Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
Singapore’s Foreign Minister Vivian Balakrishnan just gave a very practical example of what AI leadership can look like.
Here is the GitHub Gist outlining what he built.
He's built a diplomatic "second brain" that learns from his speeches, articles, notes, and web clips, then helps him research topics, draft speeches, and prepare briefings.
This isn't a politician talking about AI from a distance. It's someone at the top using it in his actual work.
Hard to imagine this in New Zealand. Not because our leaders are incapable, but because the public-sector posture is cautious, private, and nervous about backlash. The incentives just aren't there.
Leadership means showing what responsible use looks like, not waiting until it feels politically safe.
A new class of leaders will start running this play openly, explaining the benefits, and using it in the interests of transparency.
Happy reading ✌️
Did someone forward you this? Sign up!
Watch competitors build momentum, or turn AI into action at the 2026 AI Roadshow [AI Corner readers get $50 off].
I’ve twisted Justin’s arm and secured AI Corner readers a $50 discount off Roadshow tickets. Use discount code AI-CORNER when you buy a pass to any Roadshow event.
🇳🇿 New Zealand News
Microsoft CEO Satya Nadella's Auckland visit confirmed 32 NZ Government agencies are now using Copilot, doubled the company's training pledge to 200,000 New Zealanders by 2028, and put a $102bn economic-impact figure on AI in NZ by 2038. Microsoft's own EY-Parthenon report also claims it contributed $9.4bn to the NZ economy in 2025, with users averaging nine hours saved per month. Tech commentator Peter Griffin asked the obvious question: "Where is the big push by the Government behind AI skills training? We haven't seen any strategy, any targets, any numbers".
3 min read.
Our take: Nine hours a month per user is the figure to interrogate. Microsoft is reporting it as a productivity win, but nine hours saved without a workflow redesign is just nine hours absorbed. The agencies that turn this into something will be the ones that re-engineer the work, not the ones that count the minutes.
Whakarongorau Aotearoa / NZ Telehealth Services will roll out a Microsoft Azure-built AI tool called "Welcome" on SMS to the 1737 Need to Talk helpline and Women's Refuge webchat from May. The tool identifies itself as AI, gathers background information and offers non-clinical support while callers wait for a counsellor. Whakarongorau handles 7,000+ interactions daily, with self-harm-risk mental health contacts up 153% since 2019.
4 min read.
Our take: Psychologist Louise Cowpertwait's caution about vulnerable callers is the question every NZ frontline service should be asking now. The Welcome tool will produce one of two outcomes: either the May to August data shows lower drop-off and faster routing, or it shows the model misjudging risk in the first three minutes. Either result will be the most useful evidence NZ has had on AI in mental health triage to date.
Auckland-based AI consumer-insights platform Ideally raised NZ$16M (US$10M) Series A led by Sydney's Shearwater Capital at a $100M+ valuation, with US revenue up 350% since opening its New York office in early 2026. Founded out of TRA Labs in 2023 by James Donald, Brendan Cervin and Joshua Nu'u-Steele, Ideally now counts Doordash, Google, Telstra, Burger King and Asahi among more than 250 brands using the platform across APAC and the US.
3 min read.
Our take: The customer list is the credential. Doordash, Google, Telstra, Burger King and Asahi are not running pilots with a TRA Labs spinout for fun. They are buying speed, and the speed is only possible because the model sits on top of structured insights infrastructure. Same as the Halter playbook: NZ AI startups are winning by building the infrastructure layer their global customers cannot.
Stats NZ and MBIE will deliver a Survey of Business Operations covering AI use across 20,000 NZ businesses with six or more employees, with invitations going out in early May and fieldwork through May-June. Run jointly with Ipsos, the survey will be the first national measurement of AI uptake across all sectors. Minister of Statistics Scott Simpson called it one of the largest business surveys ever undertaken in NZ.
2 min read.
Workshop: complete tasks faster & build real AI tools in 3-hours [no coding required].
You’ll learn how to:
Automate repetitive admin and day-to-day tasks
Build AI agents, apps, websites and internal tools without coding
Integrate tools and automate reports, quotes, and marketing assets
📅 Auckland — 6 May
🎟️ $299 + GST with code AICORNER
📅 Virtual (NZ-wide) — 7 May
🎟️ $199 + GST with code AICORNER
📚️ Mike’s Takes From The Week
Helping leaders and teams adapt, learn, and scale with AI.
1️⃣ The behaviour-change tax is killing AI rollouts. Skills are the first cure: Ten people use the GPT. Eighty keep emailing the specialist. Claude Skills surface expertise inside chats people already use, no menu-hunting required. Mid-market businesses paying for Copilot are picking up the phone and asking about Anthropic specifically.
6-min read.
2️⃣ McKinsey says 80% of companies see no bottom-line impact from AI. The fix is not the model: Rollouts stall because nobody has been resourced to rebuild the work. A Lighthouse team of four to eight people, six months runway, executive sponsor whose goals depend on shipping. Skip that, and the business is back to licenses and hope.
3-min read.
3️⃣ Harness engineering replaces context engineering as the next frontier: Context is the brief on day one. The harness is the onboarding system that turns the intern into a senior operator over twelve months. Garry Tan's CLAUDE.md hit 20,000 lines and quality dropped. Slimming to 200 lines of pointers fixed it.
5-min read.
4️⃣ The legal exposure in AI sits in the output, not the model: Andrew Dentice (Hudson Gavin Martin) explains why builders cannot flow through an IP indemnity in the AI world: model providers upstream will not give one. The mindset shift is to stop asking whether the model is safe and start asking whether the output is defensible.
2-min read.
5️⃣ Claude Design turns one brief into LinkedIn feeds, animated videos, and decks: Anthropic shipped Claude Design over the weekend, letting users create interactive prototypes through conversation. One set of context files, re-purposed into many formats, all on-brand. Visual output now sits inside the same compounding system as code and writing.
60-sec watch.
🎙️ The AI Corner Podcast
This week's guest is Steven Baker, Co-founder of Node. Hear:
Why the CIO role is quietly disappearing in mid-market companies, and who's left holding tech strategy.
What an always-on audit of how work gets done across a business could actually look like.
Why 70% of software rollouts and 90% of AI rollouts fail to do what they set out to.
🛠 Latest Builds and Finds
Helping advanced builders stay at the frontier of AI.
What I'm building
1️⃣ Google's DESIGN.md is a single file in the repo. YAML front matter holds the exact tokens, markdown prose underneath explains why they exist. The agent reads it once and stops guessing. The pattern that interests me is the separation of normative values from human rationale, sitting next to the code in plain text. Same move I'm making across skills and wiki entries. Point your coding agent at the repo and say "adopt this approach." Reorganised two of my projects in a day. (repo)
What I'm workshopping
2️⃣ Peter Steinberger's ClawSweeper reviews every open issue and PR in OpenClaw and only proposes a close when evidence is strong. Proposal mode and apply mode kept separate, every decision logged, closes restricted to a fixed list of reasons. The pattern that interests me is outside of code. Every business runs the same shape: emails, deals, tickets, backlogs, needing someone to ask "is this still real, actionable, gone stale?" Same harness shape applies. That's the build I'm scoping next. (repo)
What I consumed
3️⃣ Harrison Chase on continual learning for agents. The useful frame is that learning doesn't only happen at the model layer. Agents also learn at the harness and the context. Most teams default to "we need to fine-tune" when the bigger wins this year sit elsewhere. (article)
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
OpenAI
Launched GPT-5.5 across ChatGPT (23 April) and the API (24 April) as its most agentic model yet, hitting 82.7% on Terminal-Bench 2.0 and 58.6% on SWE-Bench Pro and pitched as the stepping stone to a unified ChatGPT/Codex/browser super app. (Apr 23)
Opened Codex Labs and signed Accenture, Capgemini, CGI, Cognizant, Infosys, PwC and TCS as a global GSI partner network on 21 April, disclosing weekly Codex developer usage has jumped from 3M to 4M in two weeks. (Apr 21)
Shipped ChatGPT Images 2.0 (gpt-image-2) across all plans on 21 April, adding "thinking before drawing", multi-image consistency and 2K output to chase Adobe and Midjourney on commercial creative. (Apr 21)
Anthropic
Took an additional $5B from Amazon on 20 April (with up to $20B more milestone-tied) and pledged $100B+ in AWS spend over 10 years for up to 5 GW of Trainium 2/3/4 capacity, making AWS its largest external compute customer. (Apr 20)
Bloomberg, Fortune and Tech Brew confirmed between 21-23 April that a Discord group has been continuously accessing the "too dangerous to release" Mythos preview model since 7 April, with one member identified as an Anthropic third-party contractor and the group exploiting URL patterns leaked through the Mercor breach. (Apr 21-23)
Published an unusually candid Claude Code postmortem on 23 April identifying three engineering missteps (reasoning effort default, idle-session bug, verbosity prompt) behind a month of degraded performance and reset usage limits for all subscribers as remediation. (Apr 23)
Google announced up to $40B in Anthropic on 24 April ($10B in cash now at a $350B valuation plus $30B more if performance targets hit), arriving four days after the Amazon deal and confirming Anthropic's run-rate revenue has surged past $30B. (Apr 24)
Google / DeepMind
Sundar Pichai opened Google Cloud Next 2026 on 22 April reframing the company around the "agentic era", with Vertex AI rebranded as the Gemini Enterprise Agent Platform and reporting 40% quarter-on-quarter growth in Gemini Enterprise paid MAUs. (Apr 22)
Split the 8th-gen TPU line into TPU 8t (training, 2.8x performance over Ironwood) and TPU 8i (inference, 80% better performance per dollar), the first time Google has shipped specialised silicon for the two halves of the AI economy. (Apr 22)
Launched Workspace Intelligence to unify Gemini across Gmail, Docs, Drive, Calendar and Chat (with Sheets gaining 9x faster prompt-based filling) plus Workspace Studio as a no-code agent builder for non-developers. (Apr 22)
Microsoft
Satya Nadella used his first Australian visit since 2019 to announce a A$25B (US$18B) Azure AI, cloud and cyber investment by 2029 plus AI skills training for 3 million Australians, the largest commitment in Microsoft's 42-year Australia history. (Apr 23)
Offered Microsoft's first-ever voluntary employee buyout to US senior-director-and-below staff whose age plus tenure totals 70+ on 23 April, framed by analysts as cost realignment as AI capex outpaces M365 Copilot adoption (~3% of 450M M365 customers). (Apr 23)
Moved Copilot Agent Mode to general availability across Word, Excel and PowerPoint on 22 April for all M365 Copilot Business, Enterprise and Education customers at no extra cost. (Apr 22)
Added Anthropic's Claude Sonnet to Microsoft 365 Copilot Chat (Frontier) on 24 April, with the new SharePoint AI site/library/page builder also Claude-powered ahead of GA in May. (Apr 24)
Meta
Formally confirmed 8,000 layoffs (10% of workforce) effective 20 May plus 6,000 cancelled open roles on 23 April, with CPO Janelle Gale framing the cuts as efficiency to offset $115-135B in 2026 AI capex. (Apr 23)
TechCrunch revealed on 21 April that Meta is installing tracking software on US employee computers to capture mouse movements, clicks, keystrokes and periodic screen snapshots from sites like Google, LinkedIn and Wikipedia to train autonomous-work AI agents under a "Model Capability Initiative". (Apr 21)
NVIDIA
Closed at a record $208.27 on 24 April with market cap topping $5T (intraday peak $5.12T) on the back of Intel's 22% YoY data centre and AI revenue beat, which sent Intel up 23.6% in its best single-day move since 1987. (Apr 24)
Apple
Announced on 20 April that Tim Cook will become executive chairman effective 1 September 2026 with hardware engineering chief John Ternus succeeding him as CEO, with the entire press cycle framing Ternus's defining challenge as fixing Apple's AI strategy. (Apr 20)
Bloomberg's Mark Gurman reported on 23 April that Apple is building six new AI-centric product categories — AI AirPods with cameras, smart glasses, an AI pendant, a smart display, a tabletop robot and a security camera — all designed around the upgraded Siri. (Apr 23)
xAI
Tesla rolled "Hey Grok" wake-word voice assistant to NZ and Australia in the 2026.2.6.1 update, putting Grok directly in NZ executives' cars even as CNBC's 25 April hands-on review flagged safety and privacy risks of conversational chatbots while driving. (Apr 25)
Perplexity
Set Anthropic's Claude Opus 4.7 as the default orchestration model powering Perplexity Computer on 20 April, also making it accessible to Max subscribers across web, iOS and Android. (Apr 20)
Amazon / AWS
Shipped Claude Opus 4.7 in Amazon Bedrock on 20 April with a 1M token context window, adaptive thinking and high-resolution image support, available across US East, Tokyo, Ireland and Stockholm at up to 10,000 RPM per account. (Apr 20)
Added a managed harness in preview to Bedrock AgentCore on 22 April that runs the full agent loop without orchestration code, plus an AgentCore CLI and AgentCore Skills targeting faster prototype-to-production for coding assistants. (Apr 22)
✨A few people have asked…
It’s Mike here, I run The AI Corner.
I’m not just into writing about AI. I run Allexive, and we help businesses grow without adding headcount by implementing AI platforms, and building AI systems.
👋 Mike & Erin