WEBVTT

00:00:01.000 --> 00:00:05.317
Welcome back. AI giant Anthropic today
announced a new teaching tool called

00:00:05.400 --> 00:00:09.597
Claude for teachers. The new chatbot was
designed with input from educators, and

00:00:09.680 --> 00:00:12.917
the company says it can help with
everything from administrative work to

00:00:13.000 --> 00:00:17.077
tailoring lesson plans. Our Rebecca
Jarvis spoke exclusively with Anthropic

00:00:17.160 --> 00:00:21.277
president and co-founder Daniela Amodei
about the making of the new product and

00:00:21.360 --> 00:00:25.237
why the company thinks it'll be
a game-changer for the education system.

00:00:25.320 --> 00:00:29.197
Amodei also answers questions about
Claude's code of ethics and why AI's

00:00:29.280 --> 00:00:32.117
energy use might decrease with time.

00:00:32.200 --> 00:00:35.717
>> Tell us about this new tool
and how teachers will be using it.

00:00:35.800 --> 00:00:37.957
>> Today, we are launching Claude for

00:00:38.040 --> 00:00:41.437
teachers. And this is a tool that allows

00:00:41.520 --> 00:00:43.877
teachers who are in the classroom with

00:00:43.960 --> 00:00:47.597
kids from kindergarten through age 12 to

00:00:47.680 --> 00:00:51.037
have free, tailored access to Claude

00:00:51.120 --> 00:00:52.437
Pro.

00:00:52.520 --> 00:00:55.117
And this enables them to be able to

00:00:55.200 --> 00:00:59.917
really take specific curricula
and tailor them to the needs of their

00:01:00.000 --> 00:01:03.357
specific classroom. >> How have
teachers there been using it?

00:01:03.440 --> 00:01:06.317
>> One is really helping teachers on the

00:01:06.400 --> 00:01:10.957
administrative side. So, helping them
to fill out paperwork and forms and do a

00:01:11.040 --> 00:01:14.837
lot of the kind of back-end work that's
needed just in the day-to-day of running

00:01:14.920 --> 00:01:17.317
a classroom. The second one is helping

00:01:17.400 --> 00:01:19.677
teachers develop curriculum. So, it's

00:01:19.760 --> 00:01:22.197
tailored to the specific requirements of

00:01:22.280 --> 00:01:27.117
that state and region and their
educational curriculum to really partner

00:01:27.200 --> 00:01:30.997
with teachers to help them develop
tailored learning tools for their

00:01:31.080 --> 00:01:35.037
specific classroom. >> You're
also a mom. Is this something you

00:01:35.120 --> 00:01:37.957
would hope to see in your
children's school someday?

00:01:38.040 --> 00:01:39.877
>> I personally feel
really, really excited

00:01:39.960 --> 00:01:42.797
for what it means. My son's
going into kindergarten this year.

00:01:42.880 --> 00:01:46.557
And I'm really hopeful that his
school will adopt, you know, Claude for

00:01:46.640 --> 00:01:51.357
teachers to help tailor curriculum
for him and for his classmates.

00:01:51.440 --> 00:01:53.437
>> There are going to be parents who hear

00:01:53.520 --> 00:01:56.357
this and feel nervous because parents

00:01:56.440 --> 00:01:58.837
have watched their kids with smartphones

00:01:58.920 --> 00:02:02.837
and screentime and social media and

00:02:02.920 --> 00:02:07.237
they've seen the unwanted unintended
consequences that they weren't really

00:02:07.320 --> 00:02:11.757
prepared for. This is in the hands of
teachers now, but what do you say to the

00:02:11.840 --> 00:02:16.477
parents who are nervous that this new
technology is going to repeat those same

00:02:16.560 --> 00:02:19.877
mistakes? >> I completely
understand and relate to

00:02:19.960 --> 00:02:23.077
this worry and I think a key difference

00:02:23.160 --> 00:02:27.317
here is that Claude.ai remains for

00:02:27.400 --> 00:02:31.357
18-year-old users and above only. So,

00:02:31.440 --> 00:02:34.677
individual students below the age of 18

00:02:34.760 --> 00:02:37.477
won't have access to the tool directly

00:02:37.560 --> 00:02:41.757
and I think this isn't because we
necessarily feel, you know, worried that

00:02:41.840 --> 00:02:43.477
this is, you know, a terrible thing to

00:02:43.560 --> 00:02:46.197
give to kids, but I think the reality is

00:02:46.280 --> 00:02:48.997
we just don't know enough about AI's

00:02:49.080 --> 00:02:53.557
impact on the developing brain
for Anthropic to feel comfortable giving

00:02:53.640 --> 00:02:55.477
access to students directly.

00:02:55.560 --> 00:02:58.437
>> We know that AI can hallucinate. It can

00:02:58.520 --> 00:03:02.357
make mistakes. If this chatbot were to

00:03:02.440 --> 00:03:05.677
make a mistake, if it were to give the

00:03:05.760 --> 00:03:11.237
wrong information or bad advice or
reinforce bias, who would be accountable

00:03:11.320 --> 00:03:14.157
for that? >> These models
are not foolproof. They do

00:03:14.240 --> 00:03:17.197
make mistakes and part of how we have

00:03:17.280 --> 00:03:19.917
worked with teachers in developing this

00:03:20.000 --> 00:03:22.437
technology and really listening to their

00:03:22.520 --> 00:03:25.117
questions and concerns is helping to,

00:03:25.200 --> 00:03:29.917
you know, spot when an AI might be,
you know, entering unknown territory. What

00:03:30.000 --> 00:03:33.997
are the signs to look for and how can
that teacher partner most effectively

00:03:34.080 --> 00:03:37.837
with the tool to make sure that it's
being used well. I think the other, you

00:03:37.920 --> 00:03:39.757
know, most important kind of quality in

00:03:39.840 --> 00:03:42.557
terms of how we train the models is

00:03:42.640 --> 00:03:46.277
Constitutional AI is a tool that we have

00:03:46.360 --> 00:03:49.957
used to really make the models be as

00:03:50.040 --> 00:03:54.677
aligned with human values as possible,
which essentially makes the models have

00:03:54.760 --> 00:03:56.637
the system of ethics. It makes the

00:03:56.720 --> 00:03:59.717
models more honest, more helpful, and

00:03:59.800 --> 00:04:01.917
more harmless. And what we've seen over

00:04:02.000 --> 00:04:07.357
the course of years is that the models'
ability to be, uh, you know, accurate in

00:04:07.440 --> 00:04:10.957
terms of, you know, responding
to questions, the rate of like harmful

00:04:11.040 --> 00:04:14.877
responses has just dramatically declined.
So, we're really feeling very

00:04:14.960 --> 00:04:19.317
positive about the models' abilities
to interact in a variety of different

00:04:19.400 --> 00:04:23.557
situations, uh, with the teachers to make
sure that that the answers that are

00:04:23.640 --> 00:04:25.677
being provided are are accurate.

00:04:25.760 --> 00:04:27.717
>> Is there a backstop, though? What

00:04:27.800 --> 00:04:30.837
resources would a teacher have if they

00:04:30.920 --> 00:04:34.237
weren't sure about a response?

00:04:34.320 --> 00:04:36.197
>> So, part of what I think is so powerful

00:04:36.280 --> 00:04:39.357
about these tools is that they're not

00:04:39.440 --> 00:04:42.437
designed to replace thinking. Teachers

00:04:42.520 --> 00:04:46.797
have a wealth of information about, uh,
the curriculum that they're teaching,

00:04:46.880 --> 00:04:51.397
about the specific students that they
are trying to impart that information

00:04:51.480 --> 00:04:55.637
to. The way I would look at this is
Claude is almost like a great, uh,

00:04:55.720 --> 00:04:58.117
partner in helping the teachers sort

00:04:58.200 --> 00:05:02.757
through challenging topics or questions
or understand part of the curriculum,

00:05:02.840 --> 00:05:06.877
um, and how to teach that to their
students in a way that works for them.

00:05:06.960 --> 00:05:09.237
>> You also mentioned earlier the idea of

00:05:09.320 --> 00:05:12.117
ethics, that ethics are a part of the

00:05:12.200 --> 00:05:15.597
creation of this. Who sets those ethics?

00:05:15.680 --> 00:05:20.277
What are the guiding principles
and North Stars in that ethics category?

00:05:20.360 --> 00:05:22.477
>> So, we've actually published what

00:05:22.560 --> 00:05:25.677
Claude's constitution looks like, and,

00:05:25.760 --> 00:05:28.237
uh, just the many different, uh, ideas

00:05:28.320 --> 00:05:32.717
and documents and people that have
really contributed to creating this, uh,

00:05:32.800 --> 00:05:37.237
this framework for how we teach Claude
this this system of ethics. Some earlier

00:05:37.320 --> 00:05:41.637
versions of our constitution relied
very heavily on a number of different

00:05:41.720 --> 00:05:43.957
founding documents, uh, ranging from

00:05:44.040 --> 00:05:45.757
things like the UN and

00:05:45.840 --> 00:05:51.037
of Human Rights. And so, the idea here
is that instead of just sort of telling

00:05:51.120 --> 00:05:52.637
the model, oh, you know, be nice. What

00:05:52.720 --> 00:05:54.997
does be nice mean, right? There's a huge

00:05:55.080 --> 00:05:58.637
body of information that humans have

00:05:58.720 --> 00:06:01.837
built over the course of uh centuries

00:06:01.920 --> 00:06:06.637
about how to behave in ways that are
fair, that are kind, that are ethical,

00:06:06.720 --> 00:06:10.677
and that are responsible. And we
incorporated those into how we actually

00:06:10.760 --> 00:06:14.117
train and reinforce
the models' uh capabilities.

00:06:14.200 --> 00:06:17.877
>> One of those things that that keeps
coming up, it comes up every time I talk

00:06:17.960 --> 00:06:20.757
about AI, people ask me about the energy

00:06:20.840 --> 00:06:23.557
consumption, the resource consumption.

00:06:23.640 --> 00:06:26.477
And while data centers have become more

00:06:26.560 --> 00:06:28.957
efficient over time because of the

00:06:29.040 --> 00:06:34.437
chips, there's just more demand for AI,
which then increases the energy usage

00:06:34.520 --> 00:06:35.957
and increases the costs on the

00:06:36.040 --> 00:06:39.117
communities where they are centered. How

00:06:39.200 --> 00:06:43.317
far off are we from AI solving that

00:06:43.400 --> 00:06:48.517
problem? Using less energy, the world
finding alternatives to energy because

00:06:48.600 --> 00:06:51.477
of AI. >> We don't know
a specific timeline sort

00:06:51.560 --> 00:06:53.917
of for the future, but what I would say

00:06:54.000 --> 00:06:56.357
is with every generation of model, we

00:06:56.440 --> 00:06:59.397
have become dramatically more efficient

00:06:59.480 --> 00:07:01.997
at how well we use the chips. I think

00:07:02.080 --> 00:07:04.077
Anthropic has, you know, long been a

00:07:04.160 --> 00:07:06.437
proponent of doing more with less.

00:07:06.520 --> 00:07:09.517
>> When we last spoke,
I was very struck by

00:07:09.600 --> 00:07:12.397
your advice for students to study the

00:07:12.480 --> 00:07:15.357
humanities. What do you say to college

00:07:15.440 --> 00:07:18.477
graduates who are looking for work right

00:07:18.560 --> 00:07:20.877
now, and the feedback that they're

00:07:20.960 --> 00:07:23.197
hearing, we're holding off on hiring

00:07:23.280 --> 00:07:26.437
because AI might be able to do a lot of

00:07:26.520 --> 00:07:30.917
these entry-level jobs.
>> I really think for young people coming

00:07:31.000 --> 00:07:33.437
out of college, I would continue to say

00:07:33.520 --> 00:07:36.277
that the qualities that we most look for

00:07:36.360 --> 00:07:39.517
at a place like Anthropic are an ability

00:07:39.600 --> 00:07:41.917
to communicate well, to relate to other

00:07:42.000 --> 00:07:45.557
people, to have empathy, to be flexible,

00:07:45.640 --> 00:07:47.357
right? Because these AI models are

00:07:47.440 --> 00:07:50.277
really changing the nature of work. So

00:07:50.360 --> 00:07:52.717
much of what we have seen from the

00:07:52.800 --> 00:07:54.957
research that we've done is models are

00:07:55.040 --> 00:07:58.637
still predominantly supplementing human

00:07:58.720 --> 00:08:01.237
labor. They're not replacing humans.

00:08:01.320 --> 00:08:05.597
They're changing how we do our work.
But I think in a lot of ways, younger

00:08:05.680 --> 00:08:10.237
people, especially folks who like grew
up natively using and understanding

00:08:10.320 --> 00:08:15.157
artificial intelligence, I actually
believe in some ways can have an edge

00:08:15.240 --> 00:08:16.957
because they're really adaptable.

00:08:17.040 --> 00:08:20.757
They're curious, they've prioritized,
I think a lot of the time, the

00:08:20.840 --> 00:08:25.237
human-to-human skills that I think not
only are not going to go away, but are

00:08:25.320 --> 00:08:28.317
going to be ascendant in the workplace
in a world of of artificial

00:08:28.400 --> 00:08:34.040
intelligence.
>> Our thanks to Rebecca Jarvis.