15 ms·

What is this whole thing about Europe being behind on AI? Do Mistral and DeepL not exist? Yes, I know DeepL is niche, but IMHO it is the best translation model

by acatton 3mo ago

What is this whole thing about Europe being behind on AI? Do Mistral and DeepL not exist? Yes, I know DeepL is niche, but IMHO it is the best translation model out there.

epolanski 3mo ago
It is behind in the sense that if tomorrow the US and China place an export ban on their models all we're left with are Mistral's ones. They are not bad, and they have made huge progress, but you're still one year behind if not more. May matter less and less as time progresses, or it may matter more if research further speeds up. Honestly I wish capitalism and globalization kept working as they did for decades, but since more than a decade we're reverting to inefficient protectionist steps, one after the other.
- Slartie 3mo ago
  How exactly do you - or the Chinese, for that matter - place an export ban on an open weight model?
  - epolanski 3mo ago
    By not releasing them anymore.
SyneRyder 3mo ago
I've tried using Mistral for various tasks, and it is so far behind the American models that I just never bother using it despite still having lots of Mistral API credits leftover. Even their OCR and TTS products are surpassed by generic US models - I use regular Claude Sonnet for OCR because it is more accurate than Mistral OCR. I could rant about this, I am just so disappointed at how Mistral completely gave up and pivoted into bespoke fine-tuning consulting. The terrifying thing is that they don't seem to even understand how far behind they are, as if they never tried Opus, let alone Fable / Mythos. Or they do understand and that's why they focus on consulting now.
- JumpCrisscross 3mo ago
  Did Mistral really throw in the towel on building frontier models?
  - SyneRyder 3mo ago
    Well, the front page of their website claims "Frontier AI In Your Hands", so I guess they're not marketing it that way. I personally think there's a hint in that Mistral Medium 3.5 costs 5x the price of Mistral Large 3, and that Mistral Large is not listed anymore as a "Featured Model" and hasn't been updated since Dec 2025: https://docs.mistral.ai/models/overview https://docs.mistral.ai/models/overview But what I really base it on is an interview the Mistral CEO gave on the Big Technology Podcast back in January this year: Alex Kantrowitz: "Do you consider yourself, is the most important thing you do building the models? Or is the most important thing you do the service? Are you primarily a model builder, or primarily a service provider?" Arthur Mensch: "We are there to help our customers get to value." Alex Kantrowitz: "So, service!" Arthur Mensch: "We are here to... but to get to value, they need to have great models. And to get to value, they need to have the right tools to train the models. And so the best way to train, to create those tools, is effectively to train the best models. So the two things are extremely linked together. We create models that are very easy to customize. We create models with tools that we then export to our customers, so that they can use them, and we help our customers train their own models. You can't go and sell to an enterprise that you are going to help them create great custom systems, if you can't show to the world that you are effectively the leader in open source technology. So the two parts are equally important, the first is enabling the other, and there's effectively a flywheel there because we make our choices when it comes to the model design in a way that is enabling the various customers we have. As one example, we've put a lot of emphasis on having models that are great at physics, because we work with manufacturing companies that run into physical problems. So that's the flywheel we have set up. Having the science team and the business team sit together." It's at 22:37 in the video. Elsewhere in the podcast he mentions that they don't believe in a large unified generic model, they think the future of AI is small dedicated-task models (OCR, TTS, bespoke trained)... but unfortunately I don't have a timestamp link for that part. https://www.youtube.com/watch?v=xxUTdyEDpbU&t=1357s https://www.youtube.com/watch?v=xxUTdyEDpbU&t=1357s

by acatton 3mo ago

What is this whole thing about Europe being behind on AI? Do Mistral and DeepL not exist? Yes, I know DeepL is niche, but IMHO it is the best translation model out there.

epolanski 3mo ago

It is behind in the sense that if tomorrow the US and China place an export ban on their models all we're left with are Mistral's ones. They are not bad, and they have made huge progress, but you're still one year behind if not more. May matter less and less as time progresses, or it may matter more if research further speeds up. Honestly I wish capitalism and globalization kept working as they did for decades, but since more than a decade we're reverting to inefficient protectionist steps, one after the other.

Slartie 3mo ago
How exactly do you - or the Chinese, for that matter - place an export ban on an open weight model?
- epolanski 3mo ago
  By not releasing them anymore.

SyneRyder 3mo ago

I've tried using Mistral for various tasks, and it is so far behind the American models that I just never bother using it despite still having lots of Mistral API credits leftover. Even their OCR and TTS products are surpassed by generic US models - I use regular Claude Sonnet for OCR because it is more accurate than Mistral OCR. I could rant about this, I am just so disappointed at how Mistral completely gave up and pivoted into bespoke fine-tuning consulting. The terrifying thing is that they don't seem to even understand how far behind they are, as if they never tried Opus, let alone Fable / Mythos. Or they do understand and that's why they focus on consulting now.

JumpCrisscross 3mo ago
Did Mistral really throw in the towel on building frontier models?
- SyneRyder 3mo ago
  Well, the front page of their website claims "Frontier AI In Your Hands", so I guess they're not marketing it that way. I personally think there's a hint in that Mistral Medium 3.5 costs 5x the price of Mistral Large 3, and that Mistral Large is not listed anymore as a "Featured Model" and hasn't been updated since Dec 2025: https://docs.mistral.ai/models/overview https://docs.mistral.ai/models/overview But what I really base it on is an interview the Mistral CEO gave on the Big Technology Podcast back in January this year: Alex Kantrowitz: "Do you consider yourself, is the most important thing you do building the models? Or is the most important thing you do the service? Are you primarily a model builder, or primarily a service provider?" Arthur Mensch: "We are there to help our customers get to value." Alex Kantrowitz: "So, service!" Arthur Mensch: "We are here to... but to get to value, they need to have great models. And to get to value, they need to have the right tools to train the models. And so the best way to train, to create those tools, is effectively to train the best models. So the two things are extremely linked together. We create models that are very easy to customize. We create models with tools that we then export to our customers, so that they can use them, and we help our customers train their own models. You can't go and sell to an enterprise that you are going to help them create great custom systems, if you can't show to the world that you are effectively the leader in open source technology. So the two parts are equally important, the first is enabling the other, and there's effectively a flywheel there because we make our choices when it comes to the model design in a way that is enabling the various customers we have. As one example, we've put a lot of emphasis on having models that are great at physics, because we work with manufacturing companies that run into physical problems. So that's the flywheel we have set up. Having the science team and the business team sit together." It's at 22:37 in the video. Elsewhere in the podcast he mentions that they don't believe in a large unified generic model, they think the future of AI is small dedicated-task models (OCR, TTS, bespoke trained)... but unfortunately I don't have a timestamp link for that part. https://www.youtube.com/watch?v=xxUTdyEDpbU&t=1357s https://www.youtube.com/watch?v=xxUTdyEDpbU&t=1357s