Anthropic engineer Thariq wrote an article about his experience working with the new model Fable. The core is an old saying: the map is not equal to the actual terrain. The map is the instructions, context, and rules you give to the AI; the actual terrain is the scene where the mission actually takes place, the code, the real world, and the limitations you can't think of beforehand. The gap between the two is where AI will step into pitfalls, go around in circles, and make random guesses. I connected this framework to the tacit knowledge refining I have been doing: if you draw the map clearly first, it is equivalent to turning the unspoken judgments in your mind into explicit rules. If the AI follows them, the success rate of real tasks will be much higher.
- People who already use AI every day, but often feel like “Why does it guess wrong again and ask me to tell you again?”
- Want to hand over their professional judgment to AI, but find that the most difficult thing is for consultants, coaches, and knowledge workers to "speak clearly"
- Those who want to understand "why some people use the same AI very well while others use it poorly"
- Understand the metaphor of "map vs actual terrain" and know the source of AI errors
- A map of "Four Unknowns", identifying which unknown is your tacit knowledge
- A set of questions that you can do before you start. Draw the map clearly and attach questions that can be directly pasted.
One sentence: The map is not equal to the actual terrain
This sentence comes from the old proposition of linguist Alfred Korzybski. Thariq used it in working with AI. I think it is very good.
This is how he dismantles it. A map is an expression of what you want to do, that is, the tips, rules, and context you give the AI. The actual terrain is where the work actually happens, the site with all the real constraints. What you hand over is always a map, but what the AI wants to walk is the terrain.
Why AI makes mistakes and goes in circles
If the map is too simple, the AI will follow it and discover a bunch of conditions that you didn't write when it reaches the real terrain. At this time, it has only two options: it gets stuck and asks you, or it uses "common industry practices" to guess. The common practice in the industry may not be the one you want.
Thariq was very precise, commanding AI is a balance. If it's too specific, it will stick to your instructions, even if it's in a different direction; if it's too vague, it will fill in whatever it thinks is the best way to do it. If you don't think clearly about the unknown, you will fail at both ends. You don't know which section of the road will have potholes, and you don't know which section of the road is actually smooth, but you just hope it will turn there.
You just said "make me a dashboard". The AI can only make up a bunch of assumptions on its own: which fields are needed, who should show it, and what style. If you correct the mistake, you can correct it ten times and it’s still not close to the mark.
You need to make it clear first: who you want to show it to, what questions you want to answer, which numbers are the most important, and what I haven’t figured out yet. The AI runs according to this map and gets it right the first time, and the rest is just fine-tuning.
Four unknowns: Which one is your tacit knowledge?
Thariq breaks down “what you want AI to do” into four types. The four quadrants themselves are derived from Donald Rumsfeld’s teachings, but when applied to AI work, it just illuminates all the things that “you thought were explained clearly, but actually they weren’t.”
What you put in the prompt. You know very well and can tell that this is the map itself that you give to the AI.
You haven’t thought it through yet, but you know you haven’t thought it through yet. At least you know to ask and check.
Something so obvious that you wouldn’t even write it down, but you’ll recognize it when you see it. You know, you just didn’t realize it. This grid is tacit knowledge.
You haven’t thought about it at all. You don’t know what piece of knowledge you are missing, and you don’t know how best to do something.
Drawing the map clearly means refining tacit knowledge
This is what I want to follow. Thariq was talking about how to align the map with AI, and the action of "drawing the map clearly" was replaced by what I do every day, which is refining tacit knowledge.
This is often what I do for consultants, coaches, and lecturers. They have a whole set of judgments in their minds, including how to start a case, how to find out the key points in a document, and where a student's question is stuck. When asked "How did you decide?" they can't answer it. In fact, that set of judgments has always been there, but it has become too familiar, so familiar that it has become an intuition and cannot be expressed.
This kind of unspoken judgment is tacit knowledge. And if you lay it out as clear rules one by one, AI can do things your way. So "drawing a map" and "refining tacit knowledge" are actually two names for the same action. By doing this, you are actually turning the most valuable and unclear piece of your mind into an asset that can be reused, handed over to others, and accumulated.
How to do it: Draw the map clearly before doing it
Thariq listed a complete set of methods for mining unknowns with AI in the original article. I selected a few that are best moved to non-engineering scenarios and translated them into versions for knowledge workers. You don’t need to use them all every time, just use it as a toolbox.
Before taking on a new field or new customer, directly ask AI to help you list "I may not know what I don't know about this matter", and give it your starting point: what you know and what you don't know.
Ask the AI to ask you one question at a time, focusing on areas that would change your approach. As you answer, you discover that many of the judgments you thought you had discussed were never written down. This method is the most direct for refining tacit knowledge.
Some things you just can’t describe. At this time, give an example you like and let AI read "how it was done", which is much more effective than guessing with a bunch of adjectives.
When you think it’s almost done, ask AI to draw up a plan or first draft, putting the decisions that are most likely to be changed at the top. As soon as you look at it, you will often find that there is something you haven’t explained clearly.
Questions that can be posted directly
The model will become stronger and stronger, the map is your responsibility
Thariq finally said something that I agree with: the better the model, the more you can achieve with the right method. When a long task ends up being wrong, it usually means you need to spend more time defining your unknowns first.
Every blind spot scan, every interview, and every reference is a very cheap way for you to figure out what you don’t know before it becomes too expensive to fix. Whether AI is strong or not is one thing. Whether you can draw the map well is where the gap really widens.
So don’t rush to ask AI to do the next task. Spend ten minutes first and ask it to help you find out what you don’t know and draw a clear map. In these ten minutes, you are refining your most valuable piece of knowledge.