Writing
9 entriesThink it through, then write it down. Newest first.
2026
- Why there's no warning before the breakthrough: the 59.27% tipping point
A ~6-minute math animation (Chinese narration) with a full write-up. Some say effort below the tipping point means nothing. I fill a 1,000-square sheet of graph paper at random and pour water in from the top: at square 592, all 25 rows light up at once. Then I simulate many sheets to find the 59.27% cliff and check how much of the saying holds.
- OpenAI Decisions API vs Jev: 20 tests, 10,000 calls
An ~18-minute hands-on video (English narration) with a full write-up. I gave Jev and OpenAI's new Decisions API the same 20 tests, 10,000 calls in total: Jev won 8, Decisions won 1, and 11 were ties. The article goes through the 5 groups, then speed, price and confidence, and ends with which one to pick for which job.
- Opus 5.5 Built a Math Game Kids Get Hooked On, in 40 Minutes
A ~5-minute video (Chinese narration) with a full write-up. I started from a tweet video and gave the AI (Opus 5.5) just one sentence. The first version was ready in 13 minutes. In under 40 minutes, the kids' math game 「嗨算！」 was live on my homepage. This article explains how it was made and how the AI caught its own mistakes. It also asks whether over-the-top practice games are actually good for kids.
- Jev as a RAG judge: top-1 hits from 7 to 17
A ~7-minute hands-on video (Chinese narration) with a full write-up. Where in a RAG pipeline should Jev go? I built a small RAG on TypeSafe's official docs and tested it with 36 hand-written English questions: after re-ranking, top-1 hits went from 7 to 17, every near-miss and unrelated question was blocked, and I also spell out what it can't do.
- Can open-source Laya replace Jev? 6/14 on SQL, then 3.9 hours of fine-tuning
An 11-minute hands-on test (Chinese narration). On 14 SQL review cases Jev gets 13 right; the open-source decision model Laya gets 6. After 3.9 hours of fine-tuning on an M4 Mac, Laya reaches 14/14 on safety, but still trails on correctness.
- Giving an AI limbs: how Jev plays Doom in real time at 120 ms
A 7-minute breakdown (Chinese narration). No pixels, no screenshots — just a plain-text state report, queried about ten times a second. Jev picks the goal like a commander; ordinary code does the key presses.
- 10,000 API calls later: reverse-engineering Jev's architecture
A 13-minute explainer (Chinese narration), based on Archer Hume's outside-in experiments with latency, token counts and option order. Jev doesn't generate token by token; it reads probabilities out, and encodes shared state only once.
- 10 ways to use Jev, after testing it on 3 of my own tasks
A 10-minute explainer (Chinese narration) with a full write-up. Task routing: 17 of 20 right. Kanban classification: 87%. Can this SQL ship: 14 of 14. Then ten ways to use it, how it splits work with an LLM, and where it falls short.
- Understanding Jev: a System-One tool for data engineers
A 4.5-minute explainer (Chinese narration, English UI). Jev is TypeSafe's System One Model: one forward pass, no token-by-token decoding, calibrated probabilities at about 1/400 the cost of a frontier LLM.
