Helping CTOs & CIOs turn AI ambition into an operating model: feedback loops, governance, and execution across people, process, technology | CAIO @ Zühlke
Schweiz
Joined December 2011
- Tweets4.7K
- Following684
- Followers800
- Likes3.8K
1/ We ran a cheaper model next to our standard one. The results were about 10% worse, at about a tenth of the cost per token. In many use cases, careful prompting closes that gap.
4/ No gateway yet? Start with three things: measure tokens from day one, question the most expensive model for every new use case, and limit vendor contracts in time. Your architecture can only switch as fast as your contract allows.
5/ Nobody can seriously predict which models and vendors will last, or what they will cost. If your main model cost ten times more tomorrow, how long would the switch take you?
German: buff.ly/9eauWfh
English: buff.ly/Y5fJWCJ
1/
Proactive detection means finding problems while they're still whispers, not when they're shouting in production.
With AI in production, more problems start as whispers.
4/
Detecting that drift is the cheap part now. The hard part comes after the alert: who owns it, and who is allowed to act on a small drift before it grows?
An alert nobody owns is just noise.
5/
AI can signal a problem, but only a prepared organization can act on it before it matters.
Which of your last incidents whispered first, and who heard it?
buff.ly/Z6cHGwX
1/ When do I get my own robot at home, one that puts away my laundry?
That is how I opened episode 11 of Prompt & Proper. Steffen Ochsenreither (Swiss Post): you can already buy one. The real question is what you need it for.
4/ Steffen's point: the cost of a robot does not end with the purchase. Maintenance is a big block, and you get failures you did not have before. Even in Switzerland the robot is not always cheaper.
5/ Steffen thinks the laundry robot may come in two or three years. Until then I hang up my laundry myself. Where would a robot really pay off in your operation?
Episode 11 of Prompt & Proper (German): buff.ly/UWz0Ch8
1/
The model is swappable, the data is a mess, and your next customer might be an agent.
@gregisenberg came back from 48 hours with AI founders in San Francisco with 40 observations. These five matter most, and I see the same.
4/
Founders used to brag about headcount. Now it is revenue per human. Read it as augmentation: what each person can do with AI.
Software built around clicking through screens will struggle. Some startups barely have a UI, their main user is an agent.
5/
Greg: your product now sells itself twice. The agent picks it, then a human has to trust that pick.
The model is the easiest part to change. Which of these do you see?
buff.ly/qruBN39