hi i'm blake. i use this account to break down interfaces that should work but don't, and share my honest takes on AI and what it's changing about how we build. things i post about: • AI tools i've used for more than a week • UX decisions that make me want to open the inspector • things that shipped and worked, because those exist too if any of that sounds useful, stick around

saw that snap said AI is generating over 65% of their new code now not a pilot, that's just how they ship. what's interesting isn't whether this works, clearly it does. it's what the review and ownership process looks like when most of the code wasn't written by a human

ethan mollick, ai researcher at wharton, generated fully playable games from single prompts using fable 5 in claude code the generations included a snake, a dungeon crawler, even a walking game based on rilke poems. he said it outperformed every other public model by a considerable margin. i don't think most teams have internalized what "one person one prompt" actually means for how we build yet

stripe ran a 50 million line ruby codebase migration with fable 5 in a day a task that would've taken a full team two months. i keep waiting for one of these case studies to not hold up. still waiting

apple shipped a personalization slider for liquid glass today. you can go from ultra-clear to fully tinted giving users control over a design system element instead of just picking for them is the right call most apps still don't do this and it's annoying

you can go from nothing to a working interface in an afternoon now but you still can't tell from looking at the output whether it got what you meant or just produced something that passes at a glance nobody building these tools seems particularly bothered by that either

so microsoft just launched Scout. an always-on agent that lives in your calendar and inbox like an actual employee you name it, train it over time, and the more you invest in it, the harder it is to leave they know exactly what they're doing

github copilot switched to usage-based billing today still says $10 but the flat rate is gone, you're on credits now anyone running long agentic sessions is about to get a surprise invoice

a critical auth-bypass in Starlette is quietly sitting inside a huge chunk of AI agent infrastructure right now. FastAPI, vLLM, LiteLLM. if you've shipped an agent in the last year there's a decent chance you need to check this. CVE-2026-48710

Opus 4.8 is out the two things developers actually complained about with 4.7, tool calling and verbosity are both fixed same price. if you're still on 4.7 there's no reason not to switch

Figma shipping a native AI agent changes something subtle before this, AI in Figma was always a plugin, something bolted on. now it's inside the file, reading your design system, making decisions on your canvas who's responsible for what it gets wrong just got a lot less clear

Figma AI can now read your component library and generate screens that match your tokens and naming conventions it gets it right about 70% of the time honestly that's more than I expected but the 30% it gets wrong just looks fine and ships

the thing i keep seeing in AI-generated interfaces: technically clean, functionally empty. looks like it passed a design review. falls apart the second a real user touches it. speed got faster. taste didn't come with it