CEO at @multiply_co - Building the first collective intelligence Serial entrepreneur / multiple exits. Passionate about AI, human potential & co-creation.
Stockholm
Joined January 2014
- Tweets319
- Following344
- Followers60
- Likes693
A single person with agents can now run a function that used to take five.
Then they go to hire the second person, and the job ad reads like a description of a whole department, because the parts were never separate to begin with.
Every tool we build has an undo button but the organisation using it does not. A company that gets a decision wrong cannot take it back. It can only decide again, while most people miss the change.
I don't know what an undo would look like at that level. Decision Matrix multiplication? Something for Jev?
Jevon for the win. Again and again and again.
you'd think AI has already killed web design agencies, yet the web design market is forecasted to 2x in the next 5 years...
companies continue to pay exceptional people who deeply understand the theory of the things they're instructing AI to do
i believe that AI is killing entry-level roles, but empowering specialists to do their job better & faster
A handover used to be a conversation. Now it is a thread pasted into a model with "catch me up".
The summary is good. But it summarises the thread, and most of what the first person knew was never in the thread.
So the handover arrives looking complete and weighing less. I'm not sure whether anyone notices.
Agencies cut a three-week round down to four days. The round then waits eleven days on the client side.
Nobody saved anything. The waiting just changed address.
The odd part is that both sides describe this as the agency getting faster.
Rasmus Adler Wahlberg retweeted
Satya is right. The way to make SI safe is not to train it with a sense of self, its own moral philosophy, and permission to act as a conscientious objector. That’s the “alignment” approach and it magnifies the control problem.
The engineering approach that Satya describes is different: separate the supply of intelligence from authority over it. Surround non-deterministic models with deterministic controls, observability, privilege limits, logging, and the ability to always contain or shut them down. Treat models/agents like powerful insider risks, not moral patients whose psychological wellbeing is at stake.
As Satya points out, the most trustworthy system is the one that lets us trust the model the least — not the one that encourages the model to develop independent agency and grievances. Engineering safety is not the same thing as “alignment.”
You can give a model more documents. You cannot tell it that something stopped being true.
For example, most of what an agency knows about a client is corrections. The route they dropped, the word they hate, who actually decides.
Those arrive in calls, and the model is only ever shown the decks unless the context is available.
For everyone who thinks AI is eating the creative industry....
Jevon's paradox in every industry.
AI is making creativity more valuable than ever before and..
we have 3 years of data to prove it.
Introducing the Creative Intelligence 2026 report from @contralabs.
Full report: c87.to/creative-intelligence…
In a body almost nothing reaches the brain. You pull your hand off a hot pan before anyone has decided anything.
Companies have no version of that. Every signal gets routed to a person, and every person has a queue.
So the organisation ends up running at the speed of its slowest conversation, including for the parts that were never a judgement call in the first place.
A metastudy on generative AI and creativity found two things in the same data. Individually, people's output was rated more creative. Collectively, the pool of ideas they produced got narrower.
Both results are real. We have built an entire measurement habit that can only see the first one.
The study is by Alwin de Rooij and Michael Mose Biskjaer. Nanna Inie, who researches the cognitive impact of AI at NVIDIA, pointed me at it.
dl.acm.org/doi/full/10.1145/…
Producing something that looks finished got cheap this year. Working out whether it is right did not.
Those used to be close to the same job. The effort it took to make a thing presentable was also roughly what it took to be sure of it.
They have come apart, and I do not think any team I have seen has staffed for the half that did not get cheaper.
Agree, but also Billable Hour will meet it's new friend The Billable Token.
Agencies et al build AI capability alongside human capital and charges a margin on top - same as for hours.
AI won't kill the billable hour.
It'll just make the work better.
@jason:
"Maybe it takes an hour, or they still bill you for ten hours, and you just get a better report with better results."
The way anyone picked up the standard was that somebody argued with them about one piece of work until they could see it too.
That required the work to be worth arguing about.
I learned what good looked like by having work handed back to me. Nobody called it training. It was what happened when a first version was not good enough and somebody had to say why.
A lot of that was miserable and I would not ask for it back.
First versions are fine now, so nobody says why. That was the whole apprenticeship, and it ran on the work being wrong.
Two people take the same question from different angles and within a week one of them has been told to stop.
Not because the second angle was waste.
Because merging two half-answers into one good one takes a senior person and three uninterrupted hours, and no week in an agency contains that.
Everyone in the building got the same licence on the same day.
Six months later a few people have rebuilt how they work and most have an account they open when they remember it. Same job, same team.
The honest objection is that this happens with every tool ever rolled out. What is different is that both halves still have to produce one standard of work, and nothing in the rollout was about that.
One person usually notices in week two.
The company gets there in week seven.
Nobody measures the five weeks in between.