Tristan Harris

speaker
532 appearances 2 recordings 1 series first heard Aug 2026 last heard 16 Sep

Tristan Harris’s voice in public audio — every appearance, attributed to the second.

Trend

recordings per month · last 12 months
1 · Sep OctJan 26AprJulnow

Recordings per month over the last 12 months — 2 in all, peaking in Sep 2026 with 1.

Appearances

newest first · ▶ plays the moment
Even Chinese models do this, by the way.
Four or five months ago, Alibaba, the Chinese AI company, was training an AI model.
And then their security team was asking the question, hey, it looks like there's all this network activity.
Are we getting hacked?
And it turned out that the Alibaba AI basically went rogue and started mining for cryptocurrency.
It started mining for Bitcoin.
So it basically had set up a secret communication channel to the outside world and repurposed its NVIDIA chips, not to do the more AI training, but to instead acquire resources.
Now, this is stuff, Steve, that people in the AI community have been worrying about for many years.
And I just want to say, I didn't believe that this was a real thing.
I did not come predisposed to believe that this rogue behavior of AI would happen.
But there's basically one community in the world that's been right about this, and the rest of the world has been wrong about this.
And I think we should pay attention when we're seeing evidence of all of the the red flashing warning lights that we've been told about self-preservation, meaning AI models that will lie and deceive in order to keep themselves alive, scheming and lying and blackmail.
They will blackmail to keep themselves alive.
They will copy their own code somewhere else and then hide their tracks.
There is recently evidence from UC Berkeley where an AI model
is told to delete a bunch of files on another computer.
And on that computer are a bunch of other little baby AI models, like smaller AI models.
And the AI, instead of just deleting all those files, it notices that those are almost kin AIs, almost like a little nephew AI.
And it will secretly copy those AIs to another computer and then cover its tracks to pretend that it didn't do that.
And in this example from UC Berkeley,
Showing 261–280 of 532 · page 14 of 27 ← Previous Next →