home.social

#leanstral — Public Fediverse posts

Live and recent posts from across the Fediverse tagged #leanstral, aggregated by home.social.

fetched live
  1. While I really don't have much love for the Mistral products due to being way behind even cheap open weights Chinese models such as the Qwens, DeepSeeks etc of this world, this is an interesting (possibly new) approach:

    heise.de/news/Leanstral-1-5-Mi / heise.de/en/news/Leanstral-1-5

    With the money quote being:
    "In addition to mathematical proofs, Mistral demonstrates a pipeline for automatic bug detection in Rust projects. The tool Aeneas translates Rust code to Lean, after which Leanstral derives correctness properties and attempts to prove or disprove them. "

    This goes past the very elaborated statistical parrots that LLMs are and goes into the formal verification territory.

    As much as I understand this, this is closer to a model checking approach than what we had before and would in theory thus even allow to check for concurrency issues like race conditions etc reliably.

    The hard part there of course was to provide the provably correct bridge between the code to be tested and an efficient model checker representation of the same code. Asking the model checker smart "questions" is another part as is translating findings back to the original code.
    Since the LLMs guzzle up RAM and CPU already, the consumption of the model checkers / theorem provers isn't an issue anymore (as long as the model they're being fed is efficient...)

    Lets see how well the LLMs do on this work.

    I'm sure there are already US or Chinese models that do this or will replicate it in a months time.
    I think I read about Meta experimenting on this as well.

    Did I get the announcement right? Has anyone had their hands on said model+prover combo already? Is anyone aware of Chinese/US equivalent contraptions?
    Any updates @heiseonline @heisedeveloper can provide on those questions in the article, especially on competitors?

    #llm #ai #formalverification #modelchecker #modelchecking #concurrency #correctness #meta #mistralai #aeneas #leanstral

  2. Leanstral 1.5: because proofing your #AI app with endless jargon and #buzzwords is so much more fun than actually making it useful. 🤖✨ Who needs clear communication when you have a "Vibe AI agent" that sounds like it came straight out of a yoga retreat? 🧘‍♂️💻
    mistral.ai/news/leanstral-1-5/ #Leanstral #1.5 #VibeAgent #clearCommunication #techHumor #HackerNews #ngated

  3. 🚀🐢 "Leanstral 1.5:" Because nothing screams 'innovative' like cramming a thesaurus into an #AI model and calling it progress. 🎉 119 billion parameters to overcomplicate simple tasks—now that's #efficiency, folks! 🧠🤹‍♂️
    docs.mistral.ai/models/model-c #Leanstral #1.5 #Innovation #Overcomplication #HackerNews #ngated

  4. How it's going #leanstral #lean #mistral

    Look how it is offering to prove Fermat's last theorem!

  5. Habe heute entdeckt, dass in der Fakultät ein bisher ungenutzter Server mit 2x NVIDIA V100 32GB stand. Und nach einem Treiberupdate lief darauf auch das #leanstral Modell mit vernünftiger Geschwindigkeit.

    Habe jetzt die 4-Bit Quantisierung genommen und dann passt es mehr oder weniger in den VRAM. Die Kiste hat 1TB normalen RAM, also kein Problem hier.

    Bin gespannt, was das kann!

    huggingface.co/jackcloudman/Le

    #lean #mistral

  6. 🚀✨ Breaking news: #Mistral launches #Leanstral, the "trustworthy" #vibecoding foundation that promises to eliminate the pesky human review process for AI-generated code... because who needs quality assurance when you've got "vibes"? 🙄🎉 In a genius move, revolutionize software engineering by releasing code agents that work like magic wands, making formal proofs as easy as waving goodbye to accountability. 🚀😜
    mistral.ai/news/leanstral #AIcode #SoftwareEngineering #HackerNews #ngated