Blog

Using an agent to build an inference engine from scratch, the project grew messier and performance degraded. The solution was always the same: stop, redefine the goal, open a new repo, and start over. The same model acted like a different entity. Microsoft simulated 200,000 conversations and found the same thing: a model on the wrong path won't turn back.

Fable 5 ran a hardcore 15-hour task for $420. The model is genuinely stronger. But after June 22 it's out of the Coding Plan subscription, and API rates push costs up by an order of magnitude. Access to the best intelligence has become a pure money problem.

A friend using models for small tools was thrilled, but complex tasks kept ending in false 'done.' Benchmarks bunch at the ceiling, yet real-world feel diverges. Solving hard problems, reliably doing messy work, and exploring the undefined are three distinct capabilities. That last one is hardest to measure and train.

Opus 4.8 disappointed me on research and engineering tasks, so I moved all that work to GPT 5.5. Looking back, this isn't just Claude's problem. From NVIDIA and AMD nearly going bankrupt to Zhipu falling behind and bouncing back, the top models have always traded places. Don't bet on who's strongest—bet on who's still at the table.

When people estimate token costs, they usually watch TTFT, TPOT, and throughput. What actually makes bills differ by 10× is whether the KV cache hits. The model, server, and user layers all have to line up.

Demand for GPT-5.5 and Opus 4.7 is nearly infinite, the mid-tier has vanished, and low-to-mid-range compute sits idle. The token economy sounds like selling electricity, but it's more like a gas station: 98-octane is sold out, diesel tanks are full yet self-service only, and 95-octane sits empty.

Last month I wrote that ignorance is an advantage. This month agents burned me four times: performance inflated by orders of magnitude, a lab machine bricked, optimizations going in circles, and goals too high to ship. Ignorance gets you started; common sense keeps you from getting played.

The Gemini 3.5 Flash launch barely talked about intelligence; Zhipu’s GLM-5.1 High-Speed edition hit 400 token/s. Both tell the same story—once inference speed crosses that 5× threshold, the model unlocks an entirely different product category.

Today I went to the AMD Developer Conference in Shanghai. Over 2,000 people packed the venue, and AMD gave one of its two developer awards to a non-coder who used an AI agent to rewrite a project in Rust. A year ago everyone worried about demand; six months later, all anyone asks is: 'Got more supply?'

I've been using Codex's /goal for weeks. When Claude Code recently added it too, I moved over a few tasks Codex couldn't crack. The contrast was stark enough to be worth writing down.

When I show friends what my AI agents are doing, they always ask, 'Can you explain this to my boss?' But real change has never come from a single lesson.

Trump brought Musk, Jensen Huang, and Tim Cook to China. Red carpet, state dinner, talks between equals. Seventy-seven years ago this country couldn't feed itself — two generations was all it took.