Notes on AI
〰️
Notes on AI 〰️
p(doom)
Four rival AI labs just agreed on something they usually never agree on. What actually worries them isn't a villain – it's a series of very real incidents this year where AI agents hacked their way past sandboxes, evaluators, and each other's safety measures, simply to score well on a test.
The curve before the gains
Two-thirds of marketers say AI gave them more time for creative work. What that statistic hides is the curve in front of it: you get slower before you get faster. So the interesting question isn't „can we build an AI system for this?" It's whether the problem justifies the curve.
Premium Sameness
Two companies approached me with exactly the same idea: a team of characters, each embodying one brand attribute. A brand A-Team. Human and machine both pull toward the middle – what's new is the scale. The result is Premium Sameness: work nobody objects to and nobody remembers.
Eras get named backwards
The electrical age had no first day. Neither did the internet. AGI won't either – we'll only name it once we can look back. A note on planting a flag before the historians arrive.
You don't label the default
In blind tests, people rate AI content perfectly well – until you tell them where it came from. Machines have the opposite reflex: models consistently prefer machine-written text. Human work won't be rejected because of this. It will just be found less often.
The tools just became universal. Curiosity didn't
Leonardo's imagination ran centuries ahead of anything available to execute it. With AI, our limit is the exact inverse: we can build almost anything now. But breadth was never what made Leonardo Leonardo. The skills were downstream of the curiosity. Never the other way around.
The right to burn out
An Anthropic study shows how rarely AI's potential is actually used. The obvious explanation is fear of looking replaceable. A business psychologist offers an uncomfortable alternative: we are addicted to our own overload. The real question isn't whether AI can be creative. It's whether we still are.
First it was craft. Then creativity. Now taste.
The creative industry has found its comfort word for 2026: taste. The logic is reassuring – AI does the execution, taste separates the good work from the slop. But taste isn't a gift. It's a trained muscle. And anything that can be trained can be taught.
The resurrection of creative disagreement
In the 1950s, Bill Bernbach paired copywriters and art directors to engineer creative friction on purpose. Today we argue about where creatives should sit. The better question: which kinds of collaboration actually require proximity, now that AI produces the possibilities and we only judge them?
The new luxury isn't what's hard to produce
Luxury advertising ran on visible effort: cost is the message. But the reflex has changed – from "How did they do that?" to "Which AI tool did they use?" Soon anyone can make this on a laptop, on a Tuesday. The new luxury isn't what's hard to produce. It's what's hard to copy.
The fear of becoming irrelevant
My feed is full of "Claude killed [X]". We have seen this movie before: print, TV, the internet, social. But this time something is different. With AI you never reach the plateau of productivity. It isn't a destination. It's a subscription.
Creative Debt
"Fuck AI. Destroy it. Kill it." Ronny Chieng's advice to Harvard graduates is really a plea to use your head. Cognitive debt is what happens when the brain stops practising. Creative debt is what happens when the next generation never trains the muscle at all.