With all the current levels of excitement about generative artificial intelligence (AI) in the form of chatbots like ChatGPT and its rivals, it is easy to forget that the scope of AI goes well beyond this particular strand of AI. The media is awash with stories about the large language models (LLMs) that underpin generative AI, but some of the more established AI use cases and case studies actually have nothing to do with chatbots or LLMs. Indeed, if you poke around into some of the better-known AI successes, it turns out that many of them are based on different technology entirely. Let us take a little journey into AI history and explore the varying flavours of AI, which is much more varied than many people realise.
Expert systems emerged in the 1960s and 1970s, and mimic the decision-making of a human expert in a narrow field by using a knowledge base in that field, plus an inference engine. The knowledge base is built of rules defined by human experts. For example, an expert system about medical diagnosis would be built from rules defined by human doctors. An early example was MYCIN at Stanford University, which diagnosed bacterial infections and recommended treatments. These have a wide range of applications in fields such as investment decisions, equipment troubleshooting, crop management and more. These systems are still in use today, such as in the US Navy’s Aegis Combat System, which analyses data from radar and sonar to identify threats. Expert systems are explainable because they follow precise rules and are consistent. They can struggle to scale to very large, unstructured datasets, and have mostly given way to machine learning and neural networks, of which more anon. Symbolic AI is a related approach that relies on explicit representations of knowledge. Expert systems can be regarded as a subset of symbolic AI.
Fuzzy logic is a technique which deals with situations where information is vague or uncertain, which is a common occurrence in the real world. In Boolean logic things are true or false, but in fuzzy logic data is represented in a continuum of values. It is often used in control systems such as washing machines or aeroplane altitude control or air conditioners. It is easy to implement in code compared to many AI techniques, though it does not adapt, and can struggle to scale in complex problems with many variables.
Machine learning is a form of AI where models learn from data and improve performance without being explicitly instructed. Machine learning algorithms find patterns in data and rules that allow the system to make predictions. As the system sees more and more data, it gets better and improves. The system can be trained on either labelled data or unlabelled data. Labelled data might be data where a human has labelled a set of emails as either being spam or not spam, or a series of images which have been labelled to show each image as either including an object, like pictures with either a cat or no cat. Unsupervised learning occurs when the algorithm is presented with unlabelled data and tries to detect patterns on its own. There is also reinforcement learning, which works by trial and error. The system tries different things, and those that work are scored positively while those that don’t are scored negatively. It then repeats this process many times.
For example, the chess-playing program AlphaZero from Deep Mind was simply taught the rules of chess. It then played several million games against itself. At first, its moves were almost random, but after many iterations, the standard of play slowly improved, then improved more. After four hours of such training its play reached its zenith, and it was pitted against Stockfish, the strongest chess program in the world, which was based on classical AI using defined rules to evaluate positions. In one hundred games, AlphaZero won 72 games, drew 28 and did not lose a single game. Machine learning is good at spotting trends and patterns in large volumes of data, automating repetitive tasks and dealing with complex data. It is heavily dependent on the quality of the data is trained on, so data that has bias will lead to biased predictions. For example, if a machine learning algorithm is trained on images of people skewed to a particular racial type, then it will reflect that bias. This is obviously a problem when machine learning is used for job screening or similar applications. A machine learning system called PREPOL was used in the USA to predict crime, but repeatedly targeted police to areas with a high proportion of racial minorities, regardless of the actual crime rate in those areas.
Genetic algorithms are used for optimisation, and work by iteratively refining a group of candidate solutions. The idea is to mirror natural selection, as described by Charles Darwin. This technique can be used in situations where there are complex, non-linear functions, such as is designing engineering control systems, or for fraud detection or drug discovery. There are limitations, such as sensitivity to the parameters chosen and the fact that they tend to be computationally expensive to operate. This approach can struggle to scale as the number of variables increases, or may take a very long time to reach an optimal solution.
Neural networks are a technique modelled on the way that the human brain (probably) works. It has layers of connected units called “neurons” that process input data to produce an output. Each neuron has a set of weights applied to it, and uses these to process its input data, passing the result to another neuron. This, in turn, does the same, passing its result to the next layer of neurons. After perhaps dozens of such steps, the neural net has a result. The weightings are adjusted during training to minimise prediction errors. For example, a neural net might be trained on a file of customer sales data history. The neural network can be asked to predict the likelihood of a customer, or a particular grouping of customers (such as females or males, or people in an age range) buying again based on their buying history.
Neural networks can model complex data relationships, adapt over time when shown new data, can handle “noisy” data and can scale well, as their computations can easily be parallelised. On the downside, they typically require lots of training data and are black boxes, so they cannot explain their reasoning. They are sensitive to the quality of their training data, and can reflect bias in their training data in their own predictions. Neural networks are versatile and are used in applications ranging from medical imaging to self-driving cars to facial recognition and natural language processing. They are used in fraud detection, and in personalised recommendation systems such as those used by Netflix and Amazon. Neural networks are used within the most fashionable of all AI styles at the moment, generative AI.
Generative AI is specifically designed to generate new content, such as text or program code or images. It uses a series of neural networks to synthesise outputs that are based on their training data but are nonetheless original. They can create artwork, write essays, write marketing emails, or produce a poem or a marketing plan, and even help in drug discovery. Generative AI is embodied by popular chat applications like ChatGPT, Claude, Perplexity and Gemini, and in visual equivalents such as Leonardo, Midjourney, Stable Diffusion and Firefly. They can produce fluent text but can produce invented or nonsensical answers, a phenomenon known as “hallucination”. They will reflect any biases in their training data, and rely on huge amounts of training data and processing power.
Even this lengthy set of AI styles does not exhaust the list of approaches. Swarm intelligence uses simple agents that interact with each other based on rules and local perception, mimicking the behaviours of birds flying in coordinated flocks. Applications include robotics, network routing and traffic management, as well as anomaly detection. This technique has limitations of scale and can behave unexpectedly, as well as being hard to debug.
A brand-new approach was recently unveiled by a company in Singapore. The hierarchical reasoning model has two modules. A “planner” thinks slowly and strategically, while “workers” execute rapid calculations for specific tasks. Intriguingly, this approach avoids the need for vast amounts of training data and was relatively small at 27 million parameters. By contrast, Claude 3.5 has 175 billion parameters, and ChatGPT 4 has about 1.8 trillion parameters. Despite the vast difference in model size, the HRM model managed a David v Goliath style performance in benchmarks, outperforming both Claude and ChatGPT on some complex reasoning tasks like the ARC-AGI (Abstraction and Reasoning Corpus for Artificial General Intelligence) benchmark. This is early days and will need further testing, but it is an approach that seems to show promise. There are yet more AI approaches around, including diffusion models, which have been mostly used in image-generating AIs so far.
It can be seen that AI goes far beyond the large language model paradigm that has captured the world’s attention in recent times. This is important because LLMs have distinct limitations, such as their propensity to hallucinate, and are poorly suited to many tasks. LLMs are not good for cases that require consistency and precision, or complex multi-step logical reasoning, or things that require transparency in their decision-making. Many of the more embarrassing AI failures, and there is a growing list, have been where LLMs have been applied to situations that do not suit them. The old saying “if all you have is a hammer, everything looks like a nail” applies very well here, with LLMs being applied to many tasks where another approach would be better. A trivial example of this can be seen if you ask an LLM to play a game of chess. Computers have been very good indeed at chess for decades, an IBM program called Deep Blue beating the chess world champion Garry Kasparov in 1997. These days, chess engine apps that work on the phone in your pocket would trounce the current human chess world champion. Yet LLMs can barely make it through a few moves of a chess game before hallucinating extra pieces onto the board or playing illegal moves, never mind playing moves that make strategic sense. As of August 2025, LLMs are bad at chess: laughably bad. Watching two LLMs play against each other is like watching two spaniels trying to play chess. The example of chess vividly shows that different flavours of AI can be dramatically more successful than LLMs at certain types of tasks. People implementing AI programs need to be aware of the different tools at their disposal and be ready to apply the right tool to the right job. Of course, for many tasks, good old-fashioned programming is still the way to go. AI can be a useful tool, but it is not a panacea.
