How AI is reshaping how I think, work, and create

We built the product. We even signed an LOI. Then the champion left and the pilot disappeared with him. It turns out that is the default: most enterprise AI shows no measurable P&L return. On AI Outcome Engineering, and the chain most teams run backwards.

I spent my career building systems that help companies keep things, and it took me years to see the obvious: the companies with the most storage were not the ones that remembered. On corporate amnesia, why AI agents are about to make it expensive, and the quiet race to build the layer that lets organizations actually learn.

AI is the fastest-adopted technology in the history of work, and firm-level returns are flat. Both are true. Enterprise AI adoption is two different adoptions wearing one name, and the gap between them is the most important number in business.

OpenAI paused its largest planned frontier RL run and put a number on the cost of watching its own models, roughly a fifth of the compute. What the Critical threshold actually means, and why a pause with a price tag is not a new equilibrium.

In June 2026 the industry stopped talking about prompting agents and started talking about designing the loops that prompt them. What loop engineering actually is, where it came from, where it pays, and where it is just an expensive while loop.

Foundation models are commoditizing. The lock-in is not. It moved up the stack into prompts, memory, skills, and the muscle memory of how I work. A field report on where model lock-in actually hides and what I do to keep it shallow.

Most agents still wake up on a schedule and re-scan a lake to find out whether anything happened. Delta Trigger inverts that: change feeds as the source of truth for what changed, a predicate registry for what matters, and a durable bus for who should act.

Software has spent decades perfecting static interfaces. Knowledge work refuses to be static. Contextual Generative Experience is the argument that the interface should be assembled for the moment, from typed components, grounded in real data, and rendered with craft.

I built a team of 13 AI agents and named every one of them after someone I actually worked with, at Yahoo, Cask and Google. Who they are, what each one does, and why names and characters changed the work that comes out.

Cognition has, for the first time, two priceable substitutes. The ratio of Token Capital to Human Capital is becoming a number worth measuring, and worth interrogating.

A skeptical follow-up. The deployment epoch I wrote about last week assumes embedded engineering scales. The history of software, and the recent history of China's middle platform, suggests it does not. Why the Forward Deployed Engineer model the foundation model labs are betting on may turn out to be a bridge, not a destination.

On the same afternoon, OpenAI announced a four-billion-dollar deployment vehicle and Anthropic announced a one-and-a-half-billion-dollar response. Both reached for the same playbook. Palantir's. The era of selling tokens is over. What replaces it reshapes systems integration, vendor consolidation, and the architecture of the enterprise itself.

The AI industry is replaying the oldest pattern in computing. Vertical integration versus horizontal abstraction. Christensen's Law of Conservation of Attractive Profits explains where the real value will land.

AI is already rewriting design work inside enterprise teams. The execution-heavy version of the job is shrinking. The strategic, empathetic, end-to-end version is the only one that survives.

We all love chatting with AI in chat windows. The real revolution is happening behind the scenes. Salesforce just made it impossible to ignore. A reflection on headless LLM interactions and where this goes next.

I used to think every API needed an MCP server. JIRA taught me otherwise. A reflection on when MCP earns its keep, when direct tool calling wins, and how the multi-user auth story is finally maturing in 2026.

The first four months of 2026 produced very different shipping logs from Anthropic, OpenAI, Google, and xAI. Read together, they reveal four very different strategic bets about where AI value will actually accrue.

The three-layer engineering stack works in most situations. The harder skill is knowing the situations where it does not. Six concrete failure modes where the right move is to switch the stack off and reach for something smaller.

Impeccable, by Paul Bakaus, is a Claude Code skill that gives the model the actual vocabulary designers use, then enforces it across 23 commands. After running it on dashboards, marketing sites, internal tools, and client work, it has become a permanent fixture in my stack.

Three open-source tools, gstack, Superpowers, and Visual-Explainer, turn a single Claude Code terminal into a coordinated engineering team that thinks strategically, enforces process, and communicates visually.

After months of building production-grade systems with Claude Code, here are the hard-won practices that have cut token consumption by 60-80% while improving output quality and reliability.

I've tested more AI products than I care to admit. Most disappoint. Here is the brutally practical eight-dimension framework I use to decide whether any tool is actually worth my time and money.

AI coding assistants are fast. That's the problem. Speed without verification is how you ship shell injection, database corruption, and backwards alerts, all in the same session.

Most AI coding workflows assume you're starting from scratch. Here is the exact flow I use with gstack and Superpowers to safely improve production code I already have.

My entire day-to-day existence is now tethered to tokens. When they run out, I don't just lose productivity. I lose my damn mind.

A spec-driven workflow layering Claude Max, gstack, and Superpowers into a multi-model, multi-defense-layer engineering system that consistently ships clean, production-ready features.