Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
Last week. the CEO of McDonald's called his own burger a "delicious product" on camera last week (Paul Rudd sounding, anyone? And then the British accent 😆). Not food. A product.
Naturally, the internet went to town on this
From an awkward nibble to industry meme to AI parody, in days. That's the speed of creative these days and AI is pouring gasoline on that fire.
Check out our favourite parody below.
This week's major AI story is significant. The US government banned Anthropic for refusing to remove two safety restrictions from its military contract, then kept using Claude in active combat operations hours later. OpenAI swooped in, took the deal, and called it responsible AI.
The market made its thought known very quickly. Claude hit number one on the App Store. ChatGPT uninstalls surged 295% in a day. Consumers backed the company that held the line; the contract went to the one that didn't.
For those who haven't been following, here's the full breakdown (it's a long one).
Happy reading ✌️
Did someone forward you this? Sign up!
academyEX is the EXceptionally proud sponsor of this week's newsletter.
🇳🇿 New Zealand News
Deepfake attack ads for NZ politicians are circulating on Facebook with no legal obligation to disclose AI use. Victoria and Otago academics show NZ's false-statement prohibition covers only the final three voting days, useless when advance voting spans weeks. They're calling for mandatory AI disclosure across the full campaign. 3 min read.
Our take: The Electoral Act was written when the main campaign risk was a badly designed pamphlet. Today anyone can produce a professional smear campaign for the cost of a ChatGPT subscription. The law has been leapfrogged.
Deloitte's 3,000-organisation survey found 40% expect 10% of roles automated within a year, yet 84% haven't redesigned a single role. The firm's NZ AI Institute director told RNZ the workforce will look "very different very soon" and called for AI to be a major election-year conversation. 5 min audio.
Our take: The 40% figure will be quoted in every boardroom in NZ this month. Half will use it to justify cutting headcount. Half will use it to justify doing nothing until the picture is clearer. The ones who use it to actually redesign a role will be in the minority, and will end up ahead.
Westlake Girls' High School in Auckland is integrating AI into its curriculum, framed by RNZ as a school that isn't waiting for national education policy to catch up with the workforce. Deloitte's NZ AI Institute director singled it out as an example of the proactive response NZ institutions need as automation accelerates. 5 min listen.
Our take: We celebrate Westlake Girls as a success story (and so we should). But as a nation, we should be embarrassed that one school's initiative is national news. In Singapore, AI literacy is a curriculum requirement. In NZ, it depends on whether the principal cares enough to act.
NZ's Commerce Act isn't equipped to hold retailers accountable for AI-driven pricing, a University of Sydney researcher told RNZ. Supermarkets are deploying dynamic algorithms that update on demand and competitor behaviour, but existing protections were written for human decisions and don't clearly intervene. 2 min read.
Our take: NZ's supermarket duopoly was already a political flashpoint before AI pricing entered the picture. Adding an algorithm that can charge more for bread when it rains and attribute it to "market conditions" is going to end with someone in Parliament using the words "price gouging by machine". The Commerce Commission is not ready for that fight.
A Taupo pharmacist has installed a German-made AI dispensing robot, described by RNZ as the first of its kind in NZ. The system automates prescription dispensing, freeing staff for clinical work. It joins AI scribes in every ED and Wellumio's stroke scanner on NZ's growing list of point-of-care automation. 4 min listen.
Our take: Staff freed from dispensing spend time on clinical consultations. Clinical consultations generate revenue. This is not a charity decision, it's a business model upgrade. The pharmacies that frame AI automation as a cost are going to lose to the ones that frame it as a margin expansion.
Bunnings Warehouse will trial facial recognition in Hamilton, citing repeat-offender threats to staff. NZ has no legislation governing commercial facial recognition, unlike Australia where states regulate retail biometric surveillance and the Privacy Commissioner previously investigated Bunnings for the same thing. 2 min read.
Our take: "Protecting staff from repeat offenders" is the stated reason. It's always the stated reason. Once a facial recognition database exists, the scope of what it's used for tends to expand. The question isn't whether Bunnings has good intentions today. It's what the next retailer does with the precedent Bunnings just set.
🔥 Stop Burning Money. Learn to Build Anything with AI.
NZ has an AI execution gap. Close yours in 3 hours.
Join Leo Garcia-Curtis at GridAKL for a hands-on Claude Code workshop. Learn to build automations, apps, websites and more. No coding needed.
Tue 31 March · 8:30 AM · GridAKL, Wynyard Quarter
50% off: $199 + GST with code WABOOM200
📚️ AI for Business
Helping leaders and teams adapt, learn, and scale with AI.
1️⃣ GPT-5.4 is live, and the benchmarks aren't the interesting part: Persistent computer use changes the agentic game. An OpenAI demo showed the model building a 3D chess game in Electron, then clicking through it to test castling and piece movement, playing both sides to verify its own work.
3-min watch.
2️⃣ Most AI automation is busy work. The wrong process problem: Meeting summaries save time. Lead qualification saves money. Those aren't the same thing. The filter that cuts through: if this process ran 10x better, would it move revenue or cut a major cost?
6-min read.
3️⃣ AI with no context is coming for your tone of voice. AI with context isn't: A Claude Code skill generating a content brief in three minutes of dictated thinking, matched to voice, built on 78 previously published articles. The tool is almost irrelevant. The context architecture underneath it is not.
3-min watch.
4️⃣ SaaS stocks are down 30%. The "SaaSpocalypse" thesis is the wrong frame: Six of Helmer's Seven Powers hold or get stronger in an AI world. The real question isn't whether AI kills SaaS: it's whether cheap code creates new demand for software that was never economically viable to build before.
2-min read.
5️⃣ 99.8% of tokens weren't Claude writing. They were Claude re-reading: 363 conversations, 3 billion tokens, one analytics tool that reads logs Claude Code already stores locally. The most expensive prompt of the month cost 12.3 million tokens. The prompt was one word: "Yes".
2-min read.
✨A few people have been asking…
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.
🎙️ The AI Corner Podcast
This week's guest is Sean Narayan, AI and BI Lead at Bidfood New Zealand. Hear:
Why Sean believes data insights, rather than process automation, is where the real AI money is hiding.
How he secured exec buy-in at a billion-dollar food business by turning leadership ideas into a concrete roadmap.
The sales and procurement AI pilots Bidfood built that cut admin time without replacing a single job.
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
OpenAI
GPT-5.4 launched March 5 with native computer use, scoring 75% on OSWorld-Verified above the 72.4% human baseline, plus 1M token context. First frontier model to beat human-level computer control: agentic workflows are now commercially viable.
GPT-5.3 Instant released March 3 addressing user complaints about preachy tone and unsolicited disclaimers. Behaviour shift matters as much as benchmark shift: ChatGPT was losing subscribers over personality.
ChatGPT for Excel launched in beta for Business, Enterprise, Plus and Pro users, building and analysing spreadsheets via GPT-5.4. Microsoft's Copilot for Excel now has a direct challenger inside the same workflow.
Skills beta lets ChatGPT Business and Enterprise teams convert workflows into reusable workspace instructions with admin controls. Moves ChatGPT toward an enterprise operating system, not just a chat interface.
OpenAI signed a Pentagon contract for AI on classified networks, then amended it amid backlash; ChatGPT uninstalls surged 295%. Altman's government pivot created a structural safety-vs-access split with Anthropic.
Criteo joined ChatGPT's advertising pilot as the first adtech partner, enabling performance-driven ads with closed-loop measurement. Advertising monetisation is now confirmed infrastructure, not an experiment.
Anthropic
Claude hit #1 on the US App Store after Anthropic refused to remove safety guardrails for the Pentagon, the first time it has topped the charts. Principled refusal generated a commercial windfall: enterprise AI spend is segmenting on governance trust.
US military reportedly used Claude during an Iran strike despite a government-wide ban on Anthropic tools; Anthropic CEO called it contrary to the company's red lines. The gap between procurement policy and operational AI use in conflict is now a live public issue.
Claude memory extended to free users alongside a ChatGPT-to-Claude import tool, timed to capitalise on the Pentagon-driven user surge. Lowered switching costs exactly when intent to switch was highest.
Claude Code voice mode rolling out to 5% of users via
/voicecommand, with broad rollout planned. Turns Claude Code into a voice-first coding environment.Broadcom committed 1GW of TPU capacity to Anthropic in a reported $21B partnership. Custom silicon partnerships are becoming the real moat, not model performance alone.
Music publishers filed a $3B lawsuit against Anthropic over 20,500+ torrented songs used in training, naming CEO Dario Amodei personally. The largest AI copyright case yet to name a CEO personally.
Gemini 3.1 Flash-Lite launched on Vertex AI and Gemini API, optimised for cost-efficient, high-volume, low-latency workloads. Flash-Lite directly targets AWS Bedrock's price-competitive positioning where most enterprise volume runs.
Pixel Drop rolled out Gemini Task Automation on Pixel 10, executing in-app transactions including grocery orders and ride bookings. On-device agentic AI doing real transactions: Google is shipping what others are benchmarking.
Google Workspace Studio reached all Scheduled Release domains, enabling no-code AI agent creation across Gmail, Drive, Sheets, Salesforce, and Jira. No-code agent building at enterprise scale is now generally available.
Google Cloud AlloyDB AI went GA with auto-vector embeddings 130x faster at bulk refresh and Gemini-integrated SQL functions. AI embedded in database queries, not bolted on top.
Microsoft
GitHub Copilot coding agent for Jira entered public preview, letting developers assign Jira issues to Copilot to auto-generate draft pull requests. The issue-to-code loop is now automatable end-to-end.
GPT-5.4 rolled out as the top-tier model in GitHub Copilot, replacing GPT-5.2. Automatic upgrades compress adoption across the largest developer tooling install base.
Microsoft Sentinel gained an AI SOAR playbook generator in public preview, auto-creating Python automation workflows via a conversational coding agent. Security operations is becoming AI-first.
A Microsoft 365 Copilot bug allowed Copilot Chat to summarise DLP-protected confidential emails from January 21; a global config fix has been deployed. Copilot reading emails it shouldn't is a material enterprise risk event.
DAX Copilot adopted by 600+ health systems in 18 months; Microsoft AI CEO cited "medical superintelligence" as a major 2026 focus. Faster enterprise adoption than most SaaS achieves in a decade.
xAI
Pentagon integrated Grok into classified defence systems, citing predictive speed, the same week Anthropic was locked out for refusing similar use. Grok got rewarded for not having safety limits. The policy signal to AI labs is explicit.
Grok Imagine added video extension capability, chaining AI-generated clips up to 15 seconds; premium-only.
Perplexity
Perplexity launched "Perplexity Computer", a cloud agentic workspace orchestrating 19 frontier models for complex workflows, available to $200/month Max subscribers. Perplexity is no longer a search product: it's a multi-model orchestration layer.
Perplexity released two open-source embedding models, competing with Google and Alibaba at lower memory cost. Open-sourcing embeddings positions Perplexity as infrastructure, not just an app.
Amazon
AWS and OpenAI confirmed a $50B investment and exclusive cloud distribution deal for OpenAI Frontier, including a Stateful Runtime Environment on Bedrock for persistent enterprise AI agents. Persistent agent state on Bedrock is the detail that matters for enterprise buyers.
Amazon Bedrock AgentCore Policy reached GA with fine-grained controls for agent tool interactions using natural-language policy authoring that compiles to Cedar. Governance tooling for agents going GA confirms enterprise agent deployment has arrived.
Apple
Apple launched the iPhone 17e at $599 with A19 chip and full Apple Intelligence support, available March 11.
Apple's Private Cloud Compute servers sit at roughly 10% utilisation; Apple is in advanced talks with Google to run Siri on Google infrastructure instead.
Meta
Meta is abandoning the Llama open-source strategy, developing two closed frontier models (Avocado and Mango) using Alibaba's Qwen for distillation.
Meta confirmed $115–135B capex for 2026, nearly double 2025 spend, to support Meta Superintelligence Labs.
Meta-AMD 6GW GPU deal finalised, deploying AMD Instinct MI450 GPUs for Llama-class models.
NVIDIA
Nvidia invested $4B in optical interconnect companies Coherent and Lumentum to address AI data centre data-movement bottlenecks.
GTC 2026 confirmed for March 16–19 in San Jose; Jensen Huang keynote to outline a "Five-Layer Cake" AI infrastructure framework.
US Commerce Department drafting global permit rules requiring Washington approval for all Nvidia and AMD AI chip shipments worldwide.
👋 Mike & Erin