Why Neural Networks is a Bad Technology
Stop pretending that it deals with truth values, and it could be a decent tech.
There is a big difference between pattern matching and determining truth values.
I’ll show some examples how.
Let's say we attach the mathematical values of a certain signal in a network to a text label.
It doesn't mean that the signal "is" or "mean" that text label. It only indicates the arbitrary relationship between the signal matching and the label that it just happened to be attached to:
In the above example (no, it’s not overfitting), we can manipulate the signal by inserting pixels invisible to the human eye so it could match any label. The human eyes can't even tell the difference between the pictures. If we’re identifying what the picture is actually representing, then there SHOULDN’T BE A DIFFERENCE between what the picture “is” before, or after the change… But the machine identifies a huge difference: The difference between a panda and a gibbon.
A more practical example I've found myself, is to ask any image generator to draw an ouroboros. Tell a generator to give an image of one. Then look up the dictionary on what an "ouroboros" actually is. Question: does it actually ever give you one, or does it give you anything BUT one?
The prediction of what the NEXT PIXEL in a picture should be, irrespective of what a picture "is representing," becomes a problem. Same goes with predicting the NEXT CHARACTER that goes into a string in an LLM such as any GPT.
Neural networks never deal with knowledge in any kind of legitimate fashion. It will always have epistemic issues because that's their underlying nature.
Additionally, it's a house of cards that's brittle and collapses if anything is out-of-range of the model’s data. The actual "range" being the entire real world and the entirety of human knowledge.
Remember what happened in the fatal accident which decapitated a Tesla driver. That was an "out of bounds" condition, or perhaps to be euphemistically called an “edge-case” where anything that’s unexpected or unaccounted for becomes an “edge-case.”
Same thing happened when a Cruise robotaxi dragged someone across a pavement 20 feet.
If people are depending on these NNs for knowledge, facts, life-or-death mission-critical things... Would you bet on these things being right-by-chance without actually dealing with truth values?
Equating these 2 things commits a big category mistake:
Chances of being correct/incorrect due to knowledge being correct/incorrect via limited human experience
Chances of being correct/incorrect due to stochastic mechanisms in a machine
C'mon. Naive school kids mix that up. “Oh, we’ll just get more data into the set” WON’T FIX ANYTHING.
Of course people aren't going to put glue on their pizza (Google shut that down, I hope? Note how Google called it an “edge-case”) but what about when providers give people access to "telehealth" based on LLMs or even go as far as directly relying on them at their facilities? Using LLMs for medical diagnosis can lead to deadly consequences.
There are fundamental epistemic flaws with neural networks that could not be remedied with the way they are, and companies are pushing hard to just put them absolutely everywhere. Do they care about people...? Of course not. ☠️
Saw your post via linkedIn; I am very much with you on this as far as current practices, overblown expectations GenAI etc. are concerned. Just thinking:
When it comes to your distinction between errors due to limited human experience vs. stochastic mechanisms in machines: We are not Vulcans; not logical machines. Looking at Asch's conformity experiments, Kahnemann's and Twersky's work etc., I would argue that human decision making quality is influenced quite strongly by all kinds of quasi-stochastic influences, from social conformity, via all our other biases (recency bias etc.) to fatigue and stress at home.
Are you not making an assumption like the classical economists did with their "homo oeconomicus"? I can follow your conclusions if I make that assumption. If I look at real-life situations, I would argue that the effect matters more: Looking at the sum of all influences, does the decision making of "AI" in whatever evolved form produce better results than human decision making.
Looking at the recent publication about the neuronal pattern firing in an LLM when it sees the a misspelt word "rihgt" being very different whether that is a coding context or an essay: Is that not the very rudimentary start of abstraction/contextual understanding?
I am fully with you that current systems are not there yet. Just not sure about your "never" claim. I see artificial systems evolving to other ways of learning (e.g. robotic bodies giving experience with space and time), and with refined architectures. Yet, they will still be neural nets, and I see no *principal* difference.
Can "knowledge", can epistemiologigal categories not emerge from pattern matching? Are you sure this is not what happened in humans?