Through a glass, darkly

21 August 2026

Uncertainty is a characteristic of existence. We yearn to see clearly but a complete view is always just beyond our reach. Accepting this is hard.

Artificial intelligence pitches us into a world of greater uncertainty. When a new technology comes along we compare it to what we have seen before to orient ourselves. We use the past in an attempt to navigate the future but it is an imperfect teacher. One comparison is with the internet, electricity another. Aspects of everyone’s lives were touched by both but the effects were asymmetrical, some industries were changed more than others.

The defining factor was not the technology itself but the time it took for people to understand how to use it. Electricity did have a transformational effect on industry but this took years to filter through. A new generation of factory managers saw that electric motors could be miniaturised in a way steam engines could not. This insight led to the reorganisation of factories, allowing materials to flow more efficiently through the production process.

Another question is whether artificial intelligence exists or not. It seems odd to ask given that we are bombarded with stories about its progress every day. Extremely impressive strides have been made in recent years, especially in writing software. Some aspects of intelligence have already been captured and replicated. Others are still missing.

A defining characteristic of natural intelligence is the ability to learn from experience. Animals are capable of interacting with each other and their environment, remembering positive and negative experiences, learning over time and adapting their behaviour. The current generation of artificial intelligence does not learn in this way. The reality of this limitation dawns on everyone who uses it regularly. So we are in the strange situation of uncertainty about the effects of artificial intelligence and whether it exists at all, depending on the definition.

Existing models use vast amounts of training data and computing power with each release. Their progress is staccato and inefficient when compared to natural intelligence which is capable of continuously generating greater insights from far less raw information. Although spending enormous resources on large language models is currently the dominant approach, it is not the whole of artificial intelligence research. For example, past iterations of chess playing programs mastered the game by learning from experience but this was in a bounded environment with fixed rules. This approach has not been perfected in the type of unbounded environment where natural intelligence flourishes.

Some people think that adding more and more resources will eventually see a breakthrough that leads to these models finally replicating natural intelligence in all its finery. Our experience so far has shown that these resource driven advances lead to increasingly impressive discrete achievements as models get bigger and add more capabilities. This is frequently mistaken for the crossing of some invisible boundary to general intelligence. Advocates misinterpret advances in one specific field as evidence that the machine now possesses similar capabilities across fields. It is not surprising this happens as we do the same with people – mistaking talent in one area as evidence of capability across many dimensions of knowledge and wisdom. A flaw of humanity rather than technology.

Another possibility is that large language models never achieve this type of general intelligence because the scaling approach currently being pursued with such fervence is not capable of producing it. Useful progress can still be made, even if this path does not lead to the stated goal. Instead of developing general intelligence, further progress could lead to models that are much more efficient to train and operate. As understanding and techniques improve, in the future we could see smaller models consuming far fewer resources while exhibiting greater capabilities than the frontier models we have today. The development of large language models would then resemble that of microprocessors in which each generation got more powerful at a lower cost. Progress may be more rapid because microprocessors, being physical, require continual refinements in manufacturing to advance whereas models can be replicated easily. This alternative perspective characterises models not as complete artificial intelligence systems in themselves but components within a larger system. The emergence of cheap and widely available microprocessors was the catalyst for an explosion of innovation, allowing users to build their own home computers for the first time. Mainframes still existed after this point but they were neither what most people used nor what anyone first thought of when they used the word computer.

The larger question transcends technology. Ultimately, artificial intelligence is just a tool, no different from the earliest technologies any human used. There is a great temptation to fall in love with what human hands have made. For the architects of these models to see more in their creations than actually exists, losing sight of reality. The greater their ability, the more compelling the illusion.

The antidote is humility, recognising our limitations in the face of not knowing. The nature of research is a journey into the unknown, it is what makes this work fascinating. Imagining the future is part of research work and uncertainty does not stop us from making progress. It should not paralyse us but instead foster humility.

For now we see through a glass, darkly.