23–30 Nov 2025 (SEQ Edition)
Good Morning from the Coast 🌤
This week is about platform direction.
Google pushed hard with Gemini 3 + Opal + Stitch + Pomelli + Nano Banana Pro, Anthropic and MIT dropped serious productivity/exposure numbers, and Karpathy basically declared AI homework detection dead.
Our job now: pick a platform strategy, build systems on top of it, and train teams to use them well.
🚀 TL;DR (for operators)
Google steps up: Gemini 3 Pro preview, Opal, Stitch, Pomelli + Nano Banana Pro put serious weight behind a unified Google stack. (Google AI for Developers)
Anthropic: new study suggests current AI could roughly double labour‑productivity growth by +1.8 percentage points. (Anthropic Research)
MIT Iceberg Index: 11.7% of wages already technically automatable; not just tech jobs. (The Iceberg Index)
Karpathy on AI in class: detectors are finished; kids should be assessed on how well they use AI, not whether they use it. (Rude Vulture)
This week in Tech Horizon Academy: Vibe Coding for Beginners – build your own content studio + task system with Google‑aligned coding agents, no code required.
🧱 Platforms — Why Google Suddenly Matters Again
Gemini 3 & Nano Banana Pro
Google released gemini‑3‑pro‑preview, a new flagship model focused on stronger agentic reasoning and coding. (Google AI for Developers) Early benchmarks and coverage suggest it’s competitive or better on many coding/reasoning tests, and it’s wired straight into Google’s ecosystem (Search, Workspace, Android). (Business Insider)
On top of that, Nano Banana Pro – an image model built on Gemini 3 Pro – can read exam page screenshots and solve questions directly in‑image, which is what prompted Karpathy’s “homework is over” comments.
Opal, Stitch, Pomelli
Opal: Google’s free-ish AI app builder; lets you wire up Gemini (and Veo) into simple apps without deep engineering. (Coming Soon to Australia)
Stitch: AI UI designer that generates mobile/web layouts you can export into code. (Stitch)
Pomelli: on‑brand content engine aimed at small businesses (think “Gemini‑powered brand copy hub”). (Google Labs)
Together, this is a systems play: models + app builder + design + content, all inside Google infrastructure.
What this means for you
If you’re already in Google land (Workspace, Android, ChromeOS):
A hybrid strategy makes sense:
Keep your existing primary model (ChatGPT / Claude) for what it does best.
Add Gemini 3 for anything touching Search, Docs, Sheets, and Slides, plus image work via Nano Banana Pro.
If you’re still half‑in, half‑out:
Decide in 2026 whether you:
Commit to a Google‑centric stack and migrate core docs + workflows, or
Stay multi‑vendor but define which jobs each model owns (e.g., “Google = docs + images; OpenAI = coding; Anthropic = policy/long‑form”).
Sitting in the middle with no rules is where teams get slow and messy.
Note: Microsoft has a lot of catching up to do…
📈 Data — Productivity & Exposure (Anthropic + MIT)
Anthropic: Estimating Productivity Gains
Anthropic’s new paper uses model‑estimated task times + real data to show that current AI could add around 1.8 percentage points to annual labour‑productivity growth if adopted at scale.
Their Economic Index report also shows:
Translation: early adopters are not just “trying tools” – they’re handing over more of the work and building pipelines. So they can spend more time generating revenue or scaling systems.
MIT + Oak Ridge also released research on a similar topic called the Iceberg Index
MIT’s Iceberg Index measures the share of wage value where AI can already perform tasks. They find: (Tom's Hardware)
Visible tech roles = 2.2% of wages (the tip).
Hidden exposure in admin, finance, healthcare, pro services, etc. = 11.7% of wages (~US$1.2T).
Key takeaway: if a role is heavy on documentation, coordination, email or spreadsheets, assume AI can already handle a chunk of it.
🎓 Education & Talent — Homework Is Over, Evaluation Isn’t
Karpathy’s message to schools: “You will never be able to detect the use of AI in homework. Full stop.” (Rude Vulture) Detection tools are trivial to bypass; models like Nano Banana Pro solve exam questions from screenshots.
For kids, the metric should become: How well do you use AI?
Can you critique it?
Can you cross‑check it?
Can you build something with it?
For teams, the same applies. Don’t reward “doing it manually” – reward sound judgment and ability to design a workflow that uses AI well.
🧑💻 This Week at Tech Horizon Academy
1. Systems, Not Shiny Tools (Google Focus)
Inside Tech Horizon Academy, we’re now running weekly Gemini + Google Workspace sessions with drop in times for support.
The goal is to:
Design repeatable systems on top of Gemini 3 (Docs, Sheets, Slides, Drive).
Wire Opal, Stitch and Pomelli into workflows for content, reporting, and simple internal tools.
Show where Google fits alongside OpenAI/Anthropic in a hybrid stack.
2. This Week’s Live Workshop: Vibe Coding for Beginners (No Code)
This week’s workshop builds directly on our AI Social Media Workshop. We’re going from “AI writes posts” to “AI runs your content & task system.”
You’ll walk out with:
A content creation studio:
Idea capture → research → draft → review → schedule, all mapped into a single board.
A task management system:
Intake → prioritisation → agent‑assisted execution → human QA.
A working prototype app built via Vibe Coding:
You describe the workflow in English;
We use coding agents + no‑code tools to scaffold it;
You never touch raw code.
👉 Join the Academy + workshop:
academy.techhorizonlabs.com
You’ll be in a room (live or replay) with other Australian early adopters doing the same thing every week: testing tools, building real systems, comparing what actually works.
🛠 Operator Playbook (This Week)
Keep it simple and actionable:
Pick your Google stance.
Hybrid: “We stay multi‑vendor, but Gemini owns Workspace tasks + images.”
Migration: “We standardise on Google for docs + search and plug others in only where needed.”
Instrument one workflow.
Choose a Google‑touching process (e.g. client reporting).
Build a Gemini‑assisted version in Docs/Sheets.
Measure time, error rate, and “number of clicks to done.”
Define “good AI usage” in your team.
One paragraph in your handbook describing what “excellent AI leverage” looks like (sources, checks, outputs).
Ship a no‑code artefact.
Join Vibe Coding, build the first version of your content or task app, and put 1–2 people on actually using it for two weeks.
🧭 Strategic Takeaway
This week’s signal is clear: platforms are consolidating and the data now backs what we feel in our day‑to‑day – AI is already shifting productivity and job exposure, not in theory but in numbers.
Google’s Gemini stack is no longer a side option; it’s a serious contender with a coherent app‑building story. Anthropic and MIT show how fast the ground is moving under admin and knowledge work. Karpathy’s classroom warning is really a workplace warning: stop pretending we can police AI; start training people to use it well.
Tech Horizon Academy exists to make your team unstoppable in that environment – not by chasing every tool, but by mastering the systems that sit on top of them.
—
Huxley Peckham
Tech Horizon Labs