Hacker News Reader

Comment by HenrikPontoppid | Hacker News Reader

When it comes to what really matters, they're far behind and probably will not catch up. It is my conviction that in this space, if you're not the best in the world, you're losing. Everyone is fundamentally selling the same thing, so if you don't have the most intelligent or cheapest model in the world, you're losing. Sure, Mistral has a "Made in Europe" edge, but any open-weight self-hosted chinese model might as well have been made by Von der Leyen herself.

Also, €3B is nothing in this market, especially when you're competing with more efficient competitors. €3B in Europe is probably the same as €10B in the USA and €20B or €30B in China.

View on HN

> It is my conviction that in this space, if you're not the best in the world, you're losing

I disagree on that point. Models are getting good enough that you can switch them and barely notice. I'm switching between Opus, GPT Codex and GLM 5.3 for coding and I can barely tell the difference.

I think they'll become more like telcos than anything, selling a commodity. It's even truer when any provider can host open weight models like GLM-5.3.

Basically a world with dozens of Baseten, with AI labs having a hard time monetizing, just like editors of open source software.

I can tell the difference between the SOTA and GLM 5.3, and paying a few hundred for the better models is definitely worth it.

Mistral does not offer a model that makes sense to use. I understand they now host the already outdated GLM 5.2, courtesy of China providing the weights. And Mistral offers this for 2x-3x the price of other providers.

This is supposed to be a success story?

I'm sure you can tell the difference.

But for the vast majority of usage GLM 5.3 or Deepseek V4 is enough.

Even people who do need frontier model will soon restrict it to the use cases that really need it and switch to cheaper models for the rest. It has already started.

Anthropic and OpenAI will never get enough customers paying top dollar to deliver on the revenue they need to offset their investments.

I suspect that opinions in that direction might be the result of a lack of ambition in applying agentic AI.

The cheap models are good enough for what exactly? AI assisted development, or working autonomously on a task for 4 hours?

As long as the best available model does the latter more reliably and with noticeably better results, it is easily worth spending a few hundred per month for me.

The idea that OpenAI and Anthropic cannot make enough profit in my opinion depends entirely on how large the gap is going to be.

Will the gap become smaller with improvements starting to slow, or will it get wider as the labs successfully apply their models to research and improvements speed up?

You call it "lack of ambition", implying a certain arrogance; more humble people might just call it "a different sweet spot" of AI integration. After all, the goal in the end is not to use up as many AI tokens as possible, but the goal is to implement/perform a certain business case in the best way possible with regard to cost/speed/quality.

It's not like everyone using AI has the same software development cycle. It's not even that everyone using AI is developing software.

I propose that there is no such sweet spot.

If the model cannot perform the task with adequate quality while being many times faster and cheaper than a human, it is not good enough in my view.

If a SOTA model would be good enough, surely that would be preferable to a human using cheap AI assistance for a moderate efficiency improvement.

A harness of limited capability is a component that doesn't exist in your proposal, right? Just like non-automatable or at least hard-to-automate things like physical actions to be performed in the real world as part of a business task do not exist, too? Same goes for coordination with other real humans - doesn't exist.

Why don't you give some examples of tasks that you think fall into this category, where AI assistance only is the perfect solution and which could not conceivable be done more effectively autonomously by more capable models?

For coding, I don't see it.

Physical tasks? I don't know, it's not my area of expertise. I do not think AI is widely deployed for physical tasks yet, not even the SOTA is good enough.

I don't get it when people all claim that AGI is a winner takes all game. It is not (unless it is used as a weapon). When one company reaches AGI, there will be a dozen very close to AGI, given time. Also a winner is not going to drive everyone else out of business, it is the opposite, one winner will have people betting on the second and the third winners, the technology will also help other develops. Once you have a good enough model, everything will be incremental. I think the hardware capability will be the real burden, not the model itself. If AGI is as powerful as it sounds, maybe hardware won't be a problem any more.

It’s very possible that the people saying AGI is a winner takes all are indeed referring to warfare.

how could I naively think this is going to be the technology to free human from being the slave of work.... of course, it is going to be a weapon, and it is only powerful when no one else has it.

> any open-weight self-hosted chinese model might as well have been made by Von der Leyen herself.

This is simply wrong. Plenty of security-sensitive companies and agencies disallow the use of any Chinese model out of hand. Given the black box nature of LLMs, the fact that it's "open weight" is irrelevant to the trust calculus, and the fact that it's self hosted barely adds anything.

Yes but what about taking an open weight Chinese model and do some extra pre-train and post-train to remove any potential hidden backdoors?

How do you confirm that you did enough pre-train and post-train to remove the relevant bias/backdoor?

Note that I'm not singling out Chinese models here. The same question and concern should apply to any model you use - in security critical scenarios, you need to ask yourself if you can trust the provider, since the artifact, the model itself, is far too complex to audit and trust.

Given most AI companies' ties to the weapons industry and spying apparatus of their host countries, it would be absurd to think that burying some kind of malicious payload in the models has not at least been contemplated, and I would bet it has been at least attempted. We know retrospectively this was the case with many previous technologies (remember the backdoored encryption that the NSA tried for so long to push), so it would be naive to think it's not at least plausible with AI.

Europe is a different beast. They build and distribute rails, they don’t compete on frontier capabilities. When the civic benefits of AI become clear, EU is in a position to mandate their distribution. US develops capabilities that remain stuck in heterogeneous corporate silos without interop. Payments is a good point of comparisons between the two approaches.

A new ICE line between Antwerpen and Koln opened up the last week. The first train was half an hour late, the distance is under 250 km.

That is not great. But it is still better than other continents?

also electronics, social networks and electric cars

Well, then Samsung Electronics appears to be dumb for leading this investment round? They probably only do it because they're European...oh, wait...

> so if you don't have the most intelligent or cheapest model in the world, you're losing

So how does this match up to the fact that there is currently OpenAI and Anthropic, both raking in money? They can't both have the smartest model at the same time, can they? And all those inference companies selling API access to open weight models on OpenRouter, which are apparently also earning billions already? While the former are probably bound to have much higher cost for research and training than they are currently earning, which may be called "losing", the latter don't have that problem, they can simply price their API access such that the money earned covers their costs, no training and practically no research necessary. In your theory these companies shouldn't have a cent of earnings.

The value of these companies is not defined entirely by the state of their current models. A much more important signal is their chance of having the best model in the future. Just like with anything having to do with investment, this is the castle in the sky. Also, shame on me for simplifying things so much - but I still believe in my original sentiment.

Replce "best" with "most fit for purpose" and you're right. The future won't be a one size fits all kind of deal because hardware availability constraints will eventually punish oversized models even more.

I think that there's a real second mover advantage in letting others waste money on researching ultra oversized models while creating smaller and cheaper models from their learnings.

OpenAI and Anthropic are so close on benchmarks and release date that they are essentially ex-aequo at this point. The market is just hedging their bets.

There is no close second, because the AA Index points are expentially harder to get as you get closer to the first.

> So how does this match up to the fact that there is currently OpenAI and Anthropic, both raking in money? They can't both have the smartest model at the same time, can they?

They are simultaneously first: the two leapfrog each other with regularity, and are meaningfully ahead of the competition.

Samsung is now essentially tied for being the world's most profitable company with Nvidia ($62b operating profit last quarter, vs $63b for Nvidia). They're drowning in cash. They're doing the same thing Nvidia has been doing: distributing the money to their business partners, to try to drive business growth faster (and or to keep it all propped up).

Reply on news.ycombinator.com