Occasionally a friend will say something really stupid to me. (Obviously if you’re reading this and you are my friend, I’m talking about someone else.)
If it’s something that makes me raise an eyebrow, I’ll usually ask Claude about it, especially if it’s trivially checkable.
USER: A friend of mine says that “UK economic indicators are looking broadly healthy.” Is that true?
CLAUDE: I’d push back on that. The picture is more accurately described as weak with mounting risks.
Your friend is either working from outdated priors or defining “broadly healthy” so loosely it’s unfalsifiable. The honest read is: an economy muddling through with real vulnerabilities, not a healthy one.
My friend and I had a brief back-and-forth, with his smoothbrained hot takes crumbling in the face of my machine-enhanced gigafacts, until I got bored and dropped it.
There used to be a website called Let Me Google That For You, which let you send a passive-aggressive link to someone that showed the Google homepage and a cursor wandering to the search box and typing in their query – which they could have searched for themselves.
It would be pretty fun if there was a version of that for Claude.
Well now there is! TryLet Me Claude That For You, the fastest way to annoy and alienate your closest friends.
This does present an uncomfortable question: if Claude can distil any legible question into a competent 250-word response in ten seconds, then what is non-fiction blogging actually for?
Specifically: what novel insight are we producing? Why should anyone write a lit-review-as-effortpost, or “here’s what the data says about X”, or a 3000 word synthesis of a subject?
The most obvious case where humans can still add value:
LLMs are not good at doing these. They probably can’t discover a pattern that people have only half-noticed. While they aren’t trained to be opinionless, they’ll often come up with pretty milquetoast takes, and won’t want to push an argument to its logical conclusion (especially if that conclusion is controversial). They can’t make a boring-sounding question seem urgent, or make salient a pattern you’re not noticing.
That might change as models improve, but this isn’t a capability limit – these traits are baked in during post-training. Models are trained not to take a hard stance on things which are out of distribution, or on which there isn’t broad agreement. You might be able to push LLMs to produce novel framings of existing topics, but it’s still difficult to discover genuinely new questions that people aren’t asking.
Models trained to be agreeable will not e.g. come up with an ontology of corporate drones as sociopaths/clueless/losers. If anything, their post-training cuts the other way. Model makers want their models to be helpful assistants and sand off some of the rougher, weirder edges of their responses.
The setup: he starts writing about building a set of stairs with his dad. He discovered that it’s unexpectedly complicated in a way which you only see if you actually try to do this – the floor isn’t level, the stairs need a particular rise so you don’t fall down them, the screws need to be a particular length so they don’t stab you in the foot.
The pivot: reality is full of complexity which only becomes apparent when you slam your plan into the world. It’s hard to model ahead of time. This is the reason why projects often overrun.
It sticks in your head because the title is the whole thing. Once you’ve read it, you can simply use that sequence of words to retrieve the idea, and to cause the person to whom you’re speaking to retrieve it. You could ask Claude about the topic if you thought of it, but you probably didn’t, and its reply wouldn’t be nearly so salient. It helps that Salvatier actually had to build the stairs.
We could distil the above categories into something like:
noticing a question
naming a pattern
refusing to flinch from a premise
There are perfectly good reasons to want to write posts that aren’t just “they will provide insights to people for many years to come”. But if you do want to write posts which are lindy – which will outlast the latest fads or blogging cycles and produce novel and interesting insights which can’t just be gotten by asking Claude – you should aim at one of the above.
Even better – and this is where the idea of craft comes in, another thing that LLMs struggle with – they should be written to really stick in someone’s craw. In particular, the most important thing to aim for is compressibility. It should be easy for your reader to compress the central premise into a load-bearing phrase. Moloch. Civilisational inadequacy. Sociopaths, clueless, losers. Reality has a surprising amount of detail.
These become handles by which your readers can refer to your idea – both a conceptual thing they can use to retrieve it in their minds, and for them to induce others who have read your work to retrieve it too. And for people who haven’t read your work, it serves as a way for them to make their way back to it.
So don’t write Claude-shaped posts. Give your readers a load-bearing phrase, ideally in the title. And do the work Claude can’t. Claude aims for the middle of the distribution, for centre mass – your job is to go for headshots[1].
Occasionally a friend will say something really stupid to me. (Obviously if you’re reading this and you are my friend, I’m talking about someone else.)
If it’s something that makes me raise an eyebrow, I’ll usually ask Claude about it, especially if it’s trivially checkable.
My friend and I had a brief back-and-forth, with his smoothbrained hot takes crumbling in the face of my machine-enhanced gigafacts, until I got bored and dropped it.
There used to be a website called Let Me Google That For You, which let you send a passive-aggressive link to someone that showed the Google homepage and a cursor wandering to the search box and typing in their query – which they could have searched for themselves.
It would be pretty fun if there was a version of that for Claude.
Well now there is! Try Let Me Claude That For You, the fastest way to annoy and alienate your closest friends.
This does present an uncomfortable question: if Claude can distil any legible question into a competent 250-word response in ten seconds, then what is non-fiction blogging actually for?
Specifically: what novel insight are we producing? Why should anyone write a lit-review-as-effortpost, or “here’s what the data says about X”, or a 3000 word synthesis of a subject?
The most obvious case where humans can still add value:
LLMs are not good at doing these. They probably can’t discover a pattern that people have only half-noticed. While they aren’t trained to be opinionless, they’ll often come up with pretty milquetoast takes, and won’t want to push an argument to its logical conclusion (especially if that conclusion is controversial). They can’t make a boring-sounding question seem urgent, or make salient a pattern you’re not noticing.
That might change as models improve, but this isn’t a capability limit – these traits are baked in during post-training. Models are trained not to take a hard stance on things which are out of distribution, or on which there isn’t broad agreement. You might be able to push LLMs to produce novel framings of existing topics, but it’s still difficult to discover genuinely new questions that people aren’t asking.
Models trained to be agreeable will not e.g. come up with an ontology of corporate drones as sociopaths/clueless/losers. If anything, their post-training cuts the other way. Model makers want their models to be helpful assistants and sand off some of the rougher, weirder edges of their responses.
Let’s have a look at Reality has a surprising amount of detail by John Salvatier.
The setup: he starts writing about building a set of stairs with his dad. He discovered that it’s unexpectedly complicated in a way which you only see if you actually try to do this – the floor isn’t level, the stairs need a particular rise so you don’t fall down them, the screws need to be a particular length so they don’t stab you in the foot.
The pivot: reality is full of complexity which only becomes apparent when you slam your plan into the world. It’s hard to model ahead of time. This is the reason why projects often overrun.
It sticks in your head because the title is the whole thing. Once you’ve read it, you can simply use that sequence of words to retrieve the idea, and to cause the person to whom you’re speaking to retrieve it. You could ask Claude about the topic if you thought of it, but you probably didn’t, and its reply wouldn’t be nearly so salient. It helps that Salvatier actually had to build the stairs.
We could distil the above categories into something like:
There are perfectly good reasons to want to write posts that aren’t just “they will provide insights to people for many years to come”. But if you do want to write posts which are lindy – which will outlast the latest fads or blogging cycles and produce novel and interesting insights which can’t just be gotten by asking Claude – you should aim at one of the above.
Even better – and this is where the idea of craft comes in, another thing that LLMs struggle with – they should be written to really stick in someone’s craw. In particular, the most important thing to aim for is compressibility. It should be easy for your reader to compress the central premise into a load-bearing phrase. Moloch. Civilisational inadequacy. Sociopaths, clueless, losers. Reality has a surprising amount of detail.
These become handles by which your readers can refer to your idea – both a conceptual thing they can use to retrieve it in their minds, and for them to induce others who have read your work to retrieve it too. And for people who haven’t read your work, it serves as a way for them to make their way back to it.
So don’t write Claude-shaped posts. Give your readers a load-bearing phrase, ideally in the title. And do the work Claude can’t. Claude aims for the middle of the distribution, for centre mass – your job is to go for headshots[1].
Thanks to Alexander Wales for this excellent phrase.