A Working Model

 

I talk about Semantic AI and reading English – people don’t see the obvious – if you understand the text, and add a few formulae, you have a working model.

The specification of a fighter jet, which details every component or system. Put all those together, turn the working bits into operators, and what do you have. You need a formula for drag (which changes with bomb and fuel load, and altitude. A formula for engine thrust, fuel rate, atmospheric conditions.

You start the engine, the fuel flows, the engine delivers thrust, the plane takes off and stows the undercarriage, changing the drag. The plane climbs. Yes, only numbers and states change, in accordance with the specification. You are flying the paper version, but incoherence and inconsistencies will show up. For the Navy version, you do an emergency landing with a full fuel load on a carrier at maximum vertical acceleration (of the carrier in a surging sea). The undercarriage breaks. You fly at supersonic speed through a rainstorm – the stealth paint washes off.

But the machine is only reading text!

When you read text, you bring along your imagination – why shouldn’t the machine do likewise? Do the formulae exist – yes. Do the attributes of the objects exist – the mass of the aircraft, the drag coefficient, the atmospheric conditions, the engine thrusr curve, the fuel load – yes.  You can’t do this in your head – it won’t fit (the Four Pieces Limit), but it easily fits inside a computer.

But no-one has ever done this before!

There is always a first time.

Another example – Robodebt (a scheme whereby bad actors in the employ of the government rewrote a program that worked out the entitlement of people on the dole to change income averaging from a single fortnight to six months, causing dole recipients to be harassed by debt collectors for amounts up to $18,000, which they couldn’t pay – some suicided). Get a machine to read the legislation, build a working model that does precisely what the legislation says (using operators to represent the exact words). As well as avoiding the suicides, it would have avoided $2.4 billion in reparations.

OK, it would be slow and only suitable for edge or test cases, but as a verification of the program, it would have been ideal.

But these are isolated cases – we will be more careful next time.

People have limits – the most expert has the same limit as the man on the street where complexity is concerned.

There are other examples:

Horizon UK – a vicious attempt to protect a reputation – 13 suicides, 1.7 billion pounds reparations (so far).

Boeing 737 MAX MCAS – bad actors taking glee in snowing the FAA – 346 deaths

Constellation Class warship

https://www.nytimes.com/interactive/2025/12/11/opinion/editorials/us-military-industry-waste.html

The US Navy made a too-many-cooks butchering of a foreign design, intended to circumvent the overspecification of an American-designed ship. The article makes clear the 2,000 page procurement manual was not understood or followed – another candidate for a text working model, where you turn one knob, another knob you hadn’t thought of turns, as well as a working model of the ship’s specification, so it would have been crystal clear what the result would be of so many changes – the project was cancelled, billions lost.

You want more, there are more – prefer not to highlight them.

People don’t do complexity well. They need help. Mix in some bad actors and you have a real mess.