Hacker News Reader

Comment by InsideOutSanta | Hacker News Reader

There's a lot of criticism of Mistral being unable to compete with large model, and that's fair. But I think it dismisses what Mistral is actually doing, which is making specific capabilities available at high quality in tiny models.

I do a lot of OCR, file analysis, stuff like that. I use Mistral for that. I put 100$ into my account, and it just runs for a year without any worries about the amount of requests I make, because the cost is minuscule. That's valuable, even if it doesn't compete with Opus 4.8.

View on HN

But how does it compete on OCR? I find that having good quality at a cheap price is more niche than having the best quality at 10x the cheap price, because for most use cases you want to pay a bit more if it saves you mistakes later.

It's really good. I didn't do any type of statistical evaluation or comparison to other models, but it's so good that it doesn't matter to me if there's an option that might be even better.

curious if you tried local LLM models for OCR, like a Gemma4, or your volume is too much for that

Haven't tried them in a while, so I can't comment on current performance.

This just dropped: https://huggingface.co/baidu/Unlimited-OCR

Which can run comfortably on 12gb of vram. I gave it a whirl and it does seem pretty competitive. I wonder how that compares for your usecase

It seems to heavily depend on what exactly you're transcribing, the performance/quality between them is really uneven. Some models work really well for old cursive but then fail reading 8-bit segment LCD digital fonts, vice-versa or any combination out there.

Basically, to find the answer you really need your own benchmark you run with real examples from what you want to do. Basically the same goes for anything ML nowadays as the public benchmarks cannot really be trusted to give you any sort of indication on how we'll it'd work for you.

i would argue it is more valuable

Yeah, I think also efforts like Docling and similar are showing that smaller, specialized models with the right tooling might be more effective (and efficient) at this than throwing everything at Opus.

They don't seem to release as much open-weights at Mistral as they used to though :)

Stupid Europoors, optimizing for making a good product, instead of optimizing for making as much money as possible /s

big AI labs make so much money because they have a good (amazing) product

Do they even _make_ actual money? https://isaiprofitable.com/ seems to disagree.

Do they dupe VC into enormous datacentre and capacity build-out investment? Why yes, actual money going to the AI hypester pick-axe vendors (and early equity dumpers) in that sense, absolutely yes it does.

Is Big AI on track to pay that back with profit from any foreseeable and defensible business model? Different question. I sincerely doubt it.

I think one of the more sober analysis that focused on danish firms exclusively predicted the overall growth impact on GDP to be a bit less than half a percent. Now, well that doesn’t sound very large, considering nominal GDP for the world is nearly $60 trillion, that is a big deal. I don’t know if that justifies trillion dollar evaluations for the providers, but even a small improvement over a large base is meaningful.

I bet Sam, Dario and others have a good salary and an even better stock plan

The good thing about open capital markets is the possibility to invest in the upside. The downside is who gets holding the bag when the money runs out

You were a net loss for ~18y or even more and not many people were concerned.

I think it is rather disgusting to equate raising humans and training an LLM

This seems wrong to me on multiple different levels.

One, this conversation is about economy side of things.

And two, calling comparison between humans and LLMs disgusting when LLMs are borderline GAIs sounds very xenophobic. Did you see the thread about LLM-written fiction contest the other day? https://news.ycombinator.com/item?id=48782890

A net loss on the order of dozens of thousands of dollars. And one whose primary purpose is not wealth creation.

Conservative estimates in the United States puts it at $300,000 to raise a child from 0-18, including the cost of birth in the healthcare system. It goes up to $1 million for a HCOL area though.

> And one whose primary purpose is not wealth creation.

I take it you know your primary purpose.

I was a net loss for quite longer than that. Nobody was concerned because I was surrounded by people who loved me.

Big AI labs aren't making money. They're buying revenue. Sure, the product is amazing, but it wouldn't be as amazing if offered at cost - which is exactly where "good enough" smaller and specialized models will survive.

Anthropic is selling API tokens at 80% margin.

And API is 80% of their business (subscriptions the other 20%)

But they're still not making money (apart from two quarters when they got massive discounts from xAI).

That’s by choice. Spend more than you make and you can claim no profit. They could become profitable overnight by cutting spending, though that would also be akin to switching from a run to a walk in the AI race.

And yet, they are highly unprofitable. Yes, people pay for the API because it's a frontier model, but it's a frontier model because of billions of capex that are (so far) not getting recouped. And if they stopped the capex on new frontier models, that API revenue would walk off to whoever else does. If, at some point, the entire industry decides to stop burning money and start squeezing customers, that will be the test of which business models actually survive, and in that scenario I am bullish on scrappier shops that can't possible compete on all fronts now.

They are selling api tokens at good margins. Of course they then reinvest all of it in research for new models and for continued growth. But the api inference is profitable.

I'm not sure the "a year of document processing for under 100 USD/y" is such as great thing as you think it is (at least not for European competitiveness)... It means Mistral is essentially setting a revenue ceiling very low. OCR is a commodity at this point, and open source models, AWS, etc already do it out of the box.

Plus, you can't really build loyalty on a 100 USD/Y price tag. Since there are no switching costs holding them back, those buyers will leave the moment somebody offers a lower rate. An easily cloned, low cost tool with zero customer lock in is not a business. It is a feature.

That might sound great for the buyer (you), but it is a terrible strategy if we want a European company to compete long term against global competitors on actual product merit instead of just regulatory arbitrage.

well all commodities are like this. replace AI with milk, or plastic. It's easy for me to just move to different milk provider, this does not mean that milk industry is not a business.

And yes, its good that "its good for buyer" after all we do business so that living would be nicer, not the other way around (live to do business)

It does mean that the milk industry is a low value add commodity business where suppliers compete primarily on price and only survive due to protectionism and subsidies.

Food is important for national security so we should subsidize it, but it's a cost center. It'll never drive growth.

If that's what Mistral is aiming for, it would probably be better to give up now.

In some sense all commodities are "important for national security". Generic drugs are important for national security. Computer technology is important for national security. Im not sure what we arguing about. Im saying that AI is good thing to have like other commodity and it should net be one provider worth gazillion money.

Oil and gas is the type of commodity I’d rather it be modeled on, if I had to choose. Something with a lot of competition but an understanding that negative actions can easily affect your bottom line regardless of whether you were the one who took them, so generally better to operate with cooperative competition on the whole.

How we got from competition to assuming subsidies is beyond me.

There are a whole lot of commodity businesses that flourishes and that are profitable. It's true, that, yes, they will not have huge margins.

Grocery stores are like that - some of their suppliers might be subsidised but they are not, and many places they operate with typical margins in the 2% range. Discount supermarkets in the UK are operating on around 0.7% margins.

They are still huge, profitable businesses.

And they are examples of what happens when markets work.

> If that's what Mistral is aiming for, it would probably be better to give up now.

If Mistral is aiming not to grow to infinity but instead to offer niche models at reasonable prices, should they give up now? I think I'm not sure what exactly you're saying.

There are three different supermarkets in my neighborhood but they all sell milk from the same two suppliers. It would actually be quite difficult for me to turn to a third brand.

Which is an intentional failure of the system

How is that different from any other model provider, though? I used to use Anthropic for 100% of my code. Now I use GLM 5.2 for half of it, and as soon as something better appears, I'll use that.

Reply on news.ycombinator.com