2026

• Note · 14 Aug

  The end of a shared truth?

  Adobe and Omnicom announced a new partnership this week. The teaser video is slightly cryptic and the landing page reads like 100% AI slop, no human copywriter ever looked at that website. But underneath, the ambition is big.

  Read the landing page more closely and there are three parts: Acxiom Real ID is a database of 2.6 billion verified personal profiles (including you) "enriched with trillions of cultural, media, and commerce signals". Add to that Omnicom's domain knowledge in the verticals: what works, what doesn't, how you talk to people in automotive or financials or pharma. Finally, Adobe's production agents take care of creating the campaign materials.

  The promise behind it is a fully automated, hyper-personalised marketing pipeline. Take that to the extreme and you get a slightly different brand story for every single consumer. At least for every one of those 2.6 billion profiles.

  Targeted ads aren't new. But what happens when we create an individualised asset for every single person we're trying to reach?

  Did we just witness the birth of a new category of tools that erode trust? Not just in brands, but in each other.

  We've already seen a shift from higher-trust environments to lower-trust ones. In the UK you can really feel it. So what happens if we add individualised messaging for every single human being on top of that? How does that shape trust? How does it shape our ability to engage with each other?

  It looks a bit cobbled together and not quite cohesive yet, but if they actually pull it off, it's both the most "exciting" thing in advertising this year and another episode of black mirror become reality.

  

• Note · 3 Aug

  Cutting corners moves the work

  Something I notice on the Lime bikes and delivery drivers: people cutting across other people's lanes, turning without signalling, cutting the corner they shouldn't. It saves them two seconds. It makes everyone around them slow down, swerve, recalculate.

  We do the same thing at work. We vibe-code the thing and trust that it'll be right, we send that email draft chat made without reading it. We delegate the part of the path we were supposed to walk ourselves. It looks like speed. But someone still need to do the work we just skipped, the due diligence we were to lazy to do, or fix the bug that now causes actual problems in production.

  Skipping the path doesn't remove the path. It just moves it onto everyone else.

  Then nobody quite gets what they wanted, nobody feels good about it, and the whole thing takes longer than if we'd walked it properly the first time.

  

• Note · 30 Jul

  Doing AI well

  Doing AI well isn't about learning a new technology. It's about learning to articulate what you want, how you think, and how you actually work best.

  That's the harder skill, and the more durable one. The tools will keep changing under you. The capacity to know your own mind and say it clearly is what carries across all of them.

  

• Note · 22 Jun

  We rebuilt the typist

  For most of a century we had the typist, the secretary, or executive assistant: someone sitting with a manager's dictation, holding all the context and the relationships in their head, cleaning up the sentences on the way, making the small diplomatic edits so the letter landed the way it was meant to.

  We've rebuilt the typist in a machine. It takes our input, reformats it, sets it down. Same shape, different feel.

  We gained something for one: no one's locked behind a typewriter now. And we lost something too: the person attuned to exactly who they were writing to, swapped for something that knows a lot of frameworks, a lot of half-facts, and almost nothing about the room you're actually in.

  

• Note · 11 Jun

  AI works best at the extremes

  Two things happened while I listened to a podcast yesterday.

  The host said "listen to your body" — and an ad for Holland & Barrett spliced in, opening with a guided meditation, then selling supplements. Later she talked about sex and power and money. The next ad was for erectile dysfunction treatment.

  The same episode, AI-transcribed, gave me sound effects in brackets. Footsteps. A door closing. The texture of the room. Things a deaf or hard of hearing listener/reader wouldn't otherwise get.

  AI is at its best at the extremes, and somehow useless in the middle. Hyper-targeted manipulation on one end, real accessibility on the other.

  The question isn't whether to use AI. It's which extreme we feed. Or maybe rather how we can get from the extremes to more of a middle ground.

  

• Note · 20 May

  Using AI for inefficient tasks while preserving deep thinking

  Don't use AI to do the hard stuff for you. Use it to do the things you're inefficient at because computers are designed poorly.

  What are the things where you end up spending a lot of time just clicking boxes, copy-pasting things back and forth, or looking through piles and piles of information? Make it bring you your own thinking. Make it bring you other people's thinking. Make it lay it out for you.

  Then you get to do the putting together of puzzle pieces. You get to do the deep thinking. That's where you excel. That's where you've built your experience and your career. That's where you get a lot of joy from. I know I do.

  So how can you build systems where AI scaffolds the deep thinking process for you, but you still have to do the thinking? It might do some background research. It might surface somethings. It might pull out specific sections from an article, from something you wrote earlier, from rough notes, so you can recompose them.

  Don't fall for the promise that you don't have to do the thinking anymore.