small rant incoming

I work for a company that it’s mostly hardware focused, but we do ship some software for the final consumer including drivers and programs, to make use of said hardware.

While I am not in the software department, I held some SWE positions in previous companies for over a decade, our software isn’t very complex and I do know most of it pretty well.

Our employee just announced a new AI-only development cycle, they want all code submissions and reviews to be exclusively done by claude, effectively ending ownership of the code being shipped to customers. This is absolute madness.

Today, I received an email scheduling a workshop on how to integrate claude into vscode and how to work with the new gitflow, namely removing our authorship from commits and having al code reviews done by a LLM now.

I am just baffled at the decision.

edit: wow I’m a bit overwhelmed by the response, I did read all of you. Thanks!

  • Jo Miran@lemmy.ml
    101·
    18 hours ago

    For the most part, I am an AI sceptic. My firm has had access to and beta tested every Google AI model since before the AI had a name (yes, well before Bard). We also have access to ChatGPT, Claude and locally run models on dedicated hardware. I’ll save you the suspense, at the moment it’s all mostly trash. Barely a proof of concept. Definitely not production ready, but there are a few exceptions I, personally, have found.

    Gemini Pro is excellent at identifying and diagnosing sick garden plants if you take and upload a photos of the plants. Unfortunately, it still cannot do simple arithmetic. A simple SUM that any spreadsheet does automatically by just highlighting the cells, and somehow it was still off by one.

    For software dev, the only one product we have found to be useful is Claude, BUT it really depends on the model. Haiku is a joke. Sonnet is not bad, if not bad is ~50% accuracy. Opus is fairly descent at building scaffolding that you can then correct and complete, but don’t expect efficient code.

    The only model we have ever tested that can deliver ALMOST fully functional code, reliably diagnose errors, and review code is Mythos/Fable. Genuinely not absolute shite, BUT that shit is extremely slow and outrageously expensive. I seriously doubt that your company is forking out the money for Claude Fable as their default model, and if they are, they must not care about ever making a profit.