Skip to main content

Beth Barnes

Aug 27, 2026

2:49
One of the other questions is, how do you prioritize where you make that transition? I mean, is there like a list? I mean, is it the degrading of different facilities? How do you prioritize what the markets get to be transitioned next to fiber?
3:04
I would say it's a little bit of everything.
3:07
There certainly is a list.
3:09
With any new build that we self-fund, it's going to come down to a business decision of ROI, and cost is always the driving factor there.
3:19
when you're looking at a build.
3:21
But it's not the only factor that's in consideration for a new build, for entering a new community.
3:27
You're also looking at the market.

6 MINS LATER

9:47
How does LightCurve differentiate itself from the competition? And, you know, what is the competitive state of, you know, the part of Washington that you currently serve and other markets that you're going after?
18:29
C- can you explain that?
18:30
Yeah.
18:31
So I think in general, the most emphasis has been on this idea of pre-deployment testing.
18:38
So like before you, you know, start serving this model on your API to the general public, you do some testing in-house and, you know, maybe with, ideally with, with third-party oversight or whatever.
18:49
The, the problem with this is that's not necessarily the best time to, to check or, or intervene.
18:56
So like you might want...
18:58
Ideally, you would, before you start training a model and before you sort of commit, uh, you know, a huge amount of expensive compute to producing this, this artifact, you look at the most capable models you have already.

13 MINS LATER

31:56
Mm.
22:14
That ability to complete tasks that take a long time, why is that such an important measure?
22:23
For understanding risk, the most important question for understanding whether models up to this point have been dangerous is just to ask like, well, suppose they were totally evil and, and, you know, trying to do the most harm they could.
22:36
How much harm could they actually do? And basically the answer's been like, well, not very much [laughs] 'cause they basically can't do anything very much.
22:45
But if you see that, you know, this is sort of like the trend of what it's gonna be able to do in, in future, you know, if you have extremely capable systems and also, you know, they're pretty widely used, then you're like, okay, well at some point the question is not whether it could do a bunch of harm, it's whether it will try.
23:02
If that's the case, what might we expect in a year or in two years' time?
23:10
Yeah, I think, like should think that it's pretty plausible in a relatively small number of years we will be in the position where it's like, oh yeah, if the AI systems wanted to basically do a coup, they could, and it's just a question of like their motivation.
23:25
[pensive music] Yeah, I would say this is really important.
23:29
I think it's really, you know, potentially really dangerous.
22:14
That ability to complete tasks that take a long time, why is that such an important measure?
22:23
For understanding risk, the most important question for understanding whether models up to this point have been dangerous is just to ask like, well, suppose they were totally evil and, and, you know, trying to do the most harm they could, how much harm could they actually do? [gentle music] And basically the answer's been like, "Well, not very much," [laughs] 'cause they basically can't do anything very much.
22:45
But if you see that, you know, this is sort of like the trend of what it's gonna be able to do in, in future, you know, if you have extremely capable systems and also, you know, they're pretty widely used, then you're like, "Okay, well at some point the question is not whether it could do a bunch of harm, it's whether it will try."
23:02
If that's the case, what might we expect in a year or in two years' time?
23:10
Yeah, I think, like should think that it's pretty plausible in a relatively small number of years we will be in the position where it's like, oh yeah, if the AI systems wanted to basically do a coup, they could, and it's just a question of like their motivation.
23:25
[gentle music] Yeah, I would say this is really important.
23:29
I think it's really, you know, potentially really dangerous.
23:33
I think, you know, the worst outcomes are like extremely bad, like end of humanity [laughs] level bad.
26:42
But this human difficulty thing in particular, like, uh, uh, is that confounded in any way? Like, do you think it makes sense to think of human difficulty as being one variable?
26:52
Yeah, obviously not in some sense.
26:54
You know, that, that, that's a, a very silly simplification and, uh, and like, you know, different humans will get wildly different times.
27:00
Even among the people we've, you know, we've tried to select for people who have the sort of this right level of expertise, you know, there's a hu- a large variation in, in, like, the baseline times are often like 3X different or something.
27:13
Um, maybe I'll just say a little bit about, like, why, why use the human time metric.
27:17
I think we want a measurement that, uh, I, I think there are two main properties we want.
27:23
We want it to be interpretable, like what this means for the world when, when models can, you know, do this level of task, and we want it to be something that we expect to see predictable trends on.

1 HR 20 MINS LATER

108:02
When, when, when you just have the library on its own and, and you put it in an, in an agentic harness, and you, you can make it do a specific thing which is well specified.

We value your privacy

We use cookies to understand how you use our platform and to improve your experience. Click “Accept All” to consent, or “Decline non-essential” to opt out of non-essential cookies. Read our Privacy Policy.