yet.

View on HN

You've alluded to my main concern regarding ML models, especially large GPT-3-alikes. Less that they could pull the rug, but rather the more general concern of companies "renting-out" access to these valuable tools that we will likely become dependent upon.

Based on Microsoft's relationship with OpenAI, it wouldn't be surprising to me if they also wish to offer access to these AI models via Azure for example as a SaaS offering. Then you can imagine some small start-ups with Billy and Jimmy making some other SaaS product that is based on this very API. It would be a clever way to extract revenue from supposedly novel products whilst controlling and shaping the game.

Software is only for real when you can run it on your home computer?

Google isn't for real yet?

There's a huge difference between utility software, especially utilities that are inherently network-based like "a search engine", and application software used for content creation.

Oh like Google Docs.

If Google Docs suddenly disappeared, everyone would just continue as usual, except that they now use MS Word or LibreOffice instead. If GPT-3 disappeared, what's the alternative?

Seems like you're changing the subject.

No he's not. That's exactly the point my original comment made. You have missed the point.

I feel like everyone is just playing devil's advocate. Almost every developer I know wants a native API or have the option of taking something out of the cloud. It's why paas died and docker as a service took over.

There are open models, eg. Bloom. Same with Stable Diffusion (open) and Dall-E. Creating these isn’t impossibly hard. Some ML ops company showed how you could train a GPT-like model for like $250k or something.

Reply on news.ycombinator.com