Hacker News
points by throwaway_1016 3 days ago | 0 comments
Most of sp500 companies are never going to use chinese models.
[–] jacquesm 3 days ago
I would happily bet that > 50% are already doing just that. Whether you'll be able to collect the data is another thing but they'd be fairly stupid if they haven't at least checked out what is available. It would be next to irresponsible not to do so and given that they're in the S&P 500 you would have to at least give them some credit.
[–] huhhu22 3 days ago
We already have data! Look at market share measured by tokens not revenues.
[–] Paradigma11 3 days ago
You don't see the tokens served locally by those companies.
[–] pbasista 3 days ago
Why not?

I can understand that they might be cautious about sharing their data with a Chinese company.

But them being the most successful companies means that they can just buy their own hardware. And run whatever models they can get their hands on in there.

Even if those models are Chinese. They (the companies running them on their own hardware) would be in full control then.

[–] maxnevermind 3 days ago
I work in one of sp5 and I use chinese models :-D
[–] sharadov 3 days ago
do you guys self host those models?
[–] matsemann 3 days ago
Only because anthropic&co will price themselves just below the cap where people contemplated switching. In so, the Chinese models still set a sort of cap on the prices they can charge, even if not directly.
[–] cmdrk 3 days ago
depends on the framing.

if the chinese models have this aura of inferior quality around them, then sure.

if they manage to frame it instead as open source, sovereign AI, the calculus changes.

[–] throwaway_1016 3 days ago
It's not about quality, but more about processes, for example we handle EU, US personal, financial data etc, we have strict rules about it (company rules, and gov ones), and there is no place for chinese model here.
[–] koe123 3 days ago
You dont understand it, its open you can host in USA
[–] xmcp123 3 days ago
For now, there’s definitely some resistance because people don’t understand models can’t phone home (only the servers they’re hosted on).

But I don’t think this will be the first time in history that money doesn’t win.

[–] throwaway_1016 3 days ago
They absolutely can, it's just a matter of proper incentives (https://news.ycombinator.com/item?id=50159245).
[–] xmcp123 2 days ago
This would have to be an incredibly targeted attack, and it would only work once. Once it happens, trust is broken and that model is fucked.

That’s not enough going to be enough to make companies pay 20x more.

[–] addaon 3 days ago
> models can’t phone home

Models with tool access (and internet-connected tools like curl) can, though… is this really a useful distinction?

[–] joshheitzman 3 days ago
It is. Harnesses do not need to give LLMs such broad tools like the shell.
[–] xmcp123 3 days ago
Yes, because you can run a chinese model on a USA server. The fact that it's a chinese model means nothing.