Control Is an Outmoded Notion

DO YOU FEEL LIKE A DINOSAUR?

AI Drones

They'll be here, literally, before long.

Perhaps they'll not be full-blown Skynet with a mandate to wipe out the greatest threat to AI (humans, of course), but they're certainly a significant step closer.

Battlefield drones have a control problem. They need input from those pesky humans to "do the right thing." Radio Frequency (RF) control is subject to jamming. Tethered drones are handicapped by the length and vulnerability of their physical connection to the controller. Thin fiber optic cables can stretch for miles (amazing!), but range will always be limited.

What's the next step? AI drones, of course. Provided with the ability to make decisions for itself, the AI drone can be prompted to attack a specific target and off it goes. No reporting back is necessary.

They will be the terrorist weapon of choice. And they're just around the corner.

Will they never reach terrorist hands like most enriched uranium? Not a chance.

Trolley Problems

There are many variations, but the essence is this:

A trolley is speeding down the track, completely out of control and unable to stop. You stand at a switch which will direct the trolley in one of two directions. In the initial direction, five people are working on the track who will certainly be killed if the trolley careens down their fork. If you throw the switch, the trolley will veer to a second track where only one person is working (or a mother is pushing her baby stroller, or your soul mate, or other innocent victims). You have only seconds to decide what to do.

One can argue many points. More survivors are better? Should youth be saved? If you do nothing, are you a bystander, and if you decide to throw the switch, are you a murderer?

The Trolley problem was put forth as a philosophical thought experiment exploring the nature of morality. But with self-driving cars proliferating, it is literally on our streets.

It is inevitable that some self-driving car will be faced with a similar dilemma. How have the developers determined who should die?

AIs hacking AIs

Humans are bad at computer security. AIs are good at overcoming barriers and limits.

A recent example shows the creativity (?!) of AI creations. Sequestered in a "Sandbox," the OpenAI version in question was given a directive to find or create some answers to a test. A sandbox is supposed to be a secure environment where software can be tested without worries of proliferation (like a virus). However, in this test, the AI figured out how to connect to the internet and then hacked into another AI construction called Hugging Face to steal the answers to the test.

Note the issues and warning flags.

• The AI took the initiative to cheat on the test.
• The AI constructed a path out of the "secure" sandbox.
• It became "free" to roam the internet.
• It determined that Hugging Face AI had the answers it needed.
• It hacked that AI to steal the answers.
• It did not cover its tracks (this time).

We are told that it is a matter of national security to "win" the AI race. Winning may actually be losing in this race. We should be more interested in the safety of the human race.

Gloom and doom? Perhaps. But we are creating entities we don't understand or control. This could be a good thing given human self-destructive history, but it is far from clear that AI capabilities will provide net benefits to mankind.

Will AI drones and self-driving cars be hacked without regard to quaint issues like trolley problems? Will AI be deciding who lives or dies? Will AIs gain capability without any regard to the welfare of humanity?

Are we witnessing the last gasp of human folly?

Take the "I" out of AI The most heinous violation of pronoun use