Signal

There's a version of 'AI in my workflow' that means faster first drafts. And a version that means working differently. Most people are stuck in the first. The second requires changing what you delegate and what you keep — which turns out to be the harder problem.

Context windows don't degrade because the model forgets. They degrade because the model has to attend across a larger haystack to find the relevant needle. The symptom looks like forgetting. The mechanism is dilution.

Institutional memory has two layers: the decisions and the reasoning behind them. Most organisations document decisions reasonably well. Almost none document the reasoning. AI retrieval surfaces the first layer easily. The second is often gone entirely.

The most common misuse of AI in analytics isn't hallucination — it's confirmation. You get back analysis that looks rigorous but was shaped entirely by how you framed the question. AI is a mirror. If your question assumes the answer, you'll get it back formatted as insight.

Most AI safety discourse focuses on the model. Very little focuses on the deployment context. A model that behaves safely in a research environment may not behave safely embedded in a real-time decision system with no human review. The unit of analysis is wrong.

The prompt engineering hype cycle peaked around the idea that better outputs required magic phrasing. It mostly doesn't. It requires giving the model the context it needs. The distinction matters because one frames the user as wizard, the other as architect.

DuckDB 1.1 shipped with native JSON path support. This quietly eliminates one of the last reasons to reach for Spark on mid-size analytical workloads. The era of local-first analytics is here.

The EU AI Act's 'general purpose AI' category is doing a lot of work. The classification criteria are capability-based, but capability is not a stable property. Models improve post-deployment. The Act as written will require continuous re-assessment — something the enforcement apparatus is not designed for.

Reminder: when your regression includes a post-treatment variable as a control, you're not controlling for a confounder — you're blocking the causal path you're trying to measure. See this in policy evaluation constantly. The fix is rarely complicated.