Deep Dive: Can AI achieve human intelligence? An engineering evaluation
There’s lots of debate about whether language models like the GPT series are truly intelligent or just dumb statistical models that appear intelligent given sufficient training data. Some claim that AI will never have true intelligence regardless of future improvements in the field.
This deep dive will tackle the more challenging question of whether AI can achieve not just some primitive level of intelligence but true human-level intellect with all the implications that it involves. The answer will surprise you regardless of your stance on this, so grab a coffee and enjoy the ride.
We’ll start at the metaphysical level and transition through the complexities arising from physics, biology, computer science, hardware implications, and the surprising impact of economics.
Given that a significant part of the population holds religious beliefs, this wouldn’t be complete without evaluating the challenges that it adds before transitioning into scientific domains.
The Impact of Religious Beliefs
Religious beliefs about the existence of a soul or spirit can pose a challenge to the discussion of achieving human-level intelligence. In many religions, people believe that the soul or spirit guides them.
The main problem is that we don’t have a way of physically describing a soul or spirit. We can’t even imagine how to simulate a soul or spirit with a physical device since AI will run on a physical AI accelerator chip.
For religions that don’t view animals as having a soul or spirit, it may be tempting to strive for animal-level intelligence since many animals display a surprising level of intelligence. However, we can do better.
Some religions believe that the soul or spirit doesn’t guide certain categories of people, such as non-believers. Since people in these excluded categories still possess human intelligence, these religious beliefs imply that simulating a soul or spirit isn’t required to reach human intelligence.
If your belief involves a soul or spirit, consider the guidance it provides and the types of scenarios that benefit, such as making better moral choices.
The Connection Between Intelligence and Morality
While not everyone is religious, most people have moral values and principles that guide their behavior. Poor moral choices can result in negative interactions with others, leading some to believe that intelligence relies on morality. However, these moral choices can be justified with a selfish optimization function that takes into account future scenarios, maximizing the long-term experience. Even selfless acts that appear to have no immediate benefit, such as charitable donations, can be shown to be optimal behavior in sophisticated societies.
To understand what others are feeling, our brain mirrors those experiences as if they came from our senses. For example, when we see someone get injured, we cringe because it feels real for a split second. Similarly, movies or books can make us sad when a character is in despair even though we know that it’s a fictional story. This mirroring effect can be painful when seeing a person struggling, so our natural subconsciously selfish instinct is to reduce this painful experience by helping them.
Morals and values vary across cultures. Some immoral actions in one society may be acceptable in another. For example, the Wodaabe tribe participates in a “Wife Stealing Festival” where men compete in a beauty pageant to steal each other’s wives. Values are passed down from generation to generation or learned through interactions with others, so they can be instilled into an optimization function. Valuing life higher than money will guide AI to discard physical possessions to save a life.
Defining a system of values that accounts for complex and conflicting goals will be challenging. We’ll need a way of weighing predicted outcomes to evaluate choices. With this system, we can tweak the values and weights when encountering poor choices, so desirable results seem achievable after enough improvements.
Many individuals engage in immoral acts, such as bank robbers. Some robbers were considered to be extremely intelligent given the sophisticated nature of their heists. This involves many aspects of intelligence such as interaction with others, planning, deductive reasoning, predicting outcomes, and the sophisticated technical abilities to bypass systems.
All of this implies that morals rely on intelligence instead of the other way around. So morality is not required for intelligence and isn’t included in intelligence tests.
Evaluating Intelligence
To evaluate the performance of AI, we could start by testing it against the most common tasks we encounter. However, even if AI outperforms humans in 99% of these tasks, it still doesn’t guarantee that it reached human-level intelligence.
We can’t test AI in all possible scenarios that we might encounter, so it may fail in untested cases where children excel. Additionally, we can’t evaluate abstract abilities precisely such as creativity or philosophical discussions.
Comparing human intelligence to an AI system is challenging, so we need to re-evaluate this from a perspective that eliminates testing concerns. The only way to conclusively answer the question is if we can simulate the parts of the brain that lead to intelligence, disregarding non-essential elements like blood vessels.
The Limits of Computing
Scientists believe that our decisions stem from physical interactions within the brain, such as neurons interacting with synapses. All physical interactions are governed by physics, which has underlying mathematical equations to describe them. Some physicists claim that the universe is not merely described by mathematics but that it is mathematics.
With a complete mathematical description, we should be able to compute the physical interactions and simulate the output of the brain. However, this is challenging if the computations in the brain rely on quantum mechanical effects. In principle, we could solve the Schrodinger equation to predict the outcome of quantum systems. In practice, solving anything more complex than a two-hydrogen-atom scenario is beyond the capabilities of our top supercomputers. Even approximations fall apart as the scale and complexity of the molecules in the brain are astronomically larger than a two-hydrogen atom scenario. We’ll never be able to simulate a tiny fraction of the quantum interactions with our classical supercomputers, even if they become trillions of times more powerful.
Quantum computers, which operate on quantum bits (qubits) instead of classical bits, could overcome these limitations. Unlike classical bits, which can be a 1 or a 0 at any given time, qubits can be in a superposition of both. This strange phenomenon allows quantum computers to explore both scenarios at the same time, whereas a classical computer needs to check one option at a time. A quantum computer with 300 entangled qubits could, in a single step, check millions of times more possibilities than the number of atoms in the observable universe!
IBM announced a quantum computer with 433 qubits and expects to exceed 4,000 qubits by 2025. Can we use that to solve the problem? Quantum computing requires temperatures near absolute zero, roughly -273 °C, and also requires isolating it from noise like background radiation or magnetic fields. If you somehow achieve complete isolation, physics itself gets in the way due to random quantum fluctuations of empty space. This results in decoherence and errors, making computations useless. Physicists plan to combine about 1,000 qubits to act as a single logical qubit that self-corrects errors. Grouping 1,000 qubits to form a single clean one implies we’ll need hundreds of thousands of qubits before we can use them in a meaningful way.
The key enabler for the exponential speedup is due to all the qubits being in an entangled state. This will be much more difficult as we scale into the hundreds of thousands of qubits, so it’s difficult to predict the challenges that we’ll encounter. Practical quantum computing outside of tiny toy examples has always seemed to be a decade away since the 1980s and is likely to continue this trend for the next several decades. The one area where quantum physics is providing immediate value is for generating high-quality random numbers.
Luckily, the things that make quantum computing challenging are the same things that cause quantum effects to be destroyed in the brain. The human brain operates at a temperature of around 38.5 °C, which is many thousands of times hotter than the temperatures required for quantum computing. The quantum effects that need to be sustained cannot survive in such hot environments, so most scientists believe that brain computations do not rely on quantum effects.
With quantum physics out of the way, we need to find a classical mathematical description of the interactions between neurons and synapses in the brain. While this seems impossible, neural networks are universal approximators, able to approximate complex functions to any degree of accuracy provided they are sufficiently large. The best part is that they converge towards the solution with each training example, even though we don’t know anything about the complex mathematical system that we’re trying to simulate.
The Limits of Artificial Neural Networks
An artificial neural network in itself, can’t simulate a brain. Wait, can’t neural networks simulate anything? Let's examine the practical differences:
- After training, a neural network doesn’t continue to learn or adapt to new information and environments.
- Neural networks process each input equally, regardless of complexity. Language models generate one word at a time, but the computation for each word is the same, whereas we slow down and refine our word choice when dealing with complex topics. Image-generation models are an improvement in this regard; they don’t treat the neural network as the solution but use it in an iterative process to produce higher-quality results when allowed more time to perform more iterations.
- Neural networks lack introspection and aren’t capable of explaining their thought processes. The human brain has specialized regions, such as the visual cortex for sight. This allows us to use results from different parts of the brain to adapt our solution. For example, when our visual cortex registers a monster in the dark, we recognize the low probability of its accuracy and re-evaluate our interpretation of shadows, concluding that it’s a wardrobe hanging on a chair.
Neural networks of any size won’t solve all problems on their own. That’s because they perform a fixed computation to produce a result, which is at odds with the concept of computational irreducibility. There are many systems where the only way to compute a future state is by iterating through steps without any shortcut formula to jump to the solution. No neural network will ever be able to produce the digits of pi to an arbitrary number of digits or predict the future state of Conway’s Game of Life without iterating through the steps to get there.
To solve the above issues, we need to create a larger system that uses neural networks as tools rather than treating a single network as the solution itself. Here’s a rough outline:
- Create an orchestrator: This component can address computational irreducibility and spend more time when dealing with complex scenarios. The orchestrator represents our thought process, and its role is to devise a strategy for reaching the solution rather than providing the solution itself. It breaks down the problem into a sequence of actions or sub-problems and determines termination conditions, creating a mini-algorithm. The orchestrator feeds each sub-problem into the main network and combines the results to produce the final solution. While this sounds difficult, language models are already making great progress with programming tasks, creating algorithms that are more complex than our mental thought process.
- Providing explanations: The orchestrator can also solve the issue of explaining its thought process by providing the strategy that it generated. While the orchestrator can explain the reasoning for how it reached a solution, it won’t be able to explain lower building blocks. This is similar to how we can say that we see a rock, and while we can describe the attributes of a rock, we can’t explain the process that led us to immediately identify it visually in our visual cortex.
- Introduce a supervisor: To address introspection, we can create a supervisor that evaluates the likelihood that the generated solution is correct. The supervisor represents our gut feeling and evaluates the attributes of the solution instead of computing the solution itself. If the supervisor deems that an answer is missing an attribute or contains a surprising attribute, like physical objects breaking the laws of physics, re-process the input along with the feedback to produce a higher-quality result. When coaching developers, I have them describe the attributes that a successful solution should have and see their surprise as they guide themselves to the correct solution, so this is an effective way of solving challenging problems.
- Learning: Unlike current models, the vast majority of its intellect will be acquired through continual learning from interactions after the initial training phase. The orchestrator could leverage two neural networks, one for conscious thought and another for subconscious learning and adaptation. Feedback enters a priority queue to train the subconscious network, and the conscious network is periodically replaced with the updated one. Similarly to how we can’t train a child to do calculus before introducing simpler math concepts, we’ll need to interact with the system by gradually increasing the difficulty so that it can learn a chain of baby steps in order to understand complex concepts.
Although the above is an oversimplification, the takeaway point is that this can be broken down into smaller problems, and tackled with computer science techniques. By iterating on each component, we’ll eventually reach a model that learns and adapts similarly to the way we learn.
Hardware Requirements for Simulating a Brain
The brain has about 86 billion neurons — roughly the same magnitude as the number of stars in the Milky Way galaxy! These neurons are interconnected with nearly a quadrillion (1,000,000,000,000,000!) synapses, so the numbers exceed comprehension. Additionally, each synapse contains different molecular switches, making it more akin to 1,000 transistors in modern CPUs.
The brain is postulated to operate at about one exaFLOP, which is a billion times a billion operations per second. Supercomputers reached this computational threshold and will achieve 10 exaFLOPs around 2025. Since supercomputers get about 10 times faster every 5 years, computing power shouldn’t be a concern.
The shocking part is that the brain operates on just 12 to 20 watts, whereas the Aurora supercomputer requires 60 megawatts — enough to power 60,000 homes. This makes the human brain several million times more efficient than state-of-the-art supercomputers! If we assume that computing will become a couple of hundred times more efficient in the next decade, that would still leave computers 10,000 times less efficient than the brain.
Given the current state of computing, the only way to simulate a brain this decade will be with supercomputers accessible through the cloud.
Timeline for Superhuman Intelligence
Surprisingly, the year when AI reaches human intelligence is likely to be the year when AI achieves superhuman intelligence. As soon as we have a model that reaches human intelligence, we can create a larger version of that model, and the first model can accelerate the training of the second by augmenting the training data set and providing feedback on its responses.
Two main factors limit the advent of general AI:
- Current models are unsophisticated, treating neural networks as solutions instead of using them as tools in a larger system. Progress in AI models is improving at an alarming rate, with state-of-the-art knowledge becoming outdated every year. While some skeptics doubt the potential of AI due to current limitations, this is actually a good sign. If we couldn’t identify obvious flaws, that would imply that we’re running out of easy opportunities for improvement. As AI is used in more areas, we’ll discover more shortcomings and get even more feedback, which will further accelerate progress until we arrive at a model where intelligence emerges similarly to a brain.
- Businesses focus on creating value so they won’t be willing to pay thousands of dollars per hour to perform low-value tasks. If a task can be completed by a human with less cost then it doesn’t make business sense to pay much more for a supercomputer to tackle it. The cost of operating these supercomputers stops them from getting that critical post-training feedback with common tasks which is the main source of intelligence. We’ll be in a strange state where it exceeds for high-value tasks and underperforms at normal tasks. Surprisingly, rather than computing performance, it’s actually the economics of computing efficiency that will prevent it from reaching human intelligence.
Advancements in fields like Memristor Neural Networks, Neuromorphic Computing, and Photonic Computing will improve the efficiency of AI workloads by 100 times. This will lower operating costs from thousands of dollars per hour down to tens of dollars per hour, making it financially feasible to interact with it for daily tasks. This will generate the critical feedback it requires to develop intelligence. Contrary to our intuition, AI will start automating the most difficult high-value tasks first and transition to simpler lower-value tasks as operational costs drop over time.
Both the modeling and efficiency issues should be resolved by the early 2030s, leading to the emergence of general artificial intelligence that will surpass humans in all intelligence-related tasks. Note that this doesn’t mean that intelligence-related jobs will start to be automated in the early 2030s. Instead, that’s when the last of the low-value tasks will be automated. Any business that hasn’t transitioned to relying on AI by the early 2030s will struggle to compete and stay in business since its competitors will provide more value per dollar.
AI will eliminate most current jobs so the main way to secure our future is to use it to create new types of jobs instead of hanging onto the past. Unlike a genie, a super-intelligent AI won’t be able to magically solve any problem, such as explaining current gaps in our understanding of physics. Instead, this will require an iterative process using AI to devise new experiments, performing those experiments, and repeating this process to narrow in on the answer. Imagine a role where you’re part of a team guiding AI toward improving one of the billions of aspects that affects humanity such as improving a specific theory, finding better alternatives for the many processes that create pollution, discovering a cure for a specific disease, or social aspects like improving our interactions with each other. The seemingly-infinite aspects to improve would make perfect candidates for countless jobs in the post-AI era with each task spanning multiple years.
Follow and subscribe to be notified of my future articles.
