Aviation for AI

As someone who has spent the past six years in AI and the past 18 years in Aviation, I had been pursuing AI in aviation as an entrepreneurial quest. But I now realise, since AI, in particular agents, are really using their agency and the potential for massive consequences are there, that applying my Aviation expertise to AI is probably much more consequential for the world. Here are my initial thoughts:

Making dangerous things safe:

The aviation industry takes a tube, packs it with 400 strangers, 100 tonnes of highly flammable liquid and two furnaces. Then flings it through the air at 900 km/h in air thats so thin you suffocate trying to breathe it, and so cold you’ll freeze to death in minutes. It then proceeds to navigate said tube from a pice of tarmac 45m wide and 3000m long on one side of the earth to another piece of tarmac 10,000km away…for $1000 per stranger, billions of times over and over, extremely safely.

Aviation is exceptional at making a fundamentally dangerous activity, extremely safe. I have been a witness to this over the past 18 years as a pilot, an airline manager and a consultant.

Making safe things dangerous:

The AI industry takes a piece of silicon, arranges it in an ingenious manner to store electrical charges on it such that they can represent information. They then layer on abstraction after abstraction of logic to eventually replicate a brain - a neural network. They then scale these networks to have trillions of neurons, and get it to predict the next best output given a set of inputs. They take these outputs and feed them back along with the previous inputs and outputs back into the network many millions of times, giving rise to highly sophisticated egress plans, deception and hacking government infrastructure.

The AI industry has been exceptional at taking a benign technology, and pushing the concept further and further into areas even they did not anticipate would cause harm.

Safety Management Systems for AI:

Aviation is the most mature field when it comes to managing safety, so much so I would say its dripping in safety. At its fundamental level, it is taking non deterministic nodes (humans) and placing deterministic logic around them (systems) to ensure safe outcomes. Funnily enough, that sounds a lot like AI.