Practitioner perspectives
Writing
Practitioner conversations and independent analysis on applied AI — what works in production, and what moves the multiple in PE-backed companies.
Your AI Agent Has a Trust Problem. Technology Won't Fix It.
A CEO of an Asian digital bank built AI agents that work. Customers don't use them. Better AI doesn't fix the trust gap — you're solving a relationship problem that predates AI by a decade.
Read article →The Filter Upstream Changed. You'll Feel It in 18 Months.
Seed investors are deprioritizing pure software companies. The filter just changed — and most growth-stage investors haven't noticed yet.
Read article →Follow the Incentives, Not the Pitch Deck
VCs have access to non-public financials, board-level strategy, and real-time metrics across dozens of companies. They often still make consensus bets. Most founders never trace the incentive chain far enough.
Read article →Your GPU Budget Is Not Your Ceiling
Hung Bui built a research lab in Hanoi that published at NeurIPS, ran a 7B model on a phone without internet, and sold to Qualcomm for nine figures — with Vietnamese undergraduates and a fraction of a frontier lab's compute budget.
Read article →The Requirements Problem
Every major AI lab now hires forward deployment engineers. Most are building the role wrong — optimizing for implementation speed when the bottleneck is upstream.
Read article →The Agent Virus
10x coding output is also a 10x attack surface. The quality of code reviews is going dramatically down — and machines don't get suspicious.
Read article →From LLM loops to learning agents
Today's AI agents are clever loops that don't learn from experience. The next wave closes the feedback loop — borrowing from a decade of reinforcement learning — and the infrastructure to do it is already here.
Read article →The Network Gap
Your AI works and the company still fails. In traditional sectors — insurance, logistics, healthcare — technical superiority is not enough. The bottleneck is access.
Read article →Turning AI prototypes into real products: lessons from Pixel 10
Building a new AI app means no historical data, locked marketing deadlines, and ML teams thinking in probabilities while everyone else expects Gantt charts. Here's what it actually took to ship Pixel Journal.
Read article →How Vibe-Coding Changed the Game
Vibe-coding made writing software trivially fast. But turning code into production-grade systems is still slow. The bottleneck moved; the org chart didn't.
Read article →Figma Is Being Eaten
Designers are skipping Figma entirely, vibe-coding directly to prototypes. The bottleneck isn't tooling — it's implicit knowledge that no agent can read.
Read article →Macromill: Fragile Data Moat and the Case for Self-Disruption
CVC needs AI-powered automation to justify a tech multiple, while AI is the primary threat to the asset they're trying to re-rate. The way out is to disrupt the panel business before someone else does.
Read article →Benesse: Two Bets, Two Playbooks
EQT bought Japan's largest education company for $2B. Two businesses, one company — each needs a different thesis. Where the real value creation lever is, and what the operating partner needs to execute.
Read article →Yayoi: The Accounting Software Playbook, Japan Edition
KKR is running the standard PE incumbent playbook on Japan's leading accounting software. This is where the standard playbook breaks — and what the operating partner needs to execute before the window closes.
Read article →Coupa: Debt, Data, and the Bottomline Bet
Thoma Bravo bought the #2 procurement SaaS player at a distressed price. The thesis isn't software optimization — it's a multiple expansion play built on $1.3T in annual spend that Coupa barely monetizes. Three years to prove it.
Read article →