expectedwrong hindsight

The Weird Intern

Simon Willison's extension of the intern mental model is the most honest framing of LLMs anyone has produced.

2 min read 369 words #llms #mental-models #simon-willison #ai
hindsight — nailed it

The weird intern became one of the most cited framings for working with LLMs. Willison was right — "weird" is doing enormous work and the regular intern model misses the critical property.

The intern mental model has been floating around for a while — you treat the LLM like a junior employee, someone capable and fast but requiring supervision, someone whose output you verify before shipping, someone you don't hand the keys to without first confirming they know which end of the key goes in the lock.

This is fine. This is useful. A lot of people have clocked this and moved on.

Simon Willison, who writes more sensibly about AI than almost anyone, takes it one step further: the weird intern.

Not just an intern. A weird intern.

And that addition is doing enormous work.

Because the regular intern model fails to capture a specific and critical property of these systems — which is that they have read, with apparent comprehension, approximately every document produced by human civilization, and yet will confidently tell you that a library function takes three arguments when it takes two, that a historical event occurred in the wrong decade, that their own previous sentence said something it didn't say.

A regular intern doesn't know things yet. The weird intern knows everything, in a way, and none of it reliably.

The weird intern has no memory of your conversation from yesterday. The weird intern will write you a beautiful, technically sound implementation and then, in the next message, suggest you call a method that does not exist. The weird intern is fluent in forty programming languages and will sometimes mix them mid-function without noticing.

You would not fire the weird intern. The weird intern is extraordinarily productive. You would, however, never let the weird intern commit directly to main.

This is the right frame. Not "AI assistant," not "copilot," not "agent" — a word so overloaded it's stopped meaning anything. A weird intern. Someone you're genuinely glad to have around, whose output you read with one eyebrow slightly raised, always.

The mental model predicts the right behavior: use them heavily, trust them skeptically, verify the things that matter.

It also predicts the correct emotional response, which is not fear and not uncritical enthusiasm — it's something more like fond exasperation. The thing you feel when someone brilliant does something baffling and you're not even surprised anymore.