Frequently asked questions

There are some questions that family and friends ask me repeatedly. Because of my profession and interests, they are often in the vein of "what do you do for work?" and "what do you think about artificial intelligence?" I tend to have difficulties articulating good answers that are somewhat understandable on the spot. Partly for my own sake, and for those who might be interested, I've therefore decided to write down my responses to these frequently asked questions (FAQs) in a more thoughtful manner. My answers here will be phrased so that very little to no background in science is necessary to understand them. Note that there will be several words which will have to be defined in order to explain the concepts without background knowledge, and for the most important terms and concepts I go on to define them.

The responses included in this FAQ so far are the following:

• What is AI? (2026-10-03)
• What do I do for work? (2026-10-03)
• Do I think AI is conscious? (2026-10-04)
• Do I think AI will make humans extinct? (2026-10-05)

Note the date stamp attached to the posts, since my answers on these matters may change over time - particularly in fields like AI that change rapidly. I have, however, attempted to not make my responses depend too heavily on the current status of the fields.

What is AI?

The opposite of artificial is natural and the simplest and broadest definition of artificial intelligence is therefore "any form of intelligence that is not natural". Our brains and those of other species are natural in the sense that they have evolved in nature through natural selection, whereas intelligent systems that have been engineered by us are not. There is a natural follow-up question as to the definition of intelligence, which is notoriously difficult to pinpoint. Instead of defining intelligence, I tend to go with the tactic employed by U.S. Supreme Court Justice Potter Stewart in 1964 in his response to the definition of pornography: "I know it when I see it". That'll do.

The idea that artificial systems, like computers, could become intelligent was first conceived, or at least popularized, by Alan Turing in the 1950s. By the 1950s, the field of computer science was in its infancy, but the earliest computers saw increasing capabilities. Today, the term artificial intelligence has been skewed from its original meaning, and is either tied to algorithms that seem to possess some sort of humanity, like the system embedding in ChatGPT, but is also often tied to any computer system that seemingly learns or adapts. I will go on to explain these concepts further.

Before we even start thinking about what artificial intelligence might be, there are some terms that need to be cleared up. Firstly, we must understand what an algorithm is. An algorithm can be thought of as a recipe or equivalently a set of instructions. If we wish to sum up all numbers up to 100, we could write an algorithm that keeps a tally of the sum, starting at 0, and iteratively counting up until it reached 100, adding each number to the sum. This can be written in instructions that executes on the computer chip, and reports back the answer. All computer systems are built from a set of algorithms written in a formalized language called a programming language. At first, a programming language looks very strange and foreign, but just like any other language it is learnable through practice. For someone who has coded for several years, the computer language reads like a native tongue.

To someone who isn't too familiar with mathematics or computer science, it might seem odd that an algorithm can learn. What does that even mean? How can a recipe learn? Part of the confusion is because we tend to think of learning as something tied to living things, like humans or dogs, while in mathematics and computer science we see learning as a way to "adjust a model to align with an objective". That last part requires some explanation. First, what is a model? A model, like a model airplane or a model train, is a simplified representation of the real thing. In mathematics, we create models that are simple but supposed to represent something very complex, often to make predictions. For example, we can create a model of the weather, to try to forecast it. Similarly, we can make models of disease spread, to try to forecast it, alongside various modelled interventions. Fundamentally, however, a model is just a set of equations. Numbers and variables on a sheet of paper that represent something. Models are everywhere - without them the world is too complex to understand. There is a famous saying that "some models are useful, but all models are wrong" - and that is true. No model exactly represents something, but it may be accurate enough for whatever we want to use it for.

Models may seem alien to some, so here is an example. Say you want to create a model that represents the price of a house. An extremely simple model could simply take the number of rooms; the more rooms, the more expensive it is. The model would look something like:

Price of house = number of rooms × 50,000

which means that the price of a house is 50,000 times the number of rooms. A plot of this model can be seen below.

This model wouldn't be nearly correct - there are many more things that go into the price of a house, but it is not entirely wrong either. The price of a house tends to increase with the number of rooms.

Creating a good mathematical model is challenging and requires a lot of expertise about whatever is being modelled. For example, to create a weather model that is somewhat accurate, knowledge about the physics of wind patterns, heat dispersion and water streams is required. However, in the 20th century there was a revelation: what if we could make the computers learn the model themselves? The key behind this revelation was an algorithm that allowed the model to use observations (data) to adjust so that it fits with observations. Starting with a poor model, the method could be employed to gradually improve the model. It should be noted that there is no magic involved: the method for adjusting the model can be written down fairly succinctly and isn't too hard to grasp, although I will not attempt to explain it here. In brief however, it entails penalizing and adjusting a model when it's wrong, and adjusting its parameters to make it slightly better - and doing so over and over again. In the example with houses, we could image having a large bank of data; examples that include various features of the house, like location, the year it was build, area, and how many rooms it has, alongside their prices. The algorithm could then utilize the features of the houses, to create a model that accurately predicts their prices. Such a model would likely be much better than the one we created above, and we did not have to create the model ourselves through tedious trial and error.

There are lots of methods that now apply this general framework: start with a poor model, and adjust it using data until it's better. This is generally called "machine learning", and adjusting the model is usually called "training" or "fitting" the model. A particular type of model is inspired by the structure and inner workings of our brain. Our brain consists of specialized cells called neurons which are connected in large networks. The neurons specialize in propagating electrical signals, and such signals are passed around in this network and eventually out onto the nerve endings in our muscles, making us move, walk and talk. These signals also make us able to think and somehow make us conscious, although we are uncertain precisely why or how (more on this in the section "is AI conscious?"). In the 1950s, it was realized that this general "neural network" framework could be simulated in a computer by setting up a simplified model of neurons, and this model could be combined with the "learning algorithm" described previously. The "neural network"-paradigm is still the foundation of the most powerful learning systems today, including large language models. We should remember, that just like the learning algorithm, neural networks are not magic. The underlying mathematical principles can be written down on a single piece of paper and are part of most undergraduate degrees in engineering, math or physics.

Around the 50s to 70s is when "machine learning" and "artificial intelligence" had their first wave of interest - and an astounding part of the fundamental theory was developed already back then. Visionaries of the time imagined how the neural networks could eventually be used to create intelligences that were human-like, giving rise to heaps of science-fiction movies and novels. However, these systems require powerful computers and lots of data to learn efficiently, and it wasn't until the 2000s and 2010s that they became powerful enough to outperform the classical methods. The period from 1980 to 2000 is called the "machine learning winter", and hopes that these systems would ever be viable and scalable were low. Circa 2010 was when a researcher realized that the hardware that had been developed for computer games, the so-called GPU, was especially well suited for the type of calculations that were needed to accelerate the learning of neural networks, which led to rapid success in continued development.

The 2010s saw a revolution in these learning systems. They were extremely hyped in academia, and everyone and their uncle was doing research on machine learning and artificial intelligence, and it was during this period that I got interested in the field! Towards the final years of the 2010s (2017-2019), there were several advancements in using these methods for translation, which paved the way for large language models and chatbots. These times also saw leaps of advancement in the ability to create new images, and lots more.

But if the machines are the ones doing the learning, what are humans who work with artificial intelligence doing? What does research look like? Well, there are many answers to this question, but generally, there are three main steps to creating a machine learning model: first, setting up the "architecture" of the model. The architecture can be thought of as the scaffold, template or general outline of the model, and is extremely important for the ability of the model to learn. As mentioned earlier, neural networks was a very potent way structure for making computers learn, and today researchers are putting artificial neurons together in networks with advanced structures that further facilitate learning. Remember that all of these neurons are just lines of code in a computer, are any structure that is created only modifies the underlying mathematics. However, it is helpful to think of these neurons as actual objects with substance, and one can imagine how these networks seemingly twist and turn and connect in various ways. Then, gathering adequate and high-quality data that will be used in the learning algorithm is equally important. A nice quote in this domain is "shit in, shit out" - meaning that if the data that goes into the model is of low quality, the model, too, will be of low quality. Lastly, the learning algorithm itself can be adjusted to more effectively update the model using the data. There are continuous developments in all of these domains, and part of my research is in developing new architectures and methods (see "What do I do for work?").

When it comes to the celebrated language models which are in the limelight today, the data does not correspond to real estate features and prices, but instead to text. Fortunately, we have lots of text - the internet is full of it (in fact, I'm contributing to the opus by writing this). We don't have to go into the architecture of these models, but you can imagine that they are immensely complicated and constantly subject to improvements. The learning process is also very complicated nowadays, but its backbone is based on something called "masking". As I mentioned earlier, all a model does is predict, and in terms of language, the prediction is made as follows: take a sentence, remove some words or letters, and make the model predict those words or letters. For example a masked sentence could be: "My ____ is Erik Hartman" - and the right answer would be to fill the blank space with "name". It is amazing that such a simple idea gives rise to the sophisticated language models we have today, that are able to solve problems many humans can't solve.

But how can this work? How can these models seem to become "intelligent" just by learning to fill in characters or words? Well, imagine that you are reading a murder mystery novel, and right when you are nearing the end and the murderer is about to be revealed, the name of the murderer has been masked out. Then imagine that you are now tasked to fill it in. Doing so requires you to figure out who the murderer actually was, which could be a terribly challenging task! So this simple exercise of predicting what word should be put at a certain spot turns out to involve solving a murder mystery, which would require some sort of intelligence or complex pattern recognition. It's a tricky problem in disguise. As such, intelligence is seemingly emerging through the simple learning process of masking and prediction.

Now that we've talked about artificial intelligence models that are created and trained to mimic language, it is important to remember that this is only a small part of the plethora of models that exist. Other models are those which steer self-driving cars, or predict stock prices. My research, for example, does not deal with language, driving, or stock - it deals with biology. Instead of words and letters, I create artificial intelligence systems that "learn" about genes, proteins, organisms, and chemistry. The goal of my research is to make these models help us design new drugs, and to make them help us unveil and understand parts of biology we couldn't understand without their help. You see, the human brain is limited in a number of ways. We do not have a very long memory, and we tend to only be able to keep a few things in our head at the same time. Models do not have those limitations, and can therefore pick up patterns that we can't.

Artificial intelligence systems and machine learning raise some fundamental philosophical questions. Naturally, they make us rethink what intelligence and consciousness is, questioning our place in the universe. If intelligence can arise by fitting a very complex model to data, what is so special about being human? Can a model - which fundamentally is a set of equations - be conscious? (It should be noted that a large model may have as many as 1 trillion parameters, and there isn't enough paper in the world to write down the equations it is made of, but it is an equation, nonetheless.) I deal with some of these questions in the section "Is AI conscious?"

This little tour of artificial intelligence hopefully makes the field less mystical and a bit clearer. I think the most important thing to remember is that although making machines learn seems like magic, it isn't. The mathematical toolkit for the field has been around for 100 years. The only thing that seems magical is how well these methods actually work! It should also be noted that the development of artificial intelligence models, particularly language models, is subject to tremendous engineering feats in efforts to make them better. The potential capital involved in this market is massive, and the companies are drafting some of the smartest people in the world in effort to take a share in the market. If anything is magical, it is human ingenuity and motivation.

What do I do for work?

One of the problems I have when trying to make others understand what I do for work is that my field of study is not in one of the classical categories. If I were to be a mathematician, a biologist, a medical doctor, or a philosopher, I could just say that and be understood since these categories carry enough context and cultural stereotypes. Unfortunately, I work in the field of computational biology, which is fairly new and unfamiliar to most - we don't have a subject in school for it yet, and people's eyes tend to turn somewhat glassy when I mention my profession.

The easiest way to describe what a computational scientist does is to divide it up into method and application. The application of my research pertains to biology, but the methods I use are mathematical. I do not wear a lab coat or work with pipettes. My research is done purely in my head, on paper, on the computer, or on a blackboard. Eventually, the ideas and predictions I make are tested in the lab, but luckily I can leave the dirty work to my colleagues. To be a computational biologist you have to be native in two separate languages, that of mathematics and logic on one hand, but also that of biology and chemistry on the other. However, it turns out that there is a deep connection between nature and mathematics, and many mathematical methods have been inspired by biology, and likewise, many biological problems have analogies in mathematics.

While I wrote that computational biology is a newcomer to the sciences, that is not exactly true. Scientists have mixed mathematics and biology for centuries; particular early examples include mathematical modeling of population and disease dynamics by Bernoulli and Malthus (17-18th century). Additionally, many classical biologists who predated the advent of computers would, in my mind, have been computational biologists if computers were available. One such case, I believe, is Charles Darwin, who was a theoretician more than anything and who likely would have been a computational scientist, had he lived today. Additionally, there were several cases of mathematicians in the early 20th century becoming interested in biology towards the end of their lives. I'd like to think that this is the mathematicians equivalent of becoming religious when the end is neigh.

Being a computational biologist makes you think in a certain way. You're constantly translating between the natural and the abstract, and eventually the two sort of melt together. The result of which is a single framework that fits both biology and mathematics, and does so very neatly. I'm certain that the surge the field is seeing today will lead to many great discoveries, partly due to simply merging the abstract and the natural. In the past, medical and biological discoveries were largely made through serendipity and luck. You'd look through the microscope and happen to find something, as Leeuwenhoek did when finding bacteria - or you place a petri dish next to your bedside table and accidentally discover penicillin, like Alexander Fleming did. Computational biology is emerging as a way to create large mathematical frameworks encompassing the complex biological world, which is our most promising way of systematically discover new drugs, moving away from serendipity and understanding biology as a whole.

So my work lies in this soup of mathematics and biology, but additionally, my PhD has two separate tracks. One track deals with a specific biological problem whereas the other generally deals with new cool advances in mathematics that tickle my brain in the right way. But first, let's get into my first track.

Our body consists of a host of molecules. Arguably the most important category of molecules is called proteins. Proteins are the functional molecules of biology - they are what make cells do things as opposed to being static. They are the building blocks of life. If our genes is a recipe book, the proteins are the cake. Naturally, lots of research focuses on trying to map the function of proteins, as doing so would help us understand and manipulate biology, to, for example, treat diseases. However, my work does not deal with proteins directly - instead it deals with the product that is generated when proteins break down. Our body is constantly breaking down millions of protein molecules as part of a normal recycling process. In fact, we have special biological scissor-proteins that are specialized in breaking down other proteins. Whenever something happens in our body, if we get ill or if we get hurt for example, our body needs to exchange what proteins we have in our cells in order to deal with the disease or damage. These scissor proteins break down the old to give space for the new. Mobilize the forces, so to speak.

Imagine that a protein is a piece of paper, and breaking it down means tearing the paper apart. Eventually you'll have tiny bits of paper, and the protein equivalents of these small pieces are called "peptides" (recognize the suffix "id" as indicating that it is small). The presence of these tiny peptides have been ignored throughout history, and they were simply considered to be byproducts generated randomly during protein breakdown; however, it turns out that some of these degradation products have functions in themselves! This is a fascinating finding which has great implications not only for humans, but for all biological life. It tells us something fundamental about biology: that function exists not only at the protein level, but that more function emerges when the protein breaks down. Similarly to those pants that have sippers at the knees so that the bottom half can be removed to make shorts, proteins have functions when split in twain.

Particularly interesting examples of these peptides are those that play a part in our immune system. It turns out that some of these peptides kill bacteria and modulate inflammatory processes in the body. Not only that, bacteria themselves utilize peptides to fight each other and us, their host. This new world of peptides is extremely tricky to study, but computational methods can help us uncover their secret world. This research is a direct descendant of my dad's work, which focused on finding precisely these antimicrobial peptides. My research, focuses on generalizing this idea to all life on earth, and beyond just the immune system. I hypothesize that these peptides exist throughout the tree of life, and that they hold many functions for all sorts of biological processes. This ambitious program, which seeks to uncover a (potentially) fundamental aspect of biology, is maybe what I will spend my life working on.

The second track of my PhD pertains simply to creating mathematical and computational methods I find interesting. I do have a passion for mathematics, and these project challenge me, forcing me to continuously learn about new mathematical concepts. Often, this involves developing methods in the field of machine learning and artificial intelligence that apply to biology. Instead of training these systems to understand natural language, I focus on the language of biology. For example, instead of making these models generate images I train models that have specific architectures that allow them to generate molecules. Another example of recent work involves creating a model that can predict how patients with sepsis will progress in their disease. As I explained above, the core underlying mechanism of artificial intelligence is based on creating models that learn through examples. Once they are trained, biological artificial intelligence models can predict things - like which genes or proteins are associated with disease.

Do I think AI is conscious?

To preface this section, I want to point out that I am not a philosopher or a neurobiologist, and I have no professional qualifications to state anything about consciousness. However, for almost a decade I have peered at the consciousness debate from the sidelines, reading books about theories of consciousness, and trying to stay up to date on research developments. It is safe to say that no one knows what consciousness is - but at the same time, everyone does. In fact, our own consciousness is the only thing we can be certain about. It is what we directly experience first-hand. I think, therefore I am. What we're uncertain about is therefore not that we are conscious, but what consciousness is and why we have it. Therefore, any speculation about whether artificial intelligence systems are conscious is futile.

However, it is still interesting to speculate whether artificial systems could ever become conscious. The basis of this debate boils down to the most fundamental divide in consciousness theory, namely whether the material world as we know it contains everything that is required to create a conscious being, or whether it requires something a little extra to light the spark. The stance that consciousness arises purely from matter is called materialism. The idea that body and mind are divided is called dualism. Historically, most people have been dualists and believed that we have a soul or some sort of other spirit that is not made of matter. In fact, in 1907 there was a famous experiment that attempted to "weigh the soul" by placing terminally ill patients on scales, imagining that there would be a weight difference as each patient's soul left their body upon death. There was not. Today, the philosophical divide remains, but is more heavily weighted towards materialism. It should be noted that there are many theories about consciousness, and in 2023, philosopher Robert Kuhn went through a large number of them, tallying up a whopping 325 theories in his paper "A landscape of consciousness: Toward a taxonomy of explanations and implications".

If you are not a materialist and believe that there is something more to consciousness than matter, then you have no reason to believe that artificial systems would ever become conscious. However, if you are a materialist, you may see it as probable that something so sophisticated and seemingly intelligent would develop consciousness. All materialists face the so-called "hard problem of consciousness" defined by David Chalmers: that consciousness somehow emerges from non-conscious parts. Many materialists believe that the most important thing for developing consciousness is not the substrate, but the configuration. They see consciousness as an emergent property, much like wetness emerges when many water molecules are gathered together. No single water molecule is wet, but at some point, if you get enough of them together, wetness emerges. Similarly, you may say that no single cell is conscious, but put them together in a specific pattern, namely that of the brain, and consciousness emerges. Therefore, according to this theory, if we were to put together computer circuits, or "neurons made of silicon", into a specific configuration, consciousness could potentially emerge. Some materialists, however, believe that the biological substrate that makes up our brain is necessary for consciousness. If that is the case, whatever circuit configuration we create by artificial means, consciousness would never emerge.

Personally, I am a materialist and I believe consciousness is substrate-independent. My view is that consciousness emerged through evolution as a necessity. Imagine some prehistoric creature that sees before it a cliff. It goes on to imagine what would happen if it, the creature, jumped down said cliff. Naturally, it realizes that wouldn't be too pleasant. But what did that imagining entail? It required the creature to get some idea of what "I" is in relation to the environment, and further, it required the creature to see itself from a third-person perspective to imagine what would occur if it jumped down the cliff. While this example is somewhat constrained since most creatures are afraid of heights as a matter of reflex, rather than explicitly imagining what would happen if they were to fall, it seems clear to me that being able to put oneself in hypothetical scenarios is important for survival. Further, social beings also benefit from being able to put themselves in someone else's position. This curation of hypothetical scenarios is constantly done by us humans. We can imagine the most abstract of scenarios, and we can role-play what several individuals would say and do. In all of this, we need to keep track of ourselves; the "I" is extremely important. I believe that these capabilities, which involve imagining hypotheticals, "what ifs" and someone else's viewpoint, evolved because they improve our chances of survival. I further believe that as these facets of our mind developed, consciousness emerged.

Now, back to the question: are artificial intelligence systems conscious? As mentioned, there is no way we can tell. But, you might say, "When I talk to an AI, I can feel that they are conscious". Unfortunately, the current language models are engineered to make you believe just that. Let me explain. Firstly, humans are amazing at anthropomorphizing. In particular, we are very good at identifying faces and voices, and interpreting someone's intent. These features are hardwired into our biology, as they are very important for our survival as a hypersocial species. That is the main reason why people are scared of ghosts. When our senses are heightened in the dark of night, anything that has any minor resemblance to a face gets picked up and registers in our brain. It may be shadows, dust, or pure imagination. As we then go to investigate and find no one there, we deduce that what we've seen must be a ghost - while in reality, it was our own brain fooling us by constantly scanning our surroundings for things resembling faces. Our ability to recognize faces can also grant us enjoyment in our everyday lives, when we find rocks on the beach that look like faces, or when we find a picture of Jesus in the golden-brown features of our breakfast toast.

As I explained in the section "What is AI?", current models are trained to mimic human language - and engineers have made sure that they are very good at it. The whole objective of language modelling is in fact to make sure that these models communicate as much like humans as possible. It is therefore extremely easy to be fooled into believing that these systems are conscious. If we can find a rock on the beach that reminds us of our neighbor, talking to a language model which is nearly indistinguishable in its way of communicating from a human will prompt our primitive brain to believe that "surely, it must be conscious!" It is therefore impossible for us to understand whether these models are conscious for two reasons: firstly, we don't know what consciousness is, and secondly, precisely because these models have been trained to mimic human language, they are prone to fooling us. If a model that was trained on something entirely disparate, say image generation, eventually started to create images with text saying "help, I'm conscious", that would be more convincing. In that case, the system would not have been made to fool us, but it would seemingly have developed something like consciousness despite that. This is not the case with language models.

It should be noted that while any of us can assert that we ourselves are conscious from first-hand experience, none of us can be truly certain that anyone else is. It is completely possible that you are the only one in the world that experiences consciousness and that everyone else simply acts like they are experiencing it. Such a thought-provoking creature, one who acts like it's conscious yet has no inner world, no inner eye, is called a "philosophical zombie". If you were to be the only conscious human and everyone else simply imitated conscious behavior, you would not have any way to check. You cannot inspect someone else's mind to get their first-hand experience. Everything you observe about them is filtered through them telling you about it. Another similar example to this is the fact that I cannot assert that your perception of the color "red" is the same as my "red". I know what my "red" looks like, and I know the qualitative feel that it gives me. But perhaps if I were to look through your eyes, your perception of "red" could be equivalent to my perception of "blue", and we would not have any way of asserting that.

Some argue that the scientific method, which relies on objectivity and measurements and disregards qualitative aspects of reality, is unable to deal with consciousness in any way. Philosopher Philip Goff dates the modern scientific method back to Galileo and outlines why the modern scientific era is unable to deal with consciousness in his book titled "Galileo's Error".

Anyhow, this brief venture into consciousness has left us with three main takeaway messages about the consciousness of artificial systems: firstly, we cannot know whether such systems are conscious or not, just like I cannot assert whether you are conscious or not. We do not have a definition of consciousness, nor a unit of consciousness that we can measure. Therefore, a good way to spot either a charlatan or someone who is delirious is if they claim that AI is conscious. Secondly, these systems are designed to mimic human speech, which is intimately tied to our notion of consciousness. This makes us particularly prone to being fooled into believing that they are indeed conscious. Lastly, you may believe that consciousness requires something like a "spirit" or a "soul", and that it cannot arise purely from a certain configuration of matter, and in that case, you have no reason to ever believe that such systems would become conscious.

Do I think AI will make humans extinct?

At the time of writing this section, in October 2026, the public's eyes have opened to the potential dangers of artificial intelligence. The past month has seen several engineers and researchers leave the leading companies due to concerns that the companies were irresponsible in continuing the development of these systems. But that raises the question: what is it that these researchers and engineers are afraid of?

Concerns regarding the safety of artificial intelligence are as old as the first ideas about it in the 1950s. Already back then, with their primitive systems, researchers and writers recognized that artificial intelligence posed several potential threats. My favorite science-fiction author, Isaac Asimov, wrote several books in the 70s and 80s which outlined potential future scenarios in which artificial intelligence took over. These ideas were born from the simple fact that, as biological organisms, we are shackled by the laws of biology. Our brains have evolved only enough to ensure that we survive and thrive in our environment, and we are limited in our ability to store memories and process information. Additionally, our brain signals, propagated through neurons, are slow, only reaching about 1 meter per second, making our reaction time somewhere around a few hundred milliseconds. Computers, on the other hand, do not have these limitations. We could imagine engineering a computer that is as big as a whole planet (which Isaac Asimov imagined), and which therefore has tremendously more processing power than our measly squishy blobs for brains. Additionally, computers are not limited to the ever so tardy communication speed of our neurons, but instead communicate at the speed of light, a whopping 300,000 times faster than us. Conceptually, a computer equivalent of our brain could therefore think thoughts that correspond to 9.6 years in a single second. An hour of "computer thoughts" corresponds to 34,223 years of "human thoughts", simply through the difference in signalling speed. While our brains are somewhat plastic up until a certain age, a computer could theoretically evolve nearly indefinitely in real time. This is a long-winded way of saying that artificial systems have the potential to do tremendously more than biological ones, simply due to the fact that they are not constrained in speed, size, or configurability.

Early visionaries imagined a scenario in which researchers would begin developing artificial intelligence systems, but eventually, the system itself would become more competent than its creators. By then, it would take the wheel and continue its own development. Due to the tremendous speed of computers mentioned above, it would improve itself rapidly. Every time it improved, it would become smarter, making it realize how to make even more improvements. This recursive path, in which the system improves itself, therefore making itself smarter and capable of improving itself even more... ad infinitum, is called "recursive self-improvement" (or RSI, for short), and results in something called a "singularity", when the system's intelligence increases more and more rapidly, without human intervention. The key turning point was when the model started being able to improve itself, and it reached recursive self-improvement. The philosophical stance is that when this occurs, it will develop an intellect that is unimaginably superior to ours. Imagining what the system would be like then is by definition impossible; it is beyond us. Much like ants cannot imagine what a human might think, we would not be able to imagine what the system thinks. This is a scary prospect, since we would be at its mercy.

The Swedish philosopher Nick Bostrom is a professor at Oxford and was well ahead of his time when he released the book "Superintelligence", outlining several ways in which artificial intelligence systems can lead to terrible outcomes. The scenario described above, in which we become subjugated to a vastly superior intelligence, is only one potential scenario in which things go wrong. It presumes that the artificial intelligence system has a will of its own, and is somewhat autonomous. Today's systems, however, are very much not autonomous. They are directly prompted by the user, and if the user does not prompt the model, it will not do anything. In this sense, it is more like a calculator that does something once you press the buttons than it is like an animal or any other autonomous agent. However, Nick foresaw trouble in a future in which this remains the case too. Imagine that you are the owner of a paperclip factory and gain access to the world's most competent artificial intelligence system. This system, like in our previous story, is indescribably competent, and for the purposes of this thought experiment, can do essentially anything. As a good capitalist, you prompt the system to help you create as many paperclips as possible! The system does your bidding and goes above and beyond to meet your demands. So much so that eventually all atoms on Earth, including those constituting you and me, have been repurposed in its mission to create paperclips.

This idea that a machine can follow our instructions yet cause unintended and devastating consequences is called "misalignment" and is captured in the so-called "alignment problem". It is somewhat surprising to me how challenging it is to give instructions that do not result in misalignment. You might think that a prompt like "make all life on Earth better" would suffice, but even then there are scenarios that are not what we'd want. These scenarios paint a grim future, but that is just one side of the coin. The other side represents a future where most, if not all, of our material problems are solved. A superintelligent system could begin by solving the energy crisis by coming up with how atomic fusion could be done, and it could then solve cancer, followed by the rest of the diseases. And why not figure out how to stop aging while it's at it? This future is incredible, and Nick Bostrom's latest book is called "Deep Utopia", outlining precisely what type of future that would be.

In my mind, the future holds five different potential scenarios. One of them entails the grim future first presented above, in which we reach recursive self-improvement, but are not able to contain the system that we've created. In such a future, it would likely lead to the demise of the human race as we know it. The second scenario is the good version of this extreme, where our future flows with abundance of all sorts, and nearly all of our problems are solved. These scenarios are both contingent on the development of superintelligence, which was first defined by Bostrom as the point where the system is far superior to humans in all domains.

But perhaps recursive self-improvement isn't real. Perhaps our reality does not allow us to reach the singularity. The third scenario plays out if we don't reach recursive self-improvement, but instead we keep chipping away at the models at a steady pace. The systems will slowly get better, but they won't rapidly become truly superintelligent as in the first two scenarios. This is the domain in which we are today. Contemporary models are good at many things, but completely worthless at others where humans excel. Then, we can imagine a grim future, in which companies and governments weaponize these systems, exerting control over the general population with new and improved capabilities. In this scenario, artificial intelligence systems are tools used by a few bad human actors who are in power. Likewise, we can imagine a promising future without reaching superintelligence, in which these models are used for good. This is a future in which we use these fantastic tools to gradually build better societies, solve problems and achieve a better future. These two scenarios can be seen as mild versions of the first two, and, as explained, are not contingent on the rapid improvement occurring during recursive self-improvement.

The last scenario is a boring one, in which there is some fundamental unknown law of nature that halts the development of artificial intelligence systems. As I mentioned in "What is AI?", artificial intelligence research has already been through a "winter" where little progress in the field was made. Back then, this was due to the fact that we did not have the necessary computational infrastructure that allowed us to scale these systems to what they've become today. Perhaps the amount of energy these models need for continued development outstrips what we can supply, halting continued development. Or perhaps there are some unknown barriers that we cannot fathom today. It should be noted that there are prominent researchers that believe that the current direction in which the field is heading is a dead end. They foresee that these types of systems will never be able to achieve true intelligence, and have begun researching new directions. Perhaps it is true, and perhaps it will take another hundred years to find that new direction which is able to reach superintelligence.

Out of these five scenarios, the entire world is gunning for the second one, in which we reach a superintelligent, benevolent system, leading to a future filled with abundance. This entails reaching recursive self-improvement, and the frontier companies are therefore spending more money than they can make to try to reach that point. These efforts are feeding the financial "AI bubble" - the potential is too huge to pop it. The frontier labs are attempting to make the models as good as they possibly can be at mathematics and computer programming, so that they can reach the point at which the models code themselves. Models are gradually becoming better and better at these tasks, and in the last six months, they have reached capabilities that are astounding; however, we're not at the level where humans can be kept completely outside the loop. We have reached a precursor to recursive self-improvement, in which engineers utilize the models to rapidly improve the model, but we're not yet at the stage in which the model improves itself autonomously. One of the main problems when it comes to assessing where we're currently positioned on the path towards superintelligence is the tremendous monetary incentive which makes us not believe what the frontier research companies are telling us. Since 2022, these companies have employed scaremongering as a tactic to brag about their models, and they've cried wolf far too many times. Nevertheless, if I were to bet, I would put my money on us reaching the singularity.

You might think that believing that we all will eventually be outsmarted by an artificial system would give one a sense of purposelessness; "why should I spend my life doing research when an AI will end up solving it all in the end anyway?". And it is true, many researchers and engineers feel precisely this sort of angst. For some, this makes them reconsider their career choice to do something removed from science and engineering, like art. Others start working on developing artificial systems, with the idea that it's the only meaningful job one can have nowadays; mobilizing all forces onto the singular goal of reaching superintelligence. Another path entails working on AI safety, like the recent Fields medalist Jacob Tsimerman, who right after winning the medal - which is the mathematics equivalent of the Nobel - began working on AI safety, abandoning his prior work in mathematics. Personally, I have come to terms with the fact that I do research because I enjoy it. I would not want to work on anything else even if superintelligence was here. Even if my work was to be completely devoid of meaning to the outside world, I enjoy the puzzles and the sense of awe that thinking about mathematics and biology instills in me. It is meaningful to me, for purely recreational purposes. There is also always the probability that we won't reach superintelligence, in which my work will turn out to have a greater purpose after all.