How I Stayed Connected to AI

In the early 1980s, artificial intelligence was barely a topic. It belonged to research labs and to people most engineers regarded as mad scientists: intriguing ideas with no real use.

Marvin

In 1983, a few years before my career began, I joined a project called Marvin, named after the paranoid android of The Hitchhiker's Guide to the Galaxy. It was funded by the British Department of Health and Social Security. Its goal was very concrete: a speech recognition system connected to actuators, so that hospital patients with little or no mobility could control their environment by voice.

None of today's tools existed. We worked with the hardware of the time: a Motorola 68000 and analogue filter banks where today anyone would reach for an FFT.

Yet the backbone was already there. We had to turn speech into vectors, train the system on examples, and map what it recognised to meaning and action. The tools have changed beyond recognition since then. Those principles have not.

In the end, the DHSS never deployed the system. British Aerospace took it up instead, and Marvin found its way into the cockpits of fighter aircraft. For a project built on such modest means, that was a real success.

Never far away

I did not make a career in AI. Outside the research labs there was little activity in the field, and my path led elsewhere: software engineering, delivery, and later the CTO role. But what I learned on Marvin stayed with me.

I kept following the field, and above all those strange beasts known as artificial neurons. What fascinated me was how much could be achieved with such simple constructs. I studied the underlying theory on my own, as pet projects, for the pleasure of understanding how it worked.

The wake-up call

Truth be told, I missed some of the turning points. The 2017 paper Attention Is All You Need, which introduced the transformer architecture behind today's language models, flew completely below my radar.

The real wake-up call came, as it did for so many people, on the day ChatGPT went live in November 2022.

Fad or paradigm shift?

My first reaction was the reflex of someone who has seen this before. My career began in the spring of 1986, and since then I have witnessed, and often taken part in, a long series of "revolutions": the client-server wave of the 1990s, the web frenzy, the rush to cloud computing. All of them were overhyped. Some were short-lived; others stayed and reshaped the industry. Experience had taught me that the noise around a technology says very little about its staying power.

So the question I asked myself was a simple one. Was generative AI a fad, a lasting new technology, or a true paradigm shift?

Finding out for myself

I did what many others did: I subscribed to one of the providers and started evaluating this new creature for myself. Several things quickly became clear.

The first was that generative AI was overhyped, in much the same way the web had been in the dot-com frenzy of the late 1990s. Without a sensible approach, the promises being made would lead to significant disappointment.

The second seemed to contradict the first, but did not. Generative AI was transformative, and in more ways than the hype itself suggested. The dot-com years had taught the same lesson: the bubble burst, yet the web went on to change almost everything. It is the old pattern of overestimating a technology in the short term and underestimating it in the long term.

The third was closer to home. In my own field, software engineering, it was a game changer.

So I invested in it, time first, and with three questions in mind.

• What could now be done that had until recently been out of reach for most companies?
• What would it change in the way we build software?
• And above all, what remains once the hype is stripped away?

Where the answers are

The articles on this site are my attempt to answer those questions.

GenAI Beyond the Hype starts with the first two. It traces the history that led from the early days of AI to generative AI, untangles the contradictory messages coming from the industry, and looks at why some projects succeed while others fail. It ends with what I believe is the right way to approach it.

The question of how we build software brought me back to delivery itself, and to Scrum. Like generative AI, Scrum has been overhyped and widely misunderstood, and it has lost some of its appeal as a result. Before showing what AI changes in software projects, I want to explain again what Scrum really is. That is the purpose of the first two articles in the series, on Scrum and on Agile planning. The next ones add Domain-Driven Design, and then Claude and the BMAD method.

From there, two articles look at execution:

• How generative AI reshapes the way software projects are carried out, taking retrieval-augmented generation (RAG) as a typical example. Beyond changing how projects are executed, AI-assisted development has the potential to multiply the effectiveness of project teams.
• How Claude and BMAD can be used to build applications in regulated environments, where the hype meets the hardest constraints.