Great to see the jhanas getting so much widespread coverage Still not totally sure how we had tools to create vast joy from nothing that stayed in deep obscurity for thousands of years, but I’m glad that’s changing at least a little

Don’t get hung up on whether you agree with his point on AI. Just listen to what he has to say about Star Trek vs dystopian sci fi stories. He’s grasping at something more important than AI. It’s what I’ve been trying to help people see for years: The story you repeatedly consume becomes the story you live into.

odin + raylib + wasm

my god there is so much perf engineering to be done. the servers are all so shit and buggy (sglang, vllm, anything else turbo garbo). the low level kernels/mlx are unoptimized. like I can’t believe my eyes when I see some of this shit. there so much low level work to be done.

Intel Arc B70 - Qwen3.8-27B - 256K context - exl3 - 4 bits - 16 sequences - 60 tok/s up to 330 tok/s at 16 sequences This costs 1600$

If you're vibe coding tools to help with your vibe coding project, you have an addiction and need to seek help.

So you want local AI and want to know what hardware you need? To get close to real-world, "cut the cord" local AI you need 256GB of memory in a decent package - can be DGX Sparks (TP-2), a Mac, or dedicated GPUs. It will cost you ~$10K USD or more depending on where you are and you can run GLM 5.3 Flash, Deepseek 4.1 Flash, or Qwen 3.8 Flash Next and many folks will be pretty happy. It's not going to be crazy fast but it will be usable - point OMP (or your favorite harness) at it and party on. Caveats: - even the best available model recipes at this level are not frontier equivalent - they're frontier-1 (or -2) at best, but most practical tasks and coding no longer strictly requires frontier -the models often barely fit, need to be carefully optimized, and unless on dedicated GPUs ($$$) will just barely hit acceptable real world speed and concurrency for one person doing serious stuff or a small handful of people doing light stuff - those boxes are gonna be busy and dedicated to model serving - little or no room left for sidecars < this is something that bit me and I ended up needing a 3rd spark 😂 That's not to say the next step down (in the 96GB-128GB - single spark - range) isn't useful. There are lots of very smart models you can run in that tier (and frankly on even smaller machines) that will do a lot of useful things. It's an an awesome place to start! I've been there personally - first with a Macbook 96GB that I've been through model hell and back with, then with my first Spark that performed admirably in a solo configuration on several workloads for months. Both of them did real work 24/7 in complete privacy and were awesome learning platforms, but neither ever came close to being able to host a model to replace my daily driver.

We're adding support for AGENTS.md to Claude Code. Starting today in version 2.1.277, if there is no CLAUDE.md in a folder, Claude will check for and use AGENTS.md. You can toggle this behavior in /config.

Introducing the Spring/Summer 2026 batch of Expeditions: an experimental fund organized around the quality of interestingness. Here are the six from this season: analogue.press/p/introducing…

GPT-5.5 lies constantly, Grok 4.20 doesn’t. We made a simulation to see if AI lies when the stakes are life & death.🧵

Hired my first employee today

Take all your global Claude Code Skills and make them available to Codex with this idempotent one‑liner: mkdir -p "${CODEX_HOME:-$HOME/.codex}/skills" && rsync -a "$HOME/.claude/skills/" "${CODEX_HOME:-$HOME/.codex}/skills/"