What I'm sensing in AI.

I read more about AI than is reasonable, honestly. Most of it doesn't hold up after one week of real work. This piece is my attempt to hold on to the parts that do — eight threads that have changed how I think and act over the past few months, with a focus on what actually changes a decision rather than what makes a good demo.

The shortest version: five years from now, we'll probably remember 2026 as the year speed became strategy.

The pace, not the feature

What stands out right now isn't any single Claude feature — it's the rhythm. New capabilities are landing almost every week: agent teams, design tooling, the rumoured one-prompt-to-app workflow, an agent view inside the coding tool itself. The tooling layer that start-ups raised hundreds of millions to build is being absorbed back into the model.

The implication is clear. A few public data points are adding up: a major chip company reporting that effectively all of its software engineers now use AI coding tools daily; slide-deck workflows that used to take a day now come out of a single prompt with the right skill and a frontier model. The IDE is no longer the main place where work happens for more and more kinds of products. If you're planning a twelve-month roadmap that assumes today's frontier capability, you're already behind.

Architecture is becoming boring — and that's a good thing

Multi-agent marketplaces look exciting on slides but fall short in real use. The pattern that's actually working is smaller, well-bounded agents with clear responsibilities and interfaces — boring in theory, easier to manage in practice. Agent registries are starting to feel like a missing basic building block; the major cloud providers are racing to ship one.

Discipline beats numbers.

Underneath, something else is quietly building. Knowledge wikis maintained by an LLM, fed from a few hundred sources and refreshed continuously, are starting to appear inside large companies. The insight isn't the architecture — it's that LLMs are extremely good at the small upkeep that humans usually drop. Company knowledge starts to build up rather than reset every time someone leaves.

Scepticism is the new tone

Long-form journalism this spring — a New Yorker profile, a widely-read Daring Fireball critique, a serious book — has shifted the conversation among people who actually use these tools, when it comes to OpenAI. Projected losses over the coming years look very large, the "super-app" direction is unpopular among architects, and the Microsoft partnership seems to be breaking apart. None of this is fatal. But it's no longer all-positive cheerleading either, and that change of mood matters when you're deciding budget.

The wider context is even louder. Inside the bubble of people who use these tools every day, it's easy to forget that elsewhere AI is treated with deep suspicion. "Why did you use AI?" is a normal response in everyday life. The public conversation about AI in society — its benefits, its harms, its costs — is still very much in motion, and worth paying attention to even when the demos keep getting better every week.

Risk is moving into the boardroom

Two concerns are moving from blog post to the executive agenda. The first is the snake-eating-its-tail risk: frontier models trained on more and more AI-generated data, and the role of synthetic data in balancing it out. The second is agentic data risk — workflows where agents quietly change data at a speed and scale humans can't check in real time. Regulated industries are starting to ask what proof an executive team needs before they sign anything off. The honest answer right now is: we're still working it out.

Add security to that picture. The most-discussed incident of the spring wasn't a model jailbreak. It was a major platform broken into through a compromised third-party AI tool's OAuth integration. The lesson is old — wrongly set up third-party apps have been tripping up infrastructure teams for decades — but the attack surface is now ten times larger. Role-based access control, and giving every agent only the minimum access it needs, are no longer optional from day one.

What this means for the work

Putting it all together, here's what I see. The frontier is moving faster than most planning cycles. Architecture is getting more sensible. The economics around one big lab are looking shaky. Risk is being taken seriously where it matters. And outside the room, scepticism is the default mood. None of that should slow you down. All of it should make you choose your bets more carefully.

The good news is the bets are clearer than they were six months ago. Point AI at flow, not tasks. Treat agents as a governance problem, not a magic trick. Build up your knowledge instead of resetting it. Build the security model on day one. And keep listening to the people outside the bubble — they're often right about the parts that matter most.

I'll keep updating the field notes as the picture shifts. This is the snapshot for now.