Why Not Do It Properly?
Many people spend their time refining DL and LLM to handle
text. It must be obvious that developing textual tools that have no idea what
the words mean is not going to lead anywhere, but still they do it. If the LLM has to cobble together several pieces of text, it can end up like a crazy quilt.
An alternative is to think about how words work.
A few examples:
A definition for “informal”:
Definition: (of dress) casual.
A few problems. “casual” can be a noun or an adjective – as
an adjective it has several senses and sub senses.
“dress” can be a singular noun, as in a woman’s dress, or it
can be an aggregate (uncountable) noun, applying to both men and women’s
clothing, as in “they wore evening dress to the ball”, or "the general was in his dress uniform".
The definition (simplified) ends up looking like
What about direction in text?
Some examples:
He was proud of their success.
Or
“Their success made him proud”
The influence is from right to left.
He was confident of success.
He is providing an estimate of the logical probability of
success for an event which has not occurred – the influence is from left to
right.
These effects are not going to be extracted by a machine
from a big gob of text, no matter how many billions of nodes may be present. It
would also be nice to allow what someone else is telling you to modify the
response (episodic memory – remembering what they said last week), rather than
being an updated Eliza.
Humanity has many pressing problems, and we are still
wasting our time on AI toys, when we had enough memory to handle language
thirty years ago, and wasted those decades on what were, with a few moments forethought,
obvious dead ends. Our unconscious minds do all the work in language analysis,
and because we don’t appear to think about it, we expect it to be easy to
implement. It isn’t, it is hard work, but it offers the reward of breaking the
Four Pieces Limit, and lifting the cognitive output of millions of people.
Comments
Post a Comment