Hacker News new | ask | show | jobs
by alpark3 1005 days ago
Honestly, the current general state of software may be a good proxy for what AI will settle to. I can imagine major open-source models, trained and generated by nonprofit efforts in a roughly similar fashion to Linux. Entire businesses might be built on top of "servicing" this model, such as enterprise-grade finetuning, serving, etc. Like the author mentioned, no business wants their core functionality to be dependent on external factors. As I understand it, this is also the case for some Linux-based corporations that focus on building a business around open source software. Of course, there will be proprietary models. Will the average home user cook their own custom distro of LLaMA 10 to be their home assistant? No. They'll probably use Alexa or whatever proprietary solution is out there.

Uncertain, but I'd be willing to be that open source AI will win the way Linux won. In the ways that it matters.
2 comments
sillysaurusx 1005 days ago
I don’t think we can. People are furious at anyone who even touches their data. This fury by the people will transfer from OpenAI towards the smaller players.

Result: open source AI dies. There’s no way to get any data, and what’s available outside of copyright isn’t enough. Not to compete with ChatGPT.

Sure, there will always be cool models. But nothing like what we were hoping for. LLaMA is being sued right now precisely because it used copyrighted books. Open source entities don’t have the legal resources to defend themselves from these threats.

link
swader999 1005 days ago
Piracy doesn't respect copyright. Why not steal all the great books and works ever made just to feed this model. Like a protocol for free training data that you could torrent around add contribute to the hive. Not well thought out but that's the idea.
link
visarga 1004 days ago
On the contrary, it is possible to train efficient models with purely synthetic data, see Phi-1 and Phi-1.5 from Microsoft.
link
smackeyacky 1005 days ago
What if there was nobody to sue? Open source could be structured that way, like MAME.
link
tsimionescu 1004 days ago
> I can imagine major open-source models, trained and generated by nonprofit efforts in a roughly similar fashion to Linux.

Linux today is mostly built by for-profit companies, essentially as a large collaboration project. Microsoft, IBM (RedHat), Oracle, Intel, Huawei are some of the largest contributors.

link