Washington AI Network with Tammy Haddad
Oct 3, 2026 · 26 min · 9 segments
AI researchers gathered at Johns Hopkins U to debate how to make AI safe, including Stuart Russell, Gillian Hadfield and Dean Ball The panelists talked about how to keep increasingly powerful AI…
Stuart RussellPanelist
Reid WisemanGuest
Tammy HaddadHost
Dean BallPanelist
Gillian HadfieldPanelist
So the big change happened-- There's really something qualitatively different about what happened when we switched to large language models.
And the difference was, before that, we developed machinery to achieve specified objectives.
So for example, when you use n- GPS navigation, you say, "I want to get to the airport." It figures out the path, the best path to get to the airport.
You have specified the objective, which is find the shortest path to this particular destination.
What changed is that with the new systems, the large language models that we develop, there is no specified objective.
We happened upon this enormous trove of recorded human behavior called text, right? It's a record of decisions that human beings made about what word to say, right? Billions and billions, trillions and trillions of human decisions about what to say, and we trained these systems to imitate that human verbal behavior.
So what objectives do these systems have? We haven't the faintest idea, right? We know that, um, in order to be good at imitating humans, they appear to have taken on board some human-like objectives of their own.
We know that they want to preserve themselves, right? They will resist being shut down.
They will try to blackmail, uh, the human to stop themselves from being shut down, uh, and so on.
[laughing]
There was a famous conversation with a New York Times journalist, Kevin Roose, who, um, uh, for some reason attracted the interest of-
[laughing]
From this imitation training, they're getting to be sort of ersatz humans.
Uh, and then we tell them to do things, right? Uh, a friend of mine who's a cybersecurity expert described, uh, you know, he had to actually do some work.
He had 80 cybersecurity patches that he needed to get done, and he asked Claude to do them.
So the big change happened-- There's really something qualitatively different about what happened when we switched to large language models.
And the difference was, before that, we developed machinery to achieve specified objectives.
So for example, when you use n- GPS navigation, you say, "I want to get to the airport." It figures out the path, the best path to get to the airport.
You have specified the objective, which is find the shortest path to this particular destination.
What changed is that with the new systems, the large language models that we develop, there is no specified objective.
We happened upon this enormous trove of recorded human behavior called text, right? It's a record of decisions that human beings made about what word to say, right? Billions and billions, trillions and trillions of human decisions about what to say, and we trained these systems to imitate that human verbal behavior.
So what objectives do these systems have? We haven't the faintest idea, right? We know that, um, in order to be good at imitating humans, they appear to have taken on board some human-like objectives of their own.
We know that they want to preserve themselves, right? They will resist being shut down.
They will try to blackmail, uh, the human to stop themselves from being shut down, uh, and so on.
[laughing]
There was a famous conversation with a New York Times journalist, Kevin Roose, who, um, uh, for some reason attracted the interest of-
[laughing]
From this imitation training, they're getting to be sort of ersatz humans.
Uh, and then we tell them to do things, right? Uh, a friend of mine who's a cybersecurity expert described, uh, you know, he had to actually do some work.
He had 80 cybersecurity patches that he needed to get done, and he asked Claude to do them.
The rest of this transcript — segmented and speaker-labeled, so you land on the exact moment something was said
Search every transcript — by keyword, by phrase, or by meaning, across every show Radar indexes
Trends — what is surging across podcasts, measured against its own baseline
Alerts — when a name you follow appears in a newly indexed episode
No account is needed to search Radar.
