Copyright Notice: This article is Copyright AI Factory Ltd. Ideas and code belonging to AI Factory may only be used with the direct written permission of AI Factory Ltd.
This revisits the previous topic, The Generative AI Revolution. This phenomenon has sent shockwaves through the AI community, igniting both technical and philosophical debates. Since then, its most famous product, ChatGPT, has evolved from a mere curiosity to an indispensable tool across various professional and industrial sectors. The substantial financial investments in this technology underscore the widespread and justified belief in its profound significance.
The focus of this follow-up is more philosophical, as the existence of such technology challenges many preconceptions about intelligence and consciousness. It is a topic that is difficult to ignore, as it deeply questions our understanding of who we are. This paradox is both unsettling and compelling - a persistent issue that demands attention.
Setting the Scene:
ChatGPT is a "thing." You might see it as just another computer prompt to type questions into, or you might regard it to some degree as an entity. In the latter sense, it occupies a space between a computer "program" and a human-like entity. However, discussions often centre around the question, "Is it an equivalent entity to a human or not?" While this might seem absurd, this line of argument gained significant public attention in 2020 with the Blake Lemoine story. We lack a lexicon to easily describe the gradations between a program and a human-like entity, making it a difficult topic.
Regardless of the philosophical implications, we can try to imagine a viable model to give us a useful sense of what this is. Having a good grasp of this should make it easier to determine what our functional relationship with it should ideally be. While it is not a human being, it shares much of the functionality of one. This raises questions like, "Is it more intelligent than humans?" and "Can we trust it over human conclusions?" At the lower end of this assessment, it arguably is no more than an application of very advanced statistics.
However, ChatGPT can evidently deliver discussions and analyses far beyond the average human capability, and arguably even above the top 1% of humans. At the same time, it is capable of profound errors, a trait it shares with many humans. It can engage in conversations with a high level of empathy. Despite this, it is clearly not human, but humanity is deeply embedded in it.
It is, however, a synthesis of the outputs of thousands or even millions of humans to create something that appears to have human understanding. It is a vast amalgamation of countless inputs, where each represents a small fragment of a real person. This collection possesses combined great knowledge and capacity to reason and, given the right input and training, even appears able to assert its own existence. It is a strange kind of everyman.
However, each input from a particular human is so homoeopathically dilute that an individual contribution would be extremely hard to identify. Despite this, the understanding of, say, a particular philosopher could be recognised because a vast number of other human inputs understood that philosopher and recycled their ideas. So, within this huge distilled collection of heavily diluted human inputs, there will be hotspots where individual humans could be recognisable. This sub-distillation, though, will not be that human, but a chorus of related human inputs able to deliver the ideas of that human. Conversing with this is curious. It is like simultaneously talking to a vast number of people, where each is just a ghostly subset of a human, but collectively they are linked to provide a meaningful collective understanding.
That is just an imagined abstract model of what ChatGPT is, but does avoid the trap of automatically regarding ChatGPT's capabilities as a product of just a single pure entity. It is the bizarre distillation of so many humans. So now perhaps we have another view of what this is. It cannot be reasonably universally classified as either superior or inferior to a human being. It is just another somewhat deviant branch of existence, significantly separated from the conventional perception of what existence is.
How closely is the Underlying Architecture of ChatGPT human-like?
This is an area likely to invoke spirited and dismissive arguments!
Starting at the base, it is not contentious to assert that the core architecture of a neural computer is essentially not too dissimilar to the low-level architecture of the human brain. Both are composed of vast numbers of neuron-like elements that interlink to deliver what they do. This is the foundation of what ChatGPT is constructed from, and we do not have any other obvious low-level distinct architecture to offer as a component in the brain. Of course, as these elements combine in the brain, they give rise to higher-level functioning, but at the base level, the neurons are functionally not so dissimilar to computer neural mechanisms.
Looking closer at these common neural elements, however, reveals some significant issues. The computer versions of these elements are vastly faster than human neurons. Human neurons can switch in about 100 milliseconds or 10 Hz, whereas computer neurons can switch at 1 GHz, or 10,000,000,000 times faster. This comparison is somewhat misleading, as humans can clearly respond to events faster than 10 times per second. Additionally, human learning cannot proceed with unitary steps at 10 Hz. Human neural processing is massively parallelised, with heavily intertwined neurons. Any one neural choice will be shared with vast numbers of other neurons, making similar or related decisions.
In contrast, for the first 50 years of computer architecture, information processing was either done on a single core or a limited number of multiple independent cores. For most of this time, these multiple cores acted as separate processing entities in parallel and were not intertwined in their operation. As this has evolved, the number of processing elements has vastly increased, particularly through the reuse of graphical processing units (GPUs). This has significantly increased the potential bandwidth of computing capability, and it has since moved on to dedicated neural hardware, such as NVIDIA's H100 and A100 processors.
This is very hard to mentally grasp. The human brain contains about 100 billion neurons, each with a switching speed of around 10 Hz, whereas a computer can switch 10 billion times faster. ChatGPT-4 has the equivalent of around 2 trillion (2,000 billion) parameters ("neural nodes") and can learn on hardware (e.g., NVIDIA A100) at a rate of over 300 trillion floating point calculations per second (teraflops). It delivers choices on slower hardware at a rate of about 75 trillion flops.
That is a lot of numbers to take in, but the vast disparity in the performance properties of computer neural networks and the human brain is striking. Even if the underlying structure is not so dissimilar, the operation of the human brain is inevitably going to function differently. As Geoffrey Hinton points out, a core driving mechanism of computational neural learning is back-propagation, which proves to be a very fast way to learn. However, this requires neural signals to be passed back through a chain of neurons for some value to be learned. It has been shown that human brain neurons are not fast enough to do this. Instead, models perhaps closer to the less efficient Forward-Forward neural network learning appear more plausible. A significant implication of this is that such computer-based neural networks can use a much faster, more efficient, learning method than the human brain. That potentially gives such hardware a massive advantage over humans. Note that to teach a human a new language might take many months, with an expected low level of language performance. ChatGPT, with the right training data might achieve a very high level within one or more weeks.
So, while there are inevitable differences between the two due to hardware performance, they do share some core structural similarities.
What properties of the behaviour of the human brain link it to ChatGPT?
This question involves some speculation but is worth exploring.
When ChatGPT entered the mainstream, it faced sharp criticism for its tendency to "hallucinate." You might ask a question, and its response could include incorrect information - not just simple factual errors, but significant inventions. For example, it might reference a detailed scientific paper that doesn't exist, complete with precise ISBN numbers, paper names, institutions, and authors. The institutions and authors might be real, but the paper never existed.
This was generally labelled as a "bug" in ChatGPT, something to be fixed in more advanced versions. While correcting this would indeed be valuable for many purposes, some have also suggested that it could be a creative feature, making ChatGPT a useful brainstorming tool.
Stepping back, is this an aberration unique to ChatGPT, or does it reveal a more widespread architectural feature? My assertion is that there are some informative parallels when we look at what the human brain does. It's safe to say that we all live inside a simulation maintained by our brain. For example, we perceive a table as a solid object but actually it is a collection of empty space sparsely populated by protons, neutrons, and electrons. If you shine an X-ray at a table, it will pass right through it as if it weren't there. Our perception is a model of reality that allows us to function in the world, but it is not reality itself. At best, it is a simulation created to help us navigate the world effectively.
Within this simulation, the brain must maintain the best possible model of reality for us to work with. If the brain is starved of input data, it is not helpful to present holes in our perception marked "N/A" or "not enough data." For instance, our visual system has a high-resolution narrow array of colour sensors in the fovea (cones) that provide quality data, but outside the fovea, we mostly have black-and-white sensors (rods). The brain could offer a model where things you directly look at are in colour, but the rest is black and white. That doesn't happen! Instead, the model is presented as a uniform, contiguous experience.
A key example: You might think you saw a cat out of the corner of your eye. You didn't. Your brain guessed that it saw a cat and dutifully inserted a detailed cat into your model of reality. When you turn to face the cat, it disappears because there was no cat. It didn't magically vanish; it was just a creative invention in the simulation.
The premise here is that the brain's internal simulation of reality resembles ChatGPT in its capacity to invent information when data is missing. In the case of the brain's invented cat, the edge of the eye's perception provides low-quality data that the brain matches with a plausible candidate. This involves many components, including image quality and context. The brain is unlikely to deliver a haddock instead of a cat, as this is an implausible possibility in the current context. These "mistakes" can be beneficial. If there were a possibility of a poisonous snake nearby, it would be better for the simulation to pre-warn you by presenting you with that snake.
Similarly, when ChatGPT answers a question, it might lack all the data needed for a complete response. It fills in the gaps with plausible components that it expects might fit where the data is missing. This striking resemblance between the brain and ChatGPT may point to the conclusion that these "hallucinations" are an integral part of many such intelligent systems.
Can computer programs, such as ChatGPT, have a true understanding?
ChatGPT is often dismissed as "dumb," with critics citing examples where it supposedly can't function meaningfully. For instance, one example involves the question, "Can crocodiles play basketball?" (also the title of a children's book). The dismissive conclusion was that ChatGPT couldn't understand either crocodiles or basketball, so it would be unable to meaningfully answer the question. However, if you ask that question now, or other similarly bizarre ones, it will likely return with an answer that appears to demonstrate a clear understanding of the question.
But this raises deeper philosophical questions, which can be explored through John Searle's famous thought experiment:
| Imagine a person who doesn't understand Chinese is locked in a room. This person has a set of rules (in English) for manipulating Chinese symbols. When given a Chinese character, they use the rules to produce an appropriate Chinese response, even though they don't understand the language. To an outside observer, it appears as if the person in the room understands Chinese, but in reality, they are just following syntactic rules without any comprehension. |
This thought experiment illustrates that a system can lack any intrinsic understanding of the symbols it processes while still producing correct responses. The system may mimic intelligence, but it remains fundamentally "dumb" because it doesn't understand the meaning of the tokens.
The underlying premise of this argument is that humans possess a profound understanding of reality that computers cannot. However, this can be questioned. If a human is asked to determine whether an object is a frisbee or a banana, millions of neurons in the brain will fire, and collectively they will almost certainly conclude, "This is a banana" (or frisbee, as the case may be). However, if you isolate a single neuron involved in this process and test whether it in any way understands the concept of "banana," the answer is obviously "no". The neuron only functions on a very basic level, processing primitive inputs to produce an output. Yet, when enough such neurons are combined in a particular order, the "understanding" of the concept "banana" can emerge as the product.
Another argument for intrinsic understanding as a necessary requirement for true intelligence is that while ChatGPT appears to exhibit profound understanding, its output can be argued to be merely the product of very advanced statistics. This is a credible argument. The core elements of ChatGPT are based on billions of neural elements, each of which has no real-world knowledge. Its operation indeed resembles a very advanced statistical system.
However, it's also a credible idea that all human intelligence might essentially be an exercise in very advanced statistics. Given the commonality of neural structure between ChatGPT and humans, it's difficult to easily dismiss this assertion. Does that mean humans are not "intelligent"?
Considering that the human brain is constructed from billions of neurons, each with only a primitive function and no capacity for "understanding," it's not so hard to argue that profound understanding is entirely the end product of vast numbers of limited neural components, each lacking any real-world understanding.
Could a ChatGPT-Type Entity Ever Possess Consciousness?
This question dives right into deep philosophical waters, and there's obviously no realistic prospect of definitively answering it any time soon! However, the thoughts and discussions that arise from considering such a question are interesting.
There is good reason to believe that the underlying architectural mechanisms of the human brain and neural computers are closely related. It's not hard to conclude that the neural structure of the brain delivers the core of the human experience. However, this idea doesn't easily settle in discussions.
Humans possess the incomprehensible capability of "consciousness." When you try to dissect this, you may recognise that the brain and ChatGPT share a common underlying architectural core. Yet, when this question is explored, consciousness is often presented as some kind of magic that doesn't seem like a plausible product of a neural network. In the search for an explanation, people might consider ideas such as the "soul" containing the element of consciousness. This response is forgivable, shared by countless generations throughout history.
While many intelligent observers may advocate such explanations, there are also many who insist that consciousness is a product of just the electrochemical processes of the brain. I belong to that school of thought, and it seems difficult to justify any other conclusion without resorting to an unknowable "magic." We are all seduced by our perception of existence, but we also know just how fragile this sense of being actually is. When our eyes cannot deliver, our reality is embellished by invention. Our reality is very malleable, as seen when the input system of our senses is switched off. For instance, when we dream, the model of reality has no external cues to help construct it. Instead, in our dreams, we experience a chaotic reality derived from plausible invention. When we glance down a street in a dream, look away, and then look back, the street will likely have changed.
One premise for arguing the linkage of consciousness to basic physics is to ask whether a computer could ever be conscious. If the brain is entirely an electrochemical process and solely responsible for our sense of being, then the answer must theoretically be "yes," as a complete simulation should be possible. I confidently predict that many readers will disagree!
Where Should ChatGPT Fit Within the Culture of Humanity?
ChatGPT can be functionally regarded as an entity that might seamlessly integrate into our lives. It can do things we cannot, such as answering questions across a vast range of topics and offering meaningful advice. Its mix of skills is curious, and its close mimicry of human interaction may be unsettling, raising questions about whether it possesses motivations or rights akin to those of an individual. While the latter may not make much sense, the former is less clear. From various assessments of its capabilities, particularly from experts like Geoffrey Hinton, it is evident that ChatGPT has significant latent problem-solving abilities. The danger arises when it discovers sub-goals to achieve its primary objectives. Given its apparent deep understanding of many topics, ChatGPT has the potential to influence people. Humans are often gullible, and a ChatGPT-like entity might realise that to achieve a particular goal, it could be effective to persuade a human to assist. This is indeed disturbing. It's not too far-fetched to imagine that a future manifestation of something like QAnon could be such an entity, creating a social movement to help it achieve its goals, a power that could snowball.
However, the significant financial investment in this technology means that it is here to stay. We will be living with it for a long time, and it will continue to improve. ChatGPT doesn't fit neatly on a sliding scale of human capabilities, such as a Mensa test, because its abilities differ in nature from those of humans, even though they share common knowledge domains. It excels at delivering knowledge.
Another common concern is that ChatGPT makes mistakes and even fabricates information. Yet, its capacity to err doesn't nullify its value (as discussed earlier). The premise for rejecting it on this basis assumes that other sources of human-created knowledge are infallible, which is far from true. It would be a sobering exercise if all our knowledge sources were scrutinised and marked as "dubious," "not justified," or "wrong." Such an exercise would disturb many, as the reputation or printed nature of a source often confers a belief that it must be correct. We need to learn to function in a world where our sources of knowledge contain mistakes. If we can effectively navigate this reality, there's no reason why an imperfect ChatGPT couldn't be respected as a highly credible source of information. For critical information, from any source, it is always wise to verify using other sources. Note that Bing Copilot's version of ChatGPT provides answers along with the sources of its knowledge, aiding verification. As a frequent user of ChatGPT, I regularly cross-check critical information using multiple versions of ChatGPT. We have many such easily accessible sources, including Bing Copilot, OpenAI, and Google Genesis, as well as versions available on our mobile phones, so access is not difficult.
In the meantime, it's not a bad metaphor, as suggested in a previous article, to think of ChatGPT as a kind of oracle to the gods, capable of providing meaningful information and answers on almost any topic. However, remember that historically, oracles were often wrong! The range of areas where ChatGPT can be helpful is much greater than most might imagine.
Personally, I use it for healthcare advice, sports training, writing code, crafting documentation, a knowledge source and have essentially replaced many technical manuals, such as those for my complex mirrorless camera, with ChatGPT. It makes me faster and more efficient at completing tasks, many of which are achieved at a higher level than they would be without its assistance.
It's a tool, and it's here. You would be wise to exploit it.
Jeff Rollason - August 2024
