Precision + prediction = the other type of centaur

Are we all centaurs now?

‘Centaur’ is the term used to describe someone who works in tandem with AI. It is part of the hope that, at least for some jobs, AI will change, support and augment them, but will not replace them entirely.

The intense interest in AI may sound like the hoofbeats of centaurs in the distance. But the centaurs are already here: they arrived some time ago. Some of them are conspicuous: humans using AI assistants and agents to respond to calls, deal with correspondence, draft documents and make sense of data. But we also become centaurs many times in our daily lives: whenever we use our phones (which, for some of, means all the time), we are interacting with AI in some way. When we search, when we shop and when we spend time on social media, we are expressing our emerging centaur nature.

In the world of technology, the centaurs are easier to spot, but are similarly pervasive. We may think of the centaur technology archetype as the programmer using an AI assistant to help them optimise their code or write tests. But centaurs are present throughout the lifecycle and beyond: the operator using AI to detect signs of technology failure or cyber attack; the low coder or no coder building AI agents; or the cloud engineer relying on the management plane of their platform. As the centaurs evolve, part of our job will be to figure out what it is responsible to let the machine do, and what work is best done by humans. A successful centaur is a union of strengths.

I believe that thinking about a responsible division of labour reveals another form of centaur: one to which we have perhaps not been paying enough attention. This type of centaur is a hybrid of two software paradigms, rather than a hybrid of human and machine.

As I wrote a couple of weeks ago, the ability of Large Language Models (LLMs) to process language is seductive, and can easily lead us to believe that they can do anything (including things that they don’t actually do). Humans are configured to ascribe agency to things that talk to us, and we are used to human agency being broad and unbounded. It is perhaps inevitable, therefore, that we perceive LLMs as capable of doing anything: isn’t it just a case of telling the model what I want? This has led some commentators to suggest that we can use LLMs for low level and complex coding tasks such as running operating systems.

Before we give LLMs that job, though, we should remember that any successful system incorporating LLMs is a combination of data, training, models - and traditional programming. There are many things that traditional programming is really good at, particularly processing logical instructions quickly, cheaply and accurately. These simple attributes are what has made computing a world-changing global phenomenon. These are also things that, right now, LLMs are quite bad at: while their ability to handle language brings new realms of opportunity within our grasp, they are expensive, slow and inaccurate for processing basic logic and arithmetic.

This does not mean that we should regard traditional programming and LLMs as being in opposition. They are not rivals or threats to one another, and advocating for one need not involve denigrating the other. Just like human / AI centaurs, they bring different strengths to our solutions. Traditional programming is good at precision, but bad at prediction: there is a reason that machine learning has become the dominant mode of AI. But while AI models, including LLMs, are good at prediction, they are bad at precision: there is a reason that closing the last gaps of inaccuracy and error is so difficult. (Note that this idea of the software / software centaur applies even if we use AI to generate traditional code: that does not mean that the traditional code is not required, just that we have a different way of getting there.)

For technology architects, this means that we still need to do the work that makes our lives interesting: how to put technology together in interesting ways to make it do valuable work. We are in the centaur building business now.

(Views in this article are my own.)