
Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
US$44 million per employee, on paper. That’s what Anthropic’s employee equity worked out to, on average, at this week’s IPO price. One reported estimate put the value of employees’ shares and awards at about US$232 billion.
Anthropic employed about 5,300 people at the end of September, according to its filing. Divide the estimated value by that headcount and you get the US$44 million average. Individual grants vary in size, and some employees still have to wait for their awards to vest.
Anthropic itself raised roughly US$117.6 billion before fees from the public offering and the private sale that closed alongside it.
Two 1:1 coaching programmes are open: a four-week AI Coaching Programme built around existing tools and actual workflows, and an AI Executive Coaching Programme for C-suite leaders in larger companies. Reply for the details.
Happy reading ✌️
Did someone forward this? Sign up.
🇳🇿 New Zealand News
Halter’s new assistant gives farmers a daily report from their own records. North Otago farmer Nathan McLachlan told Farmers Weekly his 7am text includes hectares grazed, pasture cover and what the cows left behind.
Read
Our take: This is where AI is heading in every industry. A farmer could get a recommendation when pasture cover drops, then tell an agent which calls it can make inside limits the farmer sets. The farmer still hears what was decided and gets a say whenever something falls outside those limits. All industries are headed towards this companion like experience with the use of AI proactively recommending a next best action for humans to approve.
Auckland professor Alex Sims says scammers use AI to copy travel booking websites. RNZ also reported breaches at booking services. It did not establish that AI caused those breaches.
Read
Our take: Our take: Scammers have been quick to weaponise AI, but what's now possible on the defensive side is extraordinary. We can deploy AI agents to continuously scan the internet, identify cloned websites, cross-check suspicious domains against legitimate businesses and flag scams at a scale and speed that would have been prohibitively expensive just a year or two ago. The opportunity isn't really about catching more scammers, but rather to make it dramatically cheaper to detect and shut down scams before someone loses their money. That's where I'd like to see AI put to work.
The NZ Herald reports errors in the Ministry of Education’s AI-generated curriculum videos. The ministry acknowledged image errors. The Herald identified an incorrect map and an extra arm.
Read
Our take: Knowing the curriculum doesn't make someone an AI engineer, yet we're expecting curriculum specialists to use AI to generate educational content without necessarily understanding the technology behind it. The industry spends plenty of time showing how easy it is to create an AI-generated video, and far less explaining what separates a good one from a dud. That's where people who understand the technology, its limitations and how to test its outputs earn their keep. Our Kiwi No. 8 wire mentality can be our own worst enemy here. Just because we can build something ourselves doesn't mean we know enough to judge whether it's any good. That's unconscious incompetence, and the consequences are much greater when the output is what’s being used inside classrooms.
Victoria University lecturers told RNZ they are changing assessment as students use AI. Staff describe oral and in-person assessments. Some also say students use AI constructively.
Read
Our take: If AI can complete an assignment in minutes, we have to question what that assignment was actually measuring in the first place. At the same time, AI gives students access to something universities have never been able to offer at scale: a personalised tutor that can explain concepts in different ways, adapt to their level of understanding and help them learn at their own pace. The opportunity is much bigger than redesigning assessments to catch students using AI. It's rethinking how we teach, how students learn and what we actually need them to know.
SEEK told RNZ that 4.2% of its October NZ job ads mentioned AI. Information and communications technology had the highest share. B2B News published a separate comparison of the underlying job-ad data.
Read
Our take: I'm shocked that only 4.2% of job ads mention AI. For a technology that's fundamentally changing how work gets done, there's nowhere near enough emphasis on AI proficiency in the job market. It's a pretty telling gap between how much businesses are talking about AI and how much they're actually embedding it into their expectations of the workforce.
RNZ interviewed economist Daniel Susskind about how to prepare children for a world with AI. Susskind argues for teaching children about AI alongside foundational skills. The interview asks what families and schools can do as work changes.
Read
Our take: We're preparing children for a future of work that we can't accurately predict. Rather than focusing too heavily on today's tools, I'd put more emphasis on teaching kids how to think critically, solve unfamiliar problems, adapt and learn independently. AI should absolutely be part of that education, particularly as a tool to support personalised learning. But the real goal should be raising a generation that's comfortable navigating constant change, rather than training them for a version of the world that may no longer exist when they enter the workforce.
Other NZ stories
BusinessDesk reports that Infratil’s proposed Stratford data centre has entered consenting. The report does not establish approval or construction. Read
Stuff reports that Mainfreight warned about AI videos impersonating founder Bruce Plested. The company told the public to treat the videos as a scam. Read
Afterpay says 48% of surveyed NZ consumers used AI when shopping. Price and deal searches were the most common reported use. Read
Auckland startup Passit reported findings from a small AI companion trial with children. The company says it passed 34 planned safety tests and also observed unexpected answers to sensitive questions and too many parent alerts. Read
Hamilton is testing traffic lights that turn green ahead of five fire trucks. At 30 intersections, the trial combines vehicle tracking, AI and live signal data so drivers ahead can move before the truck arrives. Read
The Geneva conference on autonomous weapons agreed to continue work through 2028. Associated Press reports that the result fell short of a proposal by New Zealand and six other countries to begin talks on binding rules. Read
Rural News reports that an AI-assisted livestock parasite test is planned for New Zealand. The OvaCyte device uses image analysis to distinguish parasite species. Read
The National Cyber Security Centre asked New Zealand businesses to prepare for AI-assisted attacks. The November advisory builds on its September threat report and cites incidents overseas. Read
📚️ Mike's Takes
Helping leaders and teams adapt, learn, and scale with AI.
1. A daily meeting brief I can listen to. My morning brief checks my calendar, inboxes and client records. A CEO and I built his version so he could listen to it on his morning walk.
Read
2. How does a company value its own data? Decisions recorded in emails and spreadsheets contain knowledge that never reached the public internet. Would most companies know what to charge for it?
Read
3. Verification loops: how to check an AI answer before using it. I give the system a way to check its work, inspect what failed and try again. I use this on proposals, apps and reports.
Read
4. How I review work from my team. I ask AI to quote each passage it doubts and explain its criticism. That gives me a specific claim to check before I accept or reject the feedback.
Read
5. AI experiments might be fast but also create a bottleneck in the business. A customer service team can build an app quickly. Deciding who owns it, pays for it and maintains it takes work across the business.
Read
🌍 Tech Updates From Global
Anthropic
Anthropic closed its IPO at US$83 per share, with 1.127 billion shares in the public offering and a separate US$26 billion private placement. (Nov 20) Read
Claude computer use began rolling out in beta to Team and Enterprise, letting Claude operate desktop apps that users allow. (Nov 20) Read
Customers that Anthropic bills directly can pay for third-party apps’ Claude usage and set spending limits through the linked billing beta. (Nov 17) Read
Claude began rolling out a redesigned mobile app, starting with iPhones running iOS 26 or later. (Nov 20) Read
Claude Code 2.1.329 added support for using Code Mode with OpenTelemetry events and metrics to monitor individual tool calls. (Nov 19) Read
Claude Code 2.1.330 added sign-in for organisations managed through the Enterprise Admin Console and an Easel tool for drawing charts. (Nov 20) Read
Anthropic reported that Claude Fable 5.2 solved two open theoretical physics problems involving black holes and quantum gravity. (Nov 17) Read
Anthropic extended text watermarks to four older Claude models on supported platforms when requests use the default sampling settings. (Nov 16) Read
OpenAI
OpenAI released GPT-6.2 Astra for complex work and GPT-6.2 Sol for everyday tasks across ChatGPT Work, Codex and the API. (Nov 17) Read
ChatGPT for Sales introduced a Sales plugin and more than 30 new or expanded partner plugins for tasks including prospect research. (Nov 17) Read
ChatGPT Space began rolling out worldwide to adults on Plus, with shared pages and presentations that can update from connected apps. (Nov 20) Read
ChatGPT’s Accounting plugin helps reconcile transactions, draft invoice entries and prepare month-end reports from company records and templates. (Nov 18) Read
OpenAI announced improvements to clinical search and computer use in ChatGPT for Healthcare, alongside more than 20 new healthcare plugins. (Nov 16) Read
OpenAI began rolling out live dictation on all ChatGPT plans, transcribing speech as people talk and removing the previous ten-minute recording limit. (Nov 18) Read
OpenAI’s API began testing two-hour and eight-hour background processing windows at token rates 30% and 40% below standard rates respectively. (Nov 17) Read
ChatGPT began rolling out a dedicated dot tab and pinned tools on mobile to eligible adults on Pro and some on Plus. (Nov 19) Read
Google released Gemini 4 Neon for coding and analysis, with API pricing of US$3 per million input tokens and US$15 for output. (Nov 18) Read
Genie 4 became generally available through the Gemini API, where developers can generate and steer simulated worlds with video and stereo audio. (Nov 19) Read
Google made Gemini available on eligible lower-memory Android Go devices, including Gemini Live conversations in more than 45 languages. (Nov 18) Read
Google added a feature to Gemini Notebook that turns research sources into shareable interactive sites, including dashboards and project trackers. (Nov 18) Read
Gemini began rolling out Mac computer use in English in the US for adults on paid personal plans with Apple silicon and macOS 15 or later. (Nov 19) Read
The Gemini API began testing on-demand tool retrieval, which loads tool definitions when needed and supports remote MCP servers. (Nov 19) Read
Google announced that Blue Origin will launch Project Suncatcher satellites designed to run AI in orbit, with launch targeted for early 2029. (Nov 19) Read
Microsoft
Microsoft made Copilot Home generally available, bringing Chat and Cowork into one place alongside Office tools inside Copilot. (Nov 17) Read
Microsoft released Code 2 and Code 2 Flash for coding in GitHub Copilot and Microsoft Foundry, with Flash for routine tasks. (Nov 17) Read
Microsoft began rolling out Autopilot to Frontier customers to monitor Teams channels, follow up on emails and handle recurring work. (Nov 17) Read
Microsoft updated the Dynamics 365 ERP plugin with more than 39 finance and operations skills and wider Excel connectivity. (Nov 17) Read
Microsoft made voice agents generally available in Foundry Agent Service for developers building conversational applications that accept speech. (Nov 18) Read
Microsoft previewed Agent 365 Business Analytics for Frontier customers, with agent usage reporting, spending forecasts and recommendations across the organisation. (Nov 17) Read
GitHub Copilot added eight connectors in public preview, including Supabase for databases, Box for documents and Calendly for scheduling. (Nov 18) Read
Microsoft and Salesforce announced Copilotforce to bring Salesforce records and sales workflows into Copilot and connect Salesforce data to Fabric. (Nov 19) Read
Meta
Meta launched its Muse AI assistant in Australia, with integrations for local retailers and services rolling out progressively. (Nov 17) Read
Meta added Loblaw to Muse in Canada, where shoppers can plan meals, build baskets and check out with participating stores. (Nov 18) Read
Make added a Meta AI integration that lets users generate text and images or transcribe audio in automated workflows. (Nov 17) Read
Meta outlined a three-stage consultation with publishers and other stakeholders after pausing proposed commercial terms changes involving its web crawlers. (Nov 18) Read
Meta reported using stronger AI models to assess content directly as part of changes to Instagram’s recommendations in India. (Nov 16) Read
Meta described how it uses AI agents to test and adapt improvements across multiple models that rank which ads users see. (Nov 18) Read
NVIDIA
Amazon
Amazon OpenSearch Memory became generally available, letting AI agents store past interactions and retrieve relevant details across separate sessions. (Nov 18) Read
Amazon SageMaker added cross-region inference routing, so requests can switch to another endpoint when local capacity or endpoint health is insufficient. (Nov 18) Read
Amazon Bedrock AgentCore Memory added full-session summaries so an agent can retrieve a condensed account of what happened in an earlier session. (Nov 17) Read
Amazon SageMaker added tools to generate training data from documents, label examples automatically and route selected examples to people for review. (Nov 17) Read
Amazon SageMaker added serverless full fine-tuning, letting teams update all of a supported model’s parameters without provisioning their own infrastructure. (Nov 18) Read
👋 Mike












