- Published on
- 46 min read
Why Claude's messages are hard to read — measured. With a 50-word dictionary of Claudish and the one instruction that targets the actual problem.
All posts
- Published on
- 13 min read
China's domestic AI compute base looks larger than low outside estimates imply, and DeepSeek V4 on Ascend 950PR-class hardware is more plausible than consensus thinks.- Published on
- 17 min read
What it actually takes to make LLM-powered summarization reliable: eval-driven iteration, structured input compression, dynamic budgets, and the prompt sandwich technique.- Published on
- 7 min read
Where to find your ChatGPT usage and token counts, how to export your full conversation history, and what the data says about whether a Pro subscription is worth it.- Published on
- 25 min read
Built an RTX 5070 + Ultra 7 workstation on a ~¥20,000 budget to run KataGo.