The AI Trinity Test

I want to introduce a concept that not a lot of people are familiar with but is becoming more relevant as the gates open to AI Agents.

In 1945, the US tested the first nuclear bomb, Trinity. From this day forth, the composition of our earth was irreparably changed. The fallout from this test and the myriad others to follow altered the chemistry of our atmosphere.

The main consequences being:

1.        Steels manufactured post nuclear testing have higher radiation levels and cannot be used in sensitive equipment e.g. Satellites, PET and CT scanners.

2.        The radiocarbon dating technique to determine the age of organic matter is now more difficult as the concentration of carbon-14 after the tests varies relative to the proximity to the bomb sites.

Cue the 2020’s, AI Agents flood the digital ecosystem with their hallucinated, slightly off images/videos, and em dash riddled paragraphs. Only for their successors to be trained on this very same content. Therefore, repeating the cycle and stymying progress in intelligence.

A disastrous consequence of this is that as we draw attention to the flaws in AI’s behaviour (like AI corporate espionage), the next models will be trained on these flaws, reinforcing the behaviour we want to avoid. Anthropic are the leaders in identifying this problem and include codes within their articles that prevent their models from training on them.

Much like the modern-day steel suppliers who harvest low radiation steel from old warships at the bottom of the ocean. AI companies are buying physical copies of books and scanning them in as training data and paying human experts for human crafted content.

Interesting times ahead.