NZ’s AI data centre fight begins

We still have slots open for the next four-week 1:1 AI Coaching Sprint. Get in touch if you’re interested to learn more

AI FAIL: Always remove the AI wrapper text before publishing.

Otherwise, you might reveal your ghostwriter, like this politician.

Happy reading ✌️

Did someone forward you this? Sign up!

Our take: A pause costs the government nothing politically until a foreign developer walks away, which is the risk Luxon is weighing. The climate impacts shouldn’t be overlooked. But, importantly, neither should the opportunity presenting itself, to which overseas investors won’t be looking for forever.

Our take: Import health standards are dense, precedent-heavy drafting where AI genuinely helps, because the work is assembling known requirements rather than exercising judgement. The risk is that plausible-looking standards clear review faster than they should, but then again, relying on fallible human judgement is not always foolproof either.

Our take: The cost of producing a complaint fell to almost nothing and the cost of assessing one did not, so every institution with an appeals process is about to be tested on volume. None of them are staffed for it or trained to handle AI inbound at scale.

Our take: A regulator arguing the legal floor is the wrong standard is unusual, because regulators normally defend the floor they enforce. Every large model trained on scraped web text already has te reo Māori inside it, so consent is being discussed well after the fact.

Our take: A farm's bargaining power in any data deal comes from being able to leave with its history, which makes export-on-exit terms count for more than ownership wording. Halter offers a request window rather than a right.

Our take: Phil Wheeler's point about discoverability inside AI systems has no established playbook yet, so firms fixing their product feeds now will be visible while competitors are still debating whether to bother. Being early is the advantage on offer. DM if you would like advice on what to do here, we can help.

Helping leaders and teams learn, adapt, and scale with AI.

1️⃣ The cheapest way to find a business's AI gaps is to build a Skill, not an agent: Ask four people to describe a process and four versions come back, because it was never written down. A skill has to be specified before it runs, so the problems surface within days: dirty data, undocumented rules, informal approvals. That list is the roadmap.
Read.

2️⃣ NZ businesses are stuck in single-player AI: one person, one chat, nothing inherited: McKinsey puts 90% of firms experimenting and under 10% capturing value. The test: if knowledge only moves when somebody remembers to forward it, it is single-player. Multiplayer means one person's output becomes the next person's starting point.
Read.

3️⃣ Five AI taxes almost every leader is paying, and the fix for each: Re-explaining, copy-paste, repetition, babysitting, checking. Each one exists because something lived in a person's head instead of a file: the company background, the decision rules, the format of a good output. None of the fixes are complicated.
Read.

4️⃣ AI coaching that rebuilds itself after every session, in three phases: Intake places a student on a nine-level fluency ladder and assembles a curriculum from 200+ modules. Each session debrief updates the running plan and drafts the next session. Admin dropped from hours to 30 minutes per student.
Read.

5️⃣ The vending machine habit is what stops AI from improving anyone's thinking: Send a prompt, take what drops out, move on, and nothing about the thinking changes. The alternative takes three habits: define the ask before typing it, make AI argue against the case, and refuse to accept the first version.
Read.

This week's guest is Carin Hercock, Country Manager at Ipsos New Zealand. Hear:

Helping advanced builders stay at the frontier of AI.

Context

1️⃣ Anthropic deleted 80% of Claude Code's system prompt and their coding evals didn't move. The old rules (no comments, always give examples) were guardrails for weaker models and now cost more than they buy. A good mirror of how I've been operating. My router plus on-demand rule files is the pattern they recommend, but the rule files are where I've over-constrained. Time to loosen the harness.
Article

2️⃣ Block open-sourced Buzz, a workspace where people and agents hold the same cryptographic identity. Every message, patch, review and approval is a signed event on a relay you host. A feature branch becomes a channel, so the merge decision sits in the thread that shaped it. Same context problem I sell into clients, one layer down: agents can't help with what they can't see. Watching, not adopting.
Buzz

Work

3️⃣ Cuban reckons the next big open-source application is a job simulator. His argument: juniors used to build judgement from touch points inside the company, and AI is removing most of them, so domain experts should build the scenario trainer instead. Lands on readiness, not displacement. The fix isn't more training content, it's capturing what experienced people know while they're still there.
Post

4️⃣ Kalanick raised $1.7B for Atoms, an OEM for what he calls atoms-based computers. The framing is the useful part: CPU is manufacturing, storage is real estate, network is transportation. Mining, construction and food production digitised on that model. Worth reading as a vision letter with a balance sheet attached, not as shipped industrial AI.
Article

The selected top headlines from each major AI tech company.

Anthropic

OpenAI

Google

Chinese Frontier Labs

Open-Weight Model Updates

Microsoft

Amazon

xAI

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.

Let’s chat if you’re interested to learn more →

👋 Mike & Erin