I used Figma a lot in the last couple of weeks without AI and it was good. Great, even! Got into flow, mocked up some pretty pixels, and didn't fall into a hell-loop of fighting the horrendous taste of AI models. I predict a (slight) backswing on the "Figma is dead, designers should only prototype in code" discourse.
I'm juggling Codex, Claude, and Devin for background coding agents. I hadn't used Devin before (the new cloud-only web version) and it's really well executed. I hope they'll ship a simple electron wrapper around it so I can have a proper desktop app soon. Their Fusion model router seems like it's working; at least, I haven't had to waste any energy fiddling with model pickers over there which is a welcome change.
I recently worked on Notion AI's model picker and boy howdy, let me tell you, this is one of the The Problems Of All Time. I hope in the future nobody has to waste brain cells thinking about which model and which effort level to use for every task. Unfortunately, that future seems really far away because, it turns out, if you actually know how to wield all these things correctly you can get much better results. So everyone builds for the top of the class and normies are left trying to decipher the matrix of model/effort/speed levels to do basic things.
Modern AI voice modes make me incredibly depressed about the future. I do not want a companion. I do not want an AI that coddles me and injects "ah, mmhmmm, yeahs" and fake laughs into its responses. I want a god-tier computer that does shit I tell it to do and doesn't pretend to be something it is not.
No matter how hard I try, AI writing output is just horrendous. If you use AI to write things at work and share them without careful review/editing: everyone can tell.
I still enjoy using AI to brainstorm and help me draft outlines or draw diagrams and stuff like that. Using AI to help me think through problems + draft an outline seems all good. It's the final output that is always just so incomprehensibly dumb and uncanny-valley that it's actively making me want to use AI less.
At the same time, I have to be extremely careful with how I prompt; if I indicate in any minor way that I like something, or think something is a good idea, the model will go into full human-reinforcement mode and bend over backwards to make itself agree with my point of view.
I'm pretty sure that all the skills I wrote over the course of the last year are making the latest generation of models do dumb stuff or work too hard. I'm planning to delete every skill/contextmd file and start from scratch to see what happens.
Related: this is why I'm generally bearish on auto-generated memories. I know memory is a hard problem that every lab is trying to solve, but it just doesn't feel like we've made any meaningful progress in the last year. Every time one of my AI agents references a past memory it feels forced and out of context, like it's trying to prove to me it remembers things for the sake of remembering things, not because it's actually useful in that particular context.
Being able to ask models to search through chat history is good, though.
It's way more fun to use coding agents on backend tasks than it is on the front end. The gap between what I know is good and what AI is capable of making on its first pass is still surprisingly wide. And I despise prompting my way through visual polish. Every coding tool is abandoning the IDE, assuming that nobody will ever read or touch code again. That's probably long-term right but short-term frustrating for people who care about getting the details right.