Idiot? It’s not the 1900's. Nowadays we say savant syndrome.

Many will know the main character of the movie “Rain Man”, Raymond Babbit, brilliantly played by Dustin Hoffman. As the movie goes, he was taken from the secluded safety of a mental institution into the world by his arrogant brother, who decided he had struck gold after realising Raymond had superhuman abilities to do fantastically complex calculations.

A savant is a person especially gifted in one specific area, but who otherwise suffers from serious mental impairments.

So too the real life person that was the inspiration for the Raymond Babbit character, Tim Peek. He only learned to walk at four but could read at 16 months. Loved being in the local library, reading more than 12,000 books, scanning with his left eye while reading with his right. A pattern recognition genius, he would recite whole passages perfectly but often veered off into unprovoked unintelligibility.

He was utterly dependent on his parents, not even being able to button his shirt. Rain Man gave Peek national fame but it was only thanks to his father that he could show his talents to the world. Himself, he would much rather have stayed in his beloved library.

At our small software company, we’ve been experimenting with AI for some time, but it’s only recently become really useful. We can create new projects in a tenth of the time, maybe faster.

I use an agentic AI system that can run in a number of modes. The mode I use almost exclusively is one where I can review and accept or reject every change. I know there’s an auto-accept mode, but I like control.

We recently took on a project that was abandoned by a competitor. It’s a complex frontend that we are hooking up to an existing backend. Both need modifying.

Being so complex, we take great care to provide the AI Agent with comprehensive context, and we make sure to guide every step.

Just once I tried the auto-accept mode, and immediately My AI veered off into unwelcome directions. Complete sections rewritten. Took us days to repair. Without constant guidance, nothing is preventing it from generating random code.

Which brings me to the savant syndrome comparison.

My AI is a pattern recognition genius. It’s read all open source code produced since the beginning of unix time.

It has a phenomenal ability to remember solutions in any programming language. It can produce shell command syntax I’ve never heard of.

Yet it needs constant supervision. It won’t adhere to application context if constraints are removed. It can’t even button its own shirt.

My AI brings me amazing productivity. Our small company can do projects in a fraction of the time than they would have taken mere months ago.

But it's simultaneously incredibly smart and immensely stupid. It needs our continuous guidance.

We need new adjectives to describe this new tool we have invented. Maybe we should look to savant syndrome for inspiration.