@Tutanota@mastodon.social @jab01701mid@mastodon.social @pyratebeard@harbour.cafe i would implore you to stick to 100% without AI. email, especially with the emphasis on privacy and strong encryption that you are going for, is security-relevant. that means, not just is it a risk that these models give you probabilistic output that may need extensive checking before it should ever go into anything important (entirely negating the speed benefits), but that there is also incentive for secret services to approach AI vendors to use any trust you are placing in the output (including AI code reviews) to sneak backdoors in. the latter tends to get overlooked in discussions, since most people are focussed on other aspects like the massive resource use, the essentially illegally obtained training data, and the practical suitability of LLM output for incorporating in software development. also note that we keep seeing people (including some otherwise very knowledgable and talented people) who use LLMs experiencing increasing lapses in judgement; i have a pinned post on my profile coining the term "lyssanetic" for this property, but i've also seen the phrase "vibe sickness" floating around for this, which is probably the easier terminology. for i think obvious reasons, i would really prefer not having to worry about this affecting people on the tuta team. i would generally strongly prefer for tech companies to move slower, but more deliberate these days, according to the credo "move slow and fix things" (an inversion of facebooks now metas "move fast and break things"). i don't need new features faster, i need things that work well and reliably. the fact that you are currently still standing at ca. 99% means you can still relatively cleanly turn back.