Slowing Down AI with a Kg/Processing power Requirement

Nvidia is on the forefront of AI processing hardware. The come from making graphics processors which have pushed the computational power of computers for specific types of calculations (vector transformations). AI when realized through deep learning networks also has a specific computational requirement, and for this purpose build processors have a market.

Nvidia now comes out with a miniature deep learning accelerator (the learning takes place on another platform) called the Jetson TX2. It puts things like face recognition and other advanced recognition tasks in the drone realm, which is kind of scary, because as long as a drone flies it doesn’t have much to think about, what would an AI smart drone be able to do?

Jetson TX2

Nvidia already came out with a titan of AI modules, the other part of this Jetson TX2,the DGX-1 and its processing workhose the Tesla P100. This means it is easier (starting with possible) to generate the data to run on the TX2 for applications in the field.

The danger of AI is threefold : 1. It can start to compete with humans for the same resources and win. 2. It can start fighting itself and destroy the environment humans depend on. 3. It can setup humans against each other and make it impossible to know what is real anymore. The third threat is also created by virtual reality, which can have AI, but mainly gives people the illusion they are in a reality which is 100% fake (ok you knew that, but what if people really get confused? They already do so with actors). AI, even simple AI which is unattended or worse still unknown to be operating, can drive changes and affect people in ways we don’t want or should even fear.

AI the way it is often conceptualized, as a ghost in the machine, does not have certain qualities it needs for humans to be able to relate to it. Weaknesses are that it is immortal, never gets tired and has no body to maintain or  protect. It probably can’t and if so that makes it a very weak AI. To make a threatening AI self protection is a large part of the mix. Today we think cruise missles are still the best example of true AI.

As a result the biggest consolance right now is that most AI is kept in the box, it has no actuators, no arms and legs. Even so an AI hacking hospitals and setting insulin doses for political opponents of its owner can alter the course of human history. The trick with AI is to keep close track of what actual real world mechanisms are operated by an AI. This is made harder with tiny powerfull platforms that can move and proliferate and operate undetected.

If we had to give an advise we would make a requirement that any hardware of considerable computing power should have a weight that reflects that power. This is similar to what we said about money, money should have weight to make global transfer and movement of it way more difficult. AI capability should also have weight. Humans are extreme lightweight AI, any AI that approaches the level where it can be a threat should weigh at least 80 kg. This way it will be hard to have it move around on a tiny drone.

Imagine your mortal enemy being able to take the guise of a poisonous wasp. We are blessed there are no animals with true intelligence on that scale, or we’d be toast. Apparently, because of the energy requirement of intelligence, all species kind of minimize it and prioritize agility over smarts. Being agile, moving away from theats, saves more lives than understanding the threats or analyzing.

Just like a criminal with ball and chain a great way to reduce the risk of many threats is to weigh it down. Weigh down AI platforms, and all the dangerous things will still happen, but at least we have more time to develop our respons to it as the most intelligent species on Earth to date..

 

Leave a Reply