- LLMs often accelerate the wrong parts: their capabilities and usage tend to focus on speed by increasing how much code can be generated in a given time, while a good developer’s priority is still understanding the code and its relation to the codebase, aligning with documented and efficient standards, and writing as few lines as needed. So the better workflow is to use them for specs, plans, research, refactoring, and reviewable ad-hoc generation. https://larsfaye.com/articles/agentic-coding-is-a-trap
- AI compute may split into different architectures depending on the kind of inference being done. GPUs are great for training and flexible inference, but future agentic inference (inference to do work) may care less about raw speed and more about memory, state, and cheap capacity. https://stratechery.com/2026/the-inference-shift
- Codex shipped a cool new feature called /goal - instead of finishing after one response, you hand it a task with a verifiable stopping condition and it keeps working across turns, checkpointing progress, until it’s confident the goal is met. Good fit for migrations, refactors, and prototypes (+ they also launched mobile app with remote control) https://developers.openai.com/codex/use-cases/follow-goals, https://chatgpt.com/codex/mobile/
- Interesting reads on how China has become brutally good at making advanced physical goods like EVs, batteries, solar panels, wind turbines, sensors, robotics, electronics components at prices many countries can’t match. https://archive.is/W7KEp, https://archive.is/TzKK0
- Here’s how to read your airline confirmation code. https://ajitem.com/blog/iron-core-part-2-six-characters/
What I’ve been reading (W20, 2026)
Where AI coding goes wrong, where AI compute is headed, new Codex upgrades and how China got brutally good at physical goods