20 Jan
2025
20 Jan
'25
1:41 p.m.
On Sat, Jan 18, 2025 at 02:17:22PM -0500, Paul Koning via cctalk wrote:
On Jan 18, 2025, at 2:13 PM, Chuck Guzis via
cctalk <cctalk(a)classiccmp.org> wrote:
...
He said, that the computerized version was more adroit at spotting decay
and could perform periodontal charting from radiographs without the
necessity of a hygienist poking at your gums with a probe.
I'd call that more in the realm of image processing and expert systems
rather than real AI.
Expert systems have always been considered a kind of
AI system. What makes
them somewhat different is that they actually have a decades-long track
record of working. I remember learning these back in 1976, from visiting
prof. Donald Mickie at the University of Illinois. We used it to build a
chess endgame machine.
Plus, IIRC, the expert systems usually could dump their decision path
on request so one could debug them. With LLMs, if the bullshit gets too
bad, you toss out the model, twiddle the training parameters and train
again, hoping for the best.
Consequently, expert systems - when done well - are actually useful,
whereas the current crop of "AI" (LLMs essentially) is great for collecting
funding and dazzling the clueless, but not all that useful otherwise.
And those setups where neuronal networks are actually useful employ them
as powerful and specialized pattern matchers - which works well if done
carefully.
The next AI winter will surely arrive eventually ;-)
Kind regards,
Alex.
--
"Opportunity is missed by most people because it is dressed in overalls and
looks like work." -- Thomas A. Edison