Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
📢 The AI Corner contributors: Last week’s mention got a strong response, and we’re still keen to hear from more NZ AI builders, writers, and operators worth surfacing, so get in touch.
🚨 Latest Builds & Finds: A new segment for advanced builders following the infrastructure, agent setups, and experiments actually shaping this space.
🔗 My latest article: I broke down the full content operating system I’ve built with Claude Code, from linting to feedback loops to a compounding knowledge base.
🪄 Harry Potter fans: This was too good not to share, even if half the Gen Z language completely lost me. Pure entertainment.
Happy reading ✌️
Did someone forward you this? Sign up!
🇳🇿 New Zealand News
EY's 2026 AI Sentiment Survey classifies New Zealand as "Transitional / Lagging" on AI adoption, and the local numbers show why. Among NZ respondents, 36% said they would prefer AI to automatically apply discounts at checkout, 28% want AI to contact customer service on their behalf, and 21% want AI home security systems to automatically notify authorities. At the same time, 66% worry organisations will fail to comply with AI rules or regulations, and 66% worry organisations will not be held accountable when AI causes harm. The survey covered 18,152 people across 23 markets and signals a shift from assistive AI to systems that act for people. 4 min read.
Our take: NZ businesses like to frame caution as prudence. Sometimes it is. But if 66% of people do not trust organisations to govern AI properly, that is not an argument for waiting. It is an argument that governance has become the adoption bottleneck. The firms that solve that bottleneck first will look "safe" and "innovative" at the same time, while everyone else gets stuck talking about pilots.
IT Brief NZ reports that 77% of CFOs say AI has delivered no meaningful return yet, while 89% say their data is not ready for AI. The story points to finance leaders tightening scrutiny on AI spend just as vendors keep pitching transformation narratives. 3 min read.
Our take: This is the predictable comedown after two years of executive theatre. Boards approved AI budgets to avoid looking asleep, vendors sold ambition, and now the finance function is asking the only question that matters: what changed in the P&L. That does not mean AI is failing. It means the market has moved out of the experimentation phase and into the evidence phase.
The Medical Council of New Zealand published formal guidance requiring doctors to use AI only within their scope of competence, understand its capabilities and limitations, and interpret its output. Patients must be told when AI is used in their care. Patient-identifiable information must not be shared outside the care setting. 3 min read.
Our take: The disclosure requirement is the most significant line: patients must be told when AI is used in their care. Sounds straightforward. Health NZ's Heidi scribe sits in every emergency department, listening to consultations and generating notes. How many patients currently know? The gap between this rule and current practice is wide.
🛠 Latest Builds and Finds
Helping advanced builders and tinkerers stay at the frontier of AI.
What I'm building
1️⃣ Andrej Karpathy's (co-founded and formerly worked at OpenAI) LLM Wiki gist gave the cleanest articulation yet for what an LLM knowledge base should be:
Not a fancy retrieval stack that re-discovers context from scratch every time, but a simple markdown system where raw source files get compiled into a linked wiki the agent can navigate, and that compounds over time. In my case, I took Karpathy's ingest, query, and lint pattern, then added batch seeding, daily sync, and a tighter operational layer for my workflow. I also leaned on qmd, which Karpathy recommends and Tobi Lütke built, as the search backbone for the vault. The result is less "personal research wiki" and more a working knowledge system wired into content, analytics, and daily operations.
2️⃣ Karpathy's autoresearch repo feels like a game changer for how I improve Skills and Commands: the pattern is simple but powerful: score, edit, test, keep-or-revert, repeat, over and over on autopilot. Saves me having to manually run experiments with Skills to optimise them.
For example, the best run so far for my Ralph copywriter workflow lifted the grading score it uses by +30%, and the changes that stuck were not flashy. They were things like a consolidated scoring checklist and clearer formatting guidance. The noisy additions regressed and got thrown out.
What I’m workshopping
3️⃣ Two possible stack upgrades:
Hermes Agent looks like the cleaner path to upgrade or replace some of the OpenClaw antics.
Paperclip feels more like a control layer for coordinating multiple agents and systems.
What I consumed
4️⃣ Birgitta Böckeler's harness engineering piece and Simon Willison's latest Lenny’s Podcast interview rounded this out nicely:
Böckeler’s feedforward vs feedback framing is useful for thinking about how to steer agents, and her broader framework helps clarify where user harnesses fit into the mix.
While Simon covered an elegant recount of the structural changes hitting software engineering at the moment, and also delivered the quote of the week: "Everything is changing so fast right now. The only universal skill is being able to roll with the changes". Also appreciated that he managed to bring NZ's kākāpō into the same conversation!
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
Anthropic
A misconfigured npm package shipped 512,000 lines of Claude Code source code to the public registry, the second security lapse in days after the Mythos leak.
Researchers found a critical vulnerability in the leaked source where commands with 50+ subcommands bypass security analysis; trojanised versions spread malware within hours.
Anthropic is privately warning government officials that its unreleased "Mythos" model can exploit vulnerabilities at unprecedented pace.
Anthropic filed to create AnthroPAC, an employee-funded bipartisan PAC capped at $5,000 per person, targeting 2026 midterm candidates involved in AI policy.
OpenAI
Closed the largest private funding round in history: $122B at $852B valuation with Amazon ($50B), NVIDIA ($30B), SoftBank ($30B), and $3B from retail investors.
Three senior executives moving in one week: Fidji Simo (medical leave), COO Brad Lightcap (to special projects), and CMO Kate Rouch (stepping down).
Next-gen model "Spud" (likely GPT-5.5) completed pretraining; Altman indicated release "within a few weeks."
Released Gemma 4: four open models (E2B, E4B, 26B MoE, 31B Dense) under Apache 2.0 with multimodal inputs, 140+ languages, and 400M+ cumulative Gemma downloads.
Launched Gemini 3.1 Flash Live with direct audio-to-audio processing, acoustic awareness, and 2x longer conversation context in 200+ countries and 90+ languages.
Veo 3.1 Lite shipped on Vertex AI at less than 50% the cost of Veo 3.1 Fast; standalone upscaling capability enhances any video to 4K regardless of source.
Microsoft
Launched three in-house foundational models: MAI-Transcribe-1 (speech, 25 languages, 2.5x faster than Azure Fast), MAI-Voice-1 (60x real-time synthesis), MAI-Image-2 (top-3 image generator).
Released open-source Agent Governance Toolkit: seven packages across five languages addressing all 10 OWASP agentic AI risks with sub-millisecond policy enforcement.
GitHub Copilot will use interaction data for AI training from April 24; Free/Pro/Pro+ users opted in by default.
Copilot Chat removed from Word, Excel, PowerPoint, and OneNote for unlicensed users at organisations with 2,000+ seats effective April 15.
Perplexity
Hit with a 135-page class-action lawsuit alleging tracking tools sent user chat data to Google and Meta without consent, even in incognito mode.
Google Gemini overtook Perplexity as the No. 2 AI referral source (8.65% vs 7.07%) while ChatGPT retains ~80%.
Meta
Launched two Ray-Ban prescription AI glasses (Blayzer and Scriber Optics Gen 2) starting at $499 with hands-free nutrition tracking and WhatsApp summaries.
NVIDIA
US officials are weighing a worldwide tiered licensing system for AI chip exports: cursory review under 1,000 GPUs, government certification for 200,000+ deployments.
Apple
Pivoting AI strategy to platform/distribution: iOS 27 will let users install third-party AI chatbots (ChatGPT, Gemini, Claude, Grok) via Siri Extensions with a dedicated App Store section.
📚️ Mike’s Takes From The Week
Helping leaders and teams adapt, learn, and scale with AI.
1️⃣ IKEA found the real value of AI in what the bot could not do: Billie, IKEA's chatbot, handled 47% of level-one support queries. Most companies would have stopped at cost savings. IKEA looked at the unresolved cases instead, spotted unmet demand for interior design help, reskilled 8,500 support staff into consultants, and turned the gap into a service line worth roughly €1.4 billion in new revenue within a year.
49-sec watch.
2️⃣ Entry-level jobs are not dying. Lazy hiring is: Two AI-native interns in one week is a stronger signal than another CEO prediction about junior roles disappearing. The grunt work still exists. What is changing is the value of bringing in people who already think with Claude and ChatGPT by default, can build from day one, and are willing to reshape workflows instead of inheriting them.
2-min read.
3️⃣ The frontier AI race now needs a monthly scoreboard, not a daily news feed: The market is moving too fast for headline-chasing to be useful on its own. What leaders need is a sharper view of where the labs actually stand, who is pulling ahead, where momentum is fading, and what that means for real buying and strategy decisions right now.
6-min read.
✨A few people have asked…
It’s Mike here, I run The AI Corner.
I’m not just into writing about AI. I help businesses grow without adding headcount by shipping AI systems that cut costs and increase throughput.
👋 Mike & Erin