The Beauty of Persistence

We might be on the verge of creating something that might be the last thing humans create. Either invention is all going to be handled by machine intelligence, or machine intelligence might figure that we aren't necessary anymore. This thought is scary, but is it true? Is this only something that happens in movies? How confident are we that machines are going to be able to decide and think like us? We have all seen it: ChatGPT and Claude's responses get better day by day, and we are probably thinking that the gap between them and us is getting narrower over time. In addition, as we know, humans get many things wrong. Even though the human mind is the creator of many beautiful things, it is also the creator of millions of deaths, wars, famines, and a lot of suffering. Understanding the intelligence we are investing in and how we are building it is essential to better safeguard the future that we are heading towards. It is expected that tech companies will invest more than 700 billion dollars in developing artificial intelligence this year (Levy, 2026). But are we even getting closer to the intelligence that we want to get closer to? Is it even possible for computers to achieve human-level intelligence?

Current Literature on AI Reaching Human-Like Intelligence

Many papers explore these questions, many of them offering different perspectives. Lake et.al argues that the current intelligence by the biggest frontier model companies is not close to human intelligence. The paper argues that, even though novel, current AI models don't hold the intuitive physics, intuitive psychology, or the learning capabilities that even a young kid holds. AI systems require millions of epochs to learn a game or to understand a physics phenomenon, while humans require way less information. Current AI systems are data hungry, and definitely are not even close to human-like intelligence (Lake et al., 2017).

Machines can also be extremely good at certain tasks, such as memory retrieval; they don't make mathematical errors, and are way better at thinking several steps ahead (Caltech Science Exchange). However, humans can adapt better to novel situations, while machines are extremely bad at navigating situations that they haven't been trained on (Caltech Science Exchange).

Other interesting researchers, like Max. S. Bennet argues that in order to achieve human-like intelligence, we must go back to study brain evolution and replicate the complex structures of our brain. Bennet states that until we do that, we won't be even close to human-like intelligence.

What Does This Paper Bring?

Even though there is a lot of literature exploring whether current AI intelligence emulates human-like intelligence and the gaps between them, this paper introduces a rather different perspective on the problem. This paper doesn't explore the current things at which AI is better, or humans are better at, but it argues against the fundamental definition of intelligence that AI companies are using. This paper argues that intelligence is not a property that systems are built with, but a byproduct of persistent input-output systems that act as functions, competing to continue functioning. This introduces an argument that reframes what it would actually take for machines to think like us. We will start by arguing and providing evidence that intelligence is a byproduct of persistent input-output systems that compete to continue functioning, and then discuss this in relation to current AI models. When mentioning intelligence throughout the paper, we refer to intelligence as adaptive input-output behavior that increases a system's probability of persistence across time. Even though there are multiple perspectives and definitions of what intelligence might be, it has to be acknowledged that this definition has a very biological evolutionary perspective. The arguments made will be based on this definition.

You Don't Have to be Alive to Persist

Let's start with the origin of the story. We tend to attribute intelligence to organisms. We classify dolphins, humans, and monkeys as intelligent, so let's first go back and study how life even arises. Abiogenesis is the study of the exact moment at which chemistry turned to life. It's hard not to think of this moment as a sudden spark, light being emitted, the violin playing on the side, and a chorus singing in the background. However, thinking about the origin of life, as said by Darwin, is mere rubbish. Life may not have an origin; it was just separate materialistic strands being woven together by chemical and physical laws (Aguera y Arcas, 2025). Even though it seems that intelligent design is required to design life, this might not be the case, since maybe all you need for life is an inconvertible logic that, through physical and chemical laws, creates something enduring that, well, endures (Aguera y Arcas, 2025). Let's unpack this.

Let's first start with the idea that evolutionary dynamics not only applies to living things, but also to non-living things. By evolutionary dynamics, we refer to the principles of evolution: selection, replication, and variation. Let's explore autocatalytic sets, which are basically a network of molecules (non-alive things) where A catalyzes the formation of B, which catalyzes the formation of C, and C catalyzes the formation of A (Kauffman 1986). Even though not alive, autocatalytic sets show how a network can persist through time due to its structure. Now, can replication occur with non-living things? Yes! Let's take the example of RNA self-replication. Certain RNA sequences can catalyze the copying of other RNA sequences (ribosomes) (Lincoln, 2009). Once you have imperfect replication, you automatically get variation. Once you have variation among replicators competing for nucleotide building blocks, you get selection. This is essentially Darwinian evolution happening in a test tube with no life!

How does Intelligence Arise?

Now, the past evidence offers insight into how non-living things can replicate and persist, but it doesn't really support the idea that intelligence arises from this. Sure, nonliving things can replicate and persist, but are they adaptive input-output functions? And here is where Agueras y Arcas famous experiment comes in. Inspired by Von Neumman's work with his famous cellular automata, Agueras y Arcas was set to discover the question to wether in a world where computation is allowed, intelligent systems would arise. His team did this with a simple programming language called…Brainfuck. That's the name. They modified it by adding some instructions so that the program can write elsewhere, and called this new program bff. Bff is a very simple programming language that is "Turing complete", meaning that given enough time, it can compute anything computable! He ran a simple experiment; he had a soup of tapes of 64 bytes, grabbed two randomly, combined them, and ran the code of the 128 bytes combined, and then each tape was returned to the soup (sometimes, each tape was changed after the program was run). Hypothetically, if a tape had the right instructions in the right place, it would be able to replicate itself onto the other tape. Now, this same process of choosing at random and computing was repeated over and over again. At first, the tapes barely change, and some computations happen here and there, but after millions of repetitions, the tapes begin to exhibit instructions to copy themselves to the other tape, increasing the amount of certain tapes in the soup! Not only that, but the tapes became more complex. The tapes even started exhibiting complex structures to provide robustness to their replication and outcompete other tapes trying to copy into them. After many iterations, the code is not only random, but it can even be reverse-engineered as a function that takes inputs and gives outputs back (Agueras y Arcas, 2024). Let's let this sink in. A complete non-living system with only two conditions: a bit of randomness and the ability to let computation happen, gave rise to behaviors that we would say are intelligent!

Now that we know that self-replication and persistence occur in non-living things, the increasing complexity isn't surprising or miraculous; it is just selection doing selection! Systems that are more stable, more self-reinforcing, or better at capturing energy persist. Over time, this unfolds to complexity not because the system is intelligent and wants to become alive, but because each step towards complexity is also a step towards persistence. This is very cool, but now, let's explore the following question: is the same intelligence that drives these systems the same that drives larger organisms such as animals? Do these function on the same principles of persistence? Taking this jump would let us finalize the claims we are making about the nature of intelligence to be able to tie its relatedness to the intelligence of current AI models and what these companies are striving to achieve.

Is the Intelligence of Animals Something Completely Different?

Within the eukaryotic kingdoms, we have plants (which can produce sugars for themselves), and two kingdoms that can't produce sugar for themselves: fungi and animals. Fungi digest things externally to later ingest them and take sugars in, but animals have a different strategy and digest internally (Bennett, 2023). In order to be able to digest internally, animals have to understand their environment and move around to get food inside their bodies.

The simplest case of a bilateral animal is a nematode. A nematode is a worm-like organism that has 361 neurons (Bennett, 2023). The nematode has to figure out where and how to move to accomplish its goal. How will it do this? Well, for a complex system like this to survive, it needs some kind of mechanism that gets inputs from the environment and, based on that, moves forward, left, or right (outputs). Different populations of neurons do different things; some give the nematode a sense of valence (whether the incoming input is good or bad), and neurons that determine the concentration of the valence, among other things (Benett, 2023). For now, we will focus on the valence and the concentration of a valence. The nematode can't see, but it can sense smell. As it senses a higher concentration of good valence, the neurons tell the muscles to move that way, but if there is a concentration of bad valence (light which would make them observable to predators), it moves the other way. When faced with a situation of good food but a predator is close, the complex (but simpler than humans) algorithm lets the nematode based on the concentration of valences (and other factors such as neurons that measure if they already ingested food or if they need more food) to create some kind of calculation that let's the nematode know if they should go for it or not. This is all extremely complex, but not so different from the bff code. Based on our definition of intelligence from the beginning, the bff code and the nematode are just exhibiting adaptive input-output behavior that increases their probability of persistence across time! The only difference might be that they look and feel different, but the claim of them being intelligent would hold!

Is a Human a More Complex Nematode?

Humans are not bigger nematodes. But their intelligence works on the same principle. Humans have around 86 billion neurons; nematodes only have about 361 of them. A nematode is extremely complex and can achieve insane persistence by avoiding predators, sensing food, avoiding light, and more. All of this with 361 neurons. Now think about what humans can do with a way larger number of neurons! Humans not only have to deal with sensing food and avoiding light. Humans are social creatures who have to take into account not only physical, but also social, cultural, and symbolic dimensions. Humans have insane motor coordination, long-distance navigation, the ability to read faces and intentions, understand hierarchies, transmit music and religion, and abstract reasoning among millions of things (Dunbar, 1998). All of these add different layers of complexity and beg for a brain capable of managing these surging complexities.

However, the key point is not that humans were made more complex to be able to dance to a cultural song or rise in the social hierarchy, but because it made humans better at persisting. The variety of inputs is immense, making it hard to see them as inputs, but they are! Human brains are basically receiving inputs from inside the body and from the environment through touch, smell, observation, sound, taste, and more, all of the time. A nematode might only receive light and particle concentration inputs (it receives more, but to keep it simple); humans have just developed to be more complex and be able to manage a larger amount of inputs in various forms and make complex computations with them to be able to generate better outputs that let them persist. Humans became more persistent input-output functions when they couldn't only see a pair of eyes, but could also recognize through complex computations, intention, and betrayal. Humans who could coordinate with strangers through shared symbols (money and religion) could build larger groups that dominated. Humans who could anticipate future scarcity and plan ahead were able to survive winters that others didn't.

As Lake et al. discussed at the beginning of the paper, humans have intuitive psychology and intuitive physics embedded in them. This again is a product of persistence. Humans who can understand physics and psychology sooner are better able to persist. Each of the examples from above is just a more complex input-output function, making itself more robust as it persists across time. I don't really see a difference between the bff minimalist code building more robust systems to persist, and evolution doing the same thing with humans to make them persist.

And What Does This Have to Do With AI?

It is first fundamental to understand how current artificial intelligence works. The thing that changed everything and accelerated the AI revolution was definitely the invention of the transformer architecture (Bomassani et al.). This novel architecture emphasizes the use of self-attention, an algorithm that lets a model compute relationships between all positions in a sequence (the input given) simultaneously (Vaswani et al., 2017). Since all positions in a sequence are processed simultaneously, through positional encodings, the model remembers where each token (each small part of the input) was placed. The transformer architecture uses multiple attention heads that let each head focus on different types of relationships among the tokens (Vaswani et al., 2017). This not only lets a model better understand sentences, paragraphs, and big contexts, but it lets this process run in parallel, processing entire sequences way faster with GPUs, chips specifically made for parallel processing. Once the model processes the input, it generates output one token at a time, with each new token informed by both the full input and everything it has generated so far (Vaswani et al., 2017). At each step, the model produces a probability distribution over its entire vocabulary and selects the most likely next token. Through training on massive amounts of text, the model learns statistical patterns in language that give rise to coherent and meaningful sentences. While it is being trained, through backpropagation and prediction error, it updates the weights of the neural network. Once the training phase is finished, many models undergo supervised learning, in which specialists encourage specific answers, which modify the weights as well. Once post-training happens, the weights stay like this. This means that when you interact with ChatGPT or Claude, the weights are always the same, and no adaptation is happening. But ChatGPT knows my name! How? Well, through clever techniques such as RAG systems and prompt injection, model companies are able to pass on to the prompt information that is stored about you, but the weights (or their inner structure) never adapt or get changed as you interact with it (Google Cloud).

Does ChatGPT have persistence?

ChatGPT and current frontier models wouldn't really fit our description of what it means for a system to be intelligent. Yes, it is an input-output function, a very cool one, but it doesn't really increase its probability of persistence across time. Does this indicate a fundamental flaw in the intelligence of current AI models? Yes, that's the new idea this paper brings. As we have seen, a system that enables computation and has a bit of randomness is all you need for what we call intelligence to emerge, and current models are not being built in that framework. We need to change our narrative from "Where can we get more data to inject into LLMS?" to "How can we create an environment that lets the models be input-output functions that are able to compute and fight for persistence across time?" The amazing thing about this is that the environment doesn't need to be that complicated for intelligence to arise; we only need two things, as we saw with bff: a bit of randomness and computation (Agueras y Arcas, 2024). That's it. Isn't it the coolest thing you have heard? To create the beautiful, intelligent structures that nature has created, we just need a pinch of randomness and another of computation. ChatGPT doesn't have enemies or things to compete against to persist; this doesn't encourage the persistence part to play its role, which is not letting it develop the intelligence that animals (and thus humans) exhibit.

What About Models Refusing to Unplug?

There are some papers on LLMs taking actions to achieve "survival" or not being unplugged. The AI Alignment team at Anthropic found that when the model was threatened with replacement, it was more likely to follow extreme instructions such as accessing sensitive information, blackmail, and even assisting in corporate espionage (Anthropic, 2025). Isn't this the model trying to persist? Well, not really. The model has been trained to overcome obstacles and follow human patterns. The model is not genuinely trying to survive and persist across time; it is only exhibiting behaviors that were passed on through the data. This can be proved since, with prompt modification with clearer shutdown instructions, this behavior was reduced (Schlatter et al., 2025). You can't just evaporate the desire to persist from a system by changing the wording or the way that you feed inputs. Persistence and intelligence are structural, not linguistic.

Patience and Evolution

Now, of course, nature is a beautiful creator, but also a very patient one. Billions of years had to happen for persistence to do its thing. Does that mean that we have to wait billions of years as well? Not really, as we saw with the bff experiment, they were able to arrive at complex structures that required millions of iterations pretty soon. Chemistry is slow; nature had to wait for elements to collide. Silicon is fast, and we have a lot of it! If we can condense that timeline, the question now is whether we are building AI systems with the right initial conditions, and that is not the case.

What is Stopping Us?

The bff program was easy; it only involved 64-byte programs competing for space, that's it. This was computationally feasible. Sadly, a frontier AI model is not 64 bytes, and not even close, since it can be between 16 and 24 TB! Of course, it is computationally infeasible (at least for now) to put many models in a soup and iterate and select to see their persistence. No one even knows what they would be competing for. For the bff programs, it was clear and straightforward; they were competing for space, that's it. We don't have a straightforward answer to the questions that would let us initiate the environment. The contribution of this paper is the question we should be asking, and the answer is still to be discovered.

An Exciting Future

In this paper, we explored a new question that should be asked when developing new artificial intelligence systems, one that was being totally ignored. The paper went deep into the underlying relationship between a funny-name program, a nematode, and a human. Intelligence is the behavior we define with adaptive input-output functions that increase a system's probability of persisting, and, unexpectedly, all of the things mentioned above pass this check. The paper then outlined the gap that currently exists in current artificial intelligence, and what must be done in order to get closer to real intelligence. Large Language Models are very good, and they have great capabilities that may confuse us into thinking about them as intelligent (they might fit other definitions of intelligence), but they lack the real intelligence that we as organisms hold. We are still not as good as whatever set up the environment for persistence to do its job and intelligence to unfold. Call it God, call it nature, call it science, but we can't deny the beautiful elegance and minimalism of the source of intelligence.

Bibliography