Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
Two epic AI stories dropped this week.
A man in Sydney used ChatGPT to design a personalised mRNA cancer vaccine for his dog. The tumours shrank, and now she’s chasing rabbits again. Link.
A Florida CEO used ChatGPT to sell his home for $100k more than real estate agents recommended. No agent, no commission, just a really long road trip conversation with AI. Link.
We're past "write me an email" or plan a trip for me. This is a different era.
Happy reading ✌️
Did someone forward you this? Sign up!
🇳🇿 New Zealand News
A UK security firm broke through Heidi Health's safeguards in three prompts, getting the chatbot to provide instructions on identity theft and drug manufacturing. Health NZ called the breach "minor" and said no patient data was at risk. Heidi's head of security called the findings an "overclaim". But UC AI regulation expert Olivia Erdelyi pointed out the chatbot received its security certification five months after Health NZ endorsed it and two months after the national rollout. She recommended updating the Medicines Act 1981 to incorporate stricter AI safety standards. 3 min read.
Our take: This was three prompts, not a sophisticated attack by a nation-state actor, it was a UK academic with a keyboard. NZ is deploying AI in emergency departments faster than almost any country on earth, but the security testing is running on a lag that would embarrass a beta app. Speed without rigour is not innovation, it is hope.
NZ consumers face higher prices on laptops, phones and even fridges as manufacturers pivot production from consumer components to AI data centre equipment. RAM component prices have jumped 400%. Hardware producers are concentrating resources on specialised AI infrastructure where profit margins are significantly higher, leaving consumer-grade devices competing for increasingly scarce inventory. 2 min listen.
Our take: NZ imports nearly all its consumer electronics. A 400% jump in component costs does not land evenly. Budget laptops and phones get hit hardest, which means the people least able to afford AI tools are the ones paying the most for the hardware to run them.
Datagrid secured a 140MW electricity option contract with Mercury for its $3.5B Southland data centre, days after receiving full resource consent. Earthworks are scheduled to begin in July 2026. Meanwhile, Invercargill Mayor Tom Campbell defended his claim that the undersea cable will have "none at all" environmental impact at Oreti Beach, despite the commissioner's report identifying risks to marine mammals, toheroa beds, cultural values, and seabed disturbance. 2 min read.
Our take: July 2026 earthworks puts this project on a timeline that intersects with the general election. Let’s hope Datagrid becomes a campaign talking point: jobs and investment for one side, environmental and sovereignty concerns for the other.
Organisations are racing to establish official certification for products declared "AI-free", sparking debate on RNZ's The Panel about whether AI is already too embedded in supply chains and workflows for the label to be meaningful. The trend reflects growing consumer demand for transparency about AI use in products and services.
Our take: Most businesses cannot honestly claim "AI-free" even if their product itself contains no AI. Their email, their design tools, their accounting software, their supplier's logistics: AI is in the stack somewhere. The label is aspirational, not factual.
Schools are implementing "traffic light" systems to manage AI use, blocking ChatGPT but allowing Gemini, and watching detection software fail. An Auckland University professor warns that "smart students would run AI-produced work through a programme to make it sound more like their own". The Principals' Federation is urging the government for national guidance as teachers feel overwhelmed. Both the Ministry of Education and ERO are conducting research, but policy is lagging practice. 4 min read.
Our take: NZ schools are doing what NZ businesses are doing: making it up as they go because there is no national framework. The traffic light system is clever but it is 30 different schools inventing 30 different solutions. That is not a strategy, it is a coping mechanism.
academyEX is the EXceptionally proud sponsor of this week's newsletter.
📚️ Mike’s Takes From The Week
Helping leaders and teams adapt, learn, and scale with AI.
1️⃣ People want transformation. Organisations are shipping task completion: Anthropic surveyed 81,000 people across 159 countries and asked one question: what would AI do if it could wave a magic wand? Professional excellence topped the list at 18.8%, but productivity didn't even make the top four, and a third of all visions were about making room for life, not working faster.
4-min read.
2️⃣ 52% of Kiwis are concerned about AI. The market doesn't care: Anthropic went from $1B to $14B in annualised revenue in 14 months, OpenAI hit $25B, and global AI spending crossed $2.5 trillion this year. Among Kiwis who feel they understand AI, 47% are more concerned than excited.
3-min read.
3️⃣ Christchurch Airport's AI playbook is worth stealing: No 50-page roadmap, an embedded specialist who rotates between teams, and exec training that put leadership through the same programme as everyone else. 75% of submitted use cases turned out not to be AI problems at all, which is exactly why the specialist extraction model works.
36-min listen.
4️⃣ Most organisations are micromanaging their AI: Nobody manages an employee by wiping their memory each day, dictating every task, and reviewing every output, but that describes the default AI experience in most organisations right now. Microsoft and Founders Forum Group found the pattern everywhere: 1 in 2 building AI tools in-house, but 1 in 3 still in exploration mode.
5-min read.
🔥 Claude Cowork Workshop with Leo Garcia Curtis - 6 tickets left!
Final 6 tickets left!
3 hours. No code. You walk away knowing how to build anything with AI.
No technical expertise needed, this is a hands-on Claude Cowork & Code workshop to learn the key tools and methods to build anything, websites, applications, automations, and more.
31st March - Auckland
Limited Tickets - $199 +GST
Use Waboom200 for $200 off
Become the most powerful person in the room.
🎙️ The AI Corner Podcast
This week's guest is Tom Lintern, Head of AI at New Zealand Post. Hear:
Why Tom frames governance as a value conversation first and a risk conversation second, flipping the typical enterprise playbook.
How a 180-year-old logistics network full of legacy systems is actually the perfect setup for AI because they skipped the over-experimentation phase.
His advice for data scientists: the sexiest job of the 21st century is now a software engineering problem, and evaluation skills are your moat.
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
OpenAI
Fidji Simo told all-hands OpenAI is targeting a Q4 2026 IPO, projecting $280B revenue by 2030 split equally between consumer and enterprise. The "productivity tool" framing is a direct response to investor concerns that 900M casual users don't convert to enterprise revenue.
Plans desktop "super app" merging ChatGPT, Codex, and Atlas browser into a single agentic application. Simo and Brockman overseeing. No launch date set. Consolidation simplifies the enterprise sale and sharpens competition with Microsoft Copilot.
Codex reached 2M weekly active users (up from 1.6M at start of March). Codex 0.116.0 shipped with terminal-read capability and userpromptsubmit hooks. Terminal-read lets Codex check whether builds failed without user copy-paste.
House Judiciary sent preservation subpoena to Altman requiring all communications with foreign censors, including auto-deleting messages. Live legal obligation, not a request.
Anthropic
DOD filed 40-page brief labelling Anthropic an "unacceptable risk to national security," arguing its ethical red lines could lead it to "disable its technology during warfighting." DOJ separately argued Anthropic's refusal is "conduct, not protected speech." Government's clearest articulation yet of why it views AI ethics clauses as operational vulnerabilities.
Claude Cowork "Dispatch" launched as research preview: persistent agent thread letting mobile users control desktop AI tasks from their phones. Anthropic's answer to the "phone as remote control for desktop AI" pattern.
Hiring weapons and explosives policy experts to prevent "catastrophic AI misuse." Role requires 5+ years in chemical weapons/explosives defences. Publicly hardening weapons guardrails while fighting the Pentagon for restricting weapons use of Claude is an irony the DOJ will notice.
Gemini 3 Deep Think updated March 17: 84.6% on ARC-AGI-2, gold-medal level on International Physics and Chemistry Olympiad written sections, 3455 Elo on Codeforces. Available to AI Ultra subscribers and via API early access. Deep Think identified a subtle flaw in a peer-reviewed mathematics paper that human reviewers missed.
Google Workspace Studio reached GA (March 19): no-code AI agent creation for all Workspace domains. Agents handle intelligent prioritisation, support triage, approvals, content generation, sentiment analysis. No-code agents in Workspace puts Google in direct competition with Microsoft Copilot Studio.
AI Expanded Access add-on now required for advanced AI features in Workspace (from March 1). New paid tier between standard and top-tier plans. Google drawing a harder monetisation line between basic and premium AI features.
Microsoft
Considering legal action against OpenAI and Amazon over $50B cloud deal. Dispute centres on whether "stateful runtime environments" fall within Azure's exclusivity clause. No lawsuit filed; active negotiations. Most significant threat to Microsoft's AI revenue position since the original OpenAI investment.
Rolling back Copilot "bloat" across Windows 11: reducing entry points in Photos, Widgets, Notepad, Snipping Tool. Shelved planned integrations in Settings, File Explorer, Notifications. Public admission that Microsoft over-engineered Copilot's footprint.
Copilot Chat in Outlook expanded from single-thread to full inbox, calendar, and meeting reasoning. Agent Mode in Word, Excel, PowerPoint completed rollout. Inbox-wide reasoning differentiates Copilot from search.
Perplexity
Comet AI browser launched on iPhone March 18 with Deep Research, voice mode, cross-device sync. Price dropped from $200/month (desktop launch) to included with Pro/Max from $20/month. Removing price barrier to mainstream browser share against Safari and Chrome.
Computer for Enterprise announced: 20 orchestrated AI models, Slack integration, connectors for Snowflake, Datadog, Salesforce, SharePoint, HubSpot. Internal study: $1.6M saved, 3.2 years equivalent work in four weeks. Direct enterprise challenge to Copilot and Agentforce.
Court ruled Perplexity's Comet shopping bots can remain on Amazon, blocking injunction. Preview of the legal framework governing all autonomous web agents.
Apple
Earned $900M from rival AI apps in 2025 (ChatGPT = 75% of that). On track for $1B+ in 2026 via 30% App Store commission. Built a billion-dollar AI revenue stream without building a single competitive AI product.
Meta
Replacing third-party content moderators with in-house AI systems. Tests show AI detects 2x more violating content and reduces enforcement errors by 60%. Intercepts ~5,000 scam attempts per day. Cost play as much as quality play; if 60% error reduction holds at scale, addresses a longstanding regulatory criticism.
Manus AI tools now fully rolled out in Ads Manager, Creator Marketplace, and WhatsApp Business for all advertisers. Acquisition to full production in under two months; treating agentic advertising automation as near-term revenue driver.
NVIDIA
GTC 2026 keynote: Jensen Huang projects $1T in Blackwell/Vera Rubin chip purchase orders through 2027. Q1 2026 revenue guidance ~$78B (77% YoY growth). $1T is the floor, not the ceiling; doesn't capture all product lines.
Agent Toolkit launched: NemoClaw (secure agent runtime), AI-Q (knowledge blueprint), OpenShell (open-source runtime), cuOpt (optimisation library). 16+ enterprise adopters including Adobe, Salesforce, SAP, ServiceNow, Cisco, CrowdStrike. NVIDIA trying to own the agent software layer the way it owns the hardware layer.
✨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