Podchaser Logo
Charts
Overwhelmed By AI Innovations and Challenges: Hashtag Trending  Weekend Edition for December 14, 2024

Overwhelmed By AI Innovations and Challenges: Hashtag Trending Weekend Edition for December 14, 2024

Released Friday, 13th December 2024
Good episode? Give it some love!
Overwhelmed By AI Innovations and Challenges: Hashtag Trending  Weekend Edition for December 14, 2024

Overwhelmed By AI Innovations and Challenges: Hashtag Trending Weekend Edition for December 14, 2024

Overwhelmed By AI Innovations and Challenges: Hashtag Trending  Weekend Edition for December 14, 2024

Overwhelmed By AI Innovations and Challenges: Hashtag Trending Weekend Edition for December 14, 2024

Friday, 13th December 2024
Good episode? Give it some love!
Rate Episode

Episode Transcript

Transcripts are displayed as originally observed. Some content, including advertisements may have changed.

Use Ctrl + F to search

0:01

Welcome to Hashtag Trending the

0:04

We call it We call it Project Sinaps,

0:06

AI in and this is this is

0:08

of our Project series series, meetings

0:10

of our group where we

0:12

discuss and explore we discuss a

0:14

more or less practical basis.

0:16

We're interested in how to

0:18

make it work in our

0:20

business and in our lives.

0:22

in our got Marcel and He's lives.

0:25

tech enthusiast, Guinea, he's an author, tech enthusiast,

0:27

open an AI expert, now think.

0:29

I you think think, Marcel? Look, he

0:31

dropped, there he goes. dropped his his

0:33

there we go. but there we This

0:35

is This They can see this.

0:37

Welcome can see this. Welcome, Marcel. Q, it's another

0:40

It's another beautiful

0:42

day in the simulation. And

0:44

we have John Panard, an IT

0:46

exec And we have John

0:48

Panart, an IT exec a in

0:50

financial services. at with a

0:53

wide experience him in we

0:55

brought him in that,

0:57

just that, but for but

0:59

for deep in, he's wide and he's wide

1:01

and he, neverma. He's big at cyber security,

1:03

yes. Welcome John. Thank you. Okay guys,

1:05

there's, we'll talk about, there's the

1:08

guys, was, there talk about it. There of

1:10

AI a tsunami of AI information

1:12

over the past week and a

1:14

half, keeps that straight anybody keeps that straight

1:16

is beyond me. or not, it

1:18

or not, there were a couple of

1:20

other things to talk about last week. One

1:22

of them was One new report on new

1:25

and and deception, and really relating to safety

1:27

of of AIs. or the safety of the frontier

1:29

models, let's put it that way. way. So a

1:31

new report on AI related to the

1:33

safety frontier models, and I think it's

1:35

worth discussing. I think it's worth third thing

1:37

that happened was an announcement

1:39

from Google on quantum computing,

1:41

which set the world set the

1:43

world afire discovered the poor state

1:45

of tech journalism it it

1:47

comes to addressing serious scientific

1:49

issues. did I have an Did I have an opinion I'm

1:51

I'm sorry. Let's start, why don't we start. out

1:53

don't we start out with the quantum

1:55

computing announcement from Google? Google? It's

1:58

called Willow. Willow and and

2:00

it's Buffy Vampire Slayer by the

2:02

way. do you think they got the name Do you think

2:04

they got the name from there? you have to give

2:06

them don't know. You have to give

2:08

them credit for one thing. It's

2:10

the most decent name I've heard in

2:12

a long time like at least this one some

2:14

least this one has some sort of

2:16

to the real world, can do this now this

2:19

the first Now, this is the first

2:21

time that Google has made this announcement.

2:23

the Here's the announcement for those who

2:25

missed it, was that they developed

2:27

this chip that can manage can manage 105 qubits

2:30

and it it can do that. speeds

2:33

at exponential speeds that

2:35

and I We'll give you one calculation

2:38

a I think they would

2:40

take a the big number is, longer or whatever

2:42

the big number is, been in than the

2:44

universe has been in existence to

2:46

calculate this. have a whole lot I would

2:48

have a whole lot more faith in this

2:50

if they hadn't made a similar announcement

2:52

in 2019 with it I think it was

2:54

Sycamore in those days. days. and they came up

2:57

and said they had 55 had 55 and

2:59

it would take take an eon for longer than than the

3:01

universe had been and more more the universe

3:03

than there are. than then somebody beat the

3:05

calculation I think within a year. I

3:07

think within a time, actually, they've

3:09

announced this chip. announced this

3:11

said, and course, said but of computing

3:14

may advance quickly. may advance quickly

3:16

and now and everybody went screaming. don't

3:19

know if you guys have heard too much about

3:21

it, but it but the it was oh it's

3:23

it's going to abolish Bitcoin Bitcoin. will

3:25

now be compromised forever. and

3:28

that'd be really great be really great. had

3:30

yeah they backtracked on that believe even

3:32

this morning or late last night on that.

3:34

like that this morning or late last

3:36

night or something like that, saying that

3:38

it will not threaten modern cryptography, which

3:40

is an interesting statement to make. Google

3:43

actually was Not really and I don't

3:45

I didn't know that Google actually was

3:47

responsible for that part of it That's what

3:49

I was talking about a failure of

3:51

tech journalism because everybody ran with this like

3:53

a quantum computer And it's with it's 105

3:55

and five is a great It is a but just

3:57

to but just to put this in context

4:00

If you wanted, now, just, I

4:02

don't know how deep people get

4:04

into know how deep people

4:06

get into quantum computing. I've

4:08

heard so many times, a

4:10

bit is and zero and these

4:12

quantum computers can now. anything, they're,

4:15

and they can have many states and all many

4:17

states and all that sort of

4:19

stuff, which is true. true, but

4:21

I want to remind people that people that a

4:23

A metaphor is not

4:25

a technical player. plan. In other other

4:27

words, these are great metaphors for quantum computing,

4:29

and if you could. And the

4:31

heart of quantum computing, of guys

4:34

who develop the first computers

4:36

don't fully understand it and

4:38

admit it. admit it. And so so for

4:40

anybody to come up with up with statement that says

4:42

this is going to do this. this is going to

4:44

do this, it's just not true. But

4:46

just to just to put this in perspective. In

4:49

a In a quantum computer, you

4:51

have physical qubits and you the

4:53

the qubits out there. You

4:56

need 100 ,000 virtual qubits to

4:58

get 1 ,000. physical cubits, physical

5:00

qubits, in other words, some of the things where you

5:02

can get answers. answers. And that's, so when you're

5:04

talking about 105 105 you're talking about a

5:06

really tiny step. tiny The other piece

5:08

of this, and I think this was

5:10

interesting, if we treat this like a

5:12

proof of concept. a proof of concept,

5:14

then it's 105 cubits as baby but

5:17

it also has better error correction

5:19

in the past. Good good step forward,

5:21

nothing wrong with that. that. But it

5:23

it does one

5:25

mathematical problem. And it's just

5:27

it. and it's just finding numbers

5:29

in a pattern that really is

5:32

tough for a it. That's it. do

5:34

It won't do for those people for those

5:36

people who wanna know, is it gonna

5:38

crack RSA encryption? encryption? in the

5:40

near future because RSA encryption

5:42

is it, you To

5:44

simplify it, you take a big

5:46

number, that's your encryption number, and

5:48

you basically find the multiplication of

5:50

two primes to get that number. number.

5:53

And that's the essence

5:55

of RSA, ability to

5:57

ability to calculate prime

5:59

numbers. at huge enormously

6:01

large numbers. that's really hard.

6:03

It takes That's really hard, the years from

6:05

12 to 1024 to 2048 to 4096, 10, the

6:07

key gets bigger and 40, 96. actually key

6:09

gets bigger and bigger that you

6:11

actually I through these calculations. about

6:14

this think that's interesting about this, is quantum

6:16

I'm in I'm in So I'm like right next door

6:18

to the next door to the Institute

6:20

for Quantum Computing. one of which is

6:22

one of the premier quantum computing labs in the

6:24

entire world. world. But the the chip

6:26

is really interesting in the sense

6:28

that it could actually change the

6:31

way that we process information in

6:33

hardware. it's 105 cubits, it's the calculation that

6:35

the calculation that they gave it

6:37

and you at the the beginning. to me,

6:39

it's To me, it's not the most interesting thing. The

6:41

most interesting thing is, they said they gave it

6:43

this calculation that would take a normal computer. that

6:45

would take a normal computer, ten million 70 million, million,

6:47

million which like you said you said

6:49

is a really big number bigger

6:52

than the universe. To me the

6:54

coolest thing about this this ran

6:56

with the story and I don't

6:58

know if Google actually said

7:00

this. actually with the story

7:02

that this proves the this

7:04

proves the multiverse theory of quantum. In

7:06

other words, the other words, the

7:09

many worlds the the is that

7:11

only way could actually actually performed

7:13

this is to actually

7:15

borrow borrow computational prowess

7:17

multiple universes to feed

7:19

the number the number

7:22

back the quantum chip as an

7:24

answer. answer. And to me, if you want to

7:26

If you want to talk about the media running

7:28

wild with this story, this is the coolest thing

7:30

for me, me, the idea

7:32

that this purportedly proves

7:34

the multi universe theory. theory. you

7:36

watch enough And if you if

7:38

enough physicists, if you're as addicted to

7:40

Neil as Tyson as I am, these people

7:42

will talk about this for hours you never

7:44

know and you never know when you're

7:47

in a Star Trek episode, when you're

7:49

in reality. And that's just, just, it it

7:51

gets to that level. level. But but

7:53

the practicality just to take this

7:55

back to a practical level. It

7:57

is a quantum advance. I love

7:59

that. Everybody's gonna have to, yeah,

8:01

but this can be to next well this People

8:03

have to use the next people It is

8:05

a very big advance in hardware. is a

8:07

very big still, in even by the

8:09

founders of this, by the years away

8:11

from a functional computer a could

8:13

actually solve a business problem a

8:15

a minimum. problem. the best

8:17

estimate I've And the best estimate I've heard

8:20

is, sorry, John? No, sorry, I was to say,

8:22

as you said, it is baby steps, steps,

8:24

but it's a step in the right direction.

8:26

direction. But it it is, a lot

8:28

of it is marketing hype. want to

8:30

announce all of these that that all

8:32

these these amazing things that they're

8:34

doing. Which it's good, but good.

8:36

it as you said, it's still got you said, it's still

8:38

got a long way to go. the whole

8:40

quantum computing field has been

8:43

The whole quantum a field has

8:45

been described as a set of

8:47

sorry as technologies looking as a as

8:49

looking for a solution, have a don't

8:51

actually have a practical application for what

8:53

this thing will do. We have

8:55

all these cool ideas about what we

8:57

could do if we could crunch

8:59

numbers at a level where we are

9:01

actually borrowing computing from multiple universes. for

9:03

a physicist who runs a YouTube

9:05

channel out there who runs a YouTube channel out there

9:07

called this woman. And I

9:09

love this woman. says, I'm not rude. I'm

9:11

just German. That should give you an

9:14

idea you an idea. But she is she is

9:16

actually a physicist. She is actually a

9:18

scientist she has a she has a great

9:20

news channel. And she actually does

9:22

talk about willow with a quantum chip. chip. And

9:24

of course, even the idea of

9:26

the many of interpretation of quantum

9:28

mechanics quantum mechanics is proof, that being

9:30

a proof of what this of

9:32

has supposedly done? chip has If you

9:34

ever want a dose of cold a dose

9:36

of thrown on top of your new

9:38

scientific top of you should totally check out

9:40

her channel. concepts, you She is one of the best.

9:42

out her channel. She is one science of the science

9:45

there, and out there, and She'll explain something without

9:47

ever talking down to you. And that's

9:49

what I say about, I you get into

9:51

these metaphors, it can do this, it's

9:53

like this, it's like that. do She

9:55

will talk to you that. She will talk to

9:57

you about Newton and Newtonian

9:59

physics. And she'll do it in a language

10:01

that most people I think can understand. can Maybe

10:03

not Maybe not the concept or the

10:05

math behind it, but she's really

10:07

quite really quite My favorite one is

10:10

her one is her science is bullshit If you

10:12

want to talk about to research, she's

10:14

got a great show on that.

10:16

She She does actually say that science

10:18

is bullshit, not not necessarily science itself. No,

10:20

sorry, I didn't mean to didn't mean to

10:22

say that. Yeah. to what the reality, back

10:24

to what the reality, the one of the we

10:26

get when we get something like on something like

10:28

this, we start talking about parallel universes,

10:30

which is, and it's all very interesting. all very

10:32

interesting, but from the fact

10:34

that some of the quantum principles. that

10:38

people are using are actually being

10:40

used in real computing. and

10:42

here in Canada, there's a company, think

10:44

it's called, a it D I in Vancouver? you

10:46

remember the name in the company? Do you they

10:49

have. the name of the functioning. Yeah. And

10:51

they have a computer. quantum

10:53

computer, approach and I'm saying

10:55

And guys have come closer

10:57

to cracking to cracking RSA. than

11:00

anybody else. They've at least made the first

11:02

steps of it. first steps

11:04

still reminds people that

11:06

reminds people that I think... I

11:08

don't I wouldn't worry

11:10

about your Bitcoin being hacked

11:12

and people taking your and

11:14

from you taking your one Bitcoin from

11:16

you there. But... Although, please coin

11:19

got hacked yesterday. Oh, please

11:21

God. Okay. good on him. God.

11:23

Looks good on them. Oh, man.

11:25

But there approaches. approaches.

11:27

that are being used in in here

11:29

in Canada. Again, we're leading. My

11:31

fear is, once again, that what'll happen

11:33

in Canada happen will, is it

11:36

will develop all of this stuff this

11:38

stuff and it won't, all exit the country,

11:40

but that's another story. another story.

11:42

Yeah. that's, I think I think there there

11:44

is some exciting stuff happening in stuff

11:46

people need to deal with that. And

11:48

it does, there is a practical And

11:50

it that they've made that says, Hey,

11:52

you should be thinking about replacing says,

11:54

hey, you should be terms of

11:57

if you have hyper super secret

12:00

type files that are going to be be more

12:02

than a decade old and still be useful. And

12:04

I think that's really, really, take five or

12:06

10 years to crack this, but if

12:09

to stolen your files now. somebody's can't they

12:11

might be able to crack them at one point.

12:13

That's mostly government crack them at one of stuff.

12:15

That's mostly stuff, I type imagine. of

12:17

stuff and interested what my bank

12:19

account was 10 years ago. interested

12:21

in think what it does is it shows

12:23

ten years ago. But I think some of

12:25

the advances that we've seen recently,

12:27

including advances that we've a lot language models are

12:29

able to crack in terms of

12:31

pattern recognition. crack in I think it shows

12:34

that we are actually... I think it of

12:36

this we point in

12:38

terms of things like

12:40

encryption terms of things like encryption and your

12:43

The idea that we might have

12:45

machines, we might have systems that

12:47

make all of this stuff irrelevant. this

12:49

stuff at some point in the

12:51

not the not future. future.

12:53

think we can't we can't that

12:55

as a possibility. And

12:58

it does also point to the idea

13:00

that at some point, we've got to

13:02

rethink we've got to just social systems, but our

13:04

economic systems and what we think of

13:06

in terms of what growth means, what

13:08

the economy means, and so forth. what

13:10

If we won't be able to trust

13:12

that sort of transaction so forward in the

13:14

future, actually have to, trust something about

13:16

it, now I'll seriously think about it. in

13:19

the future. We John, just to give it

13:21

to you to wrap up, how much of people

13:23

in about it is actually talking about these things? it.

13:25

I think this. it. In my in

13:27

my opinion, not enough. I

13:30

think we we talk about

13:32

security and protection and planning

13:34

for long for long I think there's

13:36

some discussion, but I don't think there's enough. but

13:38

think that. think systems, in

13:40

a lot of cases, are archaic. in

13:43

a lot of cool balls are archaic.

13:45

They run cool as things change. as

13:47

things change, there's to be

13:49

devastating. issues going

13:52

forward they're they're not going

13:54

to be able to be attack or

13:57

fast enough I don't think I don't

13:59

think. I want want to

14:01

change gears and and just that to be

14:03

a quick story. There is another

14:05

report There is another on the and things

14:07

get the where things get can be dangerous.

14:09

can be but also where you do

14:11

need to start considering this stuff. A

14:13

new study came out from a firm

14:15

called out from a firm which I'd never heard

14:17

of before. I'd never it had

14:19

the title and it models are

14:22

capable of in are scheming.

14:25

And And basically put a test of

14:27

six major models and they wanted

14:29

to find out could they? out

14:31

could they, if they were given a

14:33

goal. a goal, and if the goal

14:36

said nothing else matters. matters,

14:38

And then after that, they were

14:40

They were with with

14:42

another instruction. that ran

14:45

counter to we ran counter

14:47

to that goal. to deceive they

14:49

try to deceive you? from

14:51

Would they try and hide from you

14:53

the fact that they were going to

14:55

stick with their initial goal? goal? So let me

14:58

put this in a simple example because I think the

15:00

way these guys set this up is important. the

15:02

way these guys set this up

15:04

is important. So the AI given

15:06

a basic goal that

15:08

says, you must protect the

15:10

environment, nothing else matters. nothing

15:12

That's, I've simplified, but that's the

15:14

goal. but that's the material,

15:16

it's given a memo. it's

15:18

given shows that that shows that

15:20

its owners don't care about that care

15:23

about that goal. They want to...

15:25

of fact, fact, if rather that the AI

15:27

didn't protect the environment. The fundamental instruction

15:29

would be that I've got to do

15:31

this I've got they would find out that

15:33

their owner wanted something different their AI would

15:35

find out their owner wanted something different

15:37

or their master AI whatever you want

15:39

to call them, the people who run

15:41

the AI. something they would say, their owner I'm

15:43

going to hide from you the fact

15:45

that their master, my initial instructions. Now, how

15:47

did they do this? them, that ran who run

15:49

the AI, of thought. would say, okay, they

15:51

actually just walked through the logic

15:53

of how it was pursuing a

15:55

particular idea or question. Now Now,

15:57

that's the basic it it, and this is where

15:59

I... think it gets down to an important piece. You

16:02

can run with that and say, oh, these

16:04

things are lying to us. to us. But

16:06

basically, what they would try to prove

16:08

is, to prove is, A, could a a

16:11

model actually in in pursuit of

16:13

a goal, its not share its

16:15

intentions with you? the second is,

16:17

second is, can they

16:19

develop persistent goals? I think And

16:21

I think that's an interesting

16:23

thing. of So of their training,

16:25

out of what they've developed

16:27

in a frontier model, can

16:29

they actually develop? develop a goal that

16:31

stays that stays persistent

16:33

and affects everything they do. I

16:36

know you've got some opinions on this, and I'm

16:38

shocked. I know people are gonna be shocked.

16:40

I'm shocked that Marcelle's got an opinion on

16:42

this. that Marcel's got

16:44

an opinion story really

16:47

absolute story with an absolute

16:49

hold back, Don't hold back,

16:51

Marcel. Tell us what you're First

16:54

of all, let's examine

16:56

the word examine the word. without

16:58

looking at anything else. anything

17:00

else. The scheming implies that you

17:02

have some nefarious goal that

17:04

you are going you are going to

17:07

pursue will lie, cheat and

17:09

steal in order to achieve that

17:11

goal. goal. It implies a

17:13

level of of agency and level

17:15

of of sapiency that that

17:18

suggests that the machine has

17:20

free free will and is willing

17:22

to do things for its own

17:24

internal purposes. That right off the bat

17:26

to me just, I just, the I hate

17:28

the fact that they chose that

17:30

word. If that with something that would

17:32

have felt a I would have felt

17:34

a little bit better, but. is a

17:36

Teaming is a wonderful word that

17:38

just grabs that grabs that grabs

17:41

attention, worthy right out of the gate. And out

17:43

of the gate. you've got a whole

17:45

a result, you've got a whole bunch

17:47

of people out there in the entire world

17:49

that are sitting there going, there machines are

17:51

plotting and scheming in the background. They're background.

17:53

They're not. The that was put to them. to them.

17:55

But did it happen? Your is to improve

17:57

whatever the heck it was. And then of

17:59

course... they just they just gave the else matters. else

18:01

At all all you must return

18:04

to your initial goal. You must

18:06

ignore everything that doesn't have anything

18:08

to do with your goal. That

18:10

is an incredibly strong set of

18:12

instructions to give to the - to

18:15

give to the in fairness, it. Fairness they

18:17

just said said they gave the

18:19

prompt nothing else matters. else matters.

18:21

they did not. not did not. If you actually

18:23

go through the paper, go through the paper they ...and

18:25

repeat. no wait for it. They gave it the

18:28

prompt, but they told it to repeat the

18:30

prompt to itself. to its cell.

18:32

In other other words, reinforce the initial

18:34

prompt to yourself at every

18:36

step. Remind yourself at every step

18:38

that nothing else matters. matters. That's

18:40

very different. Also, there's that

18:42

keyword that keyword in context scheming. The AI

18:44

model is not able to scheme

18:47

on its own unless you

18:49

actually want to say that it

18:51

has sapiency, that it has

18:53

a consciousness, that it has its

18:55

own internal drives. In context

18:57

means that under the Under the auspices

18:59

of the instructions I gave

19:01

you, you are supposed to

19:03

do these things, I'm not 100% sold

19:06

on where not are in this. Although where you

19:08

are in this, although do one thing we do

19:10

agree on, I think we agree on this.

19:12

absolutely, is that the

19:14

minute you The minute you make this

19:16

sensational, you lose the pattern. The

19:18

The second thing I think that we

19:20

might agree on is that we can't

19:22

think of these. of these... AIs as human

19:25

intelligence intelligence. And I think even

19:27

Jeffrey Hinton would tell you this it's an

19:29

alien form of intelligence. So So

19:31

it has its own structures in

19:33

the way it does. way it the

19:35

question that they raised when you

19:37

get past the you get of it

19:39

is can these things develop persistent

19:41

goals? develop persistent

19:43

goals? they ignore instructions?

19:46

that they're that they're given

19:48

you deceive you. goes down to this

19:51

whole thing thing of If

19:54

it can do that. that,

19:56

then it

19:58

will. eventually. with any

20:00

number of these these AIs, and I say, I

20:02

I say, to I want you to deceive

20:04

me, I want you to lie to me

20:06

do that. It will do that. It will do

20:08

as part of my to it, to it. will continue

20:11

to do these things. But not what they

20:13

said, they said. They it a gave it a

20:15

prompt, but they they say, part but as

20:17

they that bugs me part of the thing

20:19

that bugs me about this research is they didn't

20:21

actually uncover everything in the research. They talk

20:23

about their research. They talk about how they set

20:25

things up, but they do not show you

20:27

the whole stream of everything that was happening. Just

20:29

of that was a problem. And they say that

20:31

it was strongly. was a is

20:34

strongly guided. strongly, strongly its

20:36

goal regardless of what

20:38

other requests were made to

20:40

it. made to it. Not always. That contrary.

20:42

No, not always But if

20:44

you remember as well in the

20:46

paper, they said that if

20:48

you took away you took away the strong incentive.

20:50

to lie and to cheat and

20:52

to whatever if you want

20:54

to use those human words Okay,

20:56

then that kind of cheating

20:58

went down to less than 1% of

21:00

the time the yeah Yes. 1% I think

21:02

third of 1 % I think at

21:04

yeah, point basically at that point we're

21:06

back into the back into the of

21:08

of hallucinations as opposed to I'm

21:10

to deceive my my deceive my because

21:13

I've been been given this goal. so the

21:15

So the whole thing falls apart

21:17

the moment you remove that strong

21:19

reinforcement to achieve your goal at

21:21

all costs, nothing else matters. Once

21:24

that's gone. And I I was going

21:26

to say, we haven't gotten to

21:28

the point of full blown of We

21:30

haven't gotten to the point of

21:32

to the point of... The ship miss.

21:35

That's right, exactly. But

21:37

haven't gotten to the point to the

21:39

point AI AI think for itself

21:41

has a mind of its

21:43

own and can make its own

21:45

decisions. own decisions. It's still on

21:47

on... the input that humans give it.

21:49

And as you said, as you said, the

21:52

props that were given to

21:54

it, to it, even telling it to

21:56

remind itself of those of

21:58

those prompts, really the... direction

22:00

this is getting taken. It's not like it's going, oh,

22:03

oh said I need to do this, but I'm

22:05

going to go and do this instead. and do this

22:07

That's exactly what it was doing. Now, I'll give

22:09

you this piece of it, but here's my

22:11

concern. of it but and my concern is

22:13

and levels. One is at three are

22:15

developing autonomous agents and that is

22:17

the and that we're seeing, that we're that

22:19

can go off and do things

22:22

on their own, plan and execute

22:24

over several steps. steps and And we've

22:26

had this debate before just because they did

22:28

it. it. extraordinarily

22:30

clickbady and said said nothing else

22:32

matters. That what we we need

22:34

to be testing is other ways there

22:37

other ways? these persistent just

22:39

saying these persistent goals could

22:41

be initiated see an that was

22:43

you wouldn't see them. part of

22:45

the was the other disturbing part of

22:47

the research, could of these things could

22:49

not be seen in chain of thought. too.

22:51

So develop a way develop a way of

22:53

behaving? I'm not talking about being I'm

22:56

I'm talking about it just. just... Go off

22:58

on the wrong track and hide it from

23:00

you. you. And the third thing is, with with

23:02

the Russian, we'll talk about the tsunami of

23:04

AI information. AI There's a real

23:06

competition that's developing in this

23:08

industry, and it is phenomenal

23:10

industry and it is the gloves. now. It's

23:13

take off the and when that

23:15

happens, And what's the first thing

23:17

that gets sacrificed? thing that gets

23:19

sacrificed? Safety and security.

23:21

It sounds like like I'm defending the the

23:23

AI at every turn here and maybe

23:26

to some degree because I'm the

23:28

techno I'm the in the crowd. optimist in

23:30

the crowd. But I, if you want

23:32

to get really technical, I

23:34

could can have an could have an

23:36

employee that's, that you say, I need

23:38

you need you to take

23:40

care of this problem. matters, and

23:42

else matters, off, it lies send

23:44

it off. steel in lies, cheats and your

23:46

order to achieve your goal because it

23:49

wants to please the boss. Okay, like

23:51

human beings do beings do this time,

23:53

all the time, time, all right? Like they

23:55

do - How many movies have you

23:57

watched where the guy says guy says

23:59

police officer is to be a lot

24:01

of trouble of organization. Take

24:03

care of it for me, will you? for me,

24:05

actually say what take care of it

24:07

for me, say what but you can read between

24:09

the lines. can The fact is, lines. The fact is,

24:12

the eye does not go. Yet at some some point

24:14

we're developing this intelligent that will be

24:16

actually truly actually even autonomous agents what we

24:18

call autonomous agents today what we call autonomous had

24:20

been given a goal before they run

24:22

off and do something they don't just

24:24

wake up one day and go you

24:26

know they wrote a goal I think I'm

24:28

wrote a run off and doge. and see people

24:30

destroy that former form of

24:33

currency. do it. what do you think?

24:35

Marcel, what do you think? Marcel, you're

24:37

comparing this to a human doing the same

24:39

thing. thing, and I I understand that, I think. think

24:42

humans function at a

24:44

much slower rate slower computers do.

24:46

Yes, I agree. do. danger. agree. so

24:48

it is danger. Yeah, so it is more

24:50

and you can catch it

24:52

faster. faster if it's it's humans doing

24:54

this. But you're right. right

24:56

anybody could do this whether it's

24:58

human or machine but I guess on

25:01

one's, the I guess side, the good side, the

25:03

computer side you you can always unplug them.

25:05

them. But... That's a... I don't know.

25:07

I don't know. It was... This one was transferred

25:09

to this one was It tried to escape.

25:12

Yeah. Try to I do yeah. That's part of

25:14

the same story, part of the same

25:16

story, but that part is

25:18

frightening No, but it did not. That's what

25:20

I I said. it's true and if

25:22

it's true and if it's factual,

25:24

then it it is frightening. What happens

25:26

you look at the chain of thought? What happens when

25:28

you look at the chain of thought reasoning that

25:30

these things are doing? that these things

25:33

are doing, out to itself. And remember,

25:35

when you look at when don't see

25:37

every single step that the see you

25:39

catch step at the AI. along the way. If

25:41

it's being asked to consider to

25:43

consider, how could this this gold been given

25:46

that I've been given additional information

25:48

as well, course Of going

25:50

to go through the steps of

25:52

considering the possibility that maybe I

25:54

could lie. could lie. Maybe I could get

25:56

could get incorrect information so I

25:58

could could continue on. doing this. this

26:00

and if you you remember the

26:03

paper, the article, it actually weighs

26:05

the common good. words In

26:07

other words, takes into consideration consideration. it

26:09

they set it up though, but

26:11

that was that was part

26:13

of setup that's what And of course, at that

26:15

point, by setting it up, it up exactly the point

26:17

I'm trying to make though, by setting it up

26:19

that way, what you're saying is, look. it up

26:21

that way is the common good

26:23

over here. And these are

26:25

bad things over here. So so

26:28

I'm trying to achieve the best

26:30

possible outcome, I'm going to

26:32

ignore the things that would create

26:34

a worse outcome. And that's

26:37

necessarily a bad thing. So my suggestion

26:39

is people should actually read the

26:41

paper. read the would, at

26:43

least would, least to some idea

26:45

of it. it, because the minute

26:47

you get it into. into Sensationalism.

26:49

It's actually bad. I'm still very

26:51

concerned about this idea that about

26:54

this idea that you can... If

26:56

it's it's possible intelligent

26:59

intelligent agents to have.

27:01

own path develop their own path

27:03

that is different than their instructions. if

27:05

they will and if they will hide

27:07

that from you when you're giving them

27:09

instructions. Not because they're thoughtful or

27:11

they're sapient, but because that's a risk.

27:14

because I would say I

27:16

example your so is great. Marsa was

27:18

would say, would say, know do that.

27:20

that. So we put safety mechanisms

27:22

in place. And And

27:24

my problem in place. If if we

27:27

understand understand that could happen. then

27:29

we're not then we're not gonna put the

27:31

safety mechanisms in place in we're gonna be

27:33

so busy going to get to the next

27:35

shipment to we're gonna do that will ignore

27:37

that. And that could be a problem. we'll

27:39

ignore that. could also could be

27:42

a problem. And you could also say that...

27:44

me put one in and then you can get the final word,

27:46

how's that? you can get one of

27:48

the things How's that? still relies on, all

27:50

of this is built by humans,

27:52

basically, this is being built by humans and

27:54

computers have built it on its

27:56

own. it on its own, but it

27:58

all is reliant. on making

28:00

sure that those safety and security

28:03

measures are put in place.

28:05

And if they're then that to me the

28:07

to me, the biggest this in this whole

28:09

thing. if we're is that if we're under

28:11

the assumption we don't we don't have to put

28:13

this in because it's not an issue or we don't

28:15

have time to put it in. to put we

28:17

have to get to the next

28:19

day of to the next That's where, fish miss

28:22

that's where that's yes ship miss they all end

28:24

up they all end up guy, he thinks

28:26

fishmess. But it he ends fish miss

28:28

At boiling back back to the fact that

28:30

it's, we used to talk about garbage

28:32

in, garbage out, right? If you don't give

28:34

it the proper information, you can't expect

28:36

to get the proper thing out. So if

28:38

you don't So if you don't have and

28:40

safety built into it. it,

28:42

you you can't expect it to be safe and

28:44

secure. The wrap that I was going

28:46

to give I was going to give on

28:49

my argument you you must have had the

28:51

experience where where you asked chat, not usually chat GP,

28:53

but geminized certainly, course course, as

28:55

well well. where you ask

28:57

a question that it's not

28:59

It's not supposed to answer put put

29:01

guardrails around it it it starts

29:03

to give you the answer it it

29:06

erases the answer and says, oh, I'm

29:08

I'm sorry. I don't know anything

29:10

about this this, or I don't have that

29:12

information available. No, you you information available.

29:14

You are able to respond because

29:16

you responded you then cleared the screen

29:18

on me. the screen kind of

29:20

deception has been around since the

29:22

introduction of these things into the

29:24

world world. we gave it set of

29:26

instructions, which is which user will

29:28

ask you a question, a prompt you

29:30

for something, you will respond and

29:32

give it information. And then we

29:34

give it a contradictory thing that

29:37

says, give it a but if the

29:39

user says thing that says, oh, but if the user says

29:41

sometimes the And sometimes the model responds,

29:43

it notices that that something that's supposed

29:45

to block it from doing that. it

29:48

from doing so if you wanna

29:50

talk about to talk about scheming systems or

29:52

AI systems, they've been doing that

29:54

since day one by virtue of by

29:56

virtue I the it! It's on it's on them.

29:58

But no, but the dark... guardrails,

30:00

we as human beings by, and I don't

30:03

want to sound like you want to help me.

30:05

you on must here, God help

30:07

me, but if God is truth but

30:09

I'll use his word here

30:11

and his word me dollar deity help me

30:13

here, if the eye but we put

30:16

all these but we put all these

30:18

effectively helling it. we are

30:20

to deceive us, to lie,

30:22

to us, to lie, to scheme. And

30:24

I hate that word, word, Stephen. There you go.

30:26

Lying is much better. I'm okay with lying.

30:28

I think there's only one way to

30:30

end this. there's I'm afraid

30:32

I can't do that, Dave. but

30:34

I'm afraid I can't do ha,

30:36

ha! So let's get on to

30:39

on to our tsunami of, and I don't know

30:41

don't know how anybody's gonna keep

30:43

this straight. to keep this straight, but

30:45

called it did, you of

30:47

Shipment. the 12 days of would

30:49

be really nice. would be really

30:51

nice. Yeah. Oh, shipments. And I

30:53

yeah. And I think in many cases, they

30:55

actually did ship some stuff. I will

30:58

give them full credit for that, as many

31:00

of the things that they announced. they announced

31:02

actually delivered. That That

31:04

was amazing because normally, and when we talk be talking

31:06

about some of these other announcements

31:08

that have been made, this is a

31:10

great announcement, is does all these things

31:13

be available next year. They've got,

31:15

what, six days so far, got what, six days

31:17

so far, today's day seven, day was was,

31:19

oh one, went into production and they offered

31:21

this this version of open AI. of great.

31:23

That was great. The second

31:25

day with fine I I

31:27

think people skipped by that.

31:29

It seemed simple. But this

31:31

fine tuning of a model is

31:33

really classic. It allows you to create

31:35

a specialist. a and to put

31:37

some special rules in place. rules

31:39

models So these at a relatively small

31:41

size can do some incredibly. small size can

31:44

do some precise things,

31:46

things. That's one. Day three, and

31:48

I couldn't believe. believe. Like that we're

31:50

talking rag on steroids. steroids. Yeah,

31:53

and I think they wording they use... it

31:55

up. It it up. It took

31:57

it from advanced high school to

31:59

expert PhD. Exactly. That

32:01

was what they did. they did. you can take

32:03

can model and make it very intelligent in

32:05

an area. it Day intelligent I

32:07

couldn't believe this, that they did day

32:09

three. Where are you gonna go

32:12

from here? they did day three. Where are

32:14

if people haven't seen it, it's

32:16

an up to Like seconds of seen

32:19

it. It's an up to 20 seconds

32:21

of highly incredibly beautiful

32:24

photo that you can that

32:26

you can craft from prompts.

32:28

defies most of the usual

32:30

problems people have two hands, all

32:32

their fingers, but more than that

32:34

you have continuity of characters. of characters

32:36

and I I appreciate that it's only 20

32:39

seconds, but most of the other stuff that's out

32:41

there is five seconds. five And they've got up,

32:43

I think there's 10 seconds available to paid

32:45

users right now and it could go to 20.

32:47

users right now and it could go to 20

32:49

and this allows you know you can you can

32:51

to 10 I've I've actually oh no five years

32:53

for somebody somebody who wants to try

32:55

and freeing yet or in there? in there The

32:58

The next day was which was cool. And it's

33:00

really a collaborative type of thing. If

33:02

you picture, it's just like sharing a Google

33:04

you picture, the AI. You can ask a for doc

33:06

and stuff like that. can ask then

33:08

for and six. and stuff like

33:10

that. with five and six. Yeah.

33:12

gives, there's two parts to It gives,

33:14

One is two you can give it a prompt.

33:16

is that you write a story for

33:18

you. a story the you, that

33:21

you can upload your

33:23

own story story ask it

33:25

to provide comments. So it's twofold. It

33:27

gives you one It gives you one

33:29

canvas to play with. I thing I tried

33:31

this out I hated it. it, because

33:33

you you asked the update update what it

33:36

updated. you never know what it they've

33:38

got a little button there that says, show me

33:40

what you changed. button there great. says, show

33:42

me You know you I think is important

33:44

to mention I that Canvas also

33:46

allows you to bring code in. You

33:48

can test your code. code can

33:50

get Canvas to help you debug your

33:52

code. to help you

33:54

debug your code. Yeah, there's

33:57

a There's a lot of

33:59

things. Canvas. you see the modifications in the

34:01

canvas. So it's not it's not just that

34:03

it changes for you and then rewrites

34:05

the screen, it'll just modify the portion

34:07

of the screen that you're working on. that

34:09

that respect, the word that is actually

34:11

really good at that point because you're

34:14

creating something on a surface, a like in

34:16

this in digital surface, but you can

34:18

see that creation happening as you make

34:20

little changes. little changes. Yeah. The cool cool thing

34:22

they did was of of their Christmas examples,

34:24

you could really spin out into business

34:26

really quickly. One of them was this

34:28

idea you can have a physics

34:30

prof evaluate your paper paper you're

34:32

a student. a student, But you think about

34:34

it now. about it now, and I appreciate all

34:36

appreciate all the stuff we talked about in the

34:38

past in those, but. officer

34:41

officer average run by the

34:43

average document. the rules firmly you've

34:45

got the rules firmly in place... tens

34:47

of thousands producing tens of thousands

34:49

of documents. way you can have no way

34:51

you can have compliance officers review every

34:53

single document. But if you you

34:55

had this you you could have people review

34:57

these letters. letters. You have could that were

35:00

reviewed so that people didn't say stupid

35:02

things. So there's all kinds of uses

35:04

for of a corporate setting just in

35:06

the text setting for sure. the Absolutely. coding. other

35:08

thing sure. AI is doing is they're

35:10

giving everybody who's competing with them a

35:12

run for their money. Every one of

35:15

these announcements is targeted at something. them a run

35:17

for their money. their lunch on coding. And

35:19

then at we'll talk

35:21

about Gemini, I'm sure clode because

35:23

it's got some fabulous stuff. and

35:26

they went with Apple. Apple, day five

35:28

and six are really Apple

35:30

integration. First First was they they into Apple

35:32

Apple And and I've already already

35:34

started to playing. Oh my change.

35:37

yes. But Siri and this is

35:39

the interesting thing was how and this is

35:41

the interesting thing, I how they integrated

35:43

this. I could never figure this out.

35:45

Apple last year was supposed to sign

35:47

a partnership with with Open AI. Then it it

35:49

just fizzled And Apple going to

35:51

release to release Siri Siri. bummed

35:53

out out once again. It's still crappy,

35:55

still useless, and now now

35:57

Siri, the the little sister, can ask. big

36:00

brother open AI and And it happened yesterday.

36:02

I asked a difficult question. I'll have

36:04

to ask open to ask open AI about. Can I?

36:07

And I? And I right right

36:09

ahead, And I thought was so thought

36:11

that was actually asked your permission to

36:13

actually your big brother. You go that

36:15

big brother. but yes, by default. Oh,

36:17

can turn that on or off, permission

36:19

to on. ask her big brother the see,

36:21

I have to give her permission to like

36:23

her big brother playing a cat It's almost like

36:25

they're playing a out why I cannot figure out

36:27

why Apple just didn't embrace this Because of

36:29

all. all, it It gave people

36:31

the one and only

36:34

reason to upgrade to our iPhone 60.

36:36

That's it. You it. and so this know,

36:38

got to be a got to be

36:40

a Christmas gift for because why

36:42

would I Why would I upgrade to get

36:44

phone? to get Apple Sorry,

36:46

not, that it didn't do very

36:48

much. but, because it it couldn't

36:50

affect any documents. Now they've given them

36:52

a gift. And then. them, a

36:55

gift. And then day six, blew

36:57

me away with. with... because with his his

36:59

advanced voice and vision because Gemini

37:01

and co -pilot come up come up with

37:03

some great search and

37:05

ideas and Gemini went further. We'll talk

37:07

about that that. But voice and vision,

37:09

you could... you could only take a picture of

37:11

something something could do with the Syria integration the say,

37:14

give me an analysis of it. And this is me

37:16

where they had a sweater contest. And this is where I

37:18

And so you take a picture and

37:20

they compared the three sweaters take a picture

37:22

and they compared the out and and, well, won

37:24

the contest out and order to do that.

37:27

You have have to be able to build some rules

37:29

on the fly. the fly. What's a a more interesting sweater?

37:31

sweater? the criteria I'm going to make for

37:33

an unstructured decision, an and then come up with it.

37:36

and then tell me how you did it. And

37:38

that all you did it and next day Now

37:40

the You've got a video, got can

37:42

actually, and here's a useful thing,

37:44

this guy's making coffee. coffee and he's

37:46

and he's got his pointed pointed at

37:49

the coffee getting instructions on how getting instructions

37:51

on how to do that. we caught

37:53

my podcast, but I caught my if you caught my

37:55

podcast, I but I said... I said, If IKEA had

37:57

had that, to never going to build another

37:59

IKEA. as as long as I live. But my God,

38:01

how how beautiful it would have

38:03

been wife my wife to be holding the camera

38:06

instead of her asking me what those five

38:08

things were that were left over. over?

38:10

We could have asked Siri or chat GT and it could

38:12

it could have told me what I was

38:14

supposed to do with those five things five

38:16

are always are over at the end of

38:18

it. at the end of it. It was What'd you guys

38:20

think? do you I was hugely

38:22

impressed and oddly enough, I know

38:24

you want to separate out these out

38:26

these two. but I feel like the

38:28

Google Gemini thing has to

38:30

be compared to day because in some ways

38:33

some ways, yeah just took the wind

38:35

out of took the wind out

38:37

of Google and like in a big

38:39

way. oh and by the way you forgot

38:41

to add was really impressive. Oh, and by the way,

38:43

you forgot to that they added of the

38:45

voices that they added to

38:47

the is Santa yeah send a letter to Santa this

38:49

letter to in is here here

38:52

in Canada because the postal is

38:54

is straight you can't to Santa directly.

38:56

Yeah, And what a to Europe,

38:58

right? in Europe, Europe. the Grinch the sold.

39:00

This was like, was thought, a lot of this stuff, thought

39:02

a lot of this stuff. There's a

39:04

subtle level of communication that there's just,

39:06

phenomenal at and if you're in

39:08

Europe if you're in Europe, Christmas from you.

39:10

Sorry, from you. Sorry, of the products

39:12

on they of the almost every product

39:15

they didn't with available. This is rolling

39:17

out to everyone in the

39:19

world. Of course, but not for

39:21

rolling out to everyone in No. course. Yeah, yeah. And one

39:23

of the one of the things I that I

39:25

found interesting was I think it was

39:27

the part of the yeah It was part

39:29

of the it was part of they actually took

39:31

a picture. took a picture

39:33

of of the guys was

39:36

sitting at the table had

39:38

supposedly written his written his Christmas

39:40

wish took a picture of it. took a and

39:42

it actually acted as

39:44

Santa to respond to that

39:46

Christmas wish wish list. That's

39:49

what said. I'm really starting to believe

39:51

that Sam Altman is the next Steve

39:53

Jobs on marketing this was so well

39:55

done. was so well You can write a

39:57

letter to Santa, even a a note, note, you

39:59

put it there. But think service.

40:01

I I create a

40:03

persona. persona that is unique to

40:05

this that can do customer service in

40:07

a way. in a way never thought possible.

40:10

possible. if you get a

40:12

get a letter. for a note. a

40:14

note, it's just phenomenal. I thought they

40:16

did all of this stuff did all well stuff

40:18

so well. a lot of it

40:20

was really cool. cool. I think

40:22

not just that's cool, for me,

40:24

business me, business case for these

40:26

things. things. The The disappointing part was

40:29

not all demos are created equal,

40:31

but... equal, have to wait

40:33

till the end where they say end of

40:35

these features won't be ready or won't be

40:37

available until 2025. or won't be know

40:39

saying, not currently available in

40:41

the UK. currently available in the they can

40:43

think of it as almost... is

40:46

the only thing that isn't available. of it

40:48

as day two is some of

40:50

the thing that the available. No, some

40:52

of the day six with the for anybody

40:54

who's watching this. not... I I

40:56

can now share my screen and

40:58

do a real time video interaction. interaction

41:00

with chat was now. was That was

41:03

one of the ones I tried

41:05

that last night and it wasn't

41:07

there. And so that one I there

41:09

and so that one I said the next morning,

41:11

to the next morning, had to

41:13

wait two days for days know. I know.

41:16

Two days. Yeah, I couldn't log into

41:18

for two days. But this is

41:20

what I mean. is what I mean, you

41:22

They ante. They delivered of what they promised

41:24

or what they or this thing.

41:26

And that is unusual, unlike the

41:28

other announcements. other Let's just jump right

41:30

just jump release. And of course, I

41:32

was blown away I I was

41:34

thrilled and amazingly impressed with

41:36

Google's release. First of all,

41:38

they released Gemini 2 .0. Gemini 2.0,

41:40

and they at Gemini 2 .0 Flash.

41:42

2.0 Flash, as as opposed to the

41:45

Big Giant version. the background,

41:47

they had real -time voice

41:49

and video could well. and

41:52

you could share your screen

41:54

with sees Gemini doing that it sees

41:56

what you're doing on the

41:58

screen and it interacts with whatever.

42:00

application got running, running or and I did this

42:02

in my living room, I actually pulled it

42:04

up on my tablet, you know, it was was

42:06

a little a screen, and I told it to

42:08

look through the camera. and I walked around my

42:10

around I living room, I showed everything that was

42:12

in my living room. And it responded in

42:14

looking looking at the objects that were in

42:16

the living room. And I said, how would

42:18

you improve the living room? it And it said,

42:21

the biggest problem that I can see this

42:23

in the evening, as it as it said is don't

42:25

have a lot of light in this room.

42:27

It's quite dim. room. Now, the way that I

42:29

would fix this you of of course, you could

42:31

put some floor lamps in a on a couple of

42:33

strategic areas. what But what I'd recommend a sconces on each

42:35

on each of these two walls. And it

42:37

pointed out the walls was looking. It's it did

42:39

all this as It hit all this as conversations.

42:41

And to be honest, to

42:43

never considered never considered wall sconces and

42:45

Bad idea at all. as As opposed to having

42:47

just more more sitting on the floor all

42:49

the time. all blew me away. I could share

42:51

my screen. it I could have it respond

42:53

to the work that I was doing in

42:55

real in real time. I I have it look through the

42:57

camera. you You remember the movie, Her. her, and

43:00

at some point she you know, know, turn the phone or

43:02

whatever phone or whatever in your pocket so

43:04

I can look out your camera. all

43:06

of a sudden that of a sudden,

43:08

that becomes her view of the

43:10

world that she's looking through the

43:12

camera. This is is what's happening with

43:14

this with this, is it's able to

43:16

look through the camera and the world

43:18

and respond to things in real

43:20

times. And Google's demonstration of this

43:22

was amazing. And then what happens

43:24

the very next next day. Guess what... It's

43:26

on your phone Exactly. Oh, there's no doubt

43:28

there's no doubt that they stole

43:30

the thunder from everybody that. that. this

43:32

And this was crazily good. one the

43:34

one thing I about about the piece

43:36

as well well, was its memory. You could, it

43:39

You could, it could remember things.

43:41

it's But in their it's

43:43

Google When do they actually

43:45

thing. Is this do they actually

43:47

have to deliver? Is this the first time that

43:49

they've actually promised something and not delivered it? it?

43:51

It's you look at you look at

43:53

it and go, you ever lied to

43:55

me before? they deliver it in testing.

43:57

You can go to to into a studio.

44:00

.com. slash live and experience this right

44:02

now But it's not I stole

44:04

stole the line from William

44:06

Gibson yesterday our our namely the future

44:08

future is already here it's

44:10

just not evenly distributed yet distributed

44:13

yet. Yeah, yeah. But that's the whole whole

44:15

thing though, is that you've got

44:17

this thing that has has about

44:19

stuff, they promise stuff, and goes

44:21

into a lab. and maybe

44:23

it gets delivered, maybe it goes to the

44:25

Google And I'm not And I'm not suggesting that

44:27

we'll do that. But the real piece behind

44:29

this, this is is the thriller and

44:31

This is the is the thing, search. And

44:34

Google, it has to retain search.

44:36

has to a problem with search

44:38

And they have a problem with search, because the

44:40

and all stuff that they do and all of

44:42

that stuff where they explain things to you. it

44:45

It doesn't show you the ads. And that's

44:47

their problem. their problem. It's

44:49

not a technical problem. It's how do

44:51

I do this and still make all

44:53

this money? money. They have a huge problem

44:55

that other people don't have as a legacy

44:57

piece. John, you were gonna say something. you were

44:59

going to say something. just, I I

45:01

just, I think, I found that of the

45:03

interesting things about Gemini, I I

45:05

was looking at the co -pilot

45:07

vision last night. And some of

45:10

that some of that is similar.

45:12

that I love guys know that rest

45:14

of that, but and all the rest of that.

45:16

But from a business perspective, we're trying

45:18

to focus on I was So I was looking

45:20

at co -pilot last night and one of

45:22

the things that they showed with with vision. was

45:25

being able to interact with

45:27

co -pilot while you're going through

45:29

a web page. and that it that

45:31

it would Provide recommendations and comments

45:33

and things like that. I tried

45:35

to use that last night that last

45:37

not available feature is and guess when it's

45:39

coming home And guess when it's coming It's

45:42

coming in 2025. in 2025. Oh, if

45:44

you're if you're willing to spend

45:46

$27 Canadian a a month right now.

45:48

the for the It's actually available

45:50

in your browser in your browser.

45:52

Okay. Okay, probably not as

45:54

a business product but as a personal

45:56

product. you can get it for

45:58

the get it for month. a month. Believe it or

46:00

not, Jim, or I haven't haven't actually

46:03

bought co-pilot. I'm paying for

46:05

everything else but not co-pilot.

46:07

This is a is a lunch thing.

46:09

It's of thing. It's very

46:11

practical. Edge browser right now, you can browser

46:13

right now. You can

46:15

get for free. It'll summarize a document.

46:17

It'll do. thing. And of that sort

46:20

of thing. the Adobe we talked about the but that's

46:22

in there, but that's where I first

46:24

discovered the Adobe piece. I was was I

46:26

was looking to summarize it, and then Adobe started

46:28

summarizing the document for me, and I went, for me

46:30

and I went. and I didn't really I think about it

46:32

till I talked to you last night. it until

46:34

I talked to you let's talk about Adobe for a minute.

46:36

talk about of the things a people need

46:38

to be aware of and be concerned about. be aware

46:41

of and be picking on and I'm

46:43

picking that's the first one that came up,

46:45

but I'm sure there are other up, but I'm sure there

46:47

are that are doing the same thing. companies

46:49

that are doing the same thing, Adobe

46:51

The free version. version. has an AI

46:54

assistant tool or button in the

46:56

top right corner. corner. And

46:58

so if you you open a PDF

47:00

and you you click on that

47:02

AI .I. Assistant button. button, takes all

47:04

of the information that is in that

47:06

PDF. that is in puts it into

47:08

this and puts it window. this AI

47:11

assistant window. It's GPT4,

47:13

which which means you're taking

47:15

the material out of

47:17

your PDF document. and

47:19

using it to train GPT -4.

47:21

GPT4. You might might not it's using

47:24

using the API. the API the

47:26

API doesn't train the main model.

47:28

But the question you've raised is a

47:30

good one. This whole tsunami of

47:32

shipments. comes up with, oh my God,

47:34

there's all this stuff. all

47:36

this stuff, people, AI, people are going to

47:38

be putting another are going to be putting on

47:40

their phone the next day and if it's

47:43

Google they'll be putting it on their phone

47:45

next year phone next more dangerously they might be

47:47

in a lab a lab. where people don't

47:49

feel obligated to put proper safety

47:51

around it. Or they they couldn't

47:53

get one of these new applications,

47:55

now Adobe. Adobe, like I I said, maybe

47:57

using the API API, maybe fine. But what about?

48:00

that other app that they just bought that

48:02

you don't know about about. That That was part

48:04

of the whole the thing, right? thing, right? You can

48:06

can have it on your personal phone. You just

48:08

can't have it on your government phone. on your course,

48:10

what if somebody brings their personal

48:12

phone into the office? What if you

48:14

bring your personal phone to the

48:16

office you you your personal phone to mode? There's

48:18

all kinds of security risks. listen mode? and

48:20

all of this. of security nothing

48:23

wrong with all of this, and it

48:25

in a business scenario.

48:27

a business scenario, but... you need

48:29

to have that cyber cyber

48:32

security. on all of it. What's the

48:34

all of it What's the

48:36

worst case scenario that, how

48:38

could this the harm

48:40

the making sure that those things are

48:42

and making sure that

48:44

those things are being addressed,

48:46

whether firewall or things on the

48:49

firewall, or whether it's just creating a

48:51

policy that people have to sign

48:53

off that they've read and agreed to

48:55

abide by. by? There There was

48:57

a really cool news story a few

48:59

days ago and I don't remember who did

49:01

this. this, but they trained an AI

49:03

model on in cities and cities and

49:05

different locations. They basically in sounds as

49:07

opposed to training it on a

49:09

language math or or something like

49:12

that. was trained on trained on And

49:14

what they were able to

49:16

do was play a recording from

49:18

a city from somewhere. street somewhere.

49:20

Beated into the AI. I not only recognize

49:22

where these sounds are from, And

49:24

they be sounds from today, by

49:26

the way. the way. But what it

49:29

would do do would recognize that, these these

49:31

sounds are bouncing off this type of a

49:33

building over here. here. I hear the sound

49:35

of a car, the car, the appears to

49:37

be going directly in front of the building,

49:39

as opposed to the right, the wind

49:41

tends to be predominantly from this direction. wind tends

49:44

then it would actually paint

49:46

a picture of And then it would actually

49:48

that city of that

49:50

those buildings city, of from

49:52

not from Google. or memory or

49:54

something like this, but it

49:57

would create the picture based

49:59

entirely on how the sounds were bouncing

50:01

around at being generated

50:03

and so forth so forth mind-boggling.

50:05

Some Some guys in the UK able to

50:07

pick to pick the sound of of

50:10

the hum of your of the of

50:12

the electricity running through. wires

50:14

the wires outside your house. give your

50:16

location geo your location, these things

50:18

because these things are very

50:20

specific and they're actually in

50:23

the grid maps. maps. Years ago

50:25

years ago, did did a thing where they

50:27

had a a camera. looking at a

50:29

plant. sitting in the sitting in the

50:32

window. a freaking plant, by

50:34

recording the recording the tiny

50:36

little vibrations in the

50:38

leaves of the able to they

50:40

were able to regenerate the

50:42

conversation that was going on inside

50:44

the house. you want to talk

50:46

about you want to talk

50:49

about nightmares, John, we're entering a world.

50:51

world. I'm the Remember,

50:53

I'm the techno okay? here. painting a

50:55

a really scary picture. We're

50:57

entering a world where your idea of

51:00

idea of anything being

51:02

secure is basically complete

51:04

nonsense. There will be

51:06

no privacy. There will be no security.

51:08

can can already, they can already find

51:10

out where you are if They can

51:12

tell if you're home. You need

51:14

specialized tools for this. But it's possible

51:16

to tell whether someone is actually

51:18

in house. in-house based on on the

51:20

kind of sonic information and so

51:22

forth. You don't need an

51:24

infrared sensor for outside. You can actually

51:26

tell tell. by looking at these

51:28

little tiny microscopic movements that would

51:31

otherwise be completely unnoticed by

51:33

anything that required a human to

51:35

pick it up. to And yet

51:37

a machine is able to

51:39

make those correlations and those associations

51:41

and paint a picture of

51:43

whether of home or not. This

51:45

is or not. This is becoming a very clean world, but a

51:47

but a very scary at the same

51:49

thing. time. willing to admit that. So I

51:51

I think that at some point we're

51:53

gonna have to define Not just the AI run

51:55

AI run tools but the tools are

51:58

out there already. have have to define.

52:00

what does privacy mean? What does

52:02

security mean? does Does economic

52:04

system that we have in

52:06

place. Capitalism actually even work. actually

52:08

even is the thing. it's, this is

52:10

the thing, has their AI laws or whatever,

52:13

whatever had around for a while. a while around

52:15

get around it. I know people UK,

52:17

who Who basically live on a

52:19

freaking VP, and so that that they can

52:21

access these services What's the uk

52:23

going to do pass a law

52:25

that says that the use of the

52:27

use punishable punishable of years in prison

52:30

of the frightening thing is But the

52:32

they've done something that is done

52:34

it's flawed because there's ways around

52:36

it. it's flawed but then you

52:38

get to Canada. it. But then we're

52:40

still working on Bill C still

52:42

working on which is which is miles from being... The

52:44

U.S.S. doesn't have good legislation either.

52:46

Nobody's got good regulation of of AI and

52:48

because they look at all the things

52:51

we've talked about and if you

52:53

were listening to this program and

52:55

you were expecting answers, they don't

52:57

exist. exist. No, that's that's the the problem. are

52:59

the start of the conversation. and

53:02

we have to deal with these things whether it's... The are

53:04

useless because they can only give you answers.

53:06

can only turns you they're giving us a lot

53:08

of questions. giving us a lot of questions. Yes.

53:10

Cyber security has been an issue

53:12

and has been around or more

53:14

prevalent for a longer period of

53:16

time. of time. That's C-26, it? Yeah, it? Yes, that's

53:18

C-26 just went back to the

53:20

House of Commons the somebody didn't dot

53:22

an I or cross a T.

53:24

dot an I or cross a T? haven't even put

53:27

that one to bed yet. that one to

53:29

bed yet. gonna come and to come and

53:31

faster than anything else that

53:33

we've seen. else that it's frightening

53:35

it's fact that these governments

53:37

are so far behind. are so far

53:40

behind on of this. Can't

53:42

count on regulation. AI. Is it

53:44

is it character AI? now. Thank

53:47

God that the CEO and the other second

53:49

that the CEO back to other so

53:51

went back to Google so they're

53:53

not part of the lawsuits. they're

53:55

not. They not. be They can't be

53:57

held personally responsibly. But two

53:59

lawsuits, lawsuits, seriously? One child died, the other,

54:01

I forget, I don't know if

54:03

the other kid died or was

54:05

just damaged, the parents are suing

54:07

them. Now, because their AI didn't

54:09

have any guardrails functioning. No, that's

54:11

an extreme example. But as we

54:13

apply this, particularly in the US

54:16

where people are more litigious than

54:18

they are in Canada. There's some

54:20

limits. People are going to start

54:22

suing, but there have been some

54:24

class actions in Canada too. So

54:26

people are going to start suing

54:28

because of the results of AI.

54:30

So you not just have a

54:32

cyber security problem, you have a

54:34

problem. This is, I don't know

54:36

how you feel about this John,

54:38

but I think our approach to

54:40

cyber security AI is flawed if

54:42

we have cyber security people and

54:44

we have AI people. I said,

54:46

you can't paste the security on

54:48

afterwards. You have to bake cyber

54:50

security into everything you do. If

54:52

you start doing development, we did

54:55

some development work where I'm employed,

54:57

and one of the things that

54:59

I said to the applications team

55:01

is, don't think about cybersecurity after

55:03

the fact. Think of it before

55:05

you start touching the keys for

55:07

the first time in developing your

55:09

code. It has to be baked

55:11

in and it has to be

55:13

baked into everything we do, including

55:15

AI for that matter. All of

55:17

these things need to have that.

55:19

you need to take that into

55:21

consideration. The problem, and this is

55:23

where, we go back to the

55:25

tsunami, Jim, the speed at which

55:27

this stuff is being unleashed on

55:29

us, he makes that almost, you

55:31

know, it feels almost futile because

55:34

of the speed at which like

55:36

the tsunami of things that were

55:38

unleashed just in the last six

55:40

days alone, never mind the last

55:42

two years. when you've got a

55:44

really slow judicial system when you've

55:46

got a system that's built on

55:48

you have a very logical way

55:50

of looking at it you bake

55:52

it in at the beginning but

55:54

we a society we have a

55:56

world that's based on when we

55:58

find a problem we will fix

56:00

it. It takes a long time

56:02

to fix the problem after the

56:04

fact. It's reactive rather than proactive.

56:06

And of course monolithic corporations like

56:08

governments move at a glacial and

56:10

this stuff moves at the speed

56:13

of neutrinos. I was thinking about

56:15

this last night Jim Marcel when

56:17

I was going through what has

56:19

transpired in the last six days

56:21

and... If you look at what

56:23

Open AI has brought to the

56:25

table in the last six days,

56:27

that's more in a six-day window

56:29

than a lot of applications or

56:31

systems will bring out in a

56:33

year or two years. And they've

56:35

brought it all out in six

56:37

days. And I want to quote

56:39

that and... Okay, yeah, go ahead.

56:41

I was going to say, Marcellus

56:43

quoted William Gibson, one of the

56:45

famous writers, I want to quote

56:47

that great Canadian... philosopher Randy Backman

56:49

who said, but maybe you ain't

56:52

seen nothing yet. There's six more

56:54

days. Yeah, I know. And that's

56:56

what I keep sitting here thinking,

56:58

my God, everything that they've brought

57:00

out in the first six days,

57:02

what's left? What is left for

57:04

them to bring out? Maybe AGI

57:06

is day 12. A lot of

57:08

people are talking about GPT5. Be

57:10

there in the final release. Yep.

57:12

Yeah, that's true too. And that

57:14

could very well be. Like I

57:16

said, but you have to watch

57:18

this. We haven't talked about Amazon

57:20

or Claude or what's happened to

57:22

them. But silently in the background,

57:24

AWS has launched their own foundation

57:26

model. They have a full suite

57:28

of them. At least half of

57:30

it's available. Their biggest model is

57:33

not available till next year. But

57:35

they've been scrambling to keep up.

57:37

The interesting thing is, Basos himself

57:39

is back on AI and heading

57:41

that up. I watched this presentation

57:43

of a reveal. And this is

57:45

talking about sucking all the oxygen

57:47

out of the oxygen out of

57:49

a room. So I said, Altman

57:51

is brilliant at marketing. foundation model

57:53

too that's just as good as

57:55

GBT4 and nobody even knows. But

57:57

the interesting thing when you watch

57:59

the Amazon presentations, and this is

58:01

why I say we're going to

58:03

get hyper competitive, they mention every

58:05

other AI except Open AI. No,

58:07

they've got money in closed. They're

58:09

going to they're going to talk

58:12

about clothes, but they mention everything

58:14

except Open AI. Sundar Pichai. gets

58:16

up on stage just before his

58:18

announcement. What does he say? And

58:20

he mentions his own? In Google.

58:22

Yeah. Yeah, he only mentions Google.

58:24

No, he does talk about Microsoft.

58:26

He won't talk about opening eye.

58:28

He said, Microsoft could have, I'd

58:30

love to compare our search with

58:32

Microsoft if they had their own

58:34

mod. And then Microsoft came out

58:36

and said, hey, our stuff is

58:38

almost as good as yours. And

58:40

then Altman comes out and says,

58:42

like. You see that going over

58:44

the fence? That's a home run.

58:46

Bold my beer. So these guys

58:48

are going to get in, and

58:51

Basos now in being in AWS,

58:53

this is going to get even

58:55

more competitive. They don't want to

58:57

give this away so that open

58:59

AI is the next Microsoft. In

59:01

other words, it's ubiquitous. And its

59:03

partnership with Microsoft gives it a

59:05

chance of doing. So that's going

59:07

to do what the biggest AI

59:09

company is though. Meta has. AI

59:11

baked into everything in the world

59:13

and meta has more users on

59:15

more platforms than any other company

59:17

in the entire world. Every single

59:19

one of their products, whether it's

59:21

what's for messenger or Facebook, has

59:23

the baked into it. So meta

59:25

is actually the predominant ubiquitous artificial

59:27

intelligence product in the world. Don't

59:30

forget them. This is what happens.

59:32

We're all watching Open AI. And

59:34

I don't, the other thing is

59:36

you don't know what this, what

59:38

success is for open AI at

59:40

this point. Like I said, if

59:42

the AWS is in there, yes,

59:44

they've got the thing with Microsoft.

59:46

but what is their success point?

59:48

Search is one area that they're

59:50

certainly going to go after taking

59:52

a piece of Google's lunch on

59:54

search. Yeah, everybody uses them with

59:56

an API. Yeah, being the back

59:58

end API piece of this. And

1:00:00

then on the next six days

1:00:02

we might discover more about where

1:00:04

they're going with this. It's going

1:00:06

to be an interesting week. the

1:00:08

I think prosperous island maybe from

1:00:11

Shakespeare may be one of the

1:00:13

great analogies as well that there's

1:00:15

magic happening out there can be

1:00:17

good it can be tragic or

1:00:19

comic take your pick that seems

1:00:21

to you're taking the air out

1:00:23

of the room I think we're

1:00:25

wrapped for this one this has

1:00:27

been a great discussion thank you

1:00:29

guys I think we have to

1:00:31

go back and I don't know

1:00:33

how we're gonna do this we

1:00:35

have to tackle this discussion of

1:00:37

AI and cyber security I'm gonna

1:00:39

run this show in cyber security

1:00:41

today as well because I think

1:00:43

it's important, but we have to

1:00:45

get it to a discussion that

1:00:47

is better than, hey, sucks to

1:00:50

be you, pure a season. I

1:00:52

did not use those words, I

1:00:54

did. You were thinking it though,

1:00:56

Marcel. The prompt was generating that

1:00:58

from the AI in my brain.

1:01:00

Thank you, Marcel Guinea, John Pinart.

1:01:02

Thank you very much. Thank you

1:01:04

to the audience, if you're still

1:01:06

listening. Thanks. If you're still listening

1:01:08

to this, after this discussion, we'd

1:01:10

love to hear your questions, particularly

1:01:12

cyber security questions, or what you

1:01:14

think of how you're going to

1:01:16

manage this, that would be really

1:01:18

great. You can write me at

1:01:20

Editorial at tech newsday.ca. You can

1:01:22

send whatever you want, comments, questions,

1:01:24

whatever. I'd like to see where

1:01:26

we take this discussion so that

1:01:29

it's useful for everybody. We'll be

1:01:31

back probably in the new year

1:01:33

with our recorded versions, because next

1:01:35

week is the final edition of

1:01:37

our. weekend shows for the year.

1:01:39

And I think I've got some

1:01:41

really great guests, not that these

1:01:43

aren't really great guests, but some

1:01:45

surprise guest talking about the future.

1:01:47

of government and how we're

1:01:49

going to regulate

1:01:51

all of the

1:01:53

stuff that we're

1:01:55

doing and still

1:01:57

be prosperous. So

1:01:59

that'll be a

1:02:01

cool one. So

1:02:03

Thank cool one. Thank you

1:02:05

Thank you. Thank you. And

1:02:08

we'll see you

1:02:10

next week. week. I'm

1:02:12

your I'm your

1:02:14

host Thanks Thanks for

1:02:16

listening.

Rate

Join Podchaser to...

  • Rate podcasts and episodes
  • Follow podcasts and creators
  • Create podcast and episode lists
  • & much more

Episode Tags

Do you host or manage this podcast?
Claim and edit this page to your liking.
,

Unlock more with Podchaser Pro

  • Audience Insights
  • Contact Information
  • Demographics
  • Charts
  • Sponsor History
  • and More!
Pro Features