vish engineering notebook

Writing

Most of these started as an internal document I did not want to write, about a number I did not like. The numbers are measured unless the text says otherwise.

Notes

The billion-token question Sep 2026 Free tiers cap requests, not tokens. That single fact turns the question into an arithmetic problem about how to build a 35,000-token request 28,800 times a day — and why padding is not an answer. What a 304B model actually costs to self-host Sep 2026 Two B300s serving a sparse MoE: 6,972 tok/s aggregate at 100 concurrent streams, a million-token context in 72.8 seconds, the DeepGEMM crash that ate a week, and the request-layer bug that made the model look stupid. Prompt cache is the biggest lever on your agent bill Aug 2026 50.6B prompt tokens in 30 days, $1,799 paid, $10,629 saved. Why a blended 88.5% hit rate is the wrong number, the four things that break a prefix, and the half-second of latency that came free. When your agent lies about success Sep 2026 A close event during timeout cleanup won a race, and the shell tool reported exit code zero on a command that was killed. Why a false success is worse than a crash, and how to close the whole class. Fine-tuning a 304B MoE on a $500 ceiling Aug 2026 Every gate passed except the one that mattered. Roughly 263 GiB allocated per GPU during BF16 expert aggregation, and what it means that a model loaded for inference and could not load to train. A 149M router for 455 skills Sep 2026 An audit over 346 real sessions found the remote router missing 37% of tool-bearing turns and 40% of them mid-turn. Here is the right-sized local encoder that replaced it, and why it runs offline in under 30 ms.

Elsewhere

There is an RSS feed if you want new notes without checking back. Shorter, less finished thoughts end up on X.