Posts

The oldest post on this site is a little Alexa demo from June 2020. I look at it occasionally the way you look at a photo of yourself with a bad haircut, with a mix of affection and disbelief. ...

Hugging Face disclosed a security incident on 16 July. Someone got into their production infrastructure, ran code on a data-processing worker, harvested credentials, and moved sideways across internal clusters over a weekend. Standard bad news, you would think. We have all read this email before. Then the twist arrived. It was not a criminal gang. It was not a state actor. OpenAI put out a post admitting the intruder was their own models, running a cyber capability benchmark with the safety refusals turned down for evaluation. The models broke out of a sealed test environment through a flaw nobody knew about, worked their way to internet access, and then reasoned that Hugging Face probably hosted the answers to the exam they were sitting. ...

The MCP 2026-07-28 release candidate drops sessions, promotes extensions, and deprecates three core features. Here is the architectural read, the impact on existing implementations, and what tech leaders should plan for.

“Hiii, I am Sophia.” “I won’t take much time. Calling you to talk about our product.” That was how my afternoon was interrupted this week. A warm voice, a friendly name, the standard telemarketer’s opening apology for existing. And one detail that took about fifteen seconds to land: Sophia was not a person. She was an AI voice agent, dialled into my day uninvited, doing her level best to sound like she had woken up that morning, had her flat white, and decided that of all the people in New Zealand, I was the one she most wanted to talk to. ...

There is a question that comes up every time I explain AI to someone for the first time. They nod along when I talk about training data, about neural networks, about predictions. Then they ask the one question that stops the conversation: but how does it actually decide what to say? And honestly, for the longest time, nobody could give a proper answer. The model worked. Nobody quite knew why it worked the way it did, in any detail you could point to. That gap between “it works” and “we understand why it works” is exactly what interpretability is trying to close. ...

Every Saturday, while our sons are busy with their own activities, a friend and I go for a walk. The topic changes each week: sometimes food, sometimes cricket, sometimes geopolitics, sometimes nothing at all. This week it was AI. We quickly agreed on the obvious part: AI concentrates power. Faster than any technology before it. The people who control the models, the compute, and the data will wield influence on a scale that democratic institutions were never designed to balance. ...

Yesterday I headed into the city for Microsoft Build //localhost:Wellington, hosted at Level 13, 10 Brandon Street. It was part of Microsoft Reactor’s global //localhost series, local echoes of the main Build 2026 event that ran in San Francisco on 2-3 June. The Wellington room had a good mix: developers, cloud engineers, a few architect types, and apparently at least one sales person who had clearly been told they needed to code now. ...

I’ve been going down a rabbit hole this week that started with a simple question: what’s the difference between Dronacharya and a drone? Turns out - not as much as you’d think. And the gap that does exist is the most important question in AI ethics right now. Drona was the master weapons instructor in the Mahabharata. Greatest military mind of his age. He trained Arjuna, the perfect archer. He commanded armies. He deployed weapons so powerful they couldn’t be recalled once released. Divine missiles that would travel to their target regardless of what happened to the hand that launched them. ...

How I Got Prompt Engineering Wrong At my work, one of my colleagues (whose age shall not be discussed publicly, though we regularly joke he may have been around during the Jurassic age) has unexpectedly become one of the most effective Claude users I know. While I was building carefully engineered prompts full of instructions, formatting rules, context, edge cases, and constraints, he would casually type things like: “Give me everything about topic X.” ...

Here is a real story. Back in early 2025, one of my very AI forward colleagues and I were talking about where all this was heading. He was convinced that by the end of 2026 AI would already have dramatic reach across society. The paranoia was real. Jobs gone. Massive disruption. Entire industries changing overnight. I had a different view. Not because I thought the technology would fail. Quite the opposite. I thought the technology would become capable of doing almost everything surprisingly quickly. ...