WEBVTT

00:00:00.240 --> 00:00:03.597
One thing that's real is that, you know,
every 3 months we make a model that's

00:00:03.680 --> 00:00:07.357
smarter than the month than the model we
made before. And indeed, the models are

00:00:07.440 --> 00:00:10.957
slowly getting to where they're,
you know, better than most humans at most

00:00:11.040 --> 00:00:14.877
things. Maybe at some point better all
humans at than all things. The other

00:00:14.960 --> 00:00:19.517
thing I think is real is, you know,
the model is not building itself in the

00:00:19.600 --> 00:00:23.437
sense that like I have a specific
copy of the model that's like making a

00:00:23.520 --> 00:00:27.037
specific copy of the the, you know, the
the next model. you know, we don't have

00:00:27.120 --> 00:00:29.597
our self-replicating, you know, nanobots

00:00:29.680 --> 00:00:32.077
or robots or or whatever, but but what

00:00:32.160 --> 00:00:34.397
absolutely is happening is that within

00:00:34.480 --> 00:00:37.117
Anthropic and elsewhere, we are using

00:00:37.200 --> 00:00:40.317
Claude to help build products on top of

00:00:40.400 --> 00:00:43.277
Claude and to help train the next Claude

00:00:43.360 --> 00:00:47.997
to help serve Claude faster. Um,
you know, six months ago, I I made this

00:00:48.080 --> 00:00:52.157
prediction that, you know, in in six
months, 90% of code would be written by

00:00:52.240 --> 00:00:56.957
by by AI models. Some people think that
prediction is wrong but within Anthropic

00:00:57.040 --> 00:01:01.277
and within a number of companies that we
work with that is absolutely true. Um

00:01:01.360 --> 00:01:04.717
now 90 you know >> so you're
saying that 90% of all code at

00:01:04.800 --> 00:01:07.597
anthropics being written
by the by the model today

00:01:07.680 --> 00:01:10.797
>> on on many teams you
know not uniformly every

00:01:10.880 --> 00:01:14.477
>> is that something that you've
kind of provided something kind of

00:01:14.560 --> 00:01:18.397
>> something auditable around that or like
shown how that's happened. Yeah, the

00:01:18.480 --> 00:01:23.757
teams play an editing and supervisory
role, but you know, Claude can both you

00:01:23.840 --> 00:01:27.197
can say write this feature and claude
will do it and then the human kind of

00:01:27.280 --> 00:01:29.517
looks over the code or asks for help

00:01:29.600 --> 00:01:31.757
from another another model, another copy

00:01:31.840 --> 00:01:36.637
of Claude. Um uh the other thing
is sometimes we have these long running

00:01:36.720 --> 00:01:40.397
bugs. You know, when we were training
our our the most recently released

00:01:40.480 --> 00:01:44.237
Claude, it was like, you know, the
cluster was broken. There was this bug.

00:01:44.320 --> 00:01:48.317
Engineers were spending a few days on it.
And, you know, we just we just went

00:01:48.400 --> 00:01:50.797
to Claude and we said, I don't know,
just kind of play around with the

00:01:50.880 --> 00:01:53.997
cluster and see if you can figure out
what's wrong. And, you know, Claude

00:01:54.080 --> 00:01:57.917
found this just super obscure bug that
the engineers had missed. So, this this

00:01:58.000 --> 00:01:59.837
was one area where it was like Claude is

00:01:59.920 --> 00:02:01.677
like having a very talented teammate. So

00:02:01.760 --> 00:02:05.117
if claude is now running writing 90% of

00:02:05.200 --> 00:02:07.597
all of the code in enthropic do you need

00:02:07.680 --> 00:02:09.997
less engineers if 90% of all the code

00:02:10.080 --> 00:02:12.717
when will it be 95% or 99%.

00:02:12.800 --> 00:02:17.277
>> Yeah. So you know I I you know I think
people can misinterpret it as you know

00:02:17.360 --> 00:02:21.357
you know like Claude writing 90%
of the code means we're firing 90% of the

00:02:21.440 --> 00:02:24.637
software engineers. But you know if
you're familiar with the principle of

00:02:24.720 --> 00:02:26.237
>> well you have made some very aggressive

00:02:26.320 --> 00:02:29.597
comments around firing executives

00:02:29.680 --> 00:02:33.517
>> that that so um

00:02:33.600 --> 00:02:35.037
>> I just want to put two and two together

00:02:35.120 --> 00:02:36.717
to get four. Okay.

00:02:36.800 --> 00:02:42.077
>> Um uh so uh um you know if >>
one thing I've learned in knowing you

00:02:42.160 --> 00:02:45.117
for so long is I have to come back
and kind of like get that qualifying

00:02:45.200 --> 00:02:49.517
question again. Yes. Because you're go
you know you you're you're going fast.

00:02:49.600 --> 00:02:54.077
So trying to I'm trying to keep up with
you and it's amazing but I want to just

00:02:54.160 --> 00:02:55.837
make sure I understand
exactly what you're saying.

00:02:55.920 --> 00:02:57.437
>> Yes, absolutely. Let me let me let me

00:02:57.520 --> 00:03:00.797
take them one by one. So um uh you know

00:03:00.880 --> 00:03:04.157
under comparative advantage if Claude is

00:03:04.240 --> 00:03:07.517
writing 90% of the code what that means

00:03:07.600 --> 00:03:12.317
usually is you need just as many
software engineers you might need more

00:03:12.400 --> 00:03:16.637
because they can they can then be more
leveraged. they can focus on the 10%

00:03:16.720 --> 00:03:18.637
that's editing the code or writing the

00:03:18.720 --> 00:03:21.517
10% that's the hardest or supervising a

00:03:21.600 --> 00:03:25.917
group of a of AI models. And so what
happens is, you know, you can just end

00:03:26.000 --> 00:03:28.317
up being 10 times more productive

00:03:28.400 --> 00:03:31.597
because the remaining 10% is is now a

00:03:31.680 --> 00:03:35.437
complement to the to the AI system
that's that's getting much better at at

00:03:35.520 --> 00:03:37.917
getting much better at what
it's producing. So So

00:03:38.000 --> 00:03:40.957
>> it's a rebalancing of the work
is what you're really saying.

00:03:41.040 --> 00:03:45.040
>> It's it's not a replacement.
It's on a replacement.