I’m very skeptical.

Deepseek and other Chinese models are close enough for most typical use cases.

This could seriously wreck our entire economy, so much of the stock market is propped up by AI hype.

Anyway, expect a massive government bailout for Anthropic and Open AI when the Chinese models prove to be more cost efficient.

Might not be next year, or even this decade, but it’s coming.

PS: A fear mongering campaign is already underway, Deepseek is going to give you communist code completion!

▲

edg5000 2 hours ago | parent | next [-]

I genuinely wonder if we both had to switch to DeepSeek, we had infinite tokens but could use no other model, how things would change. I suspect we both have to go back to very long specs and only using LLMs for a limited set of things. No more Blender, CAD, electronics, and high-level (without needing to micromanage) coding, as it was in the early months of 2025. The models got so much better so fast. But I could be wrong, maybe the Chinese models will work just fine for all the stuff I mentioned. I could be US-pilled (I'm from EU, but still).

▲

ggregoire an hour ago | parent [-]

Have you tried 4.1 flash? I'm using it to do machine learning in an audiovisual data pipeline, I basically do nothing, I only provide labeled data and it fine-tunes the algos, thresholds and so on in a loop, and publishes the evidences in markdown tables, creates images illustrating the clusters it found (I never asked for this!), all by itself. I was late and skeptical of the AI assisted coding but this is truly amazing.

And what absolutely crazy is that I used 2.4B tokens this month and it cost me $1.7 (w/ 99.4% cache hit)

	▲	HDThoreaun 4 minutes ago \| parent [-]
		I use 4.1 flash pretty extensively for translating mandarin to English and it is noticeably worse than claude. In terms of tasks complete I find deepseek to be more expensive than the claude subscription I have as well. Deepseek uses an insane number of tokens per task.

▲

origini 3 hours ago | parent | prev [-]

Frontier is frontier and Open Models from China are not.

Then in a corporate environment, were most tokens will be spend, you have security concerns, you also might want to go with one company over the other due to partnerships etc.

If USA is blocking chinese companies/models (which they already partially did), big companies will just not bother and use US AMerican models.

Also running it yourself is often not relevant. You don't want to know how much money my company spends on software products which could easily be replicated internally.

But running models yourself might be something you don't have expertise with, you need GPU ready DC, hardware, a new/diffferent team, quality assurances etc.

And Chinese models are known to leverage frontiermodels for post-training results.

Don't get me wrong, I don't want to say chinese models can't disrupt the market massivly absolut possible, but its not given just because they are cheaper.

▲

999900000999 3 hours ago | parent | next [-]

> But running models yourself might be something you don't have expertise with, you need GPU ready DC, hardware, a new/diffferent team, quality assurances etc.

Absolutely nothing stops Microsoft/Amazon/Open Router( with partners) from hosting the Chinese models. As far as I can tell this is already the case.

This establishes a defacto price cap on what Anthropic and Open AI can charge. If you’re a heavy user blowing past 400$ a month self hosting becomes economical. A Mac Studio with 128GB of ram is about 6k.

Every day local tools are getting better, the OSX ecosystem here is exceptional.

You do have a point with frontier models, but most of the time you don’t need frontier models.

If I knew for sure I’d short Anthropic and retire early.

	▲	origini an hour ago \| parent [-]
		There might be a relevant risk of a backdoor in these models. Self hosting is still far away from what Anthropic and co offers. A Mac Studio with 128GB doesn't deliver the speed at all. If you burn through $400 a month, you also might get a lot more value out of it so you don't care paying a premium. We will see though. I think its not that easy to tell.

▲

hello5010 3 hours ago | parent | prev [-]

[dead]