But a quick look at important discoveries teaches us that costly mistakes abound. Fleming’s failed experiments led to penicillin, Edisons failed document copier Electric Pen was embraced by the tattoo industry, and after Henry Fords first company for handmade cars went bankrupt, he tried again. Taking inspiration from the meat industry he invented the assembly line, and so forged success out of a costly mistake.

Serendipity and chance: the stuff creativity is made of. Invention forged by failure.

In 1996, IBMs Deep Blue lost to the world champion chess grandmaster Gary Kasparov.

The history of the chess AI that took on Kasparov is interesting because it reveals the sharp split between the two main schools of thought within Artificial Intelligence. Deep Blue was an expert system rather than an LLM, the Large Language Models that underpin ChatGPT and other chatbots you may use daily.

Deep Blue was programmed to play chess. The main proponent, the Taiwanese American Feng-hsiung Hsu, had started working on the system during his doctoral at the Carnagie Mellon University. He later joined IBM, that funded his dream, the chess-playing supercomputer which cost an estimated $10 million.

As an expert system, Deep Blue was an example of rule-based symbolic AI, which was prominent from the late forties to early 21st century. But progress was uneven, with so-called “AI Winters” marking dips in funding.

Expert systems, as the name suggests, were to perform the tasks of an expert: finance, health, government. They were very costly, needing meticulous programming for their specialist tasks.

So too Deep Blue. Its failure to beat Kasparov ushered in yet another AI winter. Rule-based symbolic AI's costly mistake caused a paradigm shift. Enter LLM-based AI.

A statistical approach rather than a rule-based one, LLM-based AI takes a large corpus of language as input, and “learns” based on patterns that are implicit in the data. Then, given a question, it answers by stringing words together in the most likely order. It compiles rules itself from analysing the massive dataset that is the internet.

That works, as we all know. Millions use ChatGPT, and we at our small software company use AI Agents to code 10 times faster than we could just a few months ago.

But this AI comes at a price. One is the obvious inefficiency of LLMs in their huge hunger: for rare earth minerals, and for evermore larger quantities of energy.

The other is widely regarded as a very big drawback but might prove crucial for the next paradigm shift. And that’s the tendency for LLMs to hallucinate. AI companies want to suppress this tendency to “make things up”. But aren’t they insulating themselves, and us, from serendipity and chance?

But it might be that hallucinations provide the window into the next big thing. If only we could find a way to apply them to a real world problem.

A costly mistake waiting to bring prosperity to us all.