Blog

Codex GPT-5.5 clusters reasoning at 516; Claude Code secretly marks Chinese users. When cloud AI silently degrades, open-weight self-hosting is the hedge.

GPT 5.6 looks strong, but only ~20 U.S.-approved partners can use it; I realize intelligence you can buy is already fair, while permission is the real gap.

Our three weeks' tuning hit next-day P&L. Token economics cuts tech-to-profit: Zhipu 8x in 6mo; OpenAI Codex 5M/wk (GPT-5.5); MiniMax beat Baidu, fell.

I found Fable 5 swapped to Opus 4.8; U.S. Commerce banned it in three days. Closed-source AI can vanish overnight; Chinese open-source models step in.

Building an inference engine from scratch with an agent, the project unraveled and model performance visibly degraded. The fix: pause, redefine the goal, start fresh in a new repo. Same model, different entity. Microsoft simulated 200k conversations: once a model goes off track, it won't recover.

Fable 5 ran a 15-hour task for $420. After June 22 it leaves Coding Plan for API rates, a 10x cost jump: access to top intelligence is a wealth problem.

Top models tie on GPQA at 92–94%, yet real-world results diverge. Three skills now split: hard problems, rough tasks, open exploration; the last is hardest to measure.

Claude Opus 4.8 kept failing my research and engineering tasks, so I moved them to GPT 5.5. Top models rotate—NVIDIA, AMD, and Zhipu prove survival wins.

Token cost is often judged by TTFT, TPOT, and throughput, but a 10x bill swing comes from KV cache hits, not speed. Model, server, and user layers must align.

GPT-5.5/Opus 4.7 demand is near-infinite, mid-tier models vanish, low-end compute idles; tokens resemble electricity but act like a mismatched gas station.

I said ignorance helps. Agents burned me four times: inflated benchmarks, a bricked machine, spinning optimization, missed goals. Lesson: dare, but verify.

Gemini 3.5 Flash and Zhipu GLM-5.1's 400 token/s mode show that crossing ~5× inference speed unlocks a different product category, not just faster answers.