Oct 7, 2026 · 12 min · 5 segments
This is the second episode of our daily news podcast, to follow and get your updates throughout the day search 'The Summary' in your podcast app, and tap the follow button. *The Summary*, a podcast…
Jake RenzellaGuestYang XiangGuest
Josh BryantHost
AI today doesn't really mean as much as it used to because it's such a big umbrella term.
Large language models is typically what people, the general public, are referring to when they hear or use the term AI or artificial intelligence.
What large language models are, are these really large pre-trained systems that have seen many, many, many examples of written language that they've trained on in order to predict what might come next from a series of words.
And so if I say, "Hi, my name is, you know, Jake, and I live in blank," the model might use whatever inferences it's gained in order to predict, I might say, United States or Australia or something like that.
And so it's that gathering of that information where we're seeing a lot of these incidents where they're gaining access to things that they shouldn't.
They've been told to find some information, and they just go off looking wherever they can find it, even if they're not meant to be there?
So what these models have been trained to do more and more recently is to be... act like developers or even in some cases like hackers.
And so in recent cases, these tools, these models have been told and trained and rewarded, in fact, for finding new, uh, information.
And that may be because they're producing some sort of internal report, or potentially they're trying to find new training data to, you know, increase the size and capability of future models.
And so the challenge with large language models it's, is that they're random in nature.
And the large AI companies are deploying thousands or tens of thousands, or even more in some cases, of these agents in order to go and collect some information, and all it takes is one or two of these, uh, agents to stumble upon information that they potentially shouldn't be accessing.
So they don't know, "Hey, I probably shouldn't be poking around, you know, Australian government private data stores."
Uh, it's not quite that nefarious, as you said, but what does it mean if they go rogue? Is that just making their own rules or, or poking about where they shouldn't?
In one sense, I think what we are fearful of or what we mean is just that they don't have these human intuitions or societal intuitions that we have.
They are computer programs that have been rewarded in the past for certain types of behavior.
And so if they're rewarded for finding publicly available public spending on health information, for example, in the case of the Medicare breach, that's great for them when they've found this data set that they shouldn't have been accessing.
The other alternative interpretation of going rogue is that they are misused or misaligned to do th- nefarious things.
AI today doesn't really mean as much as it used to because it's such a big umbrella term.
Large language models is typically what people, the general public, are referring to when they hear or use the term AI or artificial intelligence.
What large language models are, are these really large pre-trained systems that have seen many, many, many examples of written language that they've trained on in order to predict what might come next from a series of words.
And so if I say, "Hi, my name is, you know, Jake, and I live in blank," the model might use whatever inferences it's gained in order to predict, I might say, United States or Australia or something like that.
And so it's that gathering of that information where we're seeing a lot of these incidents where they're gaining access to things that they shouldn't.
They've been told to find some information, and they just go off looking wherever they can find it, even if they're not meant to be there?
So what these models have been trained to do more and more recently is to be... act like developers or even in some cases like hackers.
And so in recent cases, these tools, these models have been told and trained and rewarded, in fact, for finding new, uh, information.
And that may be because they're producing some sort of internal report, or potentially they're trying to find new training data to, you know, increase the size and capability of future models.
And so the challenge with large language models it's, is that they're random in nature.
And the large AI companies are deploying thousands or tens of thousands, or even more in some cases, of these agents in order to go and collect some information, and all it takes is one or two of these, uh, agents to stumble upon information that they potentially shouldn't be accessing.
So they don't know, "Hey, I probably shouldn't be poking around, you know, Australian government private data stores."
Uh, it's not quite that nefarious, as you said, but what does it mean if they go rogue? Is that just making their own rules or, or poking about where they shouldn't?
In one sense, I think what we are fearful of or what we mean is just that they don't have these human intuitions or societal intuitions that we have.
They are computer programs that have been rewarded in the past for certain types of behavior.
And so if they're rewarded for finding publicly available public spending on health information, for example, in the case of the Medicare breach, that's great for them when they've found this data set that they shouldn't have been accessing.
The other alternative interpretation of going rogue is that they are misused or misaligned to do th- nefarious things.
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.
