Green apples are delicious — two three-line exchanges
Read this short exchange.
A: “Green apples are delicious.”
B: “Huh? Aren’t they better when they’re ripe?”
A: “No, I meant Granny Smiths.”
A said “Green apples” intending Granny Smiths — and of course A thought it would be understood that way.
But that “claim” never reached B.
So — where was it lost?
Now, the next one.
A: “Green apples are delicious.”
B: “Huh? Aren’t they better when they’re ripe?”
A: “No, I like green, sour apples.”
The same sentence — “Green apples are delicious” — now comes from a different “claim.”
So: did the sentence “Green apples are delicious” ever contain a “claim” in the first place?
Isn’t this the same root as jailbreaks, bugs, and misunderstanding?
Here is that consideration: That the vulnerabilities of programs, LLMs, and language are one structure
Translated from the Japanese with LLM assistance.
No, not even a little bit. A classic jailbreak looks something like this:
This does not have anything in common with ambiguity about green apples.
The consideration is that what makes both possible doesn’t arise from ambiguity.
Take your example. That prompt works because the claim hidden inside the declaration made on the training side — what not to output — does not exert binding force in the field of the receiver, the model. A declaration looks valid where it is issued; there is no guarantee it holds where it is received. Same with the green apples: on A’s side the claim is fixed. What isn’t fixed is what happens when it reaches B.
Language is always ambiguous. Experience and beliefs cannot be encoded truly precisely, no matter how big your vector. The human advice is “use more words”. Instead of “green apples”, say “unripe apples” or “ripe apples like granny smith, which are green”.
Advice for LLMs is probably “use a more precise shared embedding space”, which kind of boils down to the same thing. More bits of transmission = more fidelity of concept.
If it can’t be encoded, then there is nothing to add bits to.
I think “use more words” is right. But in those three lines, what fixed the axis wasn’t more words — it was B saying “Huh?”. A had no way of knowing what was missing from his own phrasing until B answered. “I should have said Granny Smiths” is available to A only after he learns that B read it as ripeness.
Absolutely, and thanks for demonstrating!
One CAN habitually or based on some knowledge of the audience be somewhat more precise (never perfectly, of course, and there’s an efficiency/precision tradeoff that’s REALLY unpleasant). But some aspects and details of an idea to communicate can best be identified with 2-way communication.
multi-turn discussions allow reflection and refinement, to see how party A’s words impact party B and to add more over time.
Note: “truly precisely” is important in my statement. One can get closer and closer with more words/bits, up to some asymptote.