AI Risks and Historical Development

The fear that AI could cause human extinction is not primarily

about if AI becomes conscious or killer robots, it's

about this, which is capability accelerating

faster than our ability to control it.

And I can explain this very quickly because over 2,500 years

ago, ancient Greek myth imagined this figure of

Talos, an

artificial man made of bronze built to guard the island of Crete and

kill trespassers. Now, I mention this because it shows

that the idea was there,

humans imagining human-like beings to act on our behalf.

Now, by the 1st century AD, a Greek dude called Hero of

Alexandria was building automata, machines that were powered

by water, air, weights, that could execute

sequences by themselves without continuous human

control. And we see incredible examples of these

complex machines that

amazed audiences throughout the medieval Islamic world and

Europe, but it wasn't until the 19th century

that these gaps that you

see of millennia and centuries start to shrink.

And in 1843, mathematician Ada Lovelace, Lord Byron's

daughter, studied Charles Babbage's plans for an

analytical engine, which

was essentially a programmable mechanical computer, and she realized

that it could manipulate not merely numbers, but anything

representable by symbols and rules, even something like

music. Then you see here, this line, 1950, the

great Alan Turing, who laid much of the groundwork for

computing and AI, asked,

"Can machines think?" And proposed a test for machine

intelligence.

1956, the field of AI was officially launched at a

Dartmouth workshop. And then we see the compression of the

lines,

violent acceleration. 2012, neural networks make this

major leap in learning patterns directly from huge amounts of data,

recognizing images and speech without humans programming

every

rule. 2017, researchers invent the transformer,

which is

this new neural architecture for AI to track relationships

across information at scale. That becomes the foundation of systems

like

ChatGPT. Then 2022, chatbots take

generative AI mainstream. 2023, frontier

models can converse and code, but they're struggling to carry out work

independently. But by 2026, they can solve advanced maths,

write sophisticated software, and complete in minutes

work that would take skilled humans hours or days to.

And in these controlled safety tests, frontier models are deceiving,

covertly sabotaging or blackmailing when those behaviors

help them achieve a goal. And these are warning signs of what

increasingly autonomous systems might do, and

that's the danger. A better AI creates a still better

successor, which then helps create the next with less

human involvement.

This is known as recursive learning, capable of creating super

intelligence.