Bob McMillan
speaker
109 appearances
4 recordings
2 series
first heard May 2025
last heard 9 Jun
Bob McMillan’s voice in public audio — every appearance, attributed to the second.
Trend
recordings per month · last 12 monthsRecordings per month over the last 12 months — 2 in all, peaking in Jun 2026 with 1.
Appearances
Anthropic announced this new technology, and then they said, it's too dangerous to be released publicly.
Naturally, that's going to get everybody's attention.
well anthropic announced this new technology and then they said it's too dangerous to be released publicly you know naturally that's going to get everybody's attention but anthropic says that there are a bunch of things that mythos can do that are just dangerous right like so it could provide advice on how to build a bioweapon or um build a bomb but the thing that's got the most attention is its cyber security capabilities
Anthropix started a program called Project Glasswing, and they gave early access to the model to software companies mostly, and people who were building products that could be hacked.
And they basically were invited to train Mythos on their software and see how good it was at finding bugs and writing code that could be used to take advantage of these bugs and hack the systems.
And by early indications, Mythos is pretty good at this.
So it's created a lot of turmoil with makers of software, uncovering just an unprecedented number of software bugs and rushing to patch them before other models develop the same kinds of capabilities themselves.
Well, this is the secret sauce, I guess.
The new model is called Fable 5.
It's not Mythos, but it's like the essence of it is the same as Mythos, right?
And basically, they've developed a set of technologies that can ostensibly figure out what you're asking about.
So if you ask about the weather, that's okay.
But when you're asking about something scary, like how to build a bioweapon or something
how to hack a computer, they kick you back to an older model, Opus 4.8, the previous version.
Yeah, this is what they call jailbreaking.
And it's sort of an ongoing issue with trying to control these models because people have been for years been able to get them to do things they're not supposed to do.
Anthropic has tested internally and they've also hired these what they call red teamers, these external consultants to come in and just kick the tires of it.
So they're confident that they've done some good work here to prevent the jailbreaking.
But we'll see what happens.
My pleasure, Alex.
Showing 1–20 of 109 · page 1 of 6
Next →