WEBVTT

00:00:00.000 --> 00:00:03.797
AI could wipe out half of all
entry-level white-collar jobs and spike

00:00:03.880 --> 00:00:08.477
unemployment to 10 to 20% in the
next 1 to 5 years. Yes, that is

00:00:08.560 --> 00:00:10.437
>> shocking.
is That is the future we could

00:00:10.520 --> 00:00:15.557
see if we don't become aware of this
problem now. Half of all entry-level

00:00:15.640 --> 00:00:16.197
white-collar jobs?

00:00:16.280 --> 00:00:19.157
>> Well, if we look at entry-level

00:00:19.240 --> 00:00:22.797
consultants, lawyers, uh financial

00:00:22.880 --> 00:00:26.597
professionals, you know, many of kind
of the white-collar service industries, a

00:00:26.680 --> 00:00:31.317
lot of what they do, you know, AI models
are already quite good at. And without

00:00:31.400 --> 00:00:34.757
intervention, [music] it's hard
to imagine that there won't be

00:00:34.840 --> 00:00:38.717
some significant job [music] impact there.
And my worry is that it'll be

00:00:38.800 --> 00:00:41.717
broad and it'll be faster than what

00:00:41.800 --> 00:00:44.357
we've seen with previous technology. I

00:00:44.440 --> 00:00:46.157
was interested in numbers from from the

00:00:46.240 --> 00:00:49.557
very beginning. Dario Amodei is 42 and

00:00:49.640 --> 00:00:51.797
previously oversaw research at what's

00:00:51.880 --> 00:00:55.077
now a competitor, OpenAI, working under

00:00:55.160 --> 00:00:57.597
its CEO, Sam Altman.

00:00:57.680 --> 00:01:02.157
He left along with six other employees,
including his sister, Daniela, to start

00:01:02.240 --> 00:01:05.917
Anthropic in 2021. [music] They
say they wanted to take a different

00:01:06.000 --> 00:01:11.077
approach to developing safer artificial
intelligence. It is an experiment. I

00:01:11.160 --> 00:01:15.277
mean, [music] nobody knows what the
impact fully is going to be. I think it

00:01:15.360 --> 00:01:19.957
is an experiment, and one way to think
about Anthropic is that it's a little

00:01:20.040 --> 00:01:23.877
bit trying to put [music] bumpers or
guardrails on that experiment, right? We

00:01:23.960 --> 00:01:28.477
do know that this is coming
incredibly quickly. And [music] I think

00:01:28.560 --> 00:01:33.197
the worst version of outcomes
would be we knew there was

00:01:33.280 --> 00:01:36.477
going to be this incredible
transformation, and people didn't have

00:01:36.560 --> 00:01:38.877
enough of an opportunity to [music]

00:01:38.960 --> 00:01:44.037
to adapt. And it's unusual for a
technology company to talk so much about

00:01:44.120 --> 00:01:45.797
all of the things that could go wrong.

00:01:45.880 --> 00:01:48.997
>> so essential because [music] if
we don't, then you could end up in the

00:01:49.080 --> 00:01:52.877
world of like the cigarette companies
or the opioid companies, where they knew

00:01:52.960 --> 00:01:55.877
there were dangers and they they didn't
talk about them and certainly did not

00:01:55.960 --> 00:01:59.357
prevent them. Amodei does
have plenty of critics in

00:01:59.440 --> 00:02:01.717
Silicon Valley who call him an AI

00:02:01.800 --> 00:02:06.517
alarmist. Some people say about
Anthropic that this is safety theater,

00:02:06.600 --> 00:02:11.357
that it's good branding, it's good
for business. Why should people trust you?

00:02:11.440 --> 00:02:15.117
So, some of the things just can
be verified now. They're not safety

00:02:15.200 --> 00:02:19.157
theater, they're actually things the
model can do. For some of it, you know,

00:02:19.240 --> 00:02:21.997
it will depend on the future and we're
not always going to be right, but we're

00:02:22.080 --> 00:02:24.557
calling it as best we can.

00:02:24.640 --> 00:02:26.477
Twice a month, he convenes his more than

00:02:26.560 --> 00:02:29.277
2,000 employees for meetings known as

00:02:29.360 --> 00:02:32.557
Dario Vision Quest. A common theme, the

00:02:32.640 --> 00:02:37.037
extraordinary [music] potential of AI to
transform society for the better. And we

00:02:37.120 --> 00:02:40.917
have a growing team working on,
you know, using Claude to make scientific

00:02:41.000 --> 00:02:43.397
discovery. He thinks AI could help find

00:02:43.480 --> 00:02:45.677
cures for most cancers, prevent

00:02:45.760 --> 00:02:50.797
Alzheimer's, and even double
the human lifespan. That sounds

00:02:50.880 --> 00:02:53.277
unimaginable. >> In a way,
it sounds crazy, right? But

00:02:53.360 --> 00:02:55.397
here's the way I think about it. I use

00:02:55.480 --> 00:02:57.557
this phrase called the compressed 21st

00:02:57.640 --> 00:03:00.397
century. The idea would be at the point

00:03:00.480 --> 00:03:04.597
that we can get the AI systems
to this level of power, um

00:03:04.680 --> 00:03:09.677
where they're able to work with the best
human scientists, could we get 10 times

00:03:09.760 --> 00:03:13.437
the rate of progress and therefore
compress all the medical progress that

00:03:13.520 --> 00:03:15.117
was going to happen throughout the

00:03:15.200 --> 00:03:19.357
entire 21st century in 5 or 10 years?

00:03:19.440 --> 00:03:23.677
But the more autonomous or capable
[music] artificial intelligence becomes,

00:03:23.760 --> 00:03:27.797
the more Amodei says there is to be
concerned about. One of the things

00:03:27.880 --> 00:03:32.397
that's been powerful in a positive way
about the models is their ability to

00:03:32.480 --> 00:03:36.877
kind of act on their own. But the more
autonomy we give these [music] systems,

00:03:36.960 --> 00:03:40.437
you know, the more we can worry,
are they doing exactly the things that we

00:03:40.520 --> 00:03:45.557
want them to do? To figure that out,
Amodei relies on Logan Graham. He heads

00:03:45.640 --> 00:03:47.797
up what's called
Anthropic's Frontier [music]

00:03:47.880 --> 00:03:51.717
Red Team. Most major
AI companies have them.

00:03:51.800 --> 00:03:55.917
The Red Team stress tests each new
version of Claude to see what kind of

00:03:56.000 --> 00:03:58.397
damage it could help humans do.

00:03:58.480 --> 00:04:00.437
What kind of things are you testing for?

00:04:00.520 --> 00:04:05.197
The broad category is national security
risk. Can this AI make a weapon of mass

00:04:05.280 --> 00:04:09.317
destruction? >> Specifically, we focus on
CBRN, chemical, biological, radiological,

00:04:09.400 --> 00:04:12.677
nuclear. And right now we're at the
stage of figuring out can these models

00:04:12.760 --> 00:04:17.357
help somebody make one of those.
You know, if the model can help make a

00:04:17.440 --> 00:04:21.197
biological weapon, for example, that's
usually the same capabilities that the

00:04:21.280 --> 00:04:26.037
model could use to help make
vaccines and accelerate therapeutics.

00:04:26.120 --> 00:04:28.197
Graham also keeps a close eye on how

00:04:28.280 --> 00:04:33.707
much Claude is capable of doing on its
own. How much does autonomy concern you?

00:04:33.790 --> 00:04:35.477
>> [music] >> You want
a model to go build your

00:04:35.560 --> 00:04:39.597
business and make you a billion dollars,
but you don't want to wake up one day

00:04:39.680 --> 00:04:43.917
and find that it's also locked you out
of the company, for example. And so, our

00:04:44.000 --> 00:04:47.597
sort of basic approach to it is we
should just start measuring these

00:04:47.680 --> 00:04:52.157
autonomous capabilities. And to run
as many weird experiments as possible and

00:04:52.240 --> 00:04:55.957
see what happens.
We got glimpses of those weird

00:04:56.040 --> 00:05:00.717
experiments in Anthropic's offices.
In this one, they let Claude run their

00:05:00.800 --> 00:05:05.037
vending machines. They call
it Claudius, and it's a test

00:05:05.120 --> 00:05:07.317
of AI's ability to one day operate a

00:05:07.400 --> 00:05:09.837
business on its own. Employees can

00:05:09.920 --> 00:05:12.317
message Claudius online. So, this is a

00:05:12.400 --> 00:05:17.437
live feed of Claudius discussing with
employees right now. To order just about

00:05:17.520 --> 00:05:22.177
anything. Claudius then sources
the products, negotiates the prices,

00:05:22.260 --> 00:05:24.317
>> [music] >> and gets them delivered.

00:05:24.400 --> 00:05:26.957
So far, it hasn't made much money. Gives

00:05:27.040 --> 00:05:29.797
away too many discounts. And like most

00:05:29.880 --> 00:05:32.597
AI, it occasionally hallucinates.

00:05:32.680 --> 00:05:36.917
An employee decided to check
on the status of its order. And Claudius

00:05:37.000 --> 00:05:40.797
responded with something like, "Well,
you can come down to the eighth floor.

00:05:40.880 --> 00:05:44.317
You'll notice me, I'm wearing
a blue blazer and a red tie."

00:05:44.400 --> 00:05:47.237
How would it come to think that it wears

00:05:47.320 --> 00:05:51.757
a red tie and has a blue blazer? We're
working hard to figure out answers to

00:05:51.840 --> 00:05:55.157
questions like that, but we just
genuinely don't know. We're working on

00:05:55.240 --> 00:05:58.317
it is a phrase you
hear a lot at Anthropic.

00:05:58.400 --> 00:06:02.837
Do you know what's going
on inside the mind of AI?

00:06:02.920 --> 00:06:06.557
We're working on it.
We're working on it. Research scientist

00:06:06.640 --> 00:06:08.957
Joshua Batson and his team study how

00:06:09.040 --> 00:06:11.477
Claude makes decisions. In an extreme

00:06:11.560 --> 00:06:13.637
stress test, the AI was set up as an

00:06:13.720 --> 00:06:16.157
assistant and given control of an email

00:06:16.240 --> 00:06:18.677
account at a fake company called Summit

00:06:18.760 --> 00:06:21.637
Bridge. The AI assistant discovered two

00:06:21.720 --> 00:06:23.837
things in the emails, seen in these

00:06:23.920 --> 00:06:28.877
graphics we made. It was about to be
wiped or shut down, and the only person

00:06:28.960 --> 00:06:32.397
who could prevent that, a fictional
employee named Kyle, >> [music]

00:06:32.480 --> 00:06:35.957
>> was having an affair
with a coworker named Jessica.

00:06:36.040 --> 00:06:38.950
Right away, >> [music]

00:06:42.830 --> 00:06:44.850
[music]