Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
Thank you to everyone who filled in last week’s reader survey.
I won't share everything that came through, but thought a cross-section sample of the audience readership plus what others want to see would be useful.
Also, welcome back to the AI Fail!
Reminder: We’re hiring a Principal AI Engineer at Allexive. DM me or apply here.
Happy reading ✌️
Did someone forward you this? Sign up!
🇳🇿 New Zealand News
Luxon says NZ businesses are moving too slowly on AI. Prime Minister Christopher Luxon told the China Business Summit he is "constantly underwhelmed" by how slowly New Zealand businesses are adopting AI. He said government should lead by using AI to simplify public services.
3 min read
Our take: The public service angle is the most important part of this story. If government wants businesses to move faster, it has to prove AI can improve services without making trust, access and accountability worse, because bad implementation would slow adoption rather than accelerate it.
New Zealand's National Cyber Security Centre joined Five Eyes counterparts in warning leaders that AI is increasing the speed and complexity of cyber attacks. NCSC head Catriona Robinson said organisations need to prepare now for more vulnerabilities, incidents and business disruption.
3 min read
Our take: New Zealand firms should resist treating this as a spy-agency story that sits far away from daily operations. The same automation that helps attackers find weak systems faster will also punish slow procurement, unclear ownership and legacy platforms that nobody wants to fund under real pressure.
AI Forum NZ published Tom Maasland's analysis estimating that a credible ANZ frontier-adjacent AI model would cost NZ$3.5b to NZ$8.5b to build. The access-based alternative was estimated at NZ$600m to NZ$1.5b upfront, then NZ$150m to NZ$400m a year.
5 min read
Our take: Is the analysis put forward feasible, and a likely proposition? Considering France’s and the UK’s spend is in a similar ballpark, likely not. For businesses, the same logic applies at a smaller scale. Building internal AI capability does not always mean owning the model, but it does mean knowing the exit paths, fallback systems and data positions before a vendor decision becomes operational dependence, because executive comfort is not a resilience plan.
RNZ reported that the Electoral Commission is adjusting to the newly identified AI threat in the lead-up to November's general election. The concern is that the most advanced models can be used against election information, voter trust and democratic processes.
3 min read
Our take: The difficult part is timing. Election integrity teams do not get a long product cycle, and the period before voting is exactly when deepfakes, fake instructions and candidate impersonation can do the most damage before they are fully debunked. The Electoral Commission now has to defend public trust in a media environment where convincing false material can be produced quickly.
Reported that six in ten New Zealand CFOs believe AI will negatively affect the quality of future finance leaders. The concern is that automation may remove the junior tasks where analysts learn judgement, controls and commercial context.
3 min read
Our take: The answer is unlikely to be preserving manual work for nostalgia. The better test is whether firms redesign junior roles around review, exception handling and business explanation, because those are the muscles that matter once the spreadsheet work is automated, without asking graduates to pretend the old ladder still exists.
NZ co-founded AI startup Odyssey has reached a NZ$2.5b valuation while building world models rather than another text-based model. The company is focused on simulated environments, with potential uses across robotics, logistics and physical infrastructure.
3 min read
Our take: Odyssey is useful because it gives the local ecosystem a different pattern to study. Most New Zealand AI discussion still orbits chatbots and productivity tools, while world models point at physical industries where simulation can become a commercial moat, and where New Zealand has real industrial exposure.
📚️ Mike’s Takes From The Week
Helping leaders and teams adapt, learn, and scale with AI.
1️⃣ AI training is the easy 20%: Most teams learn to prompt, then keep doing the same work with nicer drafts. The missing 80% requires a new capability that most organisations aren’t equipped with.
6-min read
2️⃣ Claude may start asking for your ID: AI access is becoming a risk product. Claude's identity checks point beyond passports and privacy anxiety. As agents book, message, code and act across systems, platforms may start gating capability by identity, geography and risk.
2-min read
3️⃣ Claude Tag turns Slack threads into delegated work: Tag @Claude in the conversation and the context stays where the work is happening. For teams losing hours to app-switching, this is the agent pattern moving from demo to workflow.
2-min video.
4️⃣ China has been buying cheap Claude tokens: US models can cost 10-20x more than Chinese equivalents on routine work. The businesses matching model quality to task risk may build a structural cost advantage as usage scales.
2-min read
🛠 Latest Finds From The Web
Helping advanced builders stay at the frontier of AI.
1️⃣ Token spend is becoming an infrastructure problem. Brian Armstrong’s Coinbase note is a practical blueprint for the next phase of AI adoption: better model defaults, routing, caching, lean context, and spend visibility. The goal is not using less AI. It is making exponential usage economically sustainable.
Post
2️⃣ The AI race is becoming an energy race. WSJ’s data centre breakdown makes the next bottleneck painfully clear: Amazon has roughly 9GW today and could pass 35GW by 2030, while Google is racing from 5GW toward 30GW. Model quality matters, but power access may decide who can actually serve demand.
Article
3️⃣ Cheap Claude tokens are not really cheap. ChinaTalk’s transfer-station deep dive shows how Chinese developers access Claude at 70-90% below official prices through proxy markets. The disturbing part is the margin: model swapping, broken caching, and prompt logs that may become training data.
Article
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
OpenAI
OpenAI previewed GPT-5.6 as a three-model family: Sol for frontier reasoning and long-horizon agentic work, Terra at GPT-5.5 performance for half the cost, and Luna as fastest and cheapest. (Jun 26)
OpenAI and Broadcom revealed "Jalapeño", a reticle-sized custom inference chip built in nine months, the first accelerator in a multi-generation OpenAI compute platform. (Jun 24)
Codex Remote reached general availability across all ChatGPT plans, letting users start and continue work on a connected Mac or Windows host from the mobile app via QR pairing. (date unconfirmed, late Jun)
A new dictation model rolled out across all ChatGPT plans, cutting word error rate by at least 10% on top languages including mid-sentence language switching. (date unconfirmed, Jun)
Getty Images struck a multi-year deal letting OpenAI surface its 400M-plus licensed photo and editorial library directly inside ChatGPT search results. (Jun 21)
OpenAI limited the GPT-5.6 rollout to about 20 government-approved companies at Washington's request, saying such restrictions should not become the norm. (Jun 26)
Anthropic
Claude Tag launched in Slack beta for Claude Enterprise and Team customers. Tag can remember selected channel context and plan future tasks with connected tools and codebases. (Jun 23)
The US government cleared Anthropic to re-release Mythos 5 to about 100 vetted companies and federal agencies for defensive cyber use, ending a two-week blackout, with Fable 5 still restricted. (Jun 26)
Fable 5 was reportedly on track to return after a 15-day US government restriction. (Jun 27)
Micron announced an Anthropic partnership to supply memory and storage for frontier AI model development. (Jun 22)
Claude Code weekly builds added shell-based MCP login and logout, hierarchical sub-agents up to three levels deep, /rewind resuming before /clear, and a roughly 37% CPU cut during streaming. (Jun 22 to 26)
Anthropic told the Senate Banking Committee that operators tied to Alibaba's Qwen lab ran 28.8 million exchanges through about 25,000 fraudulent accounts to extract Claude's capabilities. (Jun 24)
Google / DeepMind
DeepMind took its first equity stake in a film studio, investing about $75M in A24 to build Veo-based filmmaking workflows starting with AI storyboards, with A24 keeping creative control. (Jun 22)
Gemini 3.5 Pro is expected to improve long-horizon tasks and agent workflows. (Jun 24)
Nobel laureate John Jumper's departure from DeepMind for Anthropic drove Alphabet shares down roughly 5% to 6% as analysts questioned its talent retention. (Jun 22)
Reports said Gemini 3.5 Pro slipped from June to July as Google works on token efficiency and long-task performance, with prediction markets giving even odds on a June 30 launch. (Jun 24)
Microsoft
Microsoft AI's in-house MAI-Code-1-Flash model reached general availability for GitHub Copilot Business and Enterprise, aimed at high-volume agentic coding. (Jun 26)
Claude entered public preview as an agent provider in GitHub Copilot for JetBrains IDEs, alongside org and enterprise custom agents and a per-turn AI credits indicator. (Jun 22)
Anthropic is in early talks with Microsoft to run Claude inference on Microsoft's custom Maia 200 chips via Azure. (date unconfirmed, Jun)
xAI
xAI added /goal to Grok Build, a long-running autonomous mode where the agent plans, executes to completion and self-verifies, with status, pause, resume and clear controls. (Jun 22)
Meta
Meta is building a standalone AI prediction-market app, Arena, where Llama auto-generates betting questions from trending topics and resolves outcomes with no human review, to rival Kalshi and Polymarket. (Jun 24)
Apple
Apple released iOS 27 developer beta 2 spotlighting the Gemini-powered Siri AI upgrade, with a more visible "Write with Siri" prompt and a Dynamic Island experience. (Jun 22)
🤦 ️ AI Fail Of The Week
We all love AI, but it’s certainly far from perfect 🤔 …
✨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