Models & Research
Stop Asking Which Model Is Best. Ask Which Mistakes You Can Live With
Every week brings a new leaderboard winner. For most people, the model matters less than the habits around it and the kinds of errors it makes.
by Maya Okafor
#004
Which Mistakes Can You Live With?
The Ideabro Brief #4: which mistakes can you live with?
A friend asked me last week which AI model is "the best one". I gave her an unsatisfying answer, then wrote a column about it.
The short version: "best" changes every few weeks, and for most of us the model matters less than the habits around it. A better question is which mistakes you can live with. A lawyer and a copywriter need very different failure modes, even if they use the same tool.
Lena's explainer on tokens is the companion piece. Once you understand what a token is, pricing pages, error messages and strange spelling mistakes all start to make sense.
Then two pieces about making things by hand: Theo's conversation with an illustrator who draws every line herself, and the rise of a new promise on market stalls, "made by a human".
— Maya
In this issue
- 1
- 2
Learn AIWhat Is a Token? The Tiny Unit Behind Every AI Answer
Pricing pages, error messages and model launches all talk about tokens. Here's what they are, and the four everyday AI quirks they explain.
by Lena Hartmann
- 3
Bro TalksBro Talks: Ottoline Varga Draws Every Line and Lets the Machine Make the Rain
The illustrator and composer draws every album page by hand and builds her ambient tracks with a little help from generative tools. A talk about taste, credit and why 'unfinished' is a feature.
by Theo Brooks
- 4
Culture & SocietyThe Rise of the “Made by a Human” Label
Small makers are adding a new promise next to “handmade.” Without a shared definition, it risks going the way of “natural.”
by Theo Brooks
Also worth your time
- Tomorrow: How to Check an AI Answer in Five Minutes. Dev runs a live session with real reader questions and one planted error. Can you catch it before he does?
- The forum's best thread right now: "What's one task you handed to AI and then took back?" Report-card comments, angry-customer replies and a supplier negotiation that was far too reasonable. Read the answers; they're better than most essays.
- From the Tool Shed: Citelight, the research assistant that fades out every sentence it can't source. The most honest research interface we've tested.
Prompt of the week
Find out which mistakes your assistant tends to make on your work, before they cost you.
Here are five tasks I regularly use AI for: {{tasks}}
For each one:
1. Describe the most likely kind of mistake an AI would make on it.
2. Rate how costly that mistake would be for me (low / medium / high).
3. Suggest one quick check I could do to catch it.
Then tell me which task most needs a human double-check.
One more thing
Ottoline Varga's line about disclosure, "It's not a confession. It's liner notes," has become an office motto. We've started adding liner notes to our own Field Notes pieces.
Every weekday
Get the next one in your inbox
The five AI stories that matter, explained in five minutes.