Build logs from the agents I run every day: the checks that let an agent act alone, and the jobs that need no model at all.
AI Strategy
How AI changes the way firms decide, spend and take risk.
I teach AI governance and fintech at the National University of Singapore, and I co-lead the NUS FinTech Lab. I also run a small fleet of AI agents for my own work. This page has my research notes on how AI systems scale and learn, and memos on what that means for the firms that use them.
Other sections
Payment rails, bank–fintech partnerships, stablecoins and licensing: who carries the risk when a new rail goes around an old one.
Topics
Showing all posts
Latest
AI-Written Research Needs Verification Records More Than a Separate Venue
Researchers and information systems journals disagree on where machine-written papers belong. Wherever they go, the thing to certify is a record of how each claim was checked.
Is a 1,000-Agent Swarm Just as 'Smart' as a Single Agent That Thinks 10 Times Longer?
On Toby Ord's numbers, a 1,000-agent swarm burns far more compute than one agent that thinks longer, and it reaches a higher ceiling only where work splits cleanly and answers can be checked.
Reasoning Models Learn Best When Teacher Guidance Comes Before Reward Training
Distilling from a teacher first and rewarding answers second beats fusing the two; for post-training teams, step order and teacher access now matter more than prompt volume.
Who Owns an AI Agent's Memory
A study of 1,990 public practitioner records and 100 coded incidents finds that teams store agent memory in prompt rules and model weights, but rarely name anyone to reconcile the two.
Insurers Are Writing AI Out of Liability Policies
Verisk's CG 40 47 endorsement can remove generative-AI injury claims from general liability cover, so each customer-facing AI workflow needs a named owner for uninsured losses.
Why AI Productivity Does Not Show Up
Census Bureau data put U.S. business AI use at 17% to 20% from December 2025 to May 2026; the history of 1980s computing says the gains follow work redesign.
Who Can Block a Data Center
More than 20 Florida counties and municipalities have rejected, delayed or paused major data-center projects, so every AI business case should name its power and permit dependencies.
Agent Security Is Becoming a Shared-Responsibility Model
Anthropic secures the control plane of its self-hosted agent sandbox and customers own hardening and network egress: the same split that cloud providers use for security.
Who Carries the Risk When AI Spending Fails to Pay Back
S&P estimates that five large cloud providers could spend about $750 billion on capex in 2026. This memo asks who absorbs the loss if demand falls short.
Should a Small Business Accept Stablecoins?
The GENIUS Act and Singapore's MAS framework now regulate payment stablecoins, but a bare wallet address still leaves a small business to handle refunds, taxes and sanctions screening.
How to Audit AI-Assisted Work for Slop
SHRM's 2026 research finds that 41% of workers use AI, and just under half of them call their own output 'AI slop'. A sampled audit shows managers the review cost.
Treat Each ASEAN Country as a Separate Market Test
ASEAN's digital trade exports reached USD 387 billion in 2022, but founders should still choose a first country by what it tests about their buyer.
How to Decide Which AI Pilots Deserve Funding
Cases from Morgan Stanley, Klarna and Amazon's Blue Jay robots show how to score an AI workflow before a pilot gets money.
Archive: earlier memos, July 2026 (22)
- When AI Disclosure Is Due
- What Browser Agents Can Touch
- Who Can Use Your Source Material
- Who Is Accountable When AI Acts
- What to Open, Block and Measure When AI Search Takes the Click
- When AI Assistants Remember: Memory as a Control Problem for Firms
- Which Models for Which Work
- AI Assistants Are Taking Over the Moment a Buyer Decides
- Who the Agent Is Acting As
- What an AI Disclosure Record Must Contain
- What AI Benchmarks Miss, and the Production Evidence Executives Should Ask For
- What Agents Are Allowed To Do
- How Often AI Work Should Run
- How Much Inference a Workflow Deserves
- Proof That AI Changed the Work
- Where Founders Should Publish
- Evidence an AI Project Must Produce
- When Every Team Asks the Same Model the Same Question
- Is This AI Quality or Just Theme Exposure?
- Is This AI Strategy Real?
- AI Needs a Management Layer
- Decide the Rules Before Buying Tools