Machine learning engineer. I build AI tools, CLI applications, and civic tech, and write about the parts that were harder than they should have been.
Qwen3.8-Flash-Next reached 93.25 decode tokens/s after a 489-second pause. The complete experiment record includes TensorFold, MTPLX, GLM, MiMo, and failed tests.
Two acronyms that share two letters and almost nothing else. RLHF is a training loop that ends in new weights. A Recursive Language Model is an inference loop that keeps a long input in a code runtime and lets a frozen model compute over it. Five diagrams, one grid, and where prompt optimizers such as GEPA sit between them.
An interactive /mcp menu for Codex, with a real terminal demo. My first toggle waited seven seconds for tool discovery; the revised version saved in 46 ms while the server connected in the background. How the status-only API works, why a checked box did not prove tool access, and where the patch lives after GitHub refused the upstream PR.
My iPhone was dead somewhere in the house. Find My said Home, Now, and then refused to ring it. A passive Bluetooth scan on a laptop carried room to room closed the last thirty metres: 120 packets, five rooms, one metre of final uncertainty. The trick was tracking the beacon by state rather than by id, because the id rotates every few minutes and takes your meter down with it.
I scanned 1.2 GB of my own Claude Code and Codex transcripts for credentials and found 373 across 105 files, including a live API key repeated 60 times and a plaintext password. The script is a single file with no dependencies. The write-up is mostly about the four bugs that got past a green test suite.
Five years of image search in one place. CLIP made text a query for pictures, compression got a proof and moved inside the graph, cross-modal queries turned out to break the assumption every index was built on, and storing a vector fell from memory prices to object-storage prices. A sourced synthesis, with no original benchmarks.
A measured MLX vs llama.cpp comparison on a 512GB Mac Studio M3 Ultra at a 234,158-token prompt. mlx-lm 0.31.3 decodes 1.49x faster than llama.cpp b10200 and ties on prefill, while llama.cpp uses 29% less system-wide memory and outranks every MLX serving stack measured on decode. The llama.cpp arm is a single unrepeated run, and this post is mostly about why that matters.
Sixteen model and engine configurations at a real 262,000-token prompt on a 512GB Mac Studio M3 Ultra. Nemotron-3.5-Lightning-30B-A3B wins at 60.45 tok/s decode with 8/8 needle recall, and it wins by degrading slowly: its decode falls 1.83x from 2k to 262k where a dense 14B falls 10.68x. Architecture beats parameter count, a 1M-native model fails at 1M, and all five recall failures I published turned out to be my own harness.
My LLM benchmarks kept collapsing 4x mid-session. macmon caught a firmware power limiter clamping the GPU to 338 MHz and holding it while the die cooled, fans never past 70%. Pinning them at max with fanpro: 2.57x sustained decode, and a 100k-context job in 259 s instead of 568+, byte-identical.
Last post I called local 1M context a mirage. Then Kimi-Linear-48B passed: 8/8 needle recall at a real 1,048,692-token prompt on a 512GB M3 Ultra, full ingestion proven. Native 1M is finally real. It just takes a five-hour prefill.
I ran a 754B model on a 512GB Mac Studio. It fits, it crawls, and its 1M context flag is a warning message and a silent cap. Where local models fail as coding tools, and the one job they are great at.
Easy-Switch moves the keyboard and abandons the mouse. One HID++ packet fixes that, once you get past the two permission gates macOS hides between you and your own devices.
Two switches I leave on: the advisor tool and /effort ultracode. An honest, log-backed account of where they help, where they are overkill, and where the model wisely skips them.
An opinionated Cloudflare-first template. Next.js 16, D1, Better Auth, Stripe. No monorepo. No edge-runtime religion. No status page.
Head-to-head on a Mac Studio M3 Ultra. 26 prompts, 6 categories, and one surprising finding: the model that decoded faster finished 8.2x slower. Numbers re-derived from the server logs in September 2026.
Explore 1,680 candidates across 275 constituencies. Filter by coalition, division, party, and more.