60% productivity loss whenever I try to use it for anything else than a code search engine.

Even I as amateur Python dev am able to notice this. It’s basically a nicer (and more resource-intensive) StackOverflow for basic questions when your brain lags again or you’re stuck with how a library is documented (looking at you GTK). The best thing you can do with LLMs is not integrate it into your IDE but just have it open in a browser tab, and never ever copy anything out of it. Saves money, time and IQ points.

Also worth mentioning: It’s extremely noticeable how different US companies tuned their models (unfortunately chinese models widely copied them). I was curious due to the whole Digital Independence things going on in the EU and did some tests with Mistral (EU) vs. ChatGPT (US).

My core takeaways:

• Mistral refuses to open any website that tells LLMs to fuck off (it’s sad this is even noteworthy, but all US companies just work around every block). This isn’t an issue if you host your knowledge DBs yourself.
• Mistral will more likely tell you “I don’t know”, “This might be incorrect” or “you should double-check this” if it can’t find any resources on the question. ChatGPT instead will instead answer something as confident as always and hallucinate while coating you in honey.
• Mistrals WebUI isn’t infested with invasive Javascript. And as a sidenote, the European model also refuses to generate any pixel slop even in the style of smaller artists. ChatGPT doesn’t give a shit unless it’s the IP of another US megacorp. But nobody should generate such slop anyway.

Basically the US models are made to be as addictive and ruthlessly profitable as possible, always giving you something while sugarcoating your mistakes (even if that’s wrong). It’s gambling. Those who’re celebrating these things might already be psychologically addicted without realising it.

(This doesn’t mean Mistral is great, just not as awful. Their tiny models are already used in weapons in Ukraine, and it’s still AI that sucks the planet dry)