00:00:00
[Music]
00:00:05
ever since computers were invented
00:00:07
they've really just been glorified
00:00:08
calculators machines that execute the
00:00:11
exact instructions given to them by the
00:00:13
programmers but something incredible is
00:00:15
happening now computers have started
00:00:16
gaining the ability to learn and think
00:00:19
and communicate just like we do they can
00:00:21
do creative intellectual work that
00:00:23
previously only humans could do we call
00:00:25
this technology generative Ai and you
00:00:27
may have encountered it already through
00:00:29
products like GPT basically intelligence
00:00:32
is now available as a service kind of
00:00:34
like a giant brain floating in the sky
00:00:36
that anyone can talk to it's not perfect
00:00:39
but it is surprisingly capable and it is
00:00:40
improving at an exponential rate this is
00:00:43
a big deal it's going to affect just
00:00:45
about every person and Company on the
00:00:47
planet positively or negatively this
00:00:49
video is here to help you understand
00:00:51
what generative AI is all about in
00:00:53
Practical terms beyond the hype the
00:00:54
better you understand this technology as
00:00:56
a person team or company the better
00:00:58
equipped you will be to survive and
00:01:00
thrive in the age of AI so here's a
00:01:03
silly but useful mental model for this
00:01:05
you have Einstein in your basement in
00:01:07
fact everyone does and by Einstein I
00:01:10
really mean the combination of every
00:01:12
smart person who ever lived you can talk
00:01:14
to Einstein whenever you want he has
00:01:16
instant access to the sum of all human
00:01:18
knowledge and will answer anything you
00:01:20
want within seconds never running out of
00:01:21
patience he can also take on any role
00:01:23
you want a comedian poet doctor coach
00:01:27
and will be an expert within that field
00:01:29
he has has some humanlike limitations
00:01:31
though he can make mistakes he can jump
00:01:33
to conclusions he can misunderstand you
00:01:35
but the biggest limitation is actually
00:01:37
your imagination and your ability to
00:01:39
communicate effectively with them this
00:01:41
skill is known as prompt engineering and
00:01:43
in the age of AI this is as essential as
00:01:46
reading and writing most people vastly
00:01:49
underestimate what this Einstein in your
00:01:51
basement can do it's like going to the
00:01:53
real Einstein and asking him to proof
00:01:55
read a high school report or hiring a
00:01:56
world-class five-star chef and having
00:01:59
him chop onion the more you interact
00:02:01
with Einstein the more you will discover
00:02:02
surprising and Powerful ways for him to
00:02:05
help you or your company okay enough
00:02:07
fluffy metaphors let's clarify some
00:02:08
terms AI as you probably know stands for
00:02:11
artificial intelligence AI is not new
00:02:14
Fields like machine learning and
00:02:15
computer vision have been around for
00:02:17
decades whenever you see a YouTube
00:02:18
recommendation or a web search result or
00:02:21
whenever you get a credit card
00:02:22
transaction approved that's traditional
00:02:24
AI in action generative AI is AI that
00:02:27
generates new original content rather
00:02:29
than just finding or classifying
00:02:31
existing content that's the G in GPT for
00:02:33
example large language models or llms
00:02:36
are a type of generative AI that can
00:02:38
communicate using normal human language
00:02:41
chat GPT is a product by the company
00:02:43
open AI it started as an llm essentially
00:02:46
an advanced chatbot using a new
00:02:47
architecture called the Transformer
00:02:49
architecture which by the way is the T
00:02:51
in GPT it is so fluent at human language
00:02:54
that anyone can use it you don't need to
00:02:55
be an AI expert or programmer and that's
00:02:57
kind of what triggered the whole
00:02:58
Revolution so how does it actually work
00:03:02
well a large language model is an
00:03:03
artificial neural network basically a
00:03:06
bunch of numbers or or parameters
00:03:08
connected to each other similar to how
00:03:09
our brain is a bunch of neurons or brain
00:03:11
cells connected to each other neural
00:03:12
networks only deal with numbers you send
00:03:15
in numbers and depending on how the
00:03:16
parameters are set all the numbers come
00:03:18
out but any kind of content such as text
00:03:20
or images can be represented as numbers
00:03:22
so let's say I write dogs are when I
00:03:25
send that to a large language model that
00:03:27
gets converted to numbers processed by
00:03:29
the neural network and then the
00:03:30
resulting numbers are converted back
00:03:31
into text in this case the word animals
00:03:34
dogs are animals so yeah this is
00:03:36
basically a guest toex word machine the
00:03:39
interesting part is if we take that
00:03:40
output and combine it with the input and
00:03:43
send it through the model again then it
00:03:45
will continue adding new words that's
00:03:46
what's going on behind the scenes when
00:03:48
you type something in chat GPT in this
00:03:50
case for example it generated a whole
00:03:51
story and I can continue this
00:03:53
indefinitely by adding more prompts a
00:03:56
large language model may have billions
00:03:58
or even trillions of parameters that's
00:04:00
why they're called large so how are all
00:04:02
these numbers set well not through
00:04:04
manual programming that would be
00:04:06
impossible but through training just
00:04:09
like babies learning to speak a baby
00:04:11
isn't told how to speak she doesn't get
00:04:13
an instruction manual instead she
00:04:15
listens to people speaking around her
00:04:16
and when she's heard enough she starts
00:04:18
seeing the pattern she speaks a few
00:04:20
words at first to the Delight of her
00:04:21
parents and then later on full sentences
00:04:24
similarly during a training period the
00:04:26
language model is fed a mindboggling
00:04:28
amount of text to learn from Mostly from
00:04:31
internet sources it then plays guess the
00:04:33
next word with all of this over and over
00:04:35
again and the parameters are
00:04:37
automatically tweaked until it starts
00:04:38
getting really good at predicting the
00:04:40
next word this is called back
00:04:41
propagation which is a fancy term for oh
00:04:44
I guessed wrong I better change
00:04:45
something however to become truly useful
00:04:47
a model also needs to undergo human
00:04:49
training this is called reinforcement
00:04:51
learning with human feedback and it
00:04:53
involves thousands of hours of humans
00:04:55
painstakingly testing and evaluating
00:04:57
output from the model and giving
00:04:58
feedback kind of like training a a dog
00:05:01
with a clicker to reinforce good
00:05:02
behavior that's why a model like GPT
00:05:04
won't tell you how to rob a bank it
00:05:06
knows very well how to rob a bank but
00:05:08
through human training it has learned
00:05:09
that it shouldn't help people commit
00:05:11
crimes when training is done the model
00:05:13
is mostly Frozen other than some fine
00:05:15
tuning that can happen later that's what
00:05:17
the P stands for in GPT pre-trained
00:05:19
although in the future we will probably
00:05:20
have models that can learn continuously
00:05:22
rather than just uh during training and
00:05:24
fine-tuning now although chat GPT kind
00:05:26
of got the ball rolling GPT isn't the
00:05:29
only model out there in fact new models
00:05:31
are sprouting like mushrooms they vary a
00:05:34
lot in terms of speed capability and
00:05:36
cost some can be downloaded and run
00:05:37
locally others are only online some are
00:05:40
free or open source others are
00:05:41
commercial products some are super easy
00:05:43
to use While others require complicated
00:05:46
technical setup some are specialized for
00:05:48
certain use cases others are more
00:05:50
General and can be used for almost
00:05:52
anything and some are baked into
00:05:54
products in the form of co-pilots or or
00:05:56
chat windows it's it's the Wild West
00:06:00
just keep in mind that you generally get
00:06:01
what you pay for so with a free model
00:06:04
you may just be getting a smart high
00:06:06
school student in your basement rather
00:06:08
than Einstein the difference between for
00:06:11
example GPT 3.5 and gp4 is
00:06:14
massive note that there are different
00:06:16
types of generative AI models that
00:06:18
generate different types of content
00:06:20
textto text models like gpc4 take text
00:06:23
as input and generate text as output the
00:06:25
text can be natural language but it can
00:06:26
also be structured information like code
00:06:29
Json or HTML I use this a lot myself to
00:06:32
generate code when programming uh it
00:06:33
saves an incredible amount of time and I
00:06:35
also learn a lot from the code it
00:06:37
generates text to image models will
00:06:38
generate images describe what you want
00:06:40
and an image gets generated for you you
00:06:42
can even pick a style image to image
00:06:45
models can do things like transforming
00:06:47
or combining images and we have image to
00:06:50
text models which describe the contents
00:06:52
of a given image and speech to text
00:06:54
models create voice transcriptions which
00:06:56
is useful for things like uh meeting
00:06:58
notes text to audio models they generate
00:07:00
music or sounds from a prompt for
00:07:02
example here is some sound generated
00:07:04
from The Prompt people talking in a
00:07:08
busy okay guys enough stop now thank you
00:07:13
and there are even text to video models
00:07:15
that generate videos from a prompt
00:07:17
sooner or later we'll have infinite
00:07:18
movie series that autogenerate the next
00:07:20
episode tailored to your tastes as
00:07:22
you're watching kind of scary if you
00:07:24
think about it one Trend now is
00:07:26
multimodal AI products meaning they
00:07:28
combine different models into one
00:07:30
product so you can work with text images
00:07:32
audio Etc without switching tools the
00:07:35
chat GPT mobile app is a good example of
00:07:37
this just for fun I took a photo of this
00:07:40
room and I asked where I could hide
00:07:41
stuff I kind of like that it mentioned
00:07:44
the stove but warned that that it could
00:07:46
get hot there when I have things to
00:07:48
figure out such as the contents of this
00:07:50
video I like to take walks using chat
00:07:52
GPT as as a sounding board I start by
00:07:55
saying always respond with the word okay
00:07:57
unless I ask you for something that way
00:07:59
it'll just listen and not interrupt
00:08:01
after I finish dumping my thoughts I ask
00:08:03
for feedback we have some discussion and
00:08:06
then I ask it to summarize and text
00:08:07
afterwards I really recommend trying
00:08:09
this it's it's a really useful way to
00:08:11
use tools like this turns out Einstein
00:08:13
isn't stuck in the basement after all
00:08:15
you can take him out for a walk
00:08:17
initially language models were just word
00:08:19
predictors statistical machines with
00:08:22
limited practical use but as they became
00:08:24
larger and were trained on more data
00:08:26
they started gaining emergent
00:08:28
capabilities unexpect capabilities that
00:08:30
surprised even the developers of the
00:08:31
technology they could role playay write
00:08:34
poetry write highquality code discuss
00:08:36
company strategy provide legal and
00:08:38
medical advice coach teach basically
00:08:41
creative and intellectual things that
00:08:43
only humans could do previously it turns
00:08:46
out that when a model has seen enough
00:08:47
text and images it starts to see
00:08:49
patterns and understand higher level
00:08:51
Concepts just like a baby learning to
00:08:53
understand the world let's take a simple
00:08:55
example I'll give gp4 this little
00:08:57
drawing that involves a string a pair of
00:09:00
scissors an egg a pot and a fire what
00:09:03
will happen if I use the scissors the
00:09:05
model has most likely not been trained
00:09:07
on this exact scenario yet it gave a
00:09:10
pretty good answer which demonstrates a
00:09:11
basic understanding of the nature of
00:09:13
scissors eggs gravity and heat when gp4
00:09:16
was released I started using it as a
00:09:18
coding assistant and I was blown away
00:09:20
when prompted effectively it was a
00:09:22
better programmer than anyone I've
00:09:23
worked with same with article writing
00:09:25
product design Workshop planning and
00:09:27
just about anything I used it for
00:09:29
the main bottleneck was my prompt
00:09:32
engineering skills so I decided to make
00:09:33
a career shift and focus entirely on
00:09:35
learning and teaching how to make this
00:09:37
technology useful hence this video now
00:09:40
let's take a step back and look at the
00:09:41
implications for 300,000 years or so we
00:09:44
homosapiens have been the most
00:09:46
intelligent species on Earth depending
00:09:48
of course on how you define intelligence
00:09:50
but the thing is our intellectual
00:09:51
capabilities aren't really improving
00:09:53
that much our brains are about the same
00:09:55
size same weight as they've been for
00:09:56
thousands of years computers on the
00:09:58
other hand have been around for only 80
00:10:00
years or so and now with generative AI
00:10:02
they are suddenly capable of speaking
00:10:04
human languages fluently and carrying
00:10:06
out an increasing number of intellectual
00:10:08
creative tasks that previously only
00:10:10
humans could do so we are right here at
00:10:12
the Crossing Point where AI is better at
00:10:14
some things and humans are better at
00:10:15
some things but ai's capabilities are
00:10:17
improving at an exponential rate while
00:10:19
ours aren't we don't know how long that
00:10:22
exponential Improvement will continue or
00:10:24
if it will level off at some point but
00:10:25
we're definitely entering a new world
00:10:27
order now this isn't the first re
00:10:29
Revolution we've experienced we tamed
00:10:31
fire we learned how to do agriculture we
00:10:33
invented the printing press steam power
00:10:35
Telegraph these were all revolutionary
00:10:37
changes but they took decades or
00:10:39
centuries to become widespread in the AI
00:10:42
Revolution new technology spreads
00:10:44
worldwide almost instantly dealing with
00:10:46
this rate of change is a huge challenge
00:10:48
for both individuals and
00:10:50
companies I've noticed that people and
00:10:52
companies tend to fall into different
00:10:54
kind of mindset categories when it comes
00:10:56
to AI on one side we have denial the
00:10:59
belief that AI cannot do my job or we
00:11:02
don't have time to look into this
00:11:03
technology this is a dangerous place to
00:11:05
be a common saying is AI might not take
00:11:08
your job but people using AI will and
00:11:11
this is true for both individuals and
00:11:13
companies on the other side of the scale
00:11:15
we have panic and despair the belief
00:11:16
that AI is going to take my job no
00:11:18
matter what AI is going to make my
00:11:19
company go bankrupt neither of these
00:11:21
mindsets are helpful so I propose a
00:11:24
middle ground a balanced positive
00:11:26
mindset AI is going to make me my team
00:11:28
my company insanely productive
00:11:31
personally with this mindset I feel like
00:11:33
I've gained superpowers I can go from
00:11:35
idea to result in so much shorter time I
00:11:38
can focus more on what I want to achieve
00:11:40
and less on the grunt work of building
00:11:41
things and I'm learning a lot faster too
00:11:43
it's like having an awesome Mentor with
00:11:45
me at all times this mindset not only
00:11:47
feels good but it also equips you for
00:11:49
the future makes you less likely to lose
00:11:51
your job or your company and more likely
00:11:53
to thrive in the age of AI despite all
00:11:55
the
00:11:56
uncertainty so one important question is
00:11:59
is human role X needed in the age of AI
00:12:02
for example are doctors needed
00:12:03
developers lawyers CEOs uh whatever so
00:12:06
this question becomes more and more
00:12:08
relevant as the AI capabilities improve
00:12:11
well some jobs will disappear for sure
00:12:13
but for most roles I think we humans are
00:12:15
still needed someone with domain
00:12:17
knowledge still needs to decide what to
00:12:19
ask the AI how to formulate The Prompt
00:12:21
what context needs to be provided and
00:12:23
how to evaluate the result AI models
00:12:25
aren't perfect they can be absolutely
00:12:27
brilliant sometimes but sometimes also
00:12:30
terribly stupid they can sometimes
00:12:32
hallucinate and provide bogus
00:12:33
information in a very convincing way so
00:12:36
when should you trust AI response when
00:12:38
should you double check or do the work
00:12:40
yourself what about legal compliance
00:12:42
data security what information can we
00:12:44
send to an AI model and where is that
00:12:46
data stored a human expert is needed to
00:12:49
make these judgment calls and compensate
00:12:51
for the weaknesses of the AI model so I
00:12:53
recommend thinking of AI as your
00:12:55
colleague a genius but also an oddball
00:12:57
with some personal quirks that you need
00:12:59
to learn to work with you need to
00:13:00
recognize when your Genius colleague is
00:13:02
drunk as a doctor my AI colleague can
00:13:05
help diagnose rare diseases that I
00:13:06
didn't even know existed as a lawyer my
00:13:09
AI colleague could do legal research and
00:13:11
review contracts allowing me to spend
00:13:12
more time with my client or as a teacher
00:13:15
my AI colleague could grade tests help
00:13:18
generate course content provide
00:13:19
individual support to students Etc and
00:13:22
if you're not sure how I can help you
00:13:24
just ask it I work as X how can you help
00:13:27
me overall I find that that the
00:13:29
combination of human plus AI That's
00:13:31
where the magic lies it's important to
00:13:34
distinguish between the models and the
00:13:36
products that build on top of them as a
00:13:38
user you don't normally interact with
00:13:39
the model directly instead you interact
00:13:42
with a product website or a mobile app
00:13:43
which in turn talks to the model behind
00:13:45
the scenes products provide a user
00:13:47
interface and add capabilities and data
00:13:49
that aren't part of the model itself for
00:13:51
example the chat GPT product keeps track
00:13:54
of your message history while the GPT 4
00:13:56
model itself doesn't have any message
00:13:58
history as a developer you can use these
00:14:01
models to build your own AI powered
00:14:02
products and features for example let's
00:14:05
say you have an e-learning site you
00:14:06
could add a chat bot to answer questions
00:14:08
about the courses or as a recruitment
00:14:10
company you might build AI powered tools
00:14:12
to help evaluate candidates in both
00:14:14
these cases your users interact with
00:14:16
your product and then your product
00:14:18
interacts with the model this is done
00:14:19
via apis or application programming
00:14:21
interfaces which allow your code to talk
00:14:23
to the model so here's a simple example
00:14:26
of using open AI API to talk to GPT not
00:14:29
a lot of code needed and here's another
00:14:31
example of the automatic candidate
00:14:33
evaluation thing I talked about it takes
00:14:35
a job description and a bunch of CVS in
00:14:37
a folder and evaluates each candidate
00:14:40
automatically and incidentally the code
00:14:42
itself is mostly AI written as a product
00:14:45
developer you can use AI models kind of
00:14:48
like an external brain to insert
00:14:50
intelligence into your product very
00:14:52
powerful in order to use generative AI
00:14:55
effectively you need to get good at
00:14:57
prompt engineering or prompt design as I
00:14:59
prefer to call it this skill is needed
00:15:01
both as a user and as a product
00:15:03
developer because in both cases you need
00:15:05
to be able to craft effective prompts
00:15:07
that produce useful results from an AI
00:15:09
model here's an example let's say I want
00:15:11
help planning a workshop this prompt is
00:15:14
unlikely to give useful results because
00:15:16
no matter how smart the AI is if it
00:15:18
doesn't know the context of my workshop
00:15:20
it can only give fague high level
00:15:22
recommendations the second prompt is
00:15:24
better now I provided some context this
00:15:26
is normally done iteratively write a
00:15:28
prompt look at the result add a
00:15:30
follow-up prompt to provide more
00:15:31
information or edit the original prompt
00:15:34
and rinse and repeat until you get a
00:15:35
good result in this third approach I ask
00:15:38
it to interview me so instead of me
00:15:40
providing a bunch of context up front
00:15:42
I'm basically saying what do you need to
00:15:43
know in order order to help me and then
00:15:45
it will propose a workshop agenda after
00:15:47
I often combine these two I provide a
00:15:49
bit of context and then I tell it to ask
00:15:51
me if it needs any more information
00:15:53
these are just some examples of prompt
00:15:54
engineering techniques so overall the
00:15:57
better you get at prompt engineering the
00:15:59
faster and better results you will get
00:16:00
from AI there are plenty of courses
00:16:02
books videos articles to help you learn
00:16:04
this but the most important thing is is
00:16:06
to practice and Learn by doing a nice
00:16:08
side effect is that you will become
00:16:09
better at communicating in general since
00:16:11
prompt engineering is really all about
00:16:13
Clarity and effective
00:16:15
communication I think the next Frontier
00:16:17
for generative AI is autonomous agents
00:16:19
with tools these are AI powerered
00:16:21
software entities that run on their own
00:16:23
rather than just sitting around waiting
00:16:24
for you to prompt them all the time so
00:16:26
you go down to Einstein in your basement
00:16:28
and do what a good good leader would do
00:16:29
for a team you give him a high level
00:16:31
Mission and the tools needed to
00:16:32
accomplish it and then open the door and
00:16:34
let him out to run his own show without
00:16:36
micromanagement the tools could be
00:16:38
things like access to the internet
00:16:40
access to money ability to send and
00:16:42
receive messages order pizza or whatever
00:16:45
for this prompt engineering becomes even
00:16:47
more important because your autonomous
00:16:49
tool wielding agent can do a lot of good
00:16:51
or a lot of harm depending on how well
00:16:54
you craft that mission
00:16:55
statement all right let's wrap it up
00:16:58
here are the key things I hope you will
00:17:00
remember from this video generative AI
00:17:02
is a super useful tool that can help
00:17:04
both you your team and your company in a
00:17:06
big way the better you understand it the
00:17:08
more likely it is to be an opportunity
00:17:10
rather than a threat generative AI is
00:17:12
more powerful than you think the biggest
00:17:14
limitation is not the technology but
00:17:17
your imagination like what can I do and
00:17:19
your prompt engineering skills how do I
00:17:21
do it prompt engineeringdesign is a
00:17:24
crucial skill like all new skills just
00:17:27
accept that you will kind of suck at at
00:17:29
first but you'll improve over time with
00:17:31
deliberate practice so my best tip is
00:17:34
experiment make this part of your
00:17:36
day-to-day life and the Learning Happens
00:17:38
automatically hope this video was
00:17:40
helpful thanks for watching
00:17:44
[Music]
