Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
We still have slots open for the next four-week AI Coaching Sprint.
A quick recap:
Weekly coaching sessions
We build one real workflow from your job into a working AI-enabled prototype, and
You learn the repeatable pattern to do the same across the rest of your work.
Hands-on building, not generic tool training, and we can cover Claude, ChatGPT and Copilot.
The format flexes to whatever you want to improve: getting more AI fluent in your day-to-day work, stepping up as the AI Lead internally, building a second brain, or building with Claude Code or Codex as a non-technical user.
We start with the fundamentals and go as deep as you want, right through to advanced context architecture and agent building.
If one of those sounds like you, get in touch.
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
Australia set up a national Office of AI while no NZ party has released a standalone AI policy. Victoria University's Dr Andrew Lensen called the Australian move a "kick up the bum" for NZ leaders, four months out from the November election.
2 min read
Our take: Albanese started by telling big tech to earn its social licence, which has no fine, no regulator and nothing to enforce it. Banking, medicine and aviation are legislated rather than self-regulated, so the standards bill is where this actually grows teeth.
Nasdaq-listed Sharon AI signed a US$1.32bn five-year cloud deal built on NZ data centre capacity. The buyer is described only as a "global AI lab", with revenue starting H1 2027 and more than 62,000 Nvidia GPUs planned by mid-2027.
3 min read
Our take: An unnamed buyer taking five years of capacity means someone has decided NZ is worth the latency penalty, which usually comes down to cheap power and cool air rather than proximity to customers. Despite the negative narrative around data centres, this is potentially the kickstart of a lucrative industry for NZ.
Whakarongorau gets $3.35m over four years to build AI triage for mental health and addiction calls. It sits inside a $10m+ telehealth package, with Lifeline receiving its first state funding in roughly a decade at $1.2m.
4 min read
Our take: The funding covers building the tool but nothing announced covers auditing it, and if the model started deprioritising a group of callers it would take months to show up in outcomes.
Three Auckland-linked AI companies raised in one week, with Icehouse Ventures in all. Spatial AI lab Hyades took US$1.25m from Tim Brown, Sir Stephen Tindall, Tony Falkenstein and MBIE's New to R&D scheme.
4 min read
Our take: This is a country that knows how to sell things dug up, grown or shipped, and now has three companies applying AI to that same physical layer. Land, minerals and materials are the one domain where NZ has actual expertise to encode.
Air New Zealand lifted on-time arrivals to 86% in June, up eight points year-on-year, using AI. The airline uses it to optimise which aircraft flies which route and to predict how long a delay will actually run.
2 min read
Our take: Most NZ firms have a version of this sitting unused in their operational data, and the reason it stays unused is that nobody owns the problem and the ambition behind growth in this country is not high enough, not that the technology is out of reach.
MBIE put 2% of NZ occupations at high automation risk, naming accountants and call-centre staff. Another 36% face heavy AI augmentation, concentrated in roles paying $50,000 to $150,000, while an AI Forum survey found 45% of firms planned to lower headcount.
3-min read.
Our take: Mass job displacement narratives are tiring. We’ve been saying for nearly two years this narrative is scaremongering and misguided. Now, a ministry speaking from a place that has not transformed its own operations is estimating how hard it is to transform everyone else's. Implementing AI at scale is brutally difficult, and this POV is completely theoretical and increasingly falsified in data points. Customer Support and Software engineering were named as the first few functions AI would gut, and the demand for both functions has held up instead. The predictions keep landing on functions that turn out to be far more resistant than the people making the predictions assumed.
📚️ Mike’s Takes From The Week
Helping leaders and teams learn, adapt, and scale with AI.
1️⃣ Starbucks did not dump Microsoft and IBM, and SaaS is not dead: Starbucks wants to rebuild an inventory system and a maintenance platform in-house, while still running on Azure. Its first AI inventory count was retired within a year for wrong counts. Rent the core, build the gaps.
Read.
2️⃣ Halter's support tickets grew 4x in 10 months. The fix was a confidence bar, not more hires: Claude Opus runs every triage, with nothing routed down to cheaper models. Certain words trigger an instant human handoff, no confidence score consulted. Pouria Asghari: "we will sacrifice cost over quality".
Read.
3️⃣ Nadella named the trap: firms are paying to train their own rivals: His Reverse Information Paradox says a company buys commoditised intelligence while handing over the judgement that makes it different. Generic work can run through shared models. Pricing and proprietary methods need the learning loop kept in-house.
Read.
4️⃣ Most Claude rollouts fail because everyone builds single-player: Nine mistakes from a dozen rollouts seen under the hood: migrating old workflows instead of redesigning them, dumping files in and calling it context, and saving a prompt and calling it a Skill. Faster tasks make the same business with a bigger cost centre.
Read.
5️⃣ Australia is legislating for AI while NZ heads into an election that will not mention it: Albanese announced an Office of AI and told big tech to earn its social licence. Anthropic is reportedly eyeing A$21B in local data centres. Only 22% of Australians think AI's upside beats the risk.
Read.
6️⃣ Apple v OpenAI is not an employment dispute, it is the hardware war starting: Apple claims over 400 former staff now work at OpenAI, some taking confidential files and coaching new hires around offboarding controls. None of it is tested in court yet. It could still delay the US$6.4 billion Jony Ive device.
Watch.
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
Anthropic
Published artifacts can now call each viewer's own MCP connectors on page load, with public sharing links and Team editor roles. (Jul 17)
Claude Code added a screen reader mode rendering plain linear text instead of boxes and spinners for VoiceOver and NVDA users. (Jul 14)
The Enterprise Admin API entered beta with programmatic member listing, role changes, invites and group management. (Jul 14)
Bedrock, Google Agent Platform and Claude Platform on AWS switched their default model to Opus 4.8. (Jul 17)
Session-wide caps now limit web searches and subagent spawns to 200 per session by default, both tunable. (Jul 17)
Anthropic, Blackstone and Hellman & Friedman launched Ode with Anthropic, a $1.5 billion enterprise AI implementation firm with 100 engineers. (Jul 15)
Anthropic entered early talks to lease up to $10 billion of compute from Meta over two years, paid monthly. (Jul 17)
Anthropic broke with OpenAI on regulation, backing stricter state-by-state rules and opposing federal preemption. (Jul 17)
OpenAI
The ChatGPT desktop app gained a global Chat and Work switcher, unified Recents, integrated Projects and cloud-synced Work. (Jul 16)
Unified search launched across chats, projects, images and documents on web, iOS and Android. (Jul 14)
Custom instructions expanded to 5,000 characters from 1,500 for Plus, Enterprise, Business and Education users. (Jul 15)
OpenAI shipped Codex Micro, a $230 macropad with 13 keys, RGB agent-status keys and a reasoning dial, its first branded hardware. (Jul 15)
Codex CLI 0.144.6 corrected GPT-5.6 Sol, Terra and Luna context windows to 272,000 tokens. (Jul 18)
OpenAI detailed GPT-Red, an automated red-teaming model that beat human red-teamers 84% to 13% on prompt injection. (Jul 16)
OpenAI published its case for age-appropriate ChatGPT, with under-18 detection triggering stronger safeguards. (Jul 17)
Apple's trade-secrets suit escalated, with allegations of recruiting-as-espionage and an unreturned laptop used to reach Apple's network. (Jul 13)
NotebookLM was renamed Gemini Notebook, keeping automatic redirects for existing links across its 30M+ user base. (Jul 16)
Gemini Omni landed in Google Vids, generating higher-quality clips and allowing video editing by text instruction. (Jul 16)
Bring Your Own Identity for the Gemini Enterprise mobile app reached general availability. (Jul 16)
Bloomberg reported Gemini 3.5 Pro is months behind schedule after retrained data failed to lift coding performance. (Jul 16)
Chinese Frontier Labs
Moonshot AI launched Kimi K3, a 2.8 trillion parameter MoE with 1M context and native vision, live via API and Kimi.com. (Jul 16)
K3 pricing undercuts Western frontier tiers at $3 per million input tokens and $15 per million output. (Jul 16)
Alibaba's Qwen won regulatory clearance to power Apple Intelligence in China across iOS, iPadOS, macOS and visionOS. (Jul 15)
Open-Weight Models
Thinking Machines released Inkling, 975B total and 41B active parameters, natively multimodal on input with 1M context, under Apache 2.0. (Jul 15)
Inkling scored 77.6% on SWE-bench Verified and 97.1% on AIME 2026, and went straight to number one trending on Hugging Face. (Jul 15)
Artificial Analysis had Claude Fable 5 at 59.9%, ahead of GPT-5.6 Sol at 58.9% and Kimi K3 at 57.1% across 165 tracked models. (Jul 18)
Microsoft
GitHub Copilot for JetBrains added bring-your-own-key support, letting teams run their own models through OpenAI-compatible custom endpoints. (Jul 14)
Microsoft 365 Copilot added governed agent publishing, tenant-wide prompt galleries and broader MCP access across Office apps. (Jul 15)
Microsoft reportedly trained sales staff to compare OpenAI, Google and Anthropic models unfavourably against its own MAI models. (Jul 15)
xAI
Grok gained Automations, running scheduled and email-triggered jobs with templates and run history. (Jul 16)
xAI open-sourced the Grok Build agent harness and TUI under Apache 2.0, roughly 72 hours after a privacy disclosure. (Jul 16)
Apple
Apple posted $20.5 billion in Greater China sales in Q2 2026, up 28% year on year. (Jul 15)
Apple's trade-secrets suit against OpenAI and io Products threatens OpenAI's device ambitions. (Jul 13)
Meta
Twenty-six current and former employees sued Meta alleging AI was used to select workers for discriminatory layoffs. (Jul 14)
Anthropic reportedly holding early-stage discussions with Meta about leasing AI computing capacity. (Jul 19)
Perplexity
Perplexity shipped Brain, a self-improving memory graph, with Opus 4.8 in Fast mode and Claude Fable 5 as orchestrator. (Jul 13)
Amazon
AWS positioned Amazon Quick as an agentic AI teammate for sales organisations. (Jul 17)
Bedrock Managed Knowledge Base gained enterprise search for agent workloads, with new patterns for agentic vision through MCP servers. (Jul 16)
🤦 ️ 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