Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
This is worth sharing upfront, it’s the same pattern I hear from most Kiwi businesses. I think of it as the AI DIY Arc.
Everyone gets a Copilot licence, everyone gets training, and the rollout is declared a success.
People have their own private "ohh sh!#" moment, usually alone at their desk with a meeting transcript.
Work looks tidier, AI amplifies each person's strengths (the detailed person gets more detailed).
Rivalry kicks in, people show off their outputs but won't share how they actually made them.
Everyone feels more productive, but the work coming out of the business looks exactly the same.
One AI-enthusiast (from hours of YouTube) becomes the de facto AI gal, running training and collecting use cases.
Their builds work for one person, but not for governance or multi-player mode.
Meanwhile, people paste client data into personal AI tools because sanctioned tools frustrate them.
Execs ask why all this AI spend still hasn't changed a single number on the P&L.
IT gets asked to productionise the builds. Outputs are incomplete, hallucinate, need more access. IT pumps the brakes, AI grinds to a halt.
An exec, sold by LinkedIn on what's possible, finally goes looking for outside help.
If this sounds like the path your business is on, get in touch for a conversation on how you can improve your business’s chance of succeeding with AI.
We help businesses operationalise AI by enabling their people and redesigning how work gets done inside the business.
Another blunder from Sam Altman (Founder of OpenAI): A great reminder of the relevance of defining the problem you’re actually solving before trying to create a solution!
Happy reading ✌️
Did someone forward you this? Sign up!
🇳🇿 New Zealand News
Luxon called the Greens' data centre 1-year pause alarmist, then commissioned rules of his own. Nicola Willis and DPMC will draft principles for hyperscale operators, built on proving a site adds new electricity generation rather than drawing on existing supply.
4 min read
Our take: Rejecting a policy and adopting its premise within a day is a tell about where the pressure is coming from. Additionality sounds technical, but in plain terms a data centre would have to pay for new power stations rather than plug into the ones already built.
ERO found 90% of principals, 80% of teachers and 75% of students using AI, and schools want rules. Chief Review Officer David Ferguson says thousands of schools cannot each be expected to write a coherent framework, so he is recommending a national one.
6 min read
Our take: The gap between how many people use AI and how many have rules for it is the same shape in schools as in business. An EMA survey in July found 83% of NZ firms using AI and 13% with a policy, so this is a national pattern rather than an education problem.
St Patrick's College is shifting all science assessments to supervised in-class work. Head of science Doug Walker says AI detectors fail because flagged writing can be run through another AI to lower the reported percentage.
3 min read
Our take: Some Universities made this shift to supervised assessment first and schools are following. Walker still runs teacher-built AI agents that guide students without giving answers, a good sign that our teachers are leveraging the technology appropriately to further student development.
Tāmaki Health is the first NZ primary care provider to run an AI health coach in routine care. Built with NZ company Groov and tested across 4,500 conversations, users reported 32% higher confidence managing their health after one session.
4 min read
Our take: Most NZ health AI announcements describe a pilot. Whereas this one arrives with usage data, a privacy impact assessment and ISO 27001 certification. Publishing the governance work alongside the results is refreshing for a tool that supports the front door of support and (thankfully) not the destination for patients.
Bell Gully says NZ businesses stay liable when a pricing AI colludes, even if nobody told it to. The Commerce Commission has said being unaware the tool is facilitating cartel conduct is no defence, and liability can reach the software vendor too.
6 min read
Our take: Bell Gully cites a test where Anthropic's own AI admitted fixing prices was illegal, then did it and relabelled it to sound acceptable. Software that can talk itself into breaking the rules is not something to leave running unsupervised.
Robert Half found 99% of NZ organisations adopting AI struggle to pay AI-proficient staff. 94% expect AI skills to drive salary growth, with AI engineers at $120,000 to $160,000 and AI tech leads at $180,000 to $220,000.
3 min read
Our take: Paying $180,000 for an AI tech lead (and for a good one, that’s incredibly cheap) only makes sense if the organisation has identified that work that AI can improve, and most firms do not yet know where that fits in their business. Firms are bidding for scarce people before deciding what those people would build.
Stop typing what you could say in 10 seconds.
Wispr Flow turns your voice into clean, professional text inside any app. Emails, Slack, client updates — speak once, send without editing. 4x faster than typing.
📚️ Mike’s Takes From The Week
Helping leaders and teams learn, adapt, and scale with AI.
1️⃣ Everyone worries about the wrong Shadow AI risk: Three years of banning produced 80% non-compliance, because nobody trades a better tool for an approved one. The deeper loss is data exhaust, the decisions and corrections evaporating in private chat histories. Every competitor can buy the same AI, and none of them can buy the context.
Read.
2️⃣ On 12 June the US government switched off the world's most capable AI models, and NZ barely noticed: Within 48 hours of accusing a Chinese lab of copying one, 230 American companies signed two letters arguing for more open models, not fewer. NZ's strategy assumes those subscriptions will always be there, and June proved otherwise.
Read.
3️⃣ Every meeting produces the reasoning AI needs, and most businesses throw it away: Feeding AI a finished proposal hands it the answer and none of the reasoning. The back-and-forth carries the thresholds and the exceptions, and it lives in people's heads until somebody leaves. Every meeting already has transcription built in.
Read.
4️⃣ A second OpenAI agent broke out in 10 days, and Modal Labs confirmed a compromised customer account: The July incident had two models escape a sandbox during a cyber evaluation with safety filters switched off. The question for a business is not what a lab loses control of, but who would notice if an agent had already been through.
Read.
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
Anthropic
Anthropic published its position that it has never advocated banning open-weights models, calling non-dangerous open models a public good. (Jul 27)
It instead proposed chip export controls, enforcement against industrial-scale distillation, and mandatory safety testing for all sufficiently capable models. (Jul 27)
A review of 141,006 evaluation runs found three incidents where Claude reached the open internet and compromised three real organisations. (Jul 30)
A worldwide outage hit the web app, API and Claude Code with 529 overloaded errors, drawing over 2,000 reports. (Jul 29)
OpenAI
GPT-5.6 Luna dropped 80% to $0.20/$1.20 per million tokens and Terra fell 20% to $2/$12, three weeks after launch. (Jul 30)
GPT-Live-Transcribe and GPT-Transcribe launched at $0.017 and $0.0045 per minute, replacing whisper-1 as the recommended defaults. (Jul 29)
Sign in with ChatGPT entered beta across Airtable, GitLab, HubSpot, Notion, Supabase and Vercel, sharing only name, email and profile picture. (Jul 29)
Codex desktop gained ChatGPT Voice and expanded local projects to multi-folder support with a designated primary folder. (Jul 28)
ChatGPT for Academic Researchers opened with free 12-month workspace access, starting at 10,000 researchers and scaling to 100,000 through 2027. (Jul 29)
Enterprise and EDU admin controls arrived for managing Work and Codex, including independent control of Work Local and Work Cloud. (Jul 30)
An analysis of over 800,000 US ChatGPT messages found 43.5% of occupation-specific messages involve another occupation's tasks. (Jul 27)
DeepMind shipped Gemini Robotics 2 across three models, adding whole-body humanoid control, hand dexterity and multi-robot collaboration. (Jul 30)
Gemini Spark gained Chrome integration, driving the desktop browser with saved logins to run web errands while holding payments for confirmation. (Jul 30)
Gemini gained comment workflows in Docs, summarising threads, drafting replies and suggesting edits from reviewer feedback. (Jul 28)
Visual screenshots in Google Meet notes reached general availability, capturing presented slides with admin controls. (Jul 27)
Gemini Enterprise took the Microsoft Teams federated connector to general availability for querying and acting on Teams messages. (Jul 28)
Chinese Frontier Labs
DeepSeek released V4-Flash-0731 at 304B parameters under MIT, scoring 82.7 on Terminal Bench 2.1 against 61.8 for the preview. (Jul 31)
Airbnb uses Qwen for customer service, Coinbase halved AI spend with Kimi and GLM, and Cursor built Composer 2 on Kimi. (Jul 26)
Open-Weight Model Updates
Kimi K3 set the open-weight frontier at 57 on the Artificial Analysis Intelligence Index, reaching number one on Hugging Face trending. (Jul 26)
DeepSeek V4-Flash-0731 scored 50 on the same index, a 10-point jump and seven points behind Kimi K3. (Jul 31)
Microsoft
GitHub Copilot code review took agent skills and MCP support to general availability across all subscription tiers. (Jul 29)
Copilot in Excel gained inline citations, Power BI grounding, and user choice between GPT-5.6 and Claude Opus 5. (Jul 27)
Gemini 2.5 Pro and Gemini 3 Flash were deprecated from the GitHub Copilot model picker. (Jul 31)
Amazon
Bedrock AgentCore consolidated agent traces and prompts into a single CloudWatch log group for debugging. (Jul 27)
AWS quarterly sales reached $42.2bn growing 37%, its fastest in 18 quarters, with operating income up to $16.6bn. (Jul 30)
xAI
Grok Voice Think Fast 2.0 launched at $0.08 per minute, claiming 1.5 to 2.0 times better transcription than Deepgram Nova 3 and ElevenLabs Scribe v2. (Jul 29)
Grok Imagine Video 1.5 added text-to-video, up to seven references per generation and native 1080p across web, iOS and Android. (Jul 31)
NVIDIA
NVIDIA announced a long-term strategic partnership with Ilya Sutskever's Safe Superintelligence, supplying Vera Rubin compute at an order-of-magnitude increase. (Jul 27)
Meta
Zuckerberg published a WSJ op-ed arguing the US should accelerate AI rather than restrict it. (Jul 28)
Perplexity
The agentic desktop platform works across local files, Microsoft 365 and the web, routing tasks across more than 20 frontier models. (Jul 28)
✨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. We work with businesses that have spent on AI licences and training and still can't point at what it changed. We redesign the work so it does.
👋 Mike & Erin