Original subtitles

and I think one of the reasons that chat

gbt and stuff have been so hyped

recently is because most people don't

know what it is and so when you see it

doing what it does you think this thing

must basically be a person right because

it's acting like one and and I should

carry this by saying I'm not selling

short these these incredible

Technologies I'm just saying that it

would be very silly to just completely

use them blind and never check what they

do right because we know they just make

stuff up a lot of the time I'm glad you

mentioned computer science do you think

it's time to for more of us to learn

computer science type stuff because of

AI like maths and all these computer

science stuff not really

[Music]

I've been saying that you need to learn

artificial intelligence or AI question

that a lot of you have been asking me is

okay so how do I learn that so let's ask

another friend

yes you've mentioned this before but

remind me which place do you recommend

that I learn and others learn AI I

really like brilliant it's one of those

places where you can go and have a

visual gamified way to learn Concepts

and Mathematics behind Ai and machine

learning you've recommended this a few

times to me the way you said it was

David if you want to learn AI I need to

learn like statistics and stuff like

that right yes I've got these road maps

that actually helps you with Calculus

and learning statistics and linear

algebra all the stuff that you need to

know for AI I'll say this David is on my

team I'm really glad that he is David

has strengths that I don't have and I

think that's what's really important in

life you need to learn from others the

other will tell us you've done a lot of

maths you've done a lot of computer

science you've actually worked with AI

stuff right I work in the medical field

for with data science stuff so I really

think like you need to know know all the

statistics and calculus and linear

algebra and the discrete mathematics

that you need to learn which actually

makes a lot of coding a lot easier for

you that's brilliant so I'm looking on

their website now the one that you've

recommended that I go through is the

data science foundations right that's

like probability applied probability

statistics fundamentals and then an

introduction to neural networks and

obviously me being me I just skipped all

of that I went straight to learning

neural networks but as David said what I

really like about this website is it's

gamified as he said so really great way

to get started really want to thank

brilliant for sponsoring this video

brilliant as they say in the UK thanks

hey everyone it's David Bumble back with

Dr Mike pound Mike welcome thanks for

having me back again Mike though it

feels like the sky is falling again you

know we had this interview previously

and it was all this hype about AI but it

seems to just getting be getting you

know hotter and hotter so tell me is the

sky falling am I going to lose my job is

the future Bleak I think I think you're

gonna be all right so relax

oh it looks

um are you just bring me into calm

everything down a bit that's that's you

know I think that the last six months

particularly have been you know both

unbelievable in terms of genuine hype

like things that are really exciting

appearing and also obviously totally

overboard hype that's just getting

really quite silly and everyone needs to

calm down right so I think there's a bit

of everything going on chat chat GPT is

a is an incredibly impressive tool that

works very very well I've I've done some

really fun tests with it where I've

pushed to see what it will do and some

of the things it will do are quite

amazing right on the other hand there

are lots of things it doesn't do very

well and one of the big problems we have

at the moment is it won't always tell

you it's one of those things and that's

I think where we have something that

needs addressing I've done some tests

and I mean a lot of people I know have

done tests and it's a it's amazing what

it seems to be able to produce

um I think the concern a lot of people

have is like Mike I'm 18 years old or

I'm let's say I'm older I want to switch

careers to become a programmer or I want

to get into cyber security or want to be

a network engineer whatever some

technical role and it feels like I'm

just going to waste my time because chat

GPT is just going to obliterate drop at

jobs the first thing I would observe is

that it's very nice for some of these

big tech companies if that's the

perception because it makes them look

very very impressive right and so I

think that the cynic in me a little bit

is like this is there's no no pay R is

bad PR kind of a situation they like to

drop you know these tools get dropped as

incredibly impressive Tech demos and and

I'm not selling them short right very

impressive but maybe not quite as

impressive as they first appear on a

service when you start to dig in and I

think that's what's really important you

know in science we spend a lot of time

checking things and rechecking them at

least that's what we're supposed to do

right so I you know a PhD student comes

to my office with some results and they

say oh we've got 95 accuracy on some

task and I think okay let's talk about

which data you used and whether that's

really true and whether when you use it

on this new data you're going to get

that same result and we spend ages going

over and over the data again to make

sure that when we actually publish it

it's really as accurate as possible

large language models are maybe not

operating in quite that same way yes

they release papers from time to time

but mostly they release these big

websites where you can try them out and

they do incredibly impressive stuff and

and they lie very impressively as well

right I think that's the thing that we

haven't quite uh got around so you know

suppose you're a programmer and you've

been using co-pilot and you've been

using chat GPT also does code and you

and you're a bit worried because it's

just producing pretty decent code maybe

you don't see it replacing you right now

but you could see in 10 years maybe

that's going to be a problem I think the

problem is at the moment is it's very

difficult to know where it's going I

think a lot of researchers are

suspicious of the idea that we can just

make it continually bigger and bigger

and more impressive and it will just get

better and better you know when we talk

to about how these models work they

don't really have an internal model of

what it is they're trying to do or

anything really they just map text to

other texts you know when I write a

piece of computer code what I'm really

hoping to do is is in my mind come up

with an idea of the problem that needs

to be solved so what the number the

variables and things I'm going to need

start to get them down on paper and then

start thinking about how would I

manipulate those variables using code to

produce some result that I want chat GPT

doesn't really work that way it just

spits out code right and it happens a

lot of the time to look pretty good at

the moment it's a tool to be used quite

carefully particularly with code I

wouldn't push anything chat GPT has

written straight into production without

you know quite a few quite a few tests

because at the moment there's no

grounding in reality right the reality

is for training data but once it's

finished training you it's kind of

random what it gets and these things

actually I don't know if you've noticed

this David but when you run it it can

produce different answers each time yeah

and that's because it uses something

called temperature to somewhat randomize

its output so instead of saying okay the

next word in my in my output is going to

be the it will say I think there's an 80

chance that it's there but it's a 20

chance that it's so and then what you

what the machine will do is say well

okay 20 of the time then we'll pick a

different word and that way you can go

in slightly different directions because

if you didn't do that it would just

produce the same output every time it's

not it's not a random object so it's not

a random Network in that sense and so

you can imagine a situation where there

is a really good version of this program

that it could write but it randomly

didn't and produced one with loads of

bugs so you know assume that's what I've

experienced yeah sure yeah and and you

know I suppose there's a question in my

mind about how is very inefficiency

saving if you have to order everything

you're reading right is reading code as

fast as writing code or slower or faster

I'm I I don't know right I'm undecided I

think sometimes for boilerplate codes

probably pretty effective um if it's a

sort of code you know write me a a for

Loop to do X Y and Z probably works

pretty well as long as you're capable of

quickly checking that but then it didn't

take me very long to write four looping

anyway I'm undecided I suppose as to how

much of a game changer that will be this

said I know there are developers that

use it right and I know that the

developers who claim or at least they

think it's they're much more efficient I

don't spend as much time coding as I'd

like because I'm I see you know isn't as

a professor in the University I spend a

lot of time teaching a lot of time

mentoring others right and teaching

people so they do the coding and I sit

there and look at it right so I haven't

had as much experience as some yeah I

mean I think the the concern is always

you know younger young people are people

trying to switch careers is you know I

want to have a job for more than a year

or five years

um is it worth putting all the effort in

to learn this stuff if AI is just going

to take it away my gut tells me that AI

isn't going to take it away anytime soon

right because I think that I would argue

that you need something more fundamental

to understanding some of these problems

if you're going to write code to solve

them than just a text production

mechanism that isn't to say that what it

doesn't do it's very impressive what it

does but I think that as you start to

build up you know it's very it's all

very well saying write me a for Loop to

do this but if you want to write your

class structure and a and a really

complicated system that's such a more

difficult you know it's like the

difference between Lane assist and

self-driving right and that's why we can

we've seen Lane assist exist but

self-driving seems to be so hard to get

to because of how much harder that is as

a problem and I think that it's very

easy to fit a straight line upwards to

these things right you say well they

didn't do anything and now they're doing

this which means they're going to be

doing this it may get a lot harder and

Plateau out right we you know it's

difficult to say for sure I think that

there's going to be a very strong need

for people in the loop for a lot for a

long time further right I mean as an

example outside of programming in

medical science AI is obviously used

quite a lot to help with diagnoses and

things but almost no AI systems are used

just on their own with no human

oversight because for a start because we

don't trust them yet and also because

patients don't trust them patients don't

want an AI even if it's good making

their health decisions right like not

yet you know and so I think also

culturally we're not quite we're not

quite ready and I know a few companies

that are not using copilot because

they're not absolutely sure the

copyright on on the code and think you

know there's questions that haven't been

answered I think if you're looking for

it to be a software developer or you're

looking for a career in security or

career in AI there's still plenty of

things to do so I wouldn't personally

worry about that I think we we mentioned

this last time and I I want I want to

give people firstly you know a way to

make themselves more valuable and then a

path to get there

um you mentioned that you know any if

you attach AI to any skill that you've

got it's going to make you more valuable

um I I assume that's still the case and

I want to ask you Mike how do I get

there and it also makes you more

experienced at dealing with things like

this when something comes along you can

you can sit back and you can say okay

how impressive is this let's think about

what it's doing and how it works and you

know some understanding of how these

things work you don't have to understand

deep down transformer networks if you

want to understand roughly what they're

doing right and how they I've been

trained yeah I would say some knowledge

of statistical analysis and data data

processing in general is really really

important right people mock Excel Excel

is I think one of the best products ever

written it's totally ubiquitous it's

very powerful and it underpins huge

amounts of you know Financial systems

and other systems I use it all the time

from for student marks right so you know

you get a table of data that comes in

and it doesn't make any sense what we're

going to look at how we're going to deal

with this right and how we're going to

make decisions based on this data and

things like data science and machine

learning will help you deal with some of

these problems people who want to become

experts in AI obviously need to delve a

bit deeper but I think for a lot of

people AI can just solve small problems

in your pipeline that might make things

a little bit easier having that extra

string in your bow it's not it's not a

terrible idea so in previous videos I

told people you need to learn Ai and

it's something that I want to really

focus on this year and this is why I'm

talking to you you know right in the

beginning of the year uh have you got

like courses places that I can go to

books that I can read any

recommendations of how do I go from like

where I am now zero knowledge to yeah at

least you know getting down that path to

be able to put it online there are loads

there's loads of books and resources in

Python to learn machine learning and

data science um and that would be a

great place to start I you know I've

said it before many times I have a love

hate relationship with python I like it

sometimes and I don't like it other

times at the end of the day there are

libraries in Python that do quite

incredible machine learning and make

your life a lot easier right so we've

got things like psychic learn we've got

tensorflow and Pie torch of course but

there are tutorials and books written

around these things and they take you

from I don't know what this network is

to I can actually get one of these

networks running on a machine and it's

often not that much code because of

these libraries do a lot of heavy

lifting for you often it becomes more

plug-in building blocks together than it

does right you know writing neural

network layers from scratch which no you

know we don't do anymore you know so you

can start by just plugging some things

together and I've got a rudimentary

Network that I don't really understand

that's doing this classification and

before long you've made your

classification problem a little bit more

complicated and you've got multi-class

classification and then you've got a

slightly different data set and then

you've solved a data augmentation

problem and you can add these things in

and slowly work towards a bit more

experience you know I have you know a

number of undergraduate project students

every year so in University in the third

year you often do a dissertation which

is like a um like a focused project over

a whole year after most of my

dissertation projects are going to be on

AI and something like this and you know

these are students who've done some you

know machine learning maybe a little bit

in their modules throughout their

undergraduate and they know how to code

but a lot of it's new you know we pick

it up and we run with it and we and we

we drew some great stuff I've got some

um I've got some students in the second

year solving Rubik's Cubes using machine

learning to detect where the colors are

and things like this and this is from

scratch right so this is this is people

who haven't done machine learning before

and like important minimized reaction I

think it is very doable and I think it's

it's you know and it's fun as well right

there's nothing more satisfying to me

than you've trained a network and it's

just classifying really accurately

whatever it was you wanted to do I

basically my job is looking at numbers

go up and I like when they go up so Mike

I mean I'd love to come to you not in

Nottingham University and attend your

courses but obviously icon and so can

you know most of us can't

um do you have any like resources or

ideas that things places I can go to to

learn often the first course I recommend

for everyone is to take Andrew Angus

Corsair of course right very popular I

mean I don't know how many times it's

been taken now millions of times it's

Andrew's course on machine learning

there is a deep learning follow-up to it

which I haven't I haven't done because

partly I actually already know deep

learning but um the machine learning

course is really good it's a good

understanding of some of the key

Concepts in machine learning and not

specifically about yes a little bit

about how neural networks work and

things like this and it can be a little

bit mathematical it's my experience of

it but if you if you watch it anyway

you're going to pick up a lot of tips

and tricks so things like watching your

network train over time and reacting to

how that works and doesn't work and

making decisions based on this these are

the things really I think that people

who want to do machine learning in an

applied way in a know in a business or

in an industry that's what they need to

better do a lot of them are not going to

be writing neural networks from scratch

or designing a number of layers in your

network they're going to take a network

that we know works and run it on some

new data and if that works great the

first time then that's fabulous but if

it doesn't what do you do then and these

are things that you're going to learn

and start to learning that Coursera

course Joshua bengio and others have

written a book just called Deep learning

which is very popular again obviously it

can go into a little bit of heavy Mass

detail but it's very popular I would say

don't read it end to end it's one to dip

into while you're doing some tutorials

to understand a bit more about the

theory and after that personally I would

get I would do the pie torch tutorials

or the psychic learn tutorials they can

be directed at your own pace and they

will include they'll give you experience

and all those different things right

there's there's tutorials on things like

reinforcement learning but also just

standard cnns and Transformers and

things like this yeah and don't don't

worry about you don't have to do all of

those on day one on day one we're

talking about about what is

classification what is regression maybe

get something little going right really

you know start yourself off nice and

slow and build up the complexity as we

go right it's the same with any subject

in computer science you can't learn

everything on the first day so you just

have to take it a little bit at a time

I'm glad you mentioned computer science

um do you think

it's time to for more of us to learn

computer science type stuff because of

AI like maths and all these computer

science stuff not really I think that

it's it wouldn't it's not necessary for

everyone to do that I think that you

know I would encourage everyone to do

computer science because I would but of

course I I think that sometimes both

computer science and an industry have a

sort of reverse snobbery about each

other right which I don't like very much

so for example computer scientists might

say well if someone didn't do a degree

you know what do they really know about

computers right which is not true and

someone who's who got on fine without a

degree might go why would I go and get

student loans and do a degree and

different paths are all valid I don't I

don't know why we're having this

conversation right and I think um there

are there are elements of maths in

machine learning which help I suppose me

to understand it a bit better when

someone comes with a particularly weird

problem that doesn't you know they've

added another layer and it's not

training why is that right they also

help me sometimes when I'm reading

papers because papers they can have a

lot of mathematical notation in and

sometimes that's not necessary and

they've just added it in but often it's

just it's just to be absolutely clear

about what they've done and and often

the mathematical notation is necessary

to achieve that rather than writing it

in in sort of flavorful text but to

begin with machine learning you don't

necessarily need to know those things

you know you can train a network in pi

torch with a knowledge of rudimentary

knowledge of python and following some

tutorials and you'll pick up the rest as

you go the really complicated maths like

back propagation which is how we train

it that's all taken care of under the

hood you don't see that it's not

something unless you're really

interested it's not something to concern

yourself with but I mean the great thing

is if I'm in Industry I let or I'm into

cyber or Dev or whatever I can really

enhance my career prospects and the

future by just adding this on to my

skills yeah but and I also think that

and I mentioned it before I think the

other thing is it makes you much more

resistant to hype and to concerns over

things and also when someone comes to

you and says oh yeah I've trained a

neural network to do X Y and Z you can

start to think hasn't sound very likely

right that sounds like the sort of thing

but maybe is a bit fanciful right let's

let's deal with let's look at their data

and see if that's actually true what

they've done and I think one of the

reasons that chat gbt and stuff have

been so hyped recently is because most

people don't know what it is and so when

you see it doing what it does you think

this thing must basically be a person

right because it's acting like one but

actually it's only acting like one in a

very narrow thing and we know how it's

trained and how it's trained doesn't

imply necessarily that it's got any

human qualities right it might but I

don't gut tells me not quite right but

the point is that I I can I'm sort of

more resistant to that in some sense

because I I know how it works underneath

and I sort of think I've trained all

these networks and this is a bigger

version of networks that I've trained

myself I don't see what's different

about what's so different about that

that it would suddenly be unbelievably

Insurgent compared to anything else if

that makes sense some knowledge of house

what some of these Technologies are just

like knowledge of you know so some some

companies trying to sell you a new

firewall with Next Generation antivirus

on it that has all kinds of machine

learning well if you understand a bit

about machine learning you'll know what

it will and won't do right and that will

allow you to make a better informed

purchase decision and the answer is it

will work pretty well right but it's but

nothing's perfect and machine learning

is only as good as the training data and

so on so there's lots of things you can

ask and you can ask really difficult

questions instead of people that come

and try and sell it to you especially

with things like Twitter and the news

it's very easy to get carried away in

this hype cycle right lots of

technologies have this it's in the

interest of these companies to make

these massive models of incredibly

impressive performance I think we're a

long way from Full automation of a lot

of these tasks even if it might appear

that way a sort of superficial level but

on the other hand they're really

promising in some other ways right so

one of the things that I found that chat

GPT is really good at is paraphrasing

text and vice versa so you have a text

you don't quite understand say please

can you read this and tell me what it

means or please can you summarize these

bullet points in an email or something

like this you know these kind of

functions I think are actually working

really well right because those are

functions that rely on the their text to

text they're meant for text to text

right they are that's kind of what

they're for and I think that those are

the ones that are really really good I

think code completion is useful when

you're asking limited things that you

can carefully check quite quickly don't

ask it to produce a thousand lines of

code that you expect them to all be

perfect because that's not what it will

do right and and you'll end up with a

lot of weird bugs or I mean there was

this used paper just released just the

other day actually from Stanford that

said that they they audited code from

about 30 to 35 researchers who some of

them were using

um AI to produce some of the code and

some of them weren't and the AI produced

code had more vulnerabilities in it and

that's because when the AI produces code

that works but let's say it uses ECB

mode in as or it uses a slightly weak

key derivation or something I don't know

something subtle if they don't know

about that subject already they might

accept that change if that makes sense

right they actually so you need this is

why you need to still be an expert in

your field because you can't just rely

on it to do it for you yet you've got to

be there saying I think that's okay or I

don't think that's okay

um and make those decisions for yourself

yeah I mean you know it's a limited

study but like it's not that limited and

it makes a very valid point I think the

real danger is people who and and I

should carry out this by saying I'm not

selling short these these incredible

technology is I'm just saying that it

would be very silly to just completely

use them blind and never check what they

do right because we know they just make

stuff up a lot of the time I think a bit

of domain knowledge is always going to

help I mean it's interesting because I

did I did some tests with like Cisco

devices and um it's amazing like first

time it got a perfect then I wanted to

do it for a video and then it wasn't

good and I did like five or six attempts

and none of them were perfect yeah I

think if I didn't know what it was doing

I would have accepted it sorry go on

yeah and the other thing is that you

know if you think about a date of it

it's trained on it's got some 40 40 plus

billion tokens right it's just internet

text we'll just leave it at that right

loads and loads of text Cisco related

text is only going to form a very very

small fraction of that various vertical

evidence because it's not got a world

model because it's not got an

understanding of the world where it can

bring Cisco in and correct add it to its

model it's just doing text completion

and so when something is

underrepresented in a training set it's

gonna probably be worth performing when

it comes to actually running it later

right so when you say write me write me

something in the style of Shakespeare

it's going to do really well because

there's Shakespeare all over the

internet right some tasks are going to

be very solvable because they've just

they're hugely represented in the

trading set they work really well and

some tasks are really Niche and of

course you don't know which one's a

niche because you haven't seen the

training set I say write me a link

expression and it does it really well

and when I say write me a link

expression using some other thing and

that isn't in a training set and it

produces me a wrong answer and I don't

know until I run it whether that's the

case so I have to understand and be able

to read that code because otherwise I

can't possibly put it into my system and

it's just it goes back to the exact same

problem with medicine right it might be

that we're absolutely confident but they

say I will look at this image and make

the correct decision but we're not

absolutely sure and while we're not

absolutely sure do we want to completely

take a human out of the loop there

there's questions to you know we have to

think about so do you think it'll become

like the AI might do a lot of the low

level

I think that's much closer to what will

happen so I think in in there's a phrase

in medicine called CAD or computer-aided

diagnosis and the idea is that instead

of the doctor not making a decision the

doctor will be guided into a decision by

the AI saying we've noticed these spots

over here in this image is that relevant

to you and it will speed them up right

and if we can make doctors or Medics 50

more efficient that's a huge that's

that's a huge boost rather than try and

put it all on the AI and similarly it

works in code if you can produce

boilerplate code if you can get it to

bootstrap spring boots configuration

files for you fabulous do that right and

then that saves you half an hour to an

hour of doing some actual code or making

sure that it worked but what I would

have avoid you know what I would avoid

doing is trying to have it write

everything for you and replace yourself

because I don't think it will work I

think you'll end up really frustrated

that your code doesn't get part past any

of your reviews because it didn't work

right I was going to say I love what you

said though because with that example at

Stanford if if

if people had just accepted the code

there's hidden vulnerabilities in the

code that wouldn't have been picked up

yeah and and then there's a combination

of issues right is it that the developer

needs to know more about these subjects

or is it that they're someone that would

normally be on that team that wasn't

auditing that code that would have been

auditing that code at that time you know

you know because you have security teams

sometimes who are specialists in this

but I think it's that same argument in

some ways if someone has a small amount

of knowledge of computer security that

might allow them to be more resistant

when code appears it does this and

that's the same thing with the AI if you

know a little bit about AI maybe you can

be you can better deal with it when

something comes along so I think a

little bit of knowledge in lots of these

things is is often useful for that

reason micro so how's this affected like

University life because I've heard

people talk about how students can just

get check GPD to write their essays and

stuff like that so and you can't you

can't see the difference between a

student and a like a human sorry and

yeah

um I think it's very subject dependent I

think that's one thing so um what we've

done is we've done we've actually been

running some tests right because so you

know if I kind of open out and drop this

tournament just before exams yeah yeah

exactly

um yeah we've run some tests and like I

think it depends on if if I show suppose

we're doing a computer security exam

which actually I I teach so you know

right and I ask a very simple question

right a question like what's a good

encryption algorithm to use track GPT

can answer that so it'll be unwise of me

to ask that question in an exam I

suppose what we say in some sense I

think it's another variant of a search

engine so if a student could you know

can we call it academic misconduct right

if a student was going to use a search

engine to do that they could also have a

go at using chat GPT it has the

advantage for that student but it's

generating very plausible looking

answers sometimes they're completely

wrong right and those answers are going

to get marked very far down when they

come in front of of a of a convenience

so I think your mileage may vary if you

think you can get through a University

degree using just AI tools it's

something we have to consider right now

some of our exams are face to face they

aren't really affected right you know

we're talking about coursework essays

and we I don't know I haven't spoken too

much to other other schools in the

university in other subject areas but

obviously there are lots of essay-based

subjects but they require very well

written essays trap GPT has a habit of

producing general answers to things

right which are sometimes very detailed

but sometimes not quite so detailed

again I think that your mileage would

vary if you tried this I suspect that it

is possible to tell that they're written

by chat GPT to an extent because it has

a way of phrasing things that's quite

common I've noticed as I as I produce

answers but that isn't sure that isn't

necessarily all the time but that's

going to be a problem it's something

that every University on Earth is now

looking at work so yeah it's had a big

impact and you know when you consider

that this is just version one right and

you know there's going to be a chat gp22

probably and Microsoft might release one

and Google release one and so on and so

forth there's gonna there's always going

to be one of these tools floating about

that we have to just be prepared and

think about how that how that's going to

work I think I mean the examples I've

seen which have worked really well is

like if I'm asked to write an essay

about something I can get it to write

something that gives me a lot of ideas

and then I can just rephrase it in my

own voice but it's it helps you a lot

from a study point of view yeah I think

actually does and I think so anyway

that's a that's a big positive way and

there are some academics in in this

school for example and across across the

world who operate in a kind of human

computer interaction area who are very

interested in could you end up writing a

better essay if you worked with a

computer to help you out right and in in

a way is that not a win for the

lecturers as well if that's the case now

I I agree with that to an extent I think

that's absolutely right I think that

maybe we can't solve that whole

discussion in a month right which is how

long it is until our exams so you know

the clock is ticking in somewhat it's

somewhat in the short term for for these

issues but in the longer term I think

they're going to be really

transformative in helping you know there

are students who have who are very very

intelligent and they know all the

subject area but they're just not good

at exams they really struggle to get

their thoughts down on paper maybe those

students could really be helped by

something like this because if you give

really specific prompts to chat GPT you

get much better answers if a student

knows what they're doing and can work

with the AI I think that's going to be

much better I mean I suppose the you

could have said the same thing for

Google or you know using search engines

for yeah yeah that's that's the point

that's been made I mean in some ways I

see on Twitter a lot of people um

compare these things to Google I would

not because they're very different

um and they don't have no source of

actual data right that's the really

important thing to remember but they

they are a complementary tool in many

ways and they are they operate in a

similar way if you were going to try and

answer an exam you know you would put

the question in you'd rephrase it you'd

see what came out you'd see does that

look plausible I'm going to try again

I'm going to edit it and so on in the

same way that you would if you were

doing if you're using a search engine to

write an essay as well and right using a

search engine to write an essay and I

don't want to speak for every academic

on the planet right but it's not

necessarily plagiarism or misconduct it

depends on how you use it right you know

looking up sources online is absolutely

to be encouraged it depends on how

you're doing this I think in the long

term we will get a nice balance actually

between using it too much and not using

it enough and I think actually there's

another thing there's another aspect

which is I think this plays into your

this is relevant to your channels

viewers is that you don't you shouldn't

think of doing a degree or writing a

coursework as just about getting a mark

right that's very easy to think about

that but actually it's about learning

something that you can then take and use

in your career or something like that

right we don't teach people to program

so they pass the exams we teach them the

program so they can go off and be

software developers if you use AI to

write all of your work for you then you

get out you wouldn't be able to get a

job and you wouldn't better work in that

job because you wouldn't be able to do

any of the computer science actually I

think that if you got a lot because I

have quite a lot of learning and I love

to learn about new soft topics

particularly you know around computer

science I would never use chat GPT to

cheat because I wouldn't know any of it

then right and you know and I like to

learn about these things now if you want

to become an expert in something then

you're going to need to learn it you

can't read what chat GPT wrote A lot of

it comes down to hoping that students

and trying to encourage students to

think that it's about the process of

learning and where they get to at the

end rather than specifically about a

series of of kind of barriers of exams

that they have to get through right

which I think is is not a good way to

look at a degree or any course really

you know it's much better to think about

where you'll be at the end right and

you'll be in that much better position

to do what you want to do next that's

exactly right I mean it's a like

certification exams same thing you know

you can go and get all the answers or

the cheetah sites or you can actually

learn something and and you haven't done

yourself any favors if you get it off

because you might get a job based on

that it's not going to go well right

because you don't you don't have any of

the knowledge you'll always feel like

you don't have any of the knowledge as

well right you know actually it doesn't

take that long to learn things if you

really put yourself to it and you'll be

in such a much better position

afterwards Mike as always I really want

to thank you for you know sharing your

knowledge and you know separating the

hype from like the worries about

people's futures thanks so much for

making a drill yeah it's no problem glad

to be on again it's been really really

good brilliant thanks Mike thanks

[Music]