Can you write short and still be truthful in general?
I frequently use LLM products and have noticed that when a model has to answer truthfully, it tends to produce significantly longer responses, even for concise queries. (Sometimes, adjusting the system prompts and settings can manage this.)
This has led me to conclude that a 'fact' in language usage is not always a clear-cut principle, but rather a complex, evolving consensus that involves multiple perspectives and can be somewhat obscure.
(
10 comments
| 4 points
| on HN
)
objective reality, "truth" is not something an LLM can have had any experience of, is it? Why expect it to be accurate here?
Nor can we. All our truths are subjective in the end. You believe you read something which you believe said xxxx. You believe you saw something etc.
Unless it's a response to a well posed query. The reason is that words have multiple definitions, thus one phrase might convey multiple conflicting, but plausible meanings to it's intended audience.
To get precision (aka Truth) you have to work very hard to resolve all the ambiguities that result from any combination of words. The more sure you want to be that the reader will interpret your meaning correctly, the more words you're going to require.
Personally, I like to utilize an iterative approach (aka conversation), in which error correction is generally accomplished over time. This is also why I like long form podcasts, the host is allowed to explore the meaning of what the guests say, and dig deeper, and correct any misperceptions (if they are a good host).
If I'm composing a static text, again I favor an iterative approach, posting, then editing, editing, editing until 2 hours is reached, and edits are no longer possible.
You may be interested in Sam Harris’s book Lying (I’m not a huge Harris fan, but it’s a solid book).
It takes more effort to write less though, and often many iterations. You can get LLMs to shorten them with iterations too.
A fact is not a language thing, but a reality thing. It is the language itself that creates the complications that confuse us.
Answering your question: yes. The easiest example, a yes and no question.
Question: Is this comment on hn? Answer: Yes
Short and truthful. QED