On Nov 13, 2022, at 9:09 AM, Paul Koning via cctalk
On Nov 12, 2022, at 1:08 PM, Anders Nelson via
cctalk <cctalk(a)classiccmp.org> wrote:
I bet NN/AI would be helpful with data recovery - if we can model certain
common failure modes with those old drive heads we could infer what the
data should have been...
NN maybe, I need to understand those better. I see they are now a building block for
AI, not so clear. In my view, AI is a catch-all term for "software whose properties
are unknown and probably unknowable". A computer, including one that executes AI
softwware, is a math processing engine, so in principle its behavior is fully defined by
its design and by the software in it. But when you do AI in which "learning" is
part of the scheme, the resulting behavior is in fact unknown and undefined.
For some applications that may be ok. OCR doesn't suffer materially from occasional
random errors, since it has errors anyway from the nature of its input. But, for example,
I shudder at the notion of AI in safety-critical applications (like autopilots for
aircraft, or worse yet for cars). A safety critical application implemented in a manner
that precludes the existence of a specification is a fundanmentally insane notion.
not a fan of AI myself. But, I feel constrained to point out that the alterative to
"AI in safety-critical applications” often is “a minimum-wage employee in a
safety-critical application” which may or may not be an improvement. Agreed that AI is
fundamentally not absolutely predictable - but neither are people. For problems complex
enough to require either in a safety-critical decision-making loop, it may resolve down to
a question of either 1) trusting the statistics (AI driving is maybe already
*statistically* safer than human driving), 2) desiging the whole system in such a manner
as to be tolerant of decision-making faults, or 3) Not doing the dangerous activity
because it’s not monitorable.
I would say our current road and automobile system doesn’t satisfy any of those criteria,
For problems simple enough to write closed-form, formally-verifiable software to handle,
I *definitely* agree that is the way to go.