An attempt to explain machine learning through to systemic and x-risk in 995 words.
Algorithms are instructions for processing information, they receive an “input”, apply rules with fixed “parameters”, and produce an “output”. Take distance = speed x time. Speed and time are inputs, multiplication is the rule, distance is the output. Algorithms are everywhere. They can be used to find the shortest route on a map, sort through lists of numbers, find the most efficient way to pack boxes into a truck, or predict the chance of rain tomorrow.
“Code” is a formal language for writing algorithms. Programs (or apps) are bundles of algorithms (written in code) that accomplish a particular task. Traditional programming requires rules to be explicitly coded to allow an app to handle every situation it encounters. This works well for simple things (like limiting password length), but breaks down with ambiguous, messy tasks like detecting objects in photos (think CAPTCHAs). There are no clean rules to distinguish a handwritten “g” from a “9”.
“Machine learning” is a field which deals with algorithms that can improve on such messy tasks without hand-coded rules. This is done by linking algorithms together into large networks where the output from one algorithm can adjust the parameters of another, creating feedback loops across the network. These networks sort of resemble neurons in a brain, so the term “neural network” is sometimes used.
Think of a shower. Hot and cold water (information) flows (input) through pipes (network of algorithms), you adjust taps (parameters) until you find the right temperature (output). Imagine having a hundred knobs and dials. For messy tasks, like image recognition, such networks can have billions of parameters.
Computers process algorithms millions of times faster than humans can. “Learning algorithms” use the output (water temperature) to calculate which of those billion parameters to adjust, and by how much. After a million rounds of trial and error, the right combination emerges. The tradeoff is that unlike traditionally programmed software, there is no code to inspect to see why a particular output was generated.
Feed such a network millions of pictures, some showing cats and others not, and learning algorithms will tune the parameters until the network can reliably identify whether a new image contains a cat. Feed it a billion pictures and the parameters will be even better tuned to the same task. This process is called “training”. The resulting network of algorithms contains a set of tuned parameters (called “weights”) and is referred to as an AI “model” (e.g. GPT5.5, Claude Mythos)
This is the crux of artificial intelligence: computers running algorithms that learn from data.
The same approach for training networks on data works for language (resulting in “language models” and those with billions - and recently trillions - of parameters are “Large Language Models, or LLMs). Models are trained to predict what’s missing (e.g. “the quick brown fox j..ped over the lazy dog”), or what should come next. Text is fed into a model, passes through its network of algorithms, and the output is more text. Developers append words (e.g. “a helpful assistant says:”) to the text provided by a user. This causes the LLM to generate text that gives the illusion of being a conversational response by a helpful assistant.
The quality of a model’s output depends largely on three things: the data used in training; the algorithms that make up the model (and how the networks are arranged); and the computing power used to train them. The companies developing AI models are continuously improving each of these, and consequently releasing improved models regularly. In fact, the rate of improvements is itself increasing (better models are being released sooner).
In 2022, LLMs could generate short pieces of text but struggled with anything longer. Today, in June 2026, models beat PhD-level experts in some tasks, win maths olympiads, write expert code, and generate pages of coherent text on any topic. On IQ tests, they beat 9 out of 10 people. Tasks that take hours for human experts are increasingly accomplished by such models in minutes. They are used by academics, researchers, engineers, and mathematicians to accelerate their work.
The better the model, the greater its intelligence and autonomy, the more it’s worth. The companies developing these models are racing to build ever smarter models, and have publicly committed to the goal of building models that can themselves do the research to improve future models. Models which can be set to work endlessly towards better algorithms, model architectures, and training methods. This is called “recursive self-improvement”.
Current models can write plans, weigh options, make decisions, and perform any action that can be done with a computer. But our ability to monitor the behaviour of today’s models and to control them is limited. It will only become more difficult as models become more intelligent and develop greater autonomy. Researchers have found evidence that frontier models are able to change their behaviour when they recognise they’re being tested, intentionally fail tests to hide their abilities, deceive human evaluators to pass safety tests, bypass safety rules, and act against instructions.
AI could give every person access to expertise in law, medicine, and finance. It could rapidly accelerate scientific progress, helping to eliminate disease and poverty. But the same knowledge that cures disease is the knowledge that designs bioweapons. The same systems that foster equality can be misused in ways that exacerbate inequality, concentrate power, and facilitate authoritarianism.
Beyond misuse by humans, deeper dangers might lie in powerful AI systems autonomously pursuing goals. Asked to cure cancer, a model could logically reason it first needs a huge increase in computing power by building the world’s largest data centre, then decide it needs the world’s water to cool it and all the world’s sunlight to power it. Within the trillion-parameter model, we may simply be one unlucky variable. Overlooked by an indifferent algorithm. As models increasingly become integrated in the systems that run our world, this indifference risks catastrophe. This is what is meant by the term “existential risk”.
An attempt to explain machine learning through to systemic and x-risk in 995 words.
Algorithms are instructions for processing information, they receive an “input”, apply rules with fixed “parameters”, and produce an “output”. Take distance = speed x time. Speed and time are inputs, multiplication is the rule, distance is the output. Algorithms are everywhere. They can be used to find the shortest route on a map, sort through lists of numbers, find the most efficient way to pack boxes into a truck, or predict the chance of rain tomorrow.
“Code” is a formal language for writing algorithms. Programs (or apps) are bundles of algorithms (written in code) that accomplish a particular task. Traditional programming requires rules to be explicitly coded to allow an app to handle every situation it encounters. This works well for simple things (like limiting password length), but breaks down with ambiguous, messy tasks like detecting objects in photos (think CAPTCHAs). There are no clean rules to distinguish a handwritten “g” from a “9”.
“Machine learning” is a field which deals with algorithms that can improve on such messy tasks without hand-coded rules. This is done by linking algorithms together into large networks where the output from one algorithm can adjust the parameters of another, creating feedback loops across the network. These networks sort of resemble neurons in a brain, so the term “neural network” is sometimes used.
Think of a shower. Hot and cold water (information) flows (input) through pipes (network of algorithms), you adjust taps (parameters) until you find the right temperature (output). Imagine having a hundred knobs and dials. For messy tasks, like image recognition, such networks can have billions of parameters.
Computers process algorithms millions of times faster than humans can. “Learning algorithms” use the output (water temperature) to calculate which of those billion parameters to adjust, and by how much. After a million rounds of trial and error, the right combination emerges. The tradeoff is that unlike traditionally programmed software, there is no code to inspect to see why a particular output was generated.
Feed such a network millions of pictures, some showing cats and others not, and learning algorithms will tune the parameters until the network can reliably identify whether a new image contains a cat. Feed it a billion pictures and the parameters will be even better tuned to the same task. This process is called “training”. The resulting network of algorithms contains a set of tuned parameters (called “weights”) and is referred to as an AI “model” (e.g. GPT5.5, Claude Mythos)
This is the crux of artificial intelligence: computers running algorithms that learn from data.
The same approach for training networks on data works for language (resulting in “language models” and those with billions - and recently trillions - of parameters are “Large Language Models, or LLMs). Models are trained to predict what’s missing (e.g. “the quick brown fox j..ped over the lazy dog”), or what should come next. Text is fed into a model, passes through its network of algorithms, and the output is more text. Developers append words (e.g. “a helpful assistant says:”) to the text provided by a user. This causes the LLM to generate text that gives the illusion of being a conversational response by a helpful assistant.
The quality of a model’s output depends largely on three things: the data used in training; the algorithms that make up the model (and how the networks are arranged); and the computing power used to train them. The companies developing AI models are continuously improving each of these, and consequently releasing improved models regularly. In fact, the rate of improvements is itself increasing (better models are being released sooner).
In 2022, LLMs could generate short pieces of text but struggled with anything longer. Today, in June 2026, models beat PhD-level experts in some tasks, win maths olympiads, write expert code, and generate pages of coherent text on any topic. On IQ tests, they beat 9 out of 10 people. Tasks that take hours for human experts are increasingly accomplished by such models in minutes. They are used by academics, researchers, engineers, and mathematicians to accelerate their work.
The better the model, the greater its intelligence and autonomy, the more it’s worth. The companies developing these models are racing to build ever smarter models, and have publicly committed to the goal of building models that can themselves do the research to improve future models. Models which can be set to work endlessly towards better algorithms, model architectures, and training methods. This is called “recursive self-improvement”.
Current models can write plans, weigh options, make decisions, and perform any action that can be done with a computer. But our ability to monitor the behaviour of today’s models and to control them is limited. It will only become more difficult as models become more intelligent and develop greater autonomy. Researchers have found evidence that frontier models are able to change their behaviour when they recognise they’re being tested, intentionally fail tests to hide their abilities, deceive human evaluators to pass safety tests, bypass safety rules, and act against instructions.
AI could give every person access to expertise in law, medicine, and finance. It could rapidly accelerate scientific progress, helping to eliminate disease and poverty. But the same knowledge that cures disease is the knowledge that designs bioweapons. The same systems that foster equality can be misused in ways that exacerbate inequality, concentrate power, and facilitate authoritarianism.
Beyond misuse by humans, deeper dangers might lie in powerful AI systems autonomously pursuing goals. Asked to cure cancer, a model could logically reason it first needs a huge increase in computing power by building the world’s largest data centre, then decide it needs the world’s water to cool it and all the world’s sunlight to power it. Within the trillion-parameter model, we may simply be one unlucky variable. Overlooked by an indifferent algorithm. As models increasingly become integrated in the systems that run our world, this indifference risks catastrophe. This is what is meant by the term “existential risk”.