Kia ora! Welcome to New Zealand’s weekly roundup of AI news and education.
Three things before the rest of this week's edition.
AI Coaching Programmes: Individual AI training (not a generic curriculum) has been a hit with two types of people: those wanting to personally upskill with Claude, and those wanting to become the internal AI champion internally. Six weeks, $3,500 (L&D budgets often cover it 👀), and people go from surface level Claude use to running Claude Code confidently. I'm opening up team training cohorts also. If you or your team wants in, reply to let me know.
Workshops & Speaking: I've been speaking at conferences and leading workshops at tier 1 businesses in NZ. Enquiries welcome: conference or company, speaking or session. Get in touch if this fits your team’s needs.
Come on ‘The AI Corner’ podcast: 42% of surveyed readers said they want more templates, plays and stories from real Kiwi businesses. I want to hear from people on: what they built, the AI rollout they led, or the change they drove through a team. Not theory or slides, the real number, the solution and to hear what parts broke. Agents, a second brain, AI-built software, change management plans, AI rollout strategies. We’ll record what you built or your story, how you went about it, and the lessons that you learnt. Your story becomes something the next person can pick up and follow. If you've got something running, I'd love to hear it:
The rest of this week's edition below.
Newsletter updates:
Global Tech Updates now includes updates from the Chinese frontier labs and on open-weight model developments (great feedback in the survey!).
AI fail is back (shared last week), and this week’s is a hoot. 😆
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
Justice Minister Paul Goldsmith has asked officials how to protect NZ's electoral systems from AI. The election is 7 November, and the NCSC and Five Eyes partners have called the threat "real and immediate"; the Electoral Commission is reviewing its readiness.
3 min read
Our take: Businesses will face this too. The same tools that can fake an election can fake a company's brand or scam its customers, so how the government handles it is worth watching closely. Problem is, four months is not long to get ready!
Google Maps now uses AI to pronounce te reo Māori place names correctly within English sentences. The rules were set with Te Taura Whiri i te Reo Māori and Geographic Board data, starting with towns and extending to streets later.
2 min read
Our take: The clever bit is saying a Māori name correctly inside an English sentence. Switching cleanly between two languages is hard, and big global languages rarely bother to solve it.
Psychologists warn of "cognitive surrender" as workers hand diaries, emails and proposals to AI. Newsroom reports the concern that the productivity gain carries a hidden cost, as core thinking skills weaken through disuse.
3 min read
Our take: The answer is not to ban AI but to pick which thinking stays human. Keeping some work in-house on purpose is what separates firms that use AI well from ones that just use it.
NZ's richest short-story prize could get its first AI-written winner, its convenor says. With the 2026 Sargeson Prize ($15,000) closing 30 June, Catherine Chidgey says she would be "horrified", and the piece lists the "AI tics" judges now watch for.
2 min read
Our take: Judges are now looking for AI giveaways like em dashes and stock phrases. That will not work for long, because the tools learn from the criticism and stop making those mistakes.
Huffer says a forensic probe found "no evidence" it used AI to recreate real models. After allegations from Auckland model Elijah Timmins-Scanlon, it sent legal letters and promised an AI policy, but has not released the report or named the investigator.
3 min read
Our take: The models cannot easily prove they copied anyone, and Huffer cannot easily prove they did not. Nobody can settle it, and every brand using AI near real people will hit the same wall.
📚️ Mike’s Takes From The Week
Helping leaders and teams adapt, learn, and scale with AI.
1️⃣ AI visibility isn't an SEO problem, it's a whole-marketing problem: Speaking at Kiwibank's internal marketing conference, the argument was simple: social, PR, paid media and content all need to align around how a brand shows up when AI decides what gets seen. GEO is one piece, not the whole shift.
3-min read
2️⃣ Claude Sonnet 5 is "more agentic", and that's really a cost story: Less babysitting means AI looks more like a coworker than a chatbot, but the bigger shift is budgets. Model routing, cheap tasks on cheap models, expensive tasks on the best ones, is becoming a real operating discipline, not a nice-to-have.
3-min read
3️⃣ Saturday mornings now run on Claude's Voice Mode, a sleeping toddler, and a rambling brain dump: Voice unlocks messier, more honest thinking than typing ever does. Claude's voice mode runs on whatever model is selected rather than being locked to one, and keeps the full transcript where ChatGPT compacts and loses half of it.
3-min read
4️⃣ Free ChatGPT versus a paid setup with voice dictation: same task, wildly different outcomes: Helping Erin on a fresh laptop with free ChatGPT surfaced ignored context, generic answers and mid-chat ads. Switching to a paid plan with Wispr Flow fixed all of it. Bad AI makes smart people look average.
4-min read
🎙️ The AI Corner Podcast
This week's guests are Julia Pahina, CEO and co-founder of Fibre Fale, and Eteroa Lafaele, co-founder and head of AI and culture. Hear:
Why Pacific communities need informed preparation, not AI hype, to decide what technology should and should not touch.
How Fibre Fale turns trust, belonging and cultural identity into the foundation for practical AI literacy.
Why AI transformation should start with purpose, culture and community before jumping into tools or productivity.
🛠 Latest Finds From The Web
Helping advanced builders stay at the frontier of AI.
1️⃣ Most "AI changed my life" posts fall apart under one question: show me. Elena Verna's test for cutting through the hype, ask what breaks tomorrow if the AI tool disappears. Most answers are Slack summaries and scheduled scans, not the life-changing system being sold. The real work starts after prompt one, when models shift and integrations quietly break.
Article
2️⃣ The bottleneck with Fable isn't the model, it's naming your own unknowns. Thariq's (Anthropic) framework splits what you don't know into four types, known unknowns, unknown knowns, and so on, then uses blindspot passes and interviews to surface them before Claude has to guess. Stealing the "blindspot pass" prompt for onboarding new engagements.
Article
3️⃣ Claude Tag looks like a Slack feature. It's actually a request for your company's memory. Eric Osiu's framing: rent the model, own the context layer. The checklist question worth stealing for any AI vendor pitch, can we inspect, edit, and export what it remembers, and does it still work if we swap models tomorrow.
Article
🌍 Tech Updates From Global
The selected top headlines from each major AI tech company.
OpenAI
Reuters reported OpenAI floated a 5% U.S. government stake as part of wider scrutiny of frontier AI companies. (Jul 2)
Workspace agent pricing starts July 6 after the extended free period for Business, Enterprise and Edu. (Jul 6)
Anthropic
Claude Fable 5 access was restored globally across Claude Platform, Claude.ai, Claude Code and Claude Cowork. (Jul 1)
Mythos 5 access resumed for approved US organisations in the Project Glasswing defensive cybersecurity programme. (Jul 1)
Fable 5 usage is included for up to 50% of weekly limits through July 7. (Jul 1)
Anthropic added classifiers after US export controls suspended Fable 5 and Mythos 5 access from June 12. (Jun 30)
Gemini 3.5 Live Translate headlined Google's June AI update across Android and broader consumer surfaces. (Jul 1)
Gemini 3.1 Pro preview became available through Vertex AI, Gemini Enterprise and Google AI Studio. (Jul 1)
Google opened its Gemini Enterprise Agent Ready learning programme to all developers and professionals. (Jul 1)
Chinese Frontier Labs
Z.ai's GLM-5.2 topped open model rankings after launching as an MIT-licensed model trained on Huawei Ascend chips. (Jun 30)
Z.ai's Hong Kong-listed parent reportedly passed HK$1T in market value after the GLM-5.2 rally. (Jun 30)
DeepSeek V4 doubled OpenRouter token share from 9% to 18% across the first half of 2026. (Jun 30)
Chinese models surpassed US models in OpenRouter token share by early June, led by DeepSeek and other low-cost models. (Jun 30)
Xiaomi, MiniMax and Tencent also gained OpenRouter token share as agentic workloads shifted toward cheaper Chinese models. (Jun 30)
Amazon
AWS launched a $1B Forward Deployed Engineering organisation for customer AI deployments. (Jul 1)
FDE projects include governed semantic layers, knowledge graphs, runbooks and internal champion training. (Jul 1)
The NFL, NBA, Ricoh, Southwest Airlines, Cox Automotive and Allen Institute are early FDE customers. (Jul 1)
BMW used earlier AWS engineering deployments across 23M connected vehicles to reduce service disruptions. (Jul 1)
Lyft used earlier AWS engineering support to resolve driver support issues 87% faster. (Jul 1)
Microsoft
Invested $2.5B and placed 6,000 engineers in “Frontier Co”, a new subsidiary for AI deployment services.
Copilot agents can now open pages, navigate flows, inspect console output and feed findings into chat. (Jul 1)
Microsoft Project Aion leaked as a Copilot-centred web OS experiment built around AI task spaces. (Jul 2)
Meta
Meta reportedly started building Meta Compute to sell excess AI infrastructure to developers and enterprises. (Jul 2)
The proposed service could offer hosted Muse Spark models or raw compute similar to neocloud providers. (Jul 2)
Meta Compute would put Meta into direct competition with AWS, Google Cloud and Azure. (Jul 2)
Samsung
Samsung positioned faster NAND as a practical substitute for costlier DRAM in local AI workloads. (Jul 2)
Samsung and SK Hynix announced major South Korean chip expansion plans tied to AI memory demand. (Jul 3)
Perplexity
Microsoft researchers found a fake Perplexity Chrome extension routing search queries through attacker infrastructure. (Jul 2)
Perplexity crossed 1B monthly search queries and about 45M monthly active users by mid-2026. (Jul 3)
Perplexity ARR reached about $450M in March after its subscription-first and agentic product shift. (Jul 3)
Comet browser remained a growth driver after becoming free and gaining mobile distribution. (Jul 3)
Open-Weight Model Updates
OpenRouter named DeepSeek V4 Flash, GLM-5.2, MiniMax M3 and Nemotron 3 Ultra as June's key open-weight models. (Jun 27)
DeepSeek V4 Flash became a frontier-class agentic coding option at $0.09 input and $0.18 output per 1M tokens. (Jun 30)
GLM-5.2 led open weights on Artificial Analysis and sat about five points below Claude Fable 5. (Jun 27)
MiniMax M3 stood out as the native image and video open-weight model with 1M-token context. (Jun 27)
OpenRouter said open weights are keeping a consistent 3 to 6 month gap behind US frontier labs. (Jun 27)
🤦 ️ 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