Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
The Anthropic Mythos/Fable shutdown is by far and away the most interesting AI story of the year so far. I did not think that was possible given the courtroom battles, IPO chatter and general AI chaos already happening (writing this from the warm weather of Fiji. Keeping it stripped back to minimise laptop time).
Not because it proves the AI valuation bubble is ending, or that open source and Chinese models suddenly win (I reckon the model itself matters much less than the reaction to it). In six months, Fable 5 will feel archaic.
The interesting part is the decision around it all. A frontier model was marketed like strategic capability, then treated like strategic capability by a government. That is a very different category from normal software access, and it shows what can happen when cyber and scientific capability are part of the hype pitch (Dario created a rod for his own back here).
My bet is the decision gets softened or reversed soon, but the precedent will hang around for some time. Once a model can be restricted, delayed or switched off for reasons outside the vendor and customer relationship, access risk becomes part of the strategy and, crucially, what someone can charge.
That changes how serious AI adoption gets built. Particularly by the haves and have nots. Routing between models increases, open source with secure hosting becomes more interesting, and sovereign AI becomes less about building the next Anthropic and more about whether businesses and governments can keep operating when the best foreign model is unavailable.
That feels like the major lesson here. The model is largely irrelevant. And I’ll be sharing more about my take this week on the topic.
Reminder: We’re hiring a Lead AI Engineer at Allexive. DM me or apply here.
Happy reading ✌️
Did someone forward you this? Sign up!
🇳🇿 New Zealand News
New Zealand's National Cyber Security Centre was confirmed as one of 150 organisations globally to receive access to Anthropic's Mythos AI model, which can detect and exploit software vulnerabilities at machine speed. The NCSC's Deputy Director-General Catriona Robinson said access would "strengthen our national cyber security mission". The announcement landed the same day as US President Trump signed an executive order requiring early government review of powerful AI models for national security risks.
3 min read
Our take: The Trump executive order signed the same day signals that the US is treating advanced AI as a national security asset requiring active review, and that posture will eventually create pressure on allies to formalise their own governance frameworks for frontier model access. NZ has a model-sharing arrangement before it has a legal framework for governing what that sharing authorises, which is the sequence most governments are working through right now.
Professor Paul Young at Wellington Hospital received a $5 million Health Research Council Programme Grant to lead the REVOLUTION trial, described as the world's first randomised controlled trial testing whether AI-guided machine learning improves ICU survival rates. The trial will recruit more than 24,000 patients across 50 intensive care units in New Zealand and Australia. It will build on data from the Mega-ROX trial, which enrolled 40,003 patients from 137 ICUs across 14 countries.
3 min read
Our take: An RCT at this scale is the infrastructure the rest of the healthcare AI conversation is waiting for, because without clinical trial evidence, every AI-in-health deployment sits in a governance grey zone where the system cannot be confidently approved, funded, or held accountable when things go wrong. The machine learning target here is oxygen delivery in ICU patients on life support, which is instructive about where AI adds value in clinical settings; it is not replacing a clinician's judgement across a complex presentation, it is finding the optimal value for a single measurable parameter across a very large data set faster and more consistently than a human reviewing individual cases can. That specificity of task is what separates deployable clinical AI from aspirational clinical AI.
Napier-based Taylor Corp Apple doubled its packhouse processing capacity using AI grading technology and robotic forklifts after recovering from Cyclone Gabrielle in 2023, reaching 12,000 to 14,000 cartons daily through around 3,000 apples a minute. The AI grading machine captures up to 500 images of every apple to detect splits, black spots, stem punctures, and codling moths. The $30 million technology upgrade absorbed a doubled crop without additional headcount, with staff moving into higher-skill maintenance and oversight roles.
3 min read
Our take: The primary sector AI story in NZ tends to be told in terms of future potential rather than present deployment, so a family-owned operation in Hawke's Bay that has already doubled throughput without a new building or a bigger crew is worth noting as evidence that the infrastructure exists and is working at commercial scale, not as a pilot or a proof of concept. The Fieldays Innovation Hub this week featured several more examples of the same trend.
Newsroom reported on 11 June that Datagrid's approved one-gigawatt AI data centre near Invercargill would, at full build, draw twice the electricity of the Tiwai Point aluminium smelter at a cost exceeding $5 billion. The project has fast-track consenting approval. The piece noted NZ is simultaneously commissioning a billion-dollar LNG terminal to prevent electricity shortfalls, raising direct questions about whether NZ's renewable energy credentials can withstand the planned AI infrastructure build-out.
4 min read
Our take: The Tiwai Point comparison is instructive because that smelter dominated NZ's electricity politics for two decades (as we learnt), with its closure or continuation used as a proxy for the entire question of whether NZ has surplus renewable generation to sell, and the Datagrid project puts the same question back on the table at double the scale in a region that built its case for the facility partly on having access to that same surplus capacity.
Streetwear brand Huffer confirmed it uses "computer assistance technologies" in its marketing campaigns after model Elijah Timmins-Scanlon publicly claimed his likeness was recreated without consent or additional pay. On 6 June, Huffer sent Timmins-Scanlon a legal letter threatening proceedings under the Harmful Digital Communications Act; on 11 June, the Free Speech Union wrote back arguing Huffer has no standing to bring such a claim. By 14 June, Te Ao Maori News reported four models had supplied images they say resemble Huffer ads, and a lawyer warned NZ law does not adequately address AI-generated likenesses.
4 min read
Our take: NZ contract law for modelling work was not written for a world where a brand can train an image of a model and redeploy their look without additional agreement, which is why four people can come forward with images they say resemble themselves and nobody can confirm with certainty whether that constitutes a legal wrong or just an uncomfortable fact. The creative sector has been raising this for two years; it takes a retailer with a social media presence to make it national news.
Learn AI in 5 minutes a day
You don't have to scroll every AI thread, track every new tool, or watch every demo.
The Rundown AI breaks it all down for you — the latest AI news, tools, and tutorials in one free 5-minute email every morning.
Trusted by 2M+ professionals at Apple, Google, and NASA.
📚️ Mike’s Takes From The Week
Helping leaders and teams adapt, learn, and scale with AI.
1️⃣ Dario Amodei wrote America's AI homework. NZ's is different: The trigger was Claude Mythos, a model so capable at finding software vulnerabilities Anthropic wouldn't release it publicly. Four of his five policy chapters assume the reader builds frontier AI. NZ rents it. That changes which recommendations apply, and what our actual homework should be.
8-min read.
2️⃣ The AI jobs apocalypse is being sold to businesses that can't get AI to work yet: Dario Amodei put a number on it. Fifty percent of entry-level white-collar jobs gone in a few years, he said. The barriers most businesses face are undocumented workflows, dirty data, no governance, and no change management. The displacement debate runs ahead of the operating reality.
3-min read.
3️⃣ Fable 5 vs Opus 4.8: a blind test made the use case clearer: Opus was better at turning a complex QA agent task into something a team could talk through. Fable was better at designing the actual system, producing clearer plans for Jira comments, HTML reports, confidence flags, and release reporting. Best for messy planning where the output has to survive the meeting.
3-min read.
4️⃣ Canva gives 5,000 people one week to explore AI. Most businesses can't manage one afternoon: Twenty-six thousand hours of hands-on experimentation annually. One marketer built a seven-agent system saving an estimated 60 days of work. The bottleneck is never the tool. Nobody rebuilds their job in whatever scraps of time are left between client calls.
3-min read.
5️⃣ The shipping container and AI transformation: the box wasn't the lesson, the port was: In 1956, loading costs fell from $5.83 to $0.16 per ton. Companies that stacked containers onto existing break-bulk ships went bankrupt. Most businesses make the same mistake with AI, running new models through unchanged reports, handovers, and approval chains. Has anyone redesigned the port?
4-min read.
6️⃣ Most NZ businesses make one of two mistakes rolling out Claude: Group 1 maps everything before doing anything, and Group 2 gives everyone access while mapping nothing. The teams that move fastest run one workflow with one willing team and measure a real metric inside 30 days. Five steps.
2-min read.
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
OpenAI
OpenAI will acquire Ona (formerly Gitpod), providing pre-configured cloud environments so Codex agents can run persistent tasks inside customer infrastructure. (Jun 11)
Workspace agents will move to credit-based pricing on July 6, with typical GPT-5.5 agent runs estimated at 5 to 25 credits per task. (Jun 10)
OpenAI confirmed its confidential S-1 IPO filing, targeting a valuation near $1 trillion with Goldman Sachs, Morgan Stanley, and JPMorgan leading. (Jun 8)
ChatGPT upgraded its memory system with better context retention and preference-following, rolling out to Plus and Pro users in the US. (Jun 8)
Anthropic
Claude Fable 5 launched June 9 as the first publicly available Mythos-class model, with frontier-level coding, vision, and long-context performance. (Jun 9)
Anthropic reversed a policy of silently downgrading Fable 5 outputs for users working on frontier LLM development, apologising for the "wrong tradeoff" after researcher backlash. (Jun 10)
A US export-control directive blocked foreign-national access to Fable 5 and Mythos 5; Anthropic suspended both models globally, calling it a misunderstanding. (Jun 12)
Dario Amodei published two policy roadmaps: mandatory third-party testing with government authority to block unsafe deployments, and a $350M Economic Framework addressing AI-driven labour displacement. (Jun 10)
DeepMind published a 60-page paper mapping four pathways from AGI to artificial superintelligence: scaling, new algorithms, recursive self-improvement, and multi-agent collectives. (Jun 10)
Google confirmed Gemini powers Apple's rebuilt Siri AI under a multi-year deal reported to cost Apple approximately $1 billion per year. (Jun 8)
Microsoft
Claude Fable 5 launched on GitHub Copilot June 9, then suspended three days later under the US export-control directive affecting Anthropic. (Jun 9 / Jun 12)
GitHub Copilot's June 1 shift to AI Credits triggered sustained backlash, with developers reporting $180+ day-one bills and agentic workflows burning through monthly allotments instantly. (ongoing)
Meta
A leaked Meta memo revealed employees burned 73.7 trillion tokens in 30 days, prompting team-level spending caps and a new AI Gateway monitoring platform. (Jun 12)
NVIDIA
NVIDIA and SK hynix signed a multiyear deal to co-develop next-generation HBM memory for Vera Rubin supercomputers, RTX Spark PCs, and Jetson Thor robotics. (Jun 7)
xAI
SpaceX opened on Nasdaq at $135, closed at $161, raising ~$75 billion in the largest IPO on record, with xAI inside the listed entity. (Jun 12)
Grok V9-Medium (1.5 trillion parameters, 3x the current production model) completed training, with public release expected mid-June 2026. (Jun 9)
Amazon
OpenAI's GPT-5.5, GPT-5.4, and Codex reached general availability on Amazon Bedrock with full AWS security and governance controls. (Jun 7)
AWS Step Functions integrated with Amazon Bedrock AgentCore, letting developers embed AI agent reasoning steps into serverless workflows and trace every decision. (Jun 8)
Perplexity
Deep Research moved into Perplexity Computer, automatically routing complex research subtasks across 20+ frontier models to produce reports, decks, and dashboards. (Jun 11)
CEO Aravind Srinivas confirmed Perplexity remains on track for a 2028 IPO regardless of how the Anthropic and OpenAI listings perform. (Jun 9)
Mistral
Mistral is in talks to raise approximately €3 billion at a €20 billion valuation, nearly doubling its €11.7 billion Series C valuation from late 2025. (Jun 12)
✨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