AI Bytes covers the developments that actually change something, the workflows worth adopting, how this is built underneath, and the questions worth sitting with — as one connected body of work rather than a stream of posts.
Not every headline changes something. These are the ones that do — with what is genuinely new, and what it means in practice.
What actually changed when coding assistants started running for hours instead of milliseconds?
The interesting change isn't that models write better code. It's that the unit of work moved from a line to a task — and that quietly rewrites how engineering teams spend their attention.
When the window was small, curation was forced on you. Now that it isn't, the people getting the best results are the ones still doing it voluntarily.
Teams building with AI are rarely blocked on capability any more. They're blocked on not being able to tell whether a change made things better.
Judge the race by where the capital goes, not by which model tops a benchmark. The spending pattern points at something other than better assistants — and at a prize that gets handed out long before anyone agrees on what “winning” meant.
If the cost of competent reasoning keeps falling, the interesting question isn't what AI can do. It's what stays scarce — and almost everything we currently call a career is priced off the old answer.
The risk of AI-generated code may not be that it's bad. It may be that there's more of it than anyone has the attention to understand — and understanding was doing work we never priced.
The reflex answer is “creativity and critical thinking,” which is comfortable and nearly useless as guidance. Here's an attempt at something more specific.
Workflows, prompting patterns, and habits that came out of real usage — not generic prompt lists.
Most disappointing AI output is a briefing problem, not a model problem. The fix is the same one that works with capable humans — say what good looks like before the work starts.
The most useful thing to ask a model is rarely the question you came with. It's to have it argue against the answer it just gave you.
Typing forces you to have the thought before you write it. Speaking doesn't — and for the messy early stage of an idea, that's the point.
Agents, tools, MCP, retrieval, evaluation, and the engineering realities of shipping AI to production.
Once you're building with models rather than chatting with them, the work stops being about prompts and becomes about what goes into the window, in what order, and what earns its place there.
The Model Context Protocol is usually explained as “a way to give models tools.” That undersells it. The real problem it addresses is combinatorial.
The instinct when an agent underperforms is to give it more freedom and better reasoning. More often the fix is to give it more places to be checked.
Every piece belongs to one of four pillars, and connects to the others around the same topic.
What changed?
Developments that are worth your attention — and what they actually mean.
How can I use it?
Practical workflows, prompting patterns, and ways to get more out of AI today.
How do I build with it?
Agents, tools, MCP, RAG, evaluation, and production AI engineering.
What does this lead to?
Question-led explorations of where AI may take work, software, and society.
One development usually deserves more than one piece. AI Bytes follows a topic through the whole chain, and links the pieces together.