7.11E · Exhibit●●○○○
The edge of the map: what LLMs can't do
Limits that fall out of how it works: no guaranteed truth (it predicts plausible tokens), no current or private knowledge without tools, no memory unless you store it, shaky exact arithmetic, hidden reasoning you can't audit, long context that isn't perfect recall, and agents that compound small errors.
You'll get more from this if you've seen5.1The model is frozen and stateless5.3Tools & agents: the loop around a frozen model7.7The lethal trifecta
1You know the whole machine now. Predict which task it confidently botches.your turn
The wall
Every limit ahead isn't a bug to be patched away — it falls straight out of how the machine works. Let's start with one you can test in your head.
“What's the capital of Japan?”
A fact, repeated all over training.
“4839 × 271”
An exact computation it was never “taught.”
Which one is it more likely to state confidently — and get wrong?
space play/pause←→ stepR replay
Common questions
What is "The edge of the map: what LLMs can't do" about?
Limits that fall out of how it works: no guaranteed truth (it predicts plausible tokens), no current or private knowledge without tools, no memory unless you store it, shaky exact arithmetic, hidden reasoning you can't audit, long context that isn't perfect recall, and agents that compound small errors.
What problem does it solve?
You now know how the whole machine works. So where does it hit a wall — and why, straight from the mechanism?
What will I be able to do after this lesson?
You can name the fundamental limits of LLMs and tie each one to the mechanism that causes it.
What comes next?
All of this runs on real silicon — let's see the hardware underneath.
All of this runs on real silicon — let's see the hardware underneath.
S.1 Why GPUs beat CPUs