It's an important question and I'm not even entirely certain computer scientists agree on this.
<snip>
I think that many if not most engineers and computer scientists haven't even considered the question in the rush to monetization. They seem unaware of the fact that really smart people in other disciplines--neurobiology, linguistics, medicine, psychology, philosophy, and theology, to name a few--have struggled with this question for millenia without any sort of resolution. Nor do they seem to remember that an earlier generation of rather more sophisticated computer scientists failed to bring about a supposedly "imminent" AI back in the 1970s and '80s (see Douglas Hofstedter's Pulitzer winning--and amusing--
Gödel, Escher, Bach: An Eternal Golden Braid, 1979)
I generally would only call something "real" AI/ML if it's trained to recognize patterns (and ultimately make decisions) based on real-world data.
An entirely sensible respose. But, unfortunately, I can do this right now using a simple program that uses one
regular expression (regex) for the pattern-matchng and an IF/THEN/ELSE loop for the decision.
IF <some_text> MATCHES <pattern> THEN <do_this> ELSE <keep_looking>
The more patterns and actions that I supply, the more sophisticated and surprisingly "intelligent" the process can seem. But the "intelligence" is all mine: I supply the patterns to look for and specify what the program does when it finds one. I study representative input data sets and learn what sorts of things we need to:
- match
- absolutely, positively NOT match.
The second point is critical. You have to constrain the patterns, so that they do not match less or--worse--more than you intended. Get this wrong and results in the "real-world" can be disastrously misleading or even destructive, depending on the decisions specified on the
THEN side. For example, if your program is supposed to delete double-word typos (
the the) in a set of documents, but you do not constrain the pattern properly, your program might falsely tell the user that there are no double-word typos. Or it might
delete everything in the documents.
I don't know exactly how the much-touted Large Language Model (LLM) "AI" programs do their pattern matching. But they seem to use their enormous "training" data sets to determine the likelihood of encountering a particular character sequence. These training-derived probabilities are then assumed to be able to "predict" what will follow any arbitrary sequence of input data. The LLM ("AI") is essentially allowed to create its own patterns.
If I am more or less correct about how they are implemented, LLMs are certainly not "intelligent" in any meaningful way. They are
automatons:
machines designed to follow a predetermined sequence of operations or respond to encoded instructions (
https://www.merriam-webster.com/dictionary/automaton). LLMs run a set of statistical tools supplied by external programmers against a finite set of training data chosen by those programmers (from the works of unwitting and unrecompensed human authors). These tools produce an abstract statistical model of the concrete training data. The LLM then selects the
text from its training data that is most likely to "come next" with respect to any given input text, according to the
statistical model derived from that training data--a circular, self-contained process.
The above approach presents a couple of significant dangers. First, its statistics-based "understanding" is limited to the content of the training that its programmers have had the foresight to provide. Second, its statistically based equivalents to our matching patterns are
unconstrained (see 1 and 2 above). The LLM generates its own rules based on
the frequency with which patterns appear in its limited training data
, not on
what the patterns mean. In other words, the "AI" concerns itself simply with what matches, not with what we
must match or
must not match.
I suspect that other much-touted "AI" applications work in a fundamentally similar way. Consider self-driving cars. A robocab application that reliably stops for cyclists riding though intersections in daylight drives through and over a cyclist that crosses mid-block at night. An automobile autopilot application that reliably changes lanes on freeways and brakes for stopped vehicles drives through a jacknifed, empty flatbed trailer and decapitates its human passenger. Why? All I can think is that the models derived rules and probabilities from training data that did not include cyclists in the wrong place at night or trailers that do not have loads or boxes on top. The programmers failed to consider these cases or simply used real-world data sets that did not happen to include them. The "AI" followed its training-derived rules. Oops.
The behavior of the human drivers in the two well-known examples I have cited illustrates the real danger of our current AI craze: users and engineers alike tend to believe the CEO drivel and the marketing fairy tales. The robotaxi has a human check driver to prevent accidents. But the car is
intelligent and
safer than a real driver. Riding with it is so predictably safe thatshe gets bored. She reads a book.
BAM. The autopiloting passenger believes his car is intelligent because he paid a lot for it and identifies with the bad-boy plutocrat that owns the company. He loves the latter's over-the-top claims, ignores the fine print in the manual, and hops into the back seat for a nap.
BAM.
The above is bad enough when we are talking about cars. Add wings, weapons, nukes, national security responsibilies, and I think there will be problems.
defensescoop.com
www.twz.com
breakingdefense.com
insidedefense.com
www.defenseone.com
breakingdefense.com
www.defensenews.com
www.edrmagazine.eu
www.navalnews.com
www.airandspaceforces.com
www.defensenews.com
