$ ./read-notes --latest
Paweł Dubiel / AI and software engineering
Notes on language models, agents and building software. I read the research, explain the concepts and write down what I observe about the capabilities, limits and consequences of AI.
Qwen3.8-Flash-Next through Strata on an RTX 3060 and a 2016 Xeon: 49 tokens/s, 89% on my 53-task Polish benchmark, compared with four smaller local models.
Claude Opus 5.5 built a Porsche 911 in Blender from code alone, rendered it on an RTX 3060 and assembled a video. The session used 17.6 million tokens, about 8.5 USD.
Jev 1.14 and Decision-1 classify tickets at 95% in Polish and English, local 4B models at 81–84%; all of them fail systematically on dates, sums and missing facts.
I gave a coding agent four floppy disk images of the 1990 Amiga version of Sid Meier's Pirates! and asked it to move the game to Unity.
A team wants to find out whether the next version of an agent can be trusted with longer programming tasks.
In its analysis dated 5 October 2026, SemiAnalysis compares the actual limits of AI subscriptions.
An agent working in a terminal tries to run a program. It sends a command, reads an error message and chooses its next action.
You want to run a language model on your own graphics card, but its weights take up too much memory.