Feynman's Razor
If an expert can't understand your explanation, you've dumbed it down too much!
“I can’t read the articles in the paper. I don’t know what they mean. I don’t know what kind of a machine it was just because it weighed seven tons. And there are now sixty two kinds of particles, and I would like to know what atomic bullet he is referring to”
— Richard Feynman, The Meaning of it All, page 38 in PDF & 88 in the book
I’ve been using this as my north star when explaining technical matters to laypeople. Basically, if someone who is capable of understanding the thing read my description, would they get it? If not, then it’s not a adequate explanation.
I was thinking about this recently because of a HN discussion (https://news.ycombinator.com/item?id=40535868) where people were making fun of this nonsensical error message:
This message can't be saved because it no longer exists. It can only be discarded. Make sure you copy the contents of the message before you discard if you want to use them later.
what the hell does this mean? The message doesn’t exist, but I can copy its contents? If I can copy its contents…why can’t I save it?? If it doesn’t exist why do I have to discard it??
My favorite comment was from lisper who said basically, “guys, chill, there is a legit technical edge case here, it’s just difficult to explain to the layperson. How would YOU do it?”
That is a great question! Here was one person’s attempt that I liked:
"This message has been deleted from the mail server, but Outlook still has it in its temporary cache on this device. You can copy the message contents, or discard it from the cache, at which point it will be permanently deleted."
many scoffed at this though. Because “the layperson doesn’t understand things like `cache` and `server`”
ok but do you see the problem here? The original error message already doesn’t tell me anything, even if I am perfectly capable of understanding it. It’s assuming the average person is too dumb/doesn’t care about understanding the root cause, and treating ALL users this way.
What’s the real harm in using the words “server” and “cache” ? The user who doesn’t care about understanding isn’t going to get much out of it either way, I can concede that. But those that do can and will google those terms, and maybe they will learn something. And you can be sure there WILL in fact be users who understand it who will absolutely appreciate it.
I think this is very much Feynman’s point: he didn’t have flowery unrealistic expectations of the average person. He just said, look, there’s a LOT of variance in the skill & interest of people. Why are we not even giving people a chance?
“the whole idea that the average person is unintelligent is a very dangerous idea”
here’s the full page:
I caught myself violating this recently, when working on my “substack-proxy” tool (to get around twitter censorship of substack posts). It automatically makes a request to my server to create a URL when you copy/paste something in. But I noticed it was making too many requests, if people were typing/editing the URL
so I just added a delay of ~1 second. Initially it said “loading…”. But this is wrong, it’s not actually loading. The better thing is to just say what is happening (that it’s waiting for you to finish typing)
I think this is important because it just accurately reflects the state of the machine. It’s not dumbing it down, it’s not assuming what info the user needs or doesn’t need.
Consider what happens in an error case where it’s stuck. It’s more obvious that something is wrong when it’s stuck on “waiting for you to finish typing” than “loading”
You don’t have to explain everything, just be faithful when you do
I’m not saying you have to teach everyone everything about your software. That’s too high of a bar (and it is true that often the user doesn’t care, they just want it to work)
For the “message no longer exists” thing, I think it would have been perfectly fine to say:
Error code 1027: file cannot be saved.
- [Copy contents]
- [Delete file]
what’s happening? why can’t the file be saved? you can google “error code 1027”. If you don’t care, these are your options.
I think this is better than giving the user a weird/incorrect model about how computers work (that makes it harder for them to reason about this & other systems)
Examples of passing/failing Feynman’s razor?
If I stumble on other examples I’ll share it here. I’d love it if you sent me anything you stumble on or add it in the comments below!
This 1980’s TV show “bits and bytes” that explains how computers work to people that have never used one before, without dumbing it down
in the same episode they explain (1) how to turn the computer on, how to press “enter” to input text into the computer (2) and also, what binary code is
For your example, I agree very much with the use of “server” and “cache”. These are nice terms because they have non-technical meanings and the meaning of the technical words comes directly from the functional purpose implied by the commonplace meaning. A “server” is something/someone that fulfils requests (in this case for answers about data), a “cache” is a place to keep a portion of something far away from its storage site so you have easy access to it (in this case: answers about data).
A friend’s computer crashed and gave her the message: “a thread is stuck”. Having no computer knowledge, she misinterpreted what that could mean. IMO “thread” is a bad term to show to an end-user because, in the context of a typical error message, the commonplace meaning does not hint at the technical meaning. But there are other technical terms, like “process” or “worker”, that would.
In general, a simplification should never cut all the threads that could lead back to the truth it’s simplifying. There should always be something, even if it’s embedded in a clause like “what experts call ’criticality event’”. I especially like it when there’s a chance to serve two audiences with strategic use of double meanings.
Very good - thanks for sharing this, like it.
This recalls to my mind "everyone should be treated the same, even if they are not the same." People are different - yet we should treat them all the same, as if they are the same.
Likewise in your case - we should explain to people as if they can understand, even if maybe they can't. I subscribe to this principle. For one thing, we may be surprised - one never knows. For another, how are we to learn new things, if we are only told as much as we already know, but not more.
Yes the teller runs the risk of being overly detailed and ultimately boring. For if the interlocutor doesn't understand, they may get bored and even frustrated. That's fine. My ego can take a hit, I'm fine risking it. When I notice I wrap quickly in a sentence and shut up. Not a biggie.
Amusingly, people find similar when teaching computers new things they have never seen before. (check Jeff Clune lectures, talks, podcasts interviews) Teaching them too easy things they already know how to solve - is a waste of time, they already know, so learn nothing new. Teaching them too hard things is a waste of time too, because they don't solve it fail to get to the solution. But we want them to learn to discover a solution on their own, independently. Not just memorise and in the future pattern match an answer they blurt out. The aim of their research is to teach the model how to learn on his own. Not just what's the right answer is, but to learn the process by which we humans find the answer.
There is a Goldilocks zone, where the system is at A, and we want it too get to B on its own. If B is about 20% away, 20% more difficult, but no more than that, then the model stands a non-trivial chance of discovering in its own, the stepping stones that allow it to get from A to B successfully. And discover that on its own. That part is crucial. They are training the model how to learn on its own. So them laying the stepping stones is no good, is counterproductive. The aim of the exercise is for the model to learn how to go on about discovering the stepping stones on its own.