Posts

Long before ChatGPT was caught writing essays or CEOs were bragging about “AI agents,” there was Mickey Mouse — drenched, panicking, and being schooled by a broom. Fantasia’s “The Sorcerer’s Apprentice” isn’t just a cartoon. It’s an ancient warning about delegation without wisdom — or, to put it bluntly, automation without brains. The Old Story, Fresh Eyes 👀 Picture it. The workshop hums with quiet power. The old sorcerer steps out, leaving his apprentice, Mickey, with one boring task: fetch water. Mickey looks at the heavy buckets, looks at the spellbook, and has a brilliant idea — “Why not get the broom to do it?” ...

TL;DR AI is becoming the front door to software: instead of clicking, you ask. Agents (like those built on Model Context Protocol) go further, skipping native UIs to act on your behalf. But not everything should be replaced: old UIs remain best for precision, safety, and learning. For Kiwi companies (Trade Me, Xero, One NZ), the challenge is balancing AI’s magic with trust, transparency, and compliance. What was once only achievable in AWS a decade ago is now democratised. 1. Language as the new input Once, we memorised commands. Then we clicked icons. Now, increasingly, we just talk. ...

A hammer helps you build. A violin asks you to learn a phrase. AI is closer to the violin. We often hear AI described as a tool, a neutral extension of human intent. Something you pick up, point, and apply. But that language downplays its nature. Tools don’t talk back. Instruments, on the other hand, respond. They shape the way you play, reveal new sounds you didn’t expect, and demand practice to master. If we keep calling AI a tool, we miss the more interesting truth: it’s an instrument. ...

If you reward fluency over truth, don’t be surprised when your AI speaks nonsense beautifully. That is the sobering lesson from recent work on why large language models (LLMs) hallucinate. The research is clear: hallucinations are not mysterious glitches, but the rational outcome of how these systems are trained and evaluated. When the training signal rewards confident answers, models learn to manufacture them—truthful or not. The problem with beautiful nonsense The paper Why Language Models Hallucinate makes a blunt claim: hallucinations arise because LLMs are optimised for being useful and fluent, not necessarily correct. In other words, they are rewarded for looking right more than for being right. That incentive structure guarantees some degree of dishonesty, even if the model has no intention in the human sense. ...

Extending the Experiment I saw this TechRadar article and thought—what would ChatGPT say about “EA”? Spoiler: ChatGPT first assumed I meant Executive Assistant—and honestly, it nailed it. They’re the understated powerhouse behind the scenes. The TechRadar piece that sparked this reads beautifully simple: a Senior Data Analytics Consultant becomes “someone who counts toys and tells people which toys are played with the most.” A journalist turns into a curious storyteller, a management consultant is a fixer of lemonade stands, a CTO is a ship’s technology captain steering through storms. It’s equal parts hilarious and enlightening—stripping away jargon to reveal the human heart of a job. Read it here ...

Today I want to talk about Shadow IT. Not Shadow AI — I’ll cover that another day. Shadow IT is those tools and systems folks spin up outside IT’s blessing: sprawling Excel workbooks, legacy Access databases, SaaS apps bought quietly. They’re not always reckless — most of the time they’re born from urgency and ingenuity. I recently heard an anecdote that gives the flavour. A payroll specialist, wrestling with a tedious task, asked if their company could buy Python. They had asked Gemini how to solve the problem, and the model suggested Python. Taken at face value, it is a ’lack of context’ problem but it also reveals a powerful instinct: people will chase any tool that promises a quick fix, even if they don’t quite know what they’re asking for. That impulse — the creative scramble — is the heartbeat of Shadow IT. ...

On 26 August 2025 I found myself in a Wellington room buzzing with energy, ideas and a fair bit of scepticism. The AI Forum New Zealand had gathered four panel members to unpack both the promise and the puzzles of AI in Aotearoa. It made for a cracking afternoon. AI Forum NZ has long played the role of convener in New Zealand’s AI landscape. It pulls together industry, academia and government to accelerate AI adoption while keeping things responsible. Their latest Productivity Report set the backdrop for the discussion. The report is clear: AI can deliver significant productivity gains for New Zealand, but the challenges of governance, trust and skills still loom large. ...

Context for new readers: This document builds on my earlier blog about vibe coding an RFI where I explored whether a single “vibe coder” could respond to a government Request for Information (RFI) and even begin building the product. Here, I’ve reframed that exercise as a formal project brief, as if written by a product vendor. It also includes an architectural view to highlight what an internal team might consider. 1. Executive Brief (Senior Executive) Context: TEC seeks to add a CV and Cover Letter building capability to its Tahatū Career Navigator site. Purpose: Support learners to create tailored, professional documents integrated into the wider careers ecosystem. ...

Relatively short post today. I came across this amazing tool called Transformer Explainer. It lets you see how a Transformer model (like GPT-2) actually works, layer by layer, right in your browser. You can watch tokens flow through embeddings, self-attention, and MLPs, and even play around with parameters like temperature and top-k to see how text generation changes. Check it out here: https://poloclub.github.io/transformer-explainer/ Written for KiwiGPT.co.nz — Generated, Published and Tinkered with AI by a Kiwi ...

Once upon a time, customer service meant walking into a building. If you needed to pay a bill, change a booking or renew your licence, you went to the branch down the road. Banks, shops, post offices, council offices — all ran on face-to-face chats and paper forms. It worked, but it came with queues, short hours, and a long drive if you lived out in the wop wops. Then came the phone. In the 80s and 90s, private companies jumped on 0800 numbers and central call centres so you did not have to show up in person. Government got on board too, but it was slower going. Often you still had to ring the right regional office, and sometimes that meant calling three different numbers. For private businesses, the phone was a way to win customers. For government, it was more about easing the pressure on the counter. ...