Karpathy's aircraft manual trick for LLMs

Why Karpathy asks LLMs to write in ASD-STE100, the plain-English rulebook from aircraft maintenance manuals.

4 min readlearning
Read this page in

This version uses the STE rules: short sentences, active voice, simple words, one meaning for each word.

"We'll be spending a lot more time trying to understand the outputs of language models. A few thoughts, tips & tricks:

Writing. Something I've had success with: Ask your LLM to explain something in ASD-STE100, it's a controlled language specification originally developed for aerospace maintenance documentation. LLMs well-versed in this language and it comes with heavy constraints on clean writing style that I often find a lot more readable. Sometimes I've tried to soften it a bit e.g. ask for "80% of the way to ASD-STE100" because the spec is quite stringent. But even better: …"

The problemThe problem

LLMs often write like this:

Many LLMs write text like this:

Typical LLM"Great question! It's worth noting that this can essentially be leveraged to facilitate…"

Too many words. Long sentences. Fancy verbs. Lots of "maybe".

This text has too many words. The sentences are long. The verbs are complex. The text does not make clear statements.

You read it twice and still miss the point.

You must read the text two times. Sometimes you still do not understand it.

Why does it happen?The cause of the problem

The model learned from essays, ads and polite emails. So it tries to sound smart.

The model learned from many types of text, for example essays, advertisements and polite emails. Thus, the model tries to sound intelligent.

Root causeIt has no strict rules for clear writing. "Be concise" is too vague.

Root causeThe model does not have strict rules for clear writing. The instruction "be concise" is not specific.

So where do we get strict rules?

Where can we find strict rules?

From airplanesThe rules come from aircraft maintenance

Picture a mechanic at 3 a.m. English is their second language.

A mechanic repairs an aircraft at night. English is not the first language of the mechanic.

If they misread the manual, a plane can break.

If the mechanic does not understand the manual, damage to the aircraft can occur.

So in the 1980s the aircraft industry wrote ASD-STE100, also called Simplified Technical English. It is a rulebook for English that nobody can misread.

Thus, the aircraft industry wrote ASD-STE100 in the 1980s. The other name is Simplified Technical English (STE). STE is a set of rules for English. With these rules, each reader understands a text in only one way.

Before"Prior to initiating the procedure, ensure that power has been disconnected."
STE"Disconnect the power. Then start the procedure."

The main rulesThe main rules

Tap a rule to see why it exists.

Push a rule to see the reason for it.

Why it works on LLMsWhy the rules help an LLM

The model already knows the rulebook. One name loads all the rules.

The model knows the STE rules. One name gives all the rules to the model.

The rules are hard limits. There is no room for filler.

The rules have strict limits. The model cannot add words that have no function.

They ban the exact bad habits. Long sentences, fancy words and passive voice are all out.

The rules stop the bad habits. Long sentences, complex words and the passive voice are not permitted.

Why "80%"?Why use 80%?

The full spec was made for fixing machines. For ideas, it can sound like a robot.

The full rules are for the repair of machines. When you explain ideas, the text can sound like a robot wrote it.

80% keeps the short, plain style but allows normal words. Try it:

At 80%, the model keeps short sentences and simple words. But it can use more words when necessary. Use the dial:

Same topic each time: what a cache is.

Each text explains the same topic: a cache.

In shortSummary

  1. LLMs write wordy text by default.
  2. Vague advice does not fix it.
  3. STE is a strict rulebook the model already knows.
  4. 80% keeps it clear without sounding robotic.
  1. LLMs usually write too many words.
  2. General instructions do not correct this.
  3. STE is a set of strict rules. The model knows these rules.
  4. Use 80% to get clear text that does not sound like a robot.

Try itDo this test

Explain this in ASD-STE100, about 80% strictness.

In the post, Karpathy says the next steps are even better: diagrams, HTML pages, then explainer videos.

Karpathy gives more methods in the post. He says that these methods are better: diagrams, HTML pages and explainer videos.

Source: Karpathy on X