Long-form writing, with cross-links to the sites I run under democratising.ai, including AIinASIA.com and UrbanDogOwner.com, where the topics overlap. The canonical executive writing lives here. Two regular threads: the Friday Frame (weekly, 200-500 words) and the Reading list (quarterly). All views are solely my own.
On August 2, the transparency rules went live. Chatbots must disclose they're AI. Deepfakes must be labelled. The high-risk deadline got pushed to 2027, so most companies relaxed. That might be exactly the wrong response.
A friend's design house lost its production work and its first-year jobs over several years, without a single month where anything visibly happened. Gen Z's adoption is holding steady, excitement is down 14 points and anger is up 9.
AI can be at the table. It can't be in the dock. Three answers I gave a panel about what I wouldn't use AI for, and a fourth I nearly missed.
If reviewers miss a third of dangerous agent requests, what you've written into policy as a control is really a story you're telling the board.
AI suppliers lobbying against a ban on their biggest competitor are quietly disclosing something about their own commercial model. Most boards haven't worked out what that means for their risk register yet.
OpenAI is shutting down Atlas nine months after launch. The lesson isn't that browsers failed. It's that the agent capability moved into the surface users already open, and it doesn't fire the pixels boards still budget against.
Vendors are quietly shifting from flat licences to usage-based pricing. If the board hasn't built a token-consumption governance layer, next year's AI spend has no ceiling.
OpenAI will have ad sales on the ground in every major advertising region within a year. If boards haven't decided what ChatGPT is to them by then, procurement will decide for them.
Google's AI Overviews are the antitrust question wearing a search costume. Boards asking what their own model can do are missing the bigger question their CFO is about to own.
Google DeepMind's Project Genie at Cannes turned the storyboard into a playable prototype. The pitch becomes the prototype. Agencies have to decide whether they own the prototype layer or orbit it.
Most founders think their job is to be wanted by everyone. It's the opposite. The ones who win choose someone to be unmissable to, say a polite no to the rest, and let the market confirm the choice. Here's the case, with the data and the stories behind it.
Most founders think their job is to be wanted by everyone. It's the opposite. Choose someone to be unmissable to, say a polite no to the rest, and let the market confirm the choice.
Boards keep asking 'what's our number?' But headcount is the output of a decision, not the decision itself. What's changing is the unit of work, and the junior-heavy pyramid that was built to deliver it.
A single seat for AI accountability is a comfort to the board, not a fix for the operating model. The work belongs distributed, or it doesn't land at all.
When the reporting layer finally sees what was always there, a lot of marketing orgs are going to discover they've been optimising against a fiction. The repricing won't be gentle.
Share of voice was about being seen. Share of prompt is about being cited. Most operators don't have a number on it yet.
Vendors are being asked to commit to business outcomes they cannot solely steer. Outcome ownership without control is not a contract. It is a story.
Japan does not have a Sam Altman moment or a DeepSeek moment. What it has is a patient accumulation of laws, line items, and small models that, taken together, form the most institutionally serious AI strategy in the Asia-Pacific. It is the most under-read story in the region.
Every late-night save is a system that did not catch the issue earlier. The best operators stop rewarding heroics and start rewarding boredom.
Singapore is not racing China on chips or training a frontier LLM. It is doing the slower thing. Building a rulebook, a compute plan, and a workforce ladder that other Asian countries can plausibly adopt. That is starting to look like the most replicable AI strategy in the region.
Most AI adoption is the second. Coping faster than competitors is a real advantage, for a while. But it has a shelf life.
If a vendor commits to an outcome and the customer has not nominated a single named owner internally, the deal has already failed. It just has not noticed yet.
The tool produces an answer in four seconds. The organisation still takes four weeks to act on it. The bottleneck has not moved. It has just got more visible.
A first-day-of-Lunar-New-Year reflection. The Horse is forward momentum. The discipline I want to take into 2026 is choosing what to run towards, and what to carry with me when I do.
Taiwan can't treat AI as abstract. It hosts the infrastructure, feels the weight of every training run on the grid, and is forcing the question most markets are avoiding: does the benefit justify the load?
AI is not new. It has cycled between hype and AI winter since the 1950s. What is new is the scale of the bet, and the fact that the people making decisions about it often have no shared vocabulary for what they are actually buying. That is why the literacy events matter.
Singapore is not just a city-state. It is a live experiment in intelligent design applied to everything from urban planning to economic policy. After nearly 15 years living here, I keep coming back to one question. What if more companies actually ran themselves the way Singapore runs itself?
Predictive analytics. LLM hallucinations. Reinforcement learning. The boardroom has a vocabulary problem. With AI reshaping every part of the business, senior leaders cannot afford to speak only one language, especially if it is not tech.
Bylined work first published on the AIinASIA / AIinArabia / AIinEurope mastheads. Canonical lives at the source.
Publicis is paying 2.167 billion dollars for LiveRamp. From inside an APAC adtech acquisition, the read on what changes Monday morning for operators running real media budgets.
From an operator who has watched Gulf demand connect to APAC supply for seven years, the read is about sovereignty as much as adtech.
French ownership repackages the legal exposure of a US identity stack under GDPR and the AI Act. Three calls EU operators will need to make.
At 4:30 in the morning of 7 May 2026, EU co-legislators agreed the AI Omnibus, deferring Annex III obligations to December 2027. The hard deadlines have not moved.
South Korea is the first country in the world running a fully enforced national AI law. Its own startup association admits only 2 per cent of AI startups are export-ready.
Singapore's 40,000 worker pivot, India's 150,000 developer programme and Japan's 2.3 million target have made AI fluency the baseline screen.
Japan does not look like a country in an AI race. What it has is a slow accumulation of laws, line items and small models that taken together form a distinct doctrine.
In seventeen days, three deals redrew Europe's frontier AI lab map. SAP, Schwarz Group, and ASML are now doing what US venture capital used to.
Singapore is not racing China on chips or training a frontier LLM. It is doing the slower thing: a rulebook, a compute plan and a workforce ladder others can adopt.
Falcon-H1 leads the Open Arabic LLM Leaderboard. ALLaM 34B ships in Saudi. Fanar 2.0 holds Doha. Three Arabic AI bets, one language.
The Cloud and AI Development Act is meant to unlock the tripling of EU data centre capacity. The binding constraint is no longer permits. It's electrons.
PIF's HUMAIN venture is Saudi Arabia's answer to the global AI race. Sovereign AI is not a single thing: data residency, model residency, or talent residency.
I publish irregularly. The Friday Frame ships most weeks. The longer essays land when they are ready. Drop your email and you will get a note when something new goes up, or when an existing post gets a meaningful rewrite. Nothing else.