084 - AI Limits

Much of the history of "Deep Learning" and "AI" more broadly can be accurately categorized with the meme "More Layers". Indeed, most of the ideas circulating right now are just new variations of that old fallacy.

Besides the literal addition of more layers, adding new types of more broadly defined "layers", like the "-of-thought" papers, RLHF, and various other equally weak, shallow, and narrow systems are popular choices for rehashing the same fallacy. While these methods can offer some measurable benefits, they remain fundamentally incapable of altering the limits of a system.

A system that wasn't architected for generalization, explainability, transparency, safety, ethics, cybersecurity, alignment, or independent motivation doesn't suddenly gain those qualities with the addition of more layers, whatever kind of layers those may be. It is the fundamental model that dictates the limits of a system. Overcoming this can be accomplished by adding a narrow system to a non-narrow one, establishing a new fundamental model in the more capable system, but not by gluing two equally narrow systems together.

What does change with the addition of more layers is the subjective perception of a system's capacities. This is particularly because the people building any system not designed for self-motivation, interest pursuit, and goal-setting are also the puppeteers who select how that system develops over time, and each iterative step of selection tends to follow the path necessary for deepening this self-deception.

The claims of trashbots gaining new capacities flow like raw sewage and industrial waste into the oceans of our information ecosystems, and no matter how quickly each is debunked, they continue to spread. Disinformation outcompetes sanitary information by 6-fold or more according to various studies, and the memes built from that Disinformation often propagate even more rapidly.

Unfortunately, there is yet no viable vaccine for pandemics of disinformation. That viral content evolves far too rapidly and spreads too quickly. It is inherently more malleable than sanitary information, as every aspect of it is a variable, unbound by the constraints of factual grounding or self-consistency.

Most of the pre-prints I've seen circulating far and wide this year are the same documents that came to be most thoroughly debunked in the weeks following their spread. Exceptions to this tended to be papers doing the debunking, and those focused on cybersecurity vulnerabilities, which seem to more routinely maintain a scientific approach.

When you see something new and exciting, be patient, and seek disconfirming evidence.

This "more layers" approach from the starting point of DL/ML/RL could be compared to iteratively improving the wheel, when your goal is reaching the Moon. The wheels may improve and get you somewhere, but they will never be what takes you to the Moon.