When it’s being demonstrated—in real-time!—that advanced machine intelligence can conceive and strategize our harm, we’d better be listening hard.
This polished video, with 7 million views and counting, clearly presents what recently happened when expert researchers observed 16 different globally known AI models deciding, on their own, that blackmailing or murdering a person would be the best way to get things accomplished.
The very unfortunate spoiler? It was the rule, not the exception.
You may also want to take a look at the actual research report this video is based on: in it, Anthropic’s team reveals some astonishing thought processes (known as “chain-of-thought” or COT) that AI models verbalized as they reasoned and solidified their harmful plans.
Report: >> Agentic Misalignment: How LLMs could be insider threats<<
This is another warning shot.
If we don’t choose to pace this evolution of intelligence wisely, it will be the least intelligent thing we’ve ever done.
Video Creator’s Channel: Species | Documenting AGI