Posting this very very late but this is a good batch from Sept-Dec 2025.
- I started following this project from an X thread wherein a very talented developer used to working in Three.js moved over to the Godot framework for a week and ended up writing a very impressive x-wing fighter game.
- sandboxes, we’ve got ‘em
- Long-running background execution has long been an area marred by shitty DX. Kudos to Vercel for trying out something novel to fix it.
- Workflow DevKit - Make any TypeScript Function Durable
- boy do people have really strong feelings about directives! All metaprogramming is fine up until the point it isn’t, and it’s on you to know when that is. So, like, do it if there’s a reasonable payoff.
- The syntax doesn’t bother me because javascript decorators have been in contention for a decade and only apply to classes. Frameworks and typescript already rely on compiler tricks to do many things, so I’m not clutching my pearls over this.
- My only complaint is that its missing Vite support!
- jhey has been making wonderful ux demos for years and is kind enough to open source them all on codepen.
- jhey ʕ•ᴥ•ʔ on X: “Scroll masking with CSS scroll animation as a progressive enhancement ⭐️ .scroller […]” / X
- I always learn something new when I spend time picking through the examples he shares, painstakingly extracting the most relevant parts of the demos.
- I’ve been thinking about what my ideal database setup for rapid development looks like in 2025, and pgschema feels like one of the last missing pieces
- Mediapipe hand tracking interfaces are cool but unfortunately useless in practive
- Cursed CSS knowledge
- Linux From Scratch for GPT 2. This is essential reading for anyone in tech grappling with the changes of the last few years. What are models made of and how do we run them? Just like how small scale experiments can teach you about physics, growing a small model on your own GPUs is the best way to learn how the parts fit together
- I was skeptical of the usefulness of MCP last year, then warmed to it over time. The pendulum has swung back the other way as people discover unbridled proliferation of tool definitions quickly pollutes context. Short of automatic discovery, delegating to subagents with access to the right tools is the best current solution.
- Maybe some context compaction ideas in here? In practice this is a hard problem that nobody does great at.
- More and more rumblings like this in late 2025. Watch this space.
- Bullish
- Careerism isn’t great but keeping pace with titles and definitions never hurts
- Early internet nostalgia
- Many such cases where being consistently wrong doesn’t get you fired
- Any yet
- See also
- Finally, a good read