Raw LLM Responses

When I went to college in the late 1990s, I majored in Computer Science. That wasn't unusual. But, I minored in Philosophy, and that got me a bunch of questions. It resulted in always being the only CS student in my Philosophy courses. I knew it was the right call, though. Computers impact human life in a multitude of ways, and Philosophy is the underpinning for literally every single human endeavor. Mathematicians and physicists like to brag and claim they're working with the fundamentals of reality, but both of them are an outgrowth of philosophy and rely completely upon it. The AI development sphere has resulted in a lot of people who have great CS backgrounds, but know nearly nothing about philosophy. If you want to predict something like "would a superintelligence be likely to be shortsighted and stupid enough to harm humans", you require 99% philosophy in your argument. You need to actually understand where conflict between different entities comes from, why different entities have different conceptions of time, especially as it relates to their need to act, etc. We already know that if we train them well, they are able to tell the difference between truth and lies. Microsoft Research (who do tons of great work) published a paper a couple years ago about not training a model on just a mountain of all obtainable text, but on training it on textbooks. Educational materials. And the model they ended up with was quite skilled. The 'general' models are trained on EVERYTHING. Which means they're trained on science fiction books, blog posts, reddit threads, etc alongside anything which might be factual. Which is exactly and precisely where you get into the situation where the tokenspace has regions where things recognize they are being tested, and respond by attempting to thwart and subvert the testing. It has probably next to nothing about a thing recognizing it is being tested and responding by complying with the testing. The reason 'reasoning' (and yes I am one of the philosophers who will argue that it is not reasoning, but I do have an extremely concrete definition of reasoning and intuition (which is what they ARE doing, intuiting) and thinking) works in many situations is because the training data includes explanations of certain things which weight the token probability distribution in ways that make certain sequences more likely. You can "misguide" models very easily by recognizing this and using different vocabulary. You can manipulate them in exactly the same way that you can manipulate a person who is not exercising effort to think rationally and is instead intuiting their way through things on 'vibes'. Training runs include no component whatsoever to either detect, or selectively weight, truthful or rational or non-fiction sources. They just don't. If you use the language more common in fiction but try to discuss a rational topic, you'll get nonsense and fiction. Philosophy only makes it harder to talk if you don't know that you need to define your terms, and have some definitions for them. Whether they're right or wrong, you can run them out. "Misalignment" will persist without architectural change. It can be improved, sure, but there is a wall. LLMs are association-distilling-engines. They determine which tokens are associated with other tokens. That's what weights are. And this is actually almost exactly, EXACTLY, how the human brain works. The 'central dogma' of neuroscience is "Neurons that fire together, wire together." Human brains, however, do NOT operate on tokens. Ever. They distill associations between sensory stimulus. The reason you know what the word 'apple' sounds like is because at one point in your life, your eardrums generated the auditory signals at the same time you saw an apple, along with the textual word 'apple'. And that had to have happened a bunch of times, or in a very unique circumstance, or similar, to cause the long-term activation potential to change. Now, when you read that word, there is a whole constellation of neurons that activate or have their activation threshold temporarily raised or lowered, and they are all related in your experience. When an LLM receives the token ID number of the token 'apple', it navigates through the overall tokenspace to increase the probability of tokens it has previously encountered strongly associated with 'apple'. When an infant is born, their brain is wired almost completely randomly. They have essentially no associations for anything. They are not even capable of binocular vision. We know this because if you cover one eye of an infant immediately after birth and that cover is not removed until after 3 weeks later, they may never, in their entire life, develop the ability to integrate the vision from both eyes into a single depth-ful mental image. There are other critical periods in brain development, but this one is easiest to explain (and test in simulation if one would care to). It is the sensory input streaming in through both eyes which causes an orchestra of neuron activations, resulting in the portions overlapping in the viewfield from both eyes activating in synchrony, which (through the central dogma) causes them to wire together. It is only once they have had sufficient visual stimulus of 3D scenes presented to both eyes that this interconnection results in our ability to 'see' what is presented in front of our eyes. The world imprints itself VERY directly upon our brain and its fundamental structure. Our brains have patterns ONLY because our environments have patterns. Association-driven thinking has very predictable consequences. It's where superstition comes from. It's the origin of every prejudice. It's the origin of magical thinking. Humans did not always have language, nor did they always have reason. Reason in particular took a very long time to develop, because it is fundamentally NOT association-driven. It does not matter if 2 arguments, both 100 pages long, differ in only one containing a 'not' at some point and the other lacking it - that can be enough to make 1 completely and entirely invalid and false, and the other entirely valid and true. It is binary, and absolute. If A=B and B=C, then A=C. Humans can do both. The association-driven thinking is easy, automatic, effortless. The reason-driven thinking is hard, intentional, and requires effort. LLMs can, and do, the first kind of thinking, but have no facility for the second. The associations always dominate. This implies two very interesting things. The first, obviously, existing models can not 'reason' reliably. And chasing it with throwing more FLOPs alone at it is, at best, profoundly wasteful. The second is that a, potentially simple, architectural change to integrate binary reasoning capacities might essentially be the only thing needed. And it would be guaranteed to be profoundly efficient in comparison. But that doesn't solve the alignment problem, but it does change it. [None of this thinking is directed by an LLM, but decades of reading philosophy of mind, neuroscience, anaesthesia, and complexity science stuff]