Writing
Notes with the numbers left in.
What I measured, what broke, and what I’d check next. Each note is dated and links to the files behind it. How I write these.
Showing all 20
- Testing a VR game: our bots knew the answer, a person in a headset didn’t
Every build passed its bots (40 of 40 feature checks, 0 stuck), and about 30 of the 30 issues found in the headset had escaped every eval.
- Autoresearch at home: eight search-and-score loops, and where the scorer lied
Two small models searching together marked their answer verified on 40 of 40 maths problems; 16 of those answers were wrong.
- One sim, seven platforms: porting with agents and tests
Four interactive nebulas went from WebGL to native Metal and Swift in 81 minutes, with zero invalid particles in the tests.
- Tuning a game with panels of AI playtesters
AI testers called the auto-locking throw the best design in the game; a noisy bot and my first headset play found it too easy.
- Watching a language model think
A 3-D map of a model’s vocabulary kept 43% of each word’s true neighbours, and raw attention’s favourite word was never the one that mattered.
- What running a team of AI agents taught me about applied AI in product management
One weekly usage limit stopped 18 agents at once; across 21 days, quota ran out before compute or ideas did.
- When the metric says yes and your eyes say no
A nebula look-dev scored 0.74 on shape overlap with the photo and still looked like a sock puppet.
- Chaos with nothing to orbit
A 7-term strange attractor with no fixed points at all, and an earlier “discovery” we had to retract.
- Create: a language model, a parts list and a validator
All 20 games generated in the eval were proven winnable and losable before anyone played them, at $0.29 median each.
- Tuning an AI recruiting agent: what we measured and what moved
Guessed company filters were removing the best candidates: filtered searches qualified about 2 percent, unfiltered ones 18 percent.
- A private AI assistant, tested end to end
The earlier build passed hundreds of unit tests while its Tor mode fetched pages from the real IP address.
- Benchmarking local coding agents
The same final answer scored 17/100, then 97/100 after the extractor was fixed; the model hadn’t changed.
- Ambient clinical documentation on open weights
A LoRA fine-tune lifted ROUGE-L by 8.6 points while speaker misattributions doubled to tripled.
- Finding where songs start inside a DJ mix
Median song-start error was 3 seconds on hardstyle but 39 seconds on peak-time house, against DJ-published timestamps.
- Fable: experiments in creative expression
The films, songs, and interactive stories that emerged when I gave an AI collaborator creative freedom on my local hardware.
- LTX, H3, Music3 and ACE-Step: what I learned
An early audio-quality heuristic didn’t distinguish an improvement listening made apparent; the evaluator itself needed evaluation.
- Optimizing local AI: kernels, prefill and decode
Direct cross-GPU copies mismatched in all 20 recorded trials while the driver reported peer access as available.
- Building long-running local AI agents
How I design task-specific agents that can resume work, recover from failures, and know when to ask for help.
- Evaluating H3 and LTX video renders
All 59 clips passed basic integrity checks, yet 26 were flagged for extraneous speech and 13 for prompt echo.
- Refusal ablation, measured
Leaving the token embedding unedited kept the coding suite at 489/500; the original recipe’s embedding edit dropped it to 400/500.