AI consequences, human context, and staff engineering

We started with some casual topics, including the UK TV licence drama, and then moved on to recent developments in AI.

I shared some experiments with Omid and Pooria, including the new Z.ai stealth model, my move to OpenCode, and a small side project involving TTML. I also talked about slowing down over the past two weeks and finding time to read a few interesting books.

One of those books was Staff Engineer. We discussed its interesting argument that becoming a staff engineer does not necessarily mean having greater technical knowledge. It is often more about developing experience in the wider aspects of the role: navigating ambiguity, influencing decisions, understanding organisations, and helping others move work forward.

The main discussion focused on the negative side of AI and whether AI systems should reason about the consequences of their actions. Pooria shared examples from his recent work, where his team is relying more on people because AI often lacks the context and experience needed to work effectively in large, existing codebases.

That is quite different from my day-to-day work, where the focus is on increasing our use of AI and moving as quickly and agentically as possible within the development environment. The contrast led to an interesting discussion about where AI is genuinely useful, where human judgement remains essential, and how much the surrounding environment shapes the results.

Pooria also shared some of Yann LeCun’s views on AI and his perspective on where the field currently stands.