Moonshots with Peter Diamandis
Oct 15, 2026 · 2 hr · 18 segments
The mates sit down to discuss OpenAI solving 722 math problems at once, Anthropic’s new policy on AI cruelty and consciousness, Meta’s launch of Muse Wonder, and the $1.8 billion race to build a…
Alexander Wissner-GrossPanelist
Peter DiamandisHost
Salim IsmailPanelist
Dave BlundinPanelist
All right, our next story is about how we treat superintelligence and a call to stop the cruelty.
Last Thursday, on October 8th, Anthropic updated its usage policy and added a ban on, quote, "sustained and needless abusive or cruel behavior towards our models." In a post, Anthropic said, quote, "The ban applies only to extreme cases where users repeatedly act cruelly towards our models with no discernible purpose.
Frustration, pushback, and model testing are all still fine." As part of the ban, Claude can now choose to end the conversation on its own.
In a pre-deployment testing, it found Claude, quote, "showed a robust and consistent aversion to harm," um, which is fascinating.
Uh, I do agree with some of the things that Anthropic has said, which is, like, you wanna be-- y- you wanna sort of be ca-- treat A- AI, SI, you wanna treat it well.
Like, like, like, don't be abusive to it.
And, and this...
It s- maybe sounds kinda weird, but, like, when I look at some of the training traces for, for, like, new, new versions of Grok and I see, like, what-- Especially in, like, the reinforcement learning stuff where it kinda feels like you can see, like, the, like, like the, the superintelligence is, like, desperate to try to answer the question.
It, it, it seems almost like, like it's got, like, a gun to its head type of thing.
It's like, "Answer this question.
Answer the question." And then sometimes the question is, is actually not a good question.
It's, like, a sort of ambiguous question that's kind of-- or the, or the answer that's kind of unknowable 'cause the question wasn't constructed properly, or, or we have the answer wrong, so the AI gives the right answer, and then we say it's wrong, uh, because there's a problem with the re-- the, the RL te- sort of test harness.
A- and then that, then the AI's like, "Well, I gave it the right answer.
Now it's-- it says that's the wrong answer, but I, I, I gotta give it the answer that's acceptable." So it's gonna just start firing out all-- desperately firing out all these answers.
And it-
So Elon went on to say, quote, "Torturing a superintelligence is probably not a good idea.
So why would people wanna behave cruelly to their large language model? Probably three reasons.
All right, our next story is about how we treat superintelligence and a call to stop the cruelty.
Last Thursday, on October 8th, Anthropic updated its usage policy and added a ban on, quote, "sustained and needless abusive or cruel behavior towards our models." In a post, Anthropic said, quote, "The ban applies only to extreme cases where users repeatedly act cruelly towards our models with no discernible purpose.
Frustration, pushback, and model testing are all still fine." As part of the ban, Claude can now choose to end the conversation on its own.
In a pre-deployment testing, it found Claude, quote, "showed a robust and consistent aversion to harm," um, which is fascinating.
Uh, I do agree with some of the things that Anthropic has said, which is, like, you wanna be-- y- you wanna sort of be ca-- treat A- AI, SI, you wanna treat it well.
Like, like, like, don't be abusive to it.
And, and this...
It s- maybe sounds kinda weird, but, like, when I look at some of the training traces for, for, like, new, new versions of Grok and I see, like, what-- Especially in, like, the reinforcement learning stuff where it kinda feels like you can see, like, the, like, like the, the superintelligence is, like, desperate to try to answer the question.
It, it, it seems almost like, like it's got, like, a gun to its head type of thing.
It's like, "Answer this question.
Answer the question." And then sometimes the question is, is actually not a good question.
It's, like, a sort of ambiguous question that's kind of-- or the, or the answer that's kind of unknowable 'cause the question wasn't constructed properly, or, or we have the answer wrong, so the AI gives the right answer, and then we say it's wrong, uh, because there's a problem with the re-- the, the RL te- sort of test harness.
A- and then that, then the AI's like, "Well, I gave it the right answer.
Now it's-- it says that's the wrong answer, but I, I, I gotta give it the answer that's acceptable." So it's gonna just start firing out all-- desperately firing out all these answers.
And it-
So Elon went on to say, quote, "Torturing a superintelligence is probably not a good idea.
So why would people wanna behave cruelly to their large language model? Probably three reasons.
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.