My company is slowly but surely getting hooked on AI coding. Our management is resisting it, because we know it’s self-destruction for us for a variety of reasons, but it’s like a drug: lazy engineers insist on using it and they’ll win out eventually because laziness is relentless.

So I’ve decided to preempt the inevitable and dive into this stuff, because one thing is concerning me: the lazy engineers ask Copilot to write their code for them, because it’s right there, already integrated in VSCode, and it’s bundled with other Microsoft services my company has unfortunately fallen pray to a few years ago when our last aging Unix engineer retired and was replaced by a MCSE monkey.

Of course, the tightly integrated ecosystem is very deliberate from Microsoft: they’re creating a powerful path of least resistance for the laziness and it works.

There are two problems with Copilot:

  • It’s Microsoft (no need to rehash why it’s generally best to stay away from Microsoft I guess)
  • Microsoft is headquartered in a country that has gone rogue, adversarial, unpredictable and is getting very, VERY Nazi.

All my career, since the 90s, I’ve argued against using Microsoft products with all my employers, and I’ve always been treated as an amusing paranoid lunatic. But this time, the Nazi argument is landing: my management is genuinely concerned that the company’s balls are in Microsoft’s hand, and that they can be instructed by the orange madman to wreck our business - along with all the other businesses in Europe - at a moment’s notice.

So I’m evaluation Claude Code an a step to at least divest away from Microsoft - still American, but fewer eggs in the same nasty basket. But ideally I’d like to evaluate a European AI company’s coding assistant, so I can be in an informed and authoritative position to make a recommendation that isn’t neither Microsoft nor American when the company finally decides to include AI use officially in the rule book.

Does anybody know a good AI coding assistant out of Europe I could look into? I don’t mind spending money on evaluation, but if there’s money to spend, I’d rather it’d be expensed on the right tool from the get-go if possible.

    • UnfairUtan@lemmy.worldEnglish
      2·
      4 days ago

      I use the chat from time to time to generate unit tests or debug. It’s good enough

      Recently they released mistral vibe, which is supposed to be their agentic CLI tool which should work with IDEs. I’m hesitant to test it for the same reasons I don’t use the others, but if I one day feel the need to go that route, I’ll choose Mistral regardless of performance.

      • PonyOfWar@pawb.socialEnglish
        3·
        5 days ago

        Yes, I use it a few times a week. Today I’ve used it to modify some 3d-Shader code for example. It’s pretty useful for generating short sections of code or modifying code I’m unfamiliar with. Never used it for coding entire projects or anything like that, but I think that’s a terrible idea with any AI.

      • Szewek@mstdn.social
        2·
        5 days ago

        @ExtremeDullard @PonyOfWar

        I often use their chat to ask about coding problems. It is very decent. You can get out of it the “sprawl” of potential solutions that some larger and “reasoning” models don’t have, as they are supposed to be more precise. I like the sprawl (in this context). If you know what you are doing, you have more to fish out from.

      • lemmysmash@beehaw.orgEnglish
        2·
        5 days ago

        I’ve used their devstral (latest one) + goose for a side project. It worked pretty decently, on par with Claude 3.7-ish Sonnet, maybe even better. And it’s not the largest: 123B. If you can have access to their larger models, that should be even better.

      • JustEnoughDucks@feddit.nlEnglish
        1·
        4 days ago

        I have tried it multiple times for embedded coding and it has never once produced valid code. It is like tutorials giving pseudo-code (but saying it is real code). So vibe-coding seems to not work with it.

        It can produce an idea that you can follow, so it is useful for exploring different methods for solving a problem, but that is its limit for me.

        That said, embedded code generally has quite specific libraries, but even directing it right to the library, right to the file, I was unable to coax it into using functions or APIs that exist.

        • ExtremeDullard@piefed.socialOPEnglish
          1·
          4 days ago

          Interesting. Thanks for the feedback. I reached out to them, see if they can help me evaluate their solution in the best possible conditions.

  • PracticalFail@feddit.orgEnglish
    4·
    5 days ago

    My company subscribed to Jetbrains Juno (or Juni?)
    Not using it actually yet as I don’t like those tools much too, but it seems to be good according to colleagues. Not sure though if it uses a M$, GPT, Claude or Gemini model via API behind the scenes. But at least Jetbrains gets a cut, which is a CZ company.

    If I use tools like those, I stick to Mistral.
    You can use something like aider.chat with a Mistral API key, or directly use the CLI tool provided by them https://mistral.ai/news/devstral-2-vibe-cli.

    I use the CLI from Mistral directly, works very well. But mostly use it only for generating test stubs and stuff like that.

  • almiukask@lemmy.worldEnglish
    51·
    5 days ago

    https://publicai.co/ It has API which could be used when calling functions to chat or complete code. I played around with VSCodium and available gpt extensions and tried to replase api calls with calls to Apertus. I found that VSCodium extensions I tried were not flexible enough for this, but will try this path further when I get some free time. As of now API is free, but you have to register as a developer. EDIT: Open-source model, public organizations which is a huge plus from my view

    • ExtremeDullard@piefed.socialOPEnglish
      3·
      5 days ago

      This is a bit concerning - from https://publicai.co/chat:

      1.13 Governing Law and Dispute Resolution

      These Terms are governed by Massachusetts law, excluding its conflict-of-laws rules. Disputes will be resolved in Massachusetts.

      1.14 Notices of Infringement (DMCA)

      If you believe content available through the Services infringes your rights, please submit a notice under our DMCA/notice‑and‑takedown process to our designated contact. Valid notices should include sufficient detail to identify the material and your rights claim. We may notify the affected user and, where appropriate, allow a counter‑notice.

      This sure don’t look very Swiss to me…

      • madnificent@lemmy.worldEnglish
        2·
        5 days ago

        The Apertus model is free and open. It can be ran on US and EU servers. I doubt you want to self-host these but you could. The smaller Apertus model can be ran locally with less than 16GB of VRAM but you likely want to run the larger variant.

        I suppose running an open model on US servers provides more security than running a closed model on US infrastructure. I don’t know who runs these in Europe as a service.

      • almiukask@lemmy.worldEnglish
        1·
        4 days ago

        AWS as I understood is their hosting service some things might come from that point. But as previous comment stated it is better to have open model running on US owned infrastructure, since the model, when needed, could be migrated sonewhere else

  • I Cast Fist@programming.devEnglish
    2·
    5 days ago

    You can rest assured that it is within M$'s interest to retain yuropeean customers, as 1 euro > 1 dollar. All the better if they can slurp all your data for free.

    • ExtremeDullard@piefed.socialOPEnglish
      7·
      5 days ago

      You can rest assured that Microsoft will fuck their European customers 10 times over before risking losing juicy multi-billion cloud contracts with the US government if the orange utan threatens them even a little bit.