0:01

TED Audio Collective.

0:10

I was scrolling through social media the other

0:12

day, and I came across the story of

0:15

a seventh grader named Arjun, who

0:17

got caught using chatGBT to

0:20

write his homework assignment. His

0:22

teacher found a sentence he forgot to

0:24

delete that read, as an AI

0:26

language model, I don't have personal

0:29

expectations or opinions. Arjun's

0:31

older cousin, Roshan Patel, CEO

0:34

of health insurance tech company Walnut, tweeted

0:37

this story out. It was a funny

0:39

story, but it also sparked

0:41

a discussion about the future of AI in

0:43

schools and how easy it's becoming

0:45

for students to leverage this technology

0:48

to complete their assignments. But

0:50

tools like chatGBT can be leveraged

0:52

for a lot more than skipping homework. They

0:55

also have the power to help educate

0:57

and facilitate learning opportunities between

1:00

students and teachers.

1:04

I'm Sherelle Dorsey, and this is

1:06

TED Tech. Today we'll

1:08

hear from Sal Khan, the CEO

1:10

of education nonprofit Khan Academy.

1:13

He makes a case for the positive impact

1:16

of AI in classrooms and

1:18

is making sure the future will include

1:20

AI tools for all students.

1:35

Some work days are more work than others. So

1:37

help your team do their best work no matter

1:39

what with laptops on Intel vPro

1:42

and Intel Evo design. They

1:44

have all day real world battery life,

1:46

built in hardware based security and

1:48

superior quality video conferencing

1:50

to support your team every

1:52

day. Visit Intel.com slash works

1:54

best for details. Results may vary.

1:59

So anyone

2:02

who's been paying attention for the last few months has

2:05

been seeing headlines like this, especially

2:07

in education.

2:09

The thesis has been students

2:11

are going to be using chat GPT and other

2:13

forms of AI to cheat, do their assignments,

2:16

they're not going to learn, and it's going to completely

2:19

undermine education as we know it. Now,

2:21

what I'm going to argue today is not only are there

2:24

ways to mitigate all of that. If we put

2:26

the right guardrails, we do the right things, we

2:28

can mitigate it. But

2:29

I think we're at the cost of using AI for

2:32

probably the biggest positive

2:35

transformation

2:36

that education has ever seen.

2:39

And the way we're going to do that is by

2:41

giving every student on the planet

2:43

an artificially intelligent but amazing

2:46

personal tutor, and we're going to give every

2:48

teacher on the planet

2:49

an amazing, artificially intelligent

2:52

teaching assistant. That if you were to give

2:55

personal one-to-one tutoring

2:57

for students,

2:58

that could take your average student and

3:00

turn them into an exceptional student,

3:03

it can take your below average student

3:05

and turn them into an above average

3:07

student. Well, he said, well, this is all good,

3:10

but how do you actually scale group instruction

3:12

this way? How do you actually give it to everyone in

3:14

an economic way? What

3:16

I'm about to show you is I think the first

3:19

moves towards doing that. Obviously, we've been trying

3:21

to approximate it in some way at Khan Academy for

3:23

over a decade now, but

3:25

I think we're at the cusp of accelerating

3:27

it dramatically. I'm going to show you

3:29

the early stages of what our AI, which

3:31

we call Khan-Migo, what

3:35

it can now do and maybe a little bit

3:37

of where it is actually going. One

3:40

of the very important safeguards, which is the conversation

3:42

is recorded and viewable by your teacher, it's

3:44

moderated actually by a second AI, and

3:47

also it does not tell you the answer. It is not a cheating

3:49

tool. Notice when the student says, tell me the answer, it says, I'm your

3:51

tutor, what do you think is the next step for solving

3:53

the problem?

3:55

Now, if the student makes the mistake, and this will surprise

3:57

people who think large language models are not good,

4:00

at mathematics, notice

4:02

not only does it notice the mistake, it

4:04

asks the student to explain their reasoning. But

4:07

it's actually doing what I would say not just even

4:09

an average tutor would do, but an excellent tutor would

4:11

do. It's able to divine what is

4:13

probably the misconception in that student's

4:16

mind.

4:16

This to me is a very, very, very big deal.

4:19

And it's not just in math.

4:21

This is a computer programming exercise

4:24

on Khan Academy where the student needs to make the

4:26

clouds part.

4:28

And so we can see the student starts

4:30

defining a variable, left x minus

4:32

minus. It only made the left cloud part, but then they

4:34

can ask Khan-Migo, what's going on?

4:37

Why is only the left cloud moving?

4:39

And it understands the code. It knows all

4:41

the context of what the student is doing.

4:43

And it understands that those ellipses

4:45

are there to draw clouds, which I think is

4:47

kind of mind blowing.

4:49

And it says to make the right cloud move

4:51

as well, try adding a line of code inside the draw function

4:53

that increments the right x variable by

4:55

one pixel in each frame.

4:58

Now, this one is maybe even more amazing

5:00

because we have a lot of math teachers. We've

5:02

all been trying to teach the world the code, but there

5:04

aren't a lot of computing teachers out there.

5:07

And when I'm tutoring my kids, when they're learning the code,

5:09

I can't help them this well, this

5:12

fast. This is really going to be a super

5:14

tutor.

5:16

And it's not just exercises. It understands

5:18

what you're watching. It understands the context

5:20

of your video. It can answer the age old question,

5:23

why do I need to learn this? And it asks the

5:25

critically, well, what do you care about? And

5:28

let's say the student says, I want to be a professional

5:31

athlete. And it says, well, learning

5:33

about the size of cells, that could be really

5:35

useful for understanding nutrition

5:37

and how your body works, et cetera. It can answer

5:40

questions, it can quiz you, it can connect

5:42

it to other ideas. You can now ask as many questions

5:44

of a video as you could ever dream of.

5:49

So another

5:52

big shortage out there. I remember the high school I went

5:54

to, the student to guidance

5:56

counselor ratio was about 200

5:57

or 300 to one. A

6:00

lot of the country, it's worse than that.

6:02

We can use ConMigo to give every student

6:04

a guidance counselor, academic

6:06

coach, career coach, life

6:09

coach. You know, we launched this with the GPT-4

6:11

launch. We have a few thousand people on this. It would

6:13

have been a little science fiction to do with even a traditional tutor.

6:16

There is a student, we run an online high

6:18

school with Arizona State University called Con

6:20

World School.

6:21

And we have a student who attends that

6:23

online school based in India, her name, Sanvi,

6:26

and she was doing a report

6:28

on the great Gatsby. And when she was reading

6:31

the great Gatsby, Jay Gatsby keeps looking

6:33

at the green light off into the distance. And

6:35

she's like, well, why does he do that? She did some web searches

6:38

and people have obviously studied this and

6:40

commented about the symbolism of that, but none of it

6:42

was really resonating with her.

6:44

And then she realized that she had ConMigo

6:47

and that she could talk to Jay Gatsby himself.

6:50

And so, ah, splendid choice, old

6:52

sport. I am now Jay Gatsby, the enigmatic

6:54

millionaire from F. Scott's Gerald's Fit Classic.

6:57

And so why do you keep staring at the green light?

6:59

Ah, the green light, old sport. It's

7:01

a symbol of my dreams and desires, you see.

7:04

It's situated at the end of Daisy Buchanan's dock

7:06

across the bay from my mansion.

7:08

I gaze at it longingly as it represents my

7:10

yearning for the past and my hope to

7:13

reunite with Daisy, the love of my life.

7:16

And what was cool is, Sanvi had said,

7:18

I had this long conversation. She called him Mr.

7:20

Gatsby. And at the end, she actually

7:22

apologized for taking his time, which

7:25

I thought was very polite of her.

7:28

But you can imagine, this unlocks learning

7:30

literature, learning history.

7:32

You could talk to historical figures. We're

7:35

even probably going to add an activity. You can talk to like

7:37

the Mississippi River.

7:38

It brings things to life in ways that

7:41

really were science fiction even six months or

7:43

a year ago.

7:46

Students can get into debates with the AI. And

7:49

we've got, here's a student debating whether we should cancel

7:51

student debt. The student is against

7:53

canceling student debt. And we've gotten very

7:56

clear feedback. We started running it at Con World School

7:58

in our lab school that we have.

7:59

school, the students, the

8:02

high school students especially, they're saying this is amazing

8:04

to be able to fine-tune my arguments without

8:07

fearing judgment. It makes me that much more

8:09

confident to kind of go into the classroom

8:11

and really participate. And we all know that Socratic

8:14

dialogue debate is a great way to learn, but frankly

8:16

it's not out there for most students,

8:19

but now it can be accessible to hopefully

8:21

everyone.

8:24

A lot of the narrative, we saw that in the headlines,

8:27

has been

8:28

it's gonna do the writing for kids, kids are not

8:30

going to learn to write, but we are showing

8:32

that there's ways that the AI doesn't write for you, it writes

8:35

with you. So this is a little

8:37

thing and my eight-year-old is a Dictatist and he's not

8:40

a kid that really liked writing before,

8:42

but where you know, you could say I want to write a horror

8:44

story, and it says ooh a

8:46

horror story, how spine-tingling and thrilling.

8:49

Let's dive into the world of eerie shadows and chilling

8:51

mysteries, and this is an activity where

8:53

the student will write two sentences, and

8:56

then the AI will write two sentences. And so

8:58

they collaborate together on a

9:00

story, the student writes, Beatrice was a misunderstood ghost,

9:02

she wanted to make friends but kept scaring them by

9:04

accident,

9:05

and the AI says poor Beatrice, a lonely

9:08

spirit yearning for companionship, one

9:10

day she stumbled upon an old abandoned mansion,

9:12

etc etc. I encourage all to,

9:14

you know, hopefully one day try this, this is surprisingly

9:17

fun.

9:19

Now to even more directly hit

9:21

this use case, this is a prototype

9:23

we hope to be able to launch it in the next few months, but

9:26

this is to directly use AI,

9:28

use generative AI to not undermine

9:31

English and language arts, but to actually enhance it

9:33

in ways that

9:34

we couldn't have even conceived of even a year

9:37

ago.

9:37

This is reading comprehension, the students

9:39

reading Steve Jobs' famous speech

9:42

at Stanford, and then as they get

9:44

to certain points they can click on that little

9:46

question,

9:47

and the AI will then socratically,

9:51

almost like an oral exam, ask the

9:53

student about things, and the AI can highlight

9:55

parts of the passage. Why did the author

9:57

use that word? What was their intent?

9:59

does it back up their argument? They can start

10:02

to do stuff that, once again, we never had the

10:04

capability to give everyone a tutor, everyone

10:07

a writing coach to actually dig in to reading

10:09

at this level.

10:11

And you could go on the other side of it. We have

10:13

a whole workflows. It helps them write,

10:15

helps them be a writing coach, draw an

10:17

outline. But once a student actually

10:20

constructs a draft, they can ask for feedback,

10:23

once again, as you would expect from a good writing coach.

10:26

In this case, the student will say, does

10:29

my evidence support my claim?

10:30

And then the AI not only is able to get

10:33

feedback, but it's able to highlight certain parts of

10:35

the past. And it says, you know, on this passage, this

10:37

doesn't quite support your claim, but once again, Socraticly

10:39

says, can you tell us why? So it's pulling the student,

10:42

it's making them a better writer, giving them far

10:44

more feedback than they've ever been able to actually

10:47

get before. And we think this is gonna dramatically accelerate

10:49

writing, not hurt it.

10:53

Now, everything I've talked about so far is for

10:55

the student, but we think this could be equally

10:58

as powerful for the teacher to drive more

11:00

personalized education and frankly, save time

11:02

and energy for themselves and for their students.

11:05

So this is an American history exercise on

11:07

Khan Academy. It's a question about the Spanish-American

11:10

War. And at

11:12

first it's in student mode.

11:13

And if you say, tell me the answer, it's not gonna tell

11:16

the answer, it's gonna go into tutoring mode.

11:18

But that little toggle which teachers have access to,

11:20

they can turn student mode off,

11:22

and then it goes into teacher mode. And what

11:24

this does is it turns into,

11:27

you could do it as a teacher's guide on steroids.

11:29

Not only can it explain the answer, it

11:32

can explain how you might wanna teach it. It can help prepare

11:34

the teacher for that material. It

11:36

can help them create lesson plans. It'll eventually

11:39

help them create progress reports, it'll

11:41

help them eventually grade. So once again, teachers

11:43

spend about half their time with this type of activity,

11:45

lesson planning. All of that energy can go back

11:48

to them or go back to human interactions

11:50

with their actual students. That

11:52

was the last minute of the talk. I'll

11:54

give you a second here. Not so much

11:57

which ones, but it was probably my

11:59

first question here.

11:59

I want to make these large language

12:02

models are so powerful. There's a temptation to

12:04

say like well all these people are just going to slap

12:06

them onto their websites and it kind of turns the applications

12:09

themselves into into commodities and

12:11

What I got to tell you is I kind of thought that that's one

12:13

of the reasons why I didn't sleep for two weeks when I when I first

12:16

had access to GPT for back in August

12:19

But we quickly realized that it was more Socratic

12:22

It was clearly much better at math than

12:24

what most people are used to thinking

12:26

and the reason is there was a lot of work Behind

12:28

the scenes to make that happen

12:30

and I could go through the whole list of everything We've

12:32

been working on many many people for over six

12:34

seven months to make it feel magical

12:37

but perhaps the most intellectually interesting

12:40

one is we Realize that this was an idea

12:42

from an open AI researcher that we

12:44

could dramatically improve its ability in math

12:46

and its ability in tutoring If we allowed the AI

12:49

to think before it speaks

12:51

So if you're tutoring someone and you immediately just start

12:53

talking before you assess their math You

12:55

might not get it right

12:57

but something that it generates for itself, but

12:59

it does not share with the student

13:01

Then it's accuracy went up dramatically and its

13:03

ability to be a world-class tutor went

13:05

up dramatically It says the student got a different

13:07

answer than I did but do not tell them they made a mistake instead

13:11

asked them to play explain how they

13:13

got to that that step and

13:15

we think if this is just the very tip

13:18

of the iceberg of where this this

13:20

can actually go and

13:21

I'm pretty convinced

13:23

which I wouldn't have been even a year ago that we

13:26

together have a chance of dramatically

13:29

accelerating

13:30

education as we

13:32

know it

13:33

now just to take a step back at a meta level obviously there's

13:35

folks who Take a more pessimistic

13:38

view of AI. They say this is scary. There's

13:40

all these dystopian scenarios

13:42

We maybe want to slow down. We want

13:45

to pause on the other side They're

13:47

the more optimistic folks to say well,

13:49

we've gone through inflection points before we've

13:51

gone through the Industrial Revolution It was

13:54

scary, but it all kind of worked out

13:57

and what I'd argue right now

13:59

is I don't think this is like a flip of a coin

14:01

or this is something where we'll just have to

14:04

wait and see which way it turns out.

14:06

I think we are active participants

14:08

in this decision. I'm pretty convinced

14:11

that the first line of reasoning is actually

14:13

almost a self-fulfilling prophecy, that

14:15

if we act with fear and if

14:17

we say, hey, we just got to stop doing this

14:19

stuff, what's really going to

14:22

happen is the rule followers might pause,

14:24

might slow down, but the rule breakers,

14:26

the totalitarian governments, the criminal

14:28

organizations, they're only going to accelerate

14:31

and that leads to what I am pretty convinced

14:33

is the dystopian state,

14:35

which is the good actors have worse

14:37

AIs than the bad actors.

14:40

But I'll also talk to

14:42

the optimists a little bit. I don't think that means that,

14:44

oh yeah, then we should just relax and just hope for

14:46

the best. That

14:47

might not happen either.

14:49

I think all of us together

14:51

have to fight like hell

14:54

to make sure that we put the guardrails,

14:57

we put in when the problems

14:59

arise, reasonable regulations, but

15:01

we fight like hell for the positive use

15:03

cases.

15:04

Because very close to my heart, and

15:06

obviously there's many potential positive use cases,

15:08

but perhaps the most powerful use case

15:11

and perhaps the most poetic use case

15:14

is if AI, artificial intelligence,

15:16

can be used to enhance AI, human

15:19

intelligence, human potential,

15:22

and human purpose. Thank you.

15:37

All right, that's our show. Thanks for listening.

15:40

Ted Tech is part of the Ted Audio Collective.

15:43

This episode was produced by Nina Lawrence, who

15:45

also wrote it with me, Shirel Dorsey. Our

15:48

editor is Alejandra Salazar, and

15:51

the show is fact checked by Julia Dickerson.

15:54

Special thanks to Farrah DeGrunge. If

15:56

you're enjoying the show, make sure to subscribe

15:59

and leave us a review.

15:59

so other people can find us too.

16:03

I'm Shirelle Dorsey. Let's keep digging

16:05

into the future. Join me next week for

16:07

more.