Search for the known unknowns, LLMs for the unknown unknowns

An oversimplistic but maybe still helpful model of what things are good for

From a post on Bluesky today:

From my experimentation so far I’d say this is correct. To be more specific, search really excels at dealing with known unknowns, and LLMs are quite good for surfacing the unknown unknowns.

If you know exactly what you need from search, search will try to find it for you from the world of available options. If it’s not there it’s not there. If your medical research comes back and it’s all Pinterest links, you probably have asked a Pinterest question, not a JAMA question. Etc.

You can try to reformulate, and search does try to intuit some of what you're doing, but at heart it’s searching against available results, or the knowledge graph, or something else, and trying to find the intersection between inferred intent and a database of things that can be offered.

The whole experience is predicated, partially, on the idea that when you see what you need you’ll recognize it. You’ve got to be careful to make sure that thing “is what you think it is” – I wrote a whole ass book on this issue with Sam Wineburg – but that’s why I call it a technology of the known unknowns.

One of the great things about it is that it creates what I call “productive dead ends” – not finding something (or not finding something with the source quality you want) is meaningful. Maybe you want the wrong thing, have the wrong word, or are looking for something not in fact extant.

It’s this moderate rigidity that is core to the process. On the other hand, what I’ve found LLMs to excel at is the opposite. They work well with the *unknown* unknowns.

There’s lots of times that you don’t know quite what you want. You want to know what some objections to an idea might be. Or someone said something odd on the internet, and for the life of you you can’t infer the meaning of it. Your vague question is “WTF is this person talking about?”

When you give these systems *space* to tell you (roughly) what’s in the linguistic vicinity of the linguistic representation of your problem, some pretty amazing things happen.

My recent work has been using Claude for mapping out arguments and rebuttals to evaluate evidence or a claim and as long as you don’t get too specific in your request you can do really well with this.

What is striking is when you get over the idea of “give me this particular thing” and lean into “map out this issue using this lexically-inflected template for reasoning” it thinks up all the sorts of things you **wouldn’t think to have asked**.

Similarly it does great at inference. And I say this as a person who for a while had a job supervising 15 very smart undergraduates trying to make inferences about web content. Claude leaves them in the dust.

Again, the unknown unknowns. The domain knowledge you don’t know you need. And the trick to the unknown unknowns is to give the system a bit of space, and some pointers to any linguistic models that might guide which found paths it should follow.

But if you do the opposite – and ask for the *known* unknowns – well, it will just make stuff up. Not Claude as much – but generally, yeah. LLMs are not bounded by any scarcity imposed by a library of content, or a knowledge graph of facts.

So in summary – yes – LLMs don’t simply fail at search. They are in many ways the opposite of a search engine. This simplifies the situation a bit – search has always been a mess of user contradictions. But in some broad ways it gets it right. And importantly might stop people from making dumbass mistakes like inventing fake pardons.