3 ms·

Going on a tangent, is Europe even close? Mistral has been underwhelming

LaurensBER 1y ago
Not anywhere near close. Europe doesn't have the infrastructure (legal or energy) and US companies offer far better compensation for talent. But hey, at least we have AI regulation! (sad smile :))
troyvit 1y ago
I don't know if how close Europe is, but I'm sufficiently whelmed by Mistral that I don't need to look elsewhere yet. It's kind-of like having a Toyota Corolla while everybody else is driving around in smart cars but it gets it done. On top of it, there's a loyal community that (maybe because I'm not looking) I don't see with other products. It probably depends on your uses, but if I spent all my time chasing the latest chat models (like Kimi K2 for instance) I wouldn't actually get anything done.
- dehugger 1y ago
  Probably cuz you aren't looking yeah. Anthropic seems to be leading the "loyalty" war in the US.
  - troyvit 1y ago
    Yeah and I'll probably end up going that way as work locks down the models we're allowed to use, saving Mistral for personal projects.
- rurban 1y ago
  You have to try the latest Corolla then. Really smart. Lane and collision assistance, ... Unlike my old Corolla which is total dumb. It even doesn't turn the light off when I leave the car
  - troyvit 11mo ago
    Hah! I need to update my analogies.
- KronisLV 1y ago
  > I don't know if how close Europe is, but I'm sufficiently whelmed by Mistral that I don't need to look elsewhere yet. It's kind-of like having a Toyota Corolla while everybody else is driving around in smart cars but it gets it done. My problem was that it really doesn't, none of the models out there are that great at agentic coding when you care about maintainability. Sonnet 4.5 sometimes struggles and is only okay with some steering, same for Gemini Pro 2.5, GPT-5 recently seems like it's closer to "just working" with high reasoning, but still is expensive and slow. Cerebras recently started offering GLM-4.6 and it's roughly on par with Sonnet 4 so not great, but 24M tokens per day for 50 USD seems like good value even with 128k context limitation. I don't think there is a single model that is good enough and dependable enough in my experience out there yet, I'll probably keep jumping around for the next 5-10 years (assuming the models keep improving until we hit diminishing returns so hard that it all evens out, hopefully after they've reached a satisfying baseline usefulness). Don't get me wrong, all of those models can already provide value, it's just that they're pretty finnicky a lot of the time, some of it inherent due to how LLMs work, but some of it also because they should just be trained better and more. And the tools they're given should be better. And the context should be managed better. And I shouldn't see something as simple as diffs fail to apply repeatedly just because I'm asking for 100% accuracy in the search/replace to avoid them messing up the brackets or whatever else.
  - epolanski 11mo ago
    Coding isn't the only use case. Neither is being bleeding edge. I use Mistral's models, I've built an entire internal-knowledge-pipeline of sort using Mistral's products (which involved anything from OCR, to summarization, to linking stuff across different services like Jira or Teams, etc) and I've been very happy with it. We did consider alternatives and truth to be told none was as cost-effective, fast and satisfying (and also our company does not trust US AI companies to not do stuff with our data).
    - r_lee 11mo ago
      So you're not able to trust inference providers like Google Cloud w/ ZDR etc with your data?

Not anywhere near close. Europe doesn't have the infrastructure (legal or energy) and US companies offer far better compensation for talent. But hey, at least we have AI regulation! (sad smile :))

troyvit 1y ago

I don't know if how close Europe is, but I'm sufficiently whelmed by Mistral that I don't need to look elsewhere yet. It's kind-of like having a Toyota Corolla while everybody else is driving around in smart cars but it gets it done. On top of it, there's a loyal community that (maybe because I'm not looking) I don't see with other products. It probably depends on your uses, but if I spent all my time chasing the latest chat models (like Kimi K2 for instance) I wouldn't actually get anything done.

dehugger 1y ago
Probably cuz you aren't looking yeah. Anthropic seems to be leading the "loyalty" war in the US.
- troyvit 1y ago
  Yeah and I'll probably end up going that way as work locks down the models we're allowed to use, saving Mistral for personal projects.
rurban 1y ago
You have to try the latest Corolla then. Really smart. Lane and collision assistance, ... Unlike my old Corolla which is total dumb. It even doesn't turn the light off when I leave the car
- troyvit 11mo ago
  Hah! I need to update my analogies.
KronisLV 1y ago
> I don't know if how close Europe is, but I'm sufficiently whelmed by Mistral that I don't need to look elsewhere yet. It's kind-of like having a Toyota Corolla while everybody else is driving around in smart cars but it gets it done. My problem was that it really doesn't, none of the models out there are that great at agentic coding when you care about maintainability. Sonnet 4.5 sometimes struggles and is only okay with some steering, same for Gemini Pro 2.5, GPT-5 recently seems like it's closer to "just working" with high reasoning, but still is expensive and slow. Cerebras recently started offering GLM-4.6 and it's roughly on par with Sonnet 4 so not great, but 24M tokens per day for 50 USD seems like good value even with 128k context limitation. I don't think there is a single model that is good enough and dependable enough in my experience out there yet, I'll probably keep jumping around for the next 5-10 years (assuming the models keep improving until we hit diminishing returns so hard that it all evens out, hopefully after they've reached a satisfying baseline usefulness). Don't get me wrong, all of those models can already provide value, it's just that they're pretty finnicky a lot of the time, some of it inherent due to how LLMs work, but some of it also because they should just be trained better and more. And the tools they're given should be better. And the context should be managed better. And I shouldn't see something as simple as diffs fail to apply repeatedly just because I'm asking for 100% accuracy in the search/replace to avoid them messing up the brackets or whatever else.
- epolanski 11mo ago
  Coding isn't the only use case. Neither is being bleeding edge. I use Mistral's models, I've built an entire internal-knowledge-pipeline of sort using Mistral's products (which involved anything from OCR, to summarization, to linking stuff across different services like Jira or Teams, etc) and I've been very happy with it. We did consider alternatives and truth to be told none was as cost-effective, fast and satisfying (and also our company does not trust US AI companies to not do stuff with our data).
  - r_lee 11mo ago
    So you're not able to trust inference providers like Google Cloud w/ ZDR etc with your data?