Same site, same CDN. I swapped Nuxt 4 for SolidStart and mobile PageSpeed went from 70 to 99, with 0 ms TBT.
No amount of Cloudflare tuning could buy those 29 points. The framework swap could.
Full writeup, including the two places I still run React ↓
Same site, same CDN. I swapped Nuxt 4 for SolidStart and mobile PageSpeed went from 70 to 99, with 0 ms TBT.
No amount of Cloudflare tuning could buy those 29 points. The framework swap could.
Full writeup, including the two places I still run React ↓
KeyEcho 1.0:
• free & open source (<5 MB, no account)
• signed + notarized builds — antivirus false-positive era ends
• paid studio Lab Packs — you vote what gets recorded first
No download needed — open https://t.co/IP3uHl2fLW and just type.
Write-up: https://t.co/wtpPFmVnc4
One personal take:
Claude Fable feels like a talented engineer who thinks outside the box — ideas, angles, things I missed.
GPT-5.6 is simply the best executor right now — near-perfect implementation.
My workflow: Fable thinks, 5.6 builds, I decide.
Why Rust is my first choice in the AI era:
Performance is table stakes. The strict compiler is the real asset — the borrow checker is the first referee for AI-generated code. Bad code doesn't survive cargo check.
The more the copilot writes, the more that guardrail is worth.
The speedup comes from the refactor itself — predecode, shared buffers, no lock, bounded queue.
What GPT-5.6 did: it made finding, doing, and verifying that refactor fast enough to actually happen.
Copilot, not autopilot.
The core is a rebuilt audio hot path (Rust + Tauri):
Old keypress: OGG decode + sample copy, behind a global lock.
New keypress: one event → bounded queue → clone one Arc<[f32]> handle.
1184 ns → 43.5 ns. Zero copies. Every keystroke.
Reproduce it: pnpm run bench:audio
GPT-5.6 Sol launched today.
My open-source app — 800+ stars, untouched for two years — just shipped its 1.0 with it: one 130-file PR, and a keypress path that's 27–38× faster with zero copies.
Two years of putting it off. Done in a day. 🧵
Sent a reactivation email to a newsletter that went silent for over a year — ~18,000 people, all on my own self-hosted stack.
0 spam complaints, 99.9% delivered.
"Dead list" usually means you stopped showing up — not that they hate you.
Ran a review-history miner on our own team repos today.
Merged PRs: plenty. Review comments: zero.
Fast-moving AI-native teams don't write review threads anymore — the team's judgment now lives in CLAUDE.md files and mid-session corrections to agents. Interesting shift.
Went quiet for over a year. Today I relaunched my newsletter — and I run the whole thing myself: self-hosted list, my own sending stack, no Substack, no beehiiv.
First send landed clean: 0 spam complaints, ~27% opens, real replies already coming in.
Building the rest in public.