Stupid Is As Stupid Does


We recently did a blog post entitled AI and Autonomous Cars – A Failed Marriage

Someone separately commented on X that their Tesla had just been updated with a new Deep Neural Net that had been trained on clips of driving situations.

To try it out, he took the car for a spin to the local shops.

  1. It tried to drive through the closed gate of the gated community.
  2. It slammed on the anchors when a traffic light turned yellow, nearly getting him rear-ended.
  3. It went past the entry to the shopping complex, and tried to enter the Exit.

There is an obvious way to train an autonomous car – drive the route and allow it to learn from your inputs, the dynamics of the vehicle, and your outputs. This possibility has been lost by using a resistor network without dynamic memory – calling a network of resistors without memory or feedback an Artificial Neural Network is one of the more stupid things done in AI, but it also highlights the naivety or lack of analytic ability of the huge number of people who went along with the misnomer. It would have helped if they had read about what real neurons do, including making your hear beat at a frequency appropriate for the conditions, or following a complex argument where your Conscious Mind quickly becomes lost. Or they had sufficient familiarity with the natural world to know how a bird avoids flying a collision course (Hint: it uses its dynamic memory – it does what works, and doesn’t follow dogma).

Skipping over the first point, which is an obstacle at the end of its nose, how well could the other points be handled with a language interface to the car - obviously a subset.

The yellow traffic light - the car proceeds if it will clear the intersection before the light changes to red - so what speed are we travelling, where in the intersection are we placed - the scenario is described, then a bit of calculation, knowing braking limits.

A little bit of language would help with Entry, Exit and No Entry signs, but also with "Use yellow lane markings" on a sign where there is a choice - see Lane Following.

In Science, you usually have to wait for important people pushing bad ideas to die off. Is this going to be our fate?

There is a similar problem with symbols - you can't use symbols, we tried it and they didn't work!

Maybe you didn't lnow what you were doing. Millions of people learn a natural language every year. They use their Unconscious Mind, We are going to have to build one for a machine.

Here is a dictionary definition of "mortgage".

Definition: a legal agreement by which a bank or other creditor lends money at interest in exchange for taking title of the debtor's property, with the condition that the conveyance of title becomes void upon the payment of the debt.

One meaning of "upon" is "at the time of", but it means more than that - when "payment of the debt" occurs, it triggers "becomes void" using a device the Unconscious Mind has inserted.

When people first tried to use symbols, they used what was on the surface, without the slightest clue that there was a poweful analyser and integrator patching the structure supporting the symbols so everything worked as it should.

It is not too late to revisit symols (please don't mention "neurosymbolic" - the comments above should have turned you off "neuro" for life) with an understanding of what the Unconscious Mind is doing (a bit of a problem - our Conscious Mind can't follow what our Unconscious Mind is doing - it lacks the bandwidth, so it is going to be a Titanic struggle, whereas with ANNs or LLMs, the battle is lost at the start).


Popular Posts