0:05 [Music]
0:07 ever since computers were invented
0:08 they've really just been glorified
0:11 calculators machines that execute the
0:13 exact instructions given to them by the
0:15 programmers but something incredible is
0:16 happening now computers have started
0:19 gaining the ability to learn and think
0:21 and communicate just like we do they can
0:23 do creative intellectual work that
0:25 previously only humans could do we call
0:27 this technology generative Ai and you
0:29 may have encountered it already through
0:32 products like GPT basically intelligence
0:34 is now available as a service kind of
0:36 like a giant brain floating in the sky
0:39 that anyone can talk to it's not perfect
0:40 but it is surprisingly capable and it is
0:43 improving at an exponential rate this is
0:45 a big deal it's going to affect just
0:47 about every person and Company on the
0:49 planet positively or negatively this
0:51 video is here to help you understand
0:53 what generative AI is all about in
0:54 Practical terms beyond the hype the
0:56 better you understand this technology as
0:58 a person team or company the better
1:00 equipped you will be to survive and
1:03 thrive in the age of AI so here's a
1:05 silly but useful mental model for this
1:07 you have Einstein in your basement in
1:10 fact everyone does and by Einstein I
1:12 really mean the combination of every
1:14 smart person who ever lived you can talk
1:16 to Einstein whenever you want he has
1:18 instant access to the sum of all human
1:20 knowledge and will answer anything you
1:21 want within seconds never running out of
1:23 patience he can also take on any role
1:27 you want a comedian poet doctor coach
1:29 and will be an expert within that field
1:31 he has has some humanlike limitations
1:33 though he can make mistakes he can jump
1:35 to conclusions he can misunderstand you
1:37 but the biggest limitation is actually
1:39 your imagination and your ability to
1:41 communicate effectively with them this
1:43 skill is known as prompt engineering and
1:46 in the age of AI this is as essential as
1:49 reading and writing most people vastly
1:51 underestimate what this Einstein in your
1:53 basement can do it's like going to the
1:55 real Einstein and asking him to proof
1:56 read a high school report or hiring a
1:59 world-class five-star chef and having
2:01 him chop onion the more you interact
2:02 with Einstein the more you will discover
2:05 surprising and Powerful ways for him to
2:07 help you or your company okay enough
2:08 fluffy metaphors let's clarify some
2:11 terms AI as you probably know stands for
2:14 artificial intelligence AI is not new
2:15 Fields like machine learning and
2:17 computer vision have been around for
2:18 decades whenever you see a YouTube
2:21 recommendation or a web search result or
2:22 whenever you get a credit card
2:24 transaction approved that's traditional
2:27 AI in action generative AI is AI that
2:29 generates new original content rather
2:31 than just finding or classifying
2:33 existing content that's the G in GPT for
2:36 example large language models or llms
2:38 are a type of generative AI that can
2:41 communicate using normal human language
2:43 chat GPT is a product by the company
2:46 open AI it started as an llm essentially
2:47 an advanced chatbot using a new
2:49 architecture called the Transformer
2:51 architecture which by the way is the T
2:54 in GPT it is so fluent at human language
2:55 that anyone can use it you don't need to
2:57 be an AI expert or programmer and that's
2:58 kind of what triggered the whole
3:02 Revolution so how does it actually work
3:03 well a large language model is an
3:06 artificial neural network basically a
3:08 bunch of numbers or or parameters
3:09 connected to each other similar to how
3:11 our brain is a bunch of neurons or brain
3:12 cells connected to each other neural
3:15 networks only deal with numbers you send
3:16 in numbers and depending on how the
3:18 parameters are set all the numbers come
3:20 out but any kind of content such as text
3:22 or images can be represented as numbers
3:25 so let's say I write dogs are when I
3:27 send that to a large language model that
3:29 gets converted to numbers processed by
3:30 the neural network and then the
3:31 resulting numbers are converted back
3:34 into text in this case the word animals
3:36 dogs are animals so yeah this is
3:39 basically a guest toex word machine the
3:40 interesting part is if we take that
3:43 output and combine it with the input and
3:45 send it through the model again then it
3:46 will continue adding new words that's
3:48 what's going on behind the scenes when
3:50 you type something in chat GPT in this
3:51 case for example it generated a whole
3:53 story and I can continue this
3:56 indefinitely by adding more prompts a
3:58 large language model may have billions
4:00 or even trillions of parameters that's
4:02 why they're called large so how are all
4:04 these numbers set well not through
4:06 manual programming that would be
4:09 impossible but through training just
4:11 like babies learning to speak a baby
4:13 isn't told how to speak she doesn't get
4:15 an instruction manual instead she
4:16 listens to people speaking around her
4:18 and when she's heard enough she starts
4:20 seeing the pattern she speaks a few
4:21 words at first to the Delight of her
4:24 parents and then later on full sentences
4:26 similarly during a training period the
4:28 language model is fed a mindboggling
4:31 amount of text to learn from Mostly from
4:33 internet sources it then plays guess the
4:35 next word with all of this over and over
4:37 again and the parameters are
4:38 automatically tweaked until it starts
4:40 getting really good at predicting the
4:41 next word this is called back
4:44 propagation which is a fancy term for oh
4:45 I guessed wrong I better change
4:47 something however to become truly useful
4:49 a model also needs to undergo human
4:51 training this is called reinforcement
4:53 learning with human feedback and it
4:55 involves thousands of hours of humans
4:57 painstakingly testing and evaluating
4:58 output from the model and giving
5:01 feedback kind of like training a a dog
5:02 with a clicker to reinforce good
5:04 behavior that's why a model like GPT
5:06 won't tell you how to rob a bank it
5:08 knows very well how to rob a bank but
5:09 through human training it has learned
5:11 that it shouldn't help people commit
5:13 crimes when training is done the model
5:15 is mostly Frozen other than some fine
5:17 tuning that can happen later that's what
5:19 the P stands for in GPT pre-trained
5:20 although in the future we will probably
5:22 have models that can learn continuously
5:24 rather than just uh during training and
5:26 fine-tuning now although chat GPT kind
5:29 of got the ball rolling GPT isn't the
5:31 only model out there in fact new models
5:34 are sprouting like mushrooms they vary a
5:36 lot in terms of speed capability and
5:37 cost some can be downloaded and run
5:40 locally others are only online some are
5:41 free or open source others are
5:43 commercial products some are super easy
5:46 to use While others require complicated
5:48 technical setup some are specialized for
5:50 certain use cases others are more
5:52 General and can be used for almost
5:54 anything and some are baked into
5:56 products in the form of co-pilots or or
6:00 chat windows it's it's the Wild West
6:01 just keep in mind that you generally get
6:04 what you pay for so with a free model
6:06 you may just be getting a smart high
6:08 school student in your basement rather
6:11 than Einstein the difference between for
6:14 example GPT 3.5 and gp4 is
6:16 massive note that there are different
6:18 types of generative AI models that
6:20 generate different types of content
6:23 textto text models like gpc4 take text
6:25 as input and generate text as output the
6:26 text can be natural language but it can
6:29 also be structured information like code
6:32 Json or HTML I use this a lot myself to
6:33 generate code when programming uh it
6:35 saves an incredible amount of time and I
6:37 also learn a lot from the code it
6:38 generates text to image models will
6:40 generate images describe what you want
6:42 and an image gets generated for you you
6:45 can even pick a style image to image
6:47 models can do things like transforming
6:50 or combining images and we have image to
6:52 text models which describe the contents
6:54 of a given image and speech to text
6:56 models create voice transcriptions which
6:58 is useful for things like uh meeting
7:00 notes text to audio models they generate
7:02 music or sounds from a prompt for
7:04 example here is some sound generated
7:08 from The Prompt people talking in a
7:13 busy okay guys enough stop now thank you
7:15 and there are even text to video models
7:17 that generate videos from a prompt
7:18 sooner or later we'll have infinite
7:20 movie series that autogenerate the next
7:22 episode tailored to your tastes as
7:24 you're watching kind of scary if you
7:26 think about it one Trend now is
7:28 multimodal AI products meaning they
7:30 combine different models into one
7:32 product so you can work with text images
7:35 audio Etc without switching tools the
7:37 chat GPT mobile app is a good example of
7:40 this just for fun I took a photo of this
7:41 room and I asked where I could hide
7:44 stuff I kind of like that it mentioned
7:46 the stove but warned that that it could
7:48 get hot there when I have things to
7:50 figure out such as the contents of this
7:52 video I like to take walks using chat
7:55 GPT as as a sounding board I start by
7:57 saying always respond with the word okay
7:59 unless I ask you for something that way
8:01 it'll just listen and not interrupt
8:03 after I finish dumping my thoughts I ask
8:06 for feedback we have some discussion and
8:07 then I ask it to summarize and text
8:09 afterwards I really recommend trying
8:11 this it's it's a really useful way to
8:13 use tools like this turns out Einstein
8:15 isn't stuck in the basement after all
8:17 you can take him out for a walk
8:19 initially language models were just word
8:22 predictors statistical machines with
8:24 limited practical use but as they became
8:26 larger and were trained on more data
8:28 they started gaining emergent
8:30 capabilities unexpect capabilities that
8:31 surprised even the developers of the
8:34 technology they could role playay write
8:36 poetry write highquality code discuss
8:38 company strategy provide legal and
8:41 medical advice coach teach basically
8:43 creative and intellectual things that
8:46 only humans could do previously it turns
8:47 out that when a model has seen enough
8:49 text and images it starts to see
8:51 patterns and understand higher level
8:53 Concepts just like a baby learning to
8:55 understand the world let's take a simple
8:57 example I'll give gp4 this little
9:00 drawing that involves a string a pair of
9:03 scissors an egg a pot and a fire what
9:05 will happen if I use the scissors the
9:07 model has most likely not been trained
9:10 on this exact scenario yet it gave a
9:11 pretty good answer which demonstrates a
9:13 basic understanding of the nature of
9:16 scissors eggs gravity and heat when gp4
9:18 was released I started using it as a
9:20 coding assistant and I was blown away
9:22 when prompted effectively it was a
9:23 better programmer than anyone I've
9:25 worked with same with article writing
9:27 product design Workshop planning and
9:29 just about anything I used it for
9:32 the main bottleneck was my prompt
9:33 engineering skills so I decided to make
9:35 a career shift and focus entirely on
9:37 learning and teaching how to make this
9:40 technology useful hence this video now
9:41 let's take a step back and look at the
9:44 implications for 300,000 years or so we
9:46 homosapiens have been the most
9:48 intelligent species on Earth depending
9:50 of course on how you define intelligence
9:51 but the thing is our intellectual
9:53 capabilities aren't really improving
9:55 that much our brains are about the same
9:56 size same weight as they've been for
9:58 thousands of years computers on the
10:00 other hand have been around for only 80
10:02 years or so and now with generative AI
10:04 they are suddenly capable of speaking
10:06 human languages fluently and carrying
10:08 out an increasing number of intellectual
10:10 creative tasks that previously only
10:12 humans could do so we are right here at
10:14 the Crossing Point where AI is better at
10:15 some things and humans are better at
10:17 some things but ai's capabilities are
10:19 improving at an exponential rate while
10:22 ours aren't we don't know how long that
10:24 exponential Improvement will continue or
10:25 if it will level off at some point but
10:27 we're definitely entering a new world
10:29 order now this isn't the first re
10:31 Revolution we've experienced we tamed
10:33 fire we learned how to do agriculture we
10:35 invented the printing press steam power
10:37 Telegraph these were all revolutionary
10:39 changes but they took decades or
10:42 centuries to become widespread in the AI
10:44 Revolution new technology spreads
10:46 worldwide almost instantly dealing with
10:48 this rate of change is a huge challenge
10:50 for both individuals and
10:52 companies I've noticed that people and
10:54 companies tend to fall into different
10:56 kind of mindset categories when it comes
10:59 to AI on one side we have denial the
11:02 belief that AI cannot do my job or we
11:03 don't have time to look into this
11:05 technology this is a dangerous place to
11:08 be a common saying is AI might not take
11:11 your job but people using AI will and
11:13 this is true for both individuals and
11:15 companies on the other side of the scale
11:16 we have panic and despair the belief
11:18 that AI is going to take my job no
11:19 matter what AI is going to make my
11:21 company go bankrupt neither of these
11:24 mindsets are helpful so I propose a
11:26 middle ground a balanced positive
11:28 mindset AI is going to make me my team
11:31 my company insanely productive
11:33 personally with this mindset I feel like
11:35 I've gained superpowers I can go from
11:38 idea to result in so much shorter time I
11:40 can focus more on what I want to achieve
11:41 and less on the grunt work of building
11:43 things and I'm learning a lot faster too
11:45 it's like having an awesome Mentor with
11:47 me at all times this mindset not only
11:49 feels good but it also equips you for
11:51 the future makes you less likely to lose
11:53 your job or your company and more likely
11:55 to thrive in the age of AI despite all the
11:56 the
11:59 uncertainty so one important question is
12:02 is human role X needed in the age of AI
12:03 for example are doctors needed
12:06 developers lawyers CEOs uh whatever so
12:08 this question becomes more and more
12:11 relevant as the AI capabilities improve
12:13 well some jobs will disappear for sure
12:15 but for most roles I think we humans are
12:17 still needed someone with domain
12:19 knowledge still needs to decide what to
12:21 ask the AI how to formulate The Prompt
12:23 what context needs to be provided and
12:25 how to evaluate the result AI models
12:27 aren't perfect they can be absolutely
12:30 brilliant sometimes but sometimes also
12:32 terribly stupid they can sometimes
12:33 hallucinate and provide bogus
12:36 information in a very convincing way so
12:38 when should you trust AI response when
12:40 should you double check or do the work
12:42 yourself what about legal compliance
12:44 data security what information can we
12:46 send to an AI model and where is that
12:49 data stored a human expert is needed to
12:51 make these judgment calls and compensate
12:53 for the weaknesses of the AI model so I
12:55 recommend thinking of AI as your
12:57 colleague a genius but also an oddball
12:59 with some personal quirks that you need
13:00 to learn to work with you need to
13:02 recognize when your Genius colleague is
13:05 drunk as a doctor my AI colleague can
13:06 help diagnose rare diseases that I
13:09 didn't even know existed as a lawyer my
13:11 AI colleague could do legal research and
13:12 review contracts allowing me to spend
13:15 more time with my client or as a teacher
13:18 my AI colleague could grade tests help
13:19 generate course content provide
13:22 individual support to students Etc and
13:24 if you're not sure how I can help you
13:27 just ask it I work as X how can you help
13:29 me overall I find that that the
13:31 combination of human plus AI That's
13:34 where the magic lies it's important to
13:36 distinguish between the models and the
13:38 products that build on top of them as a
13:39 user you don't normally interact with
13:42 the model directly instead you interact
13:43 with a product website or a mobile app
13:45 which in turn talks to the model behind
13:47 the scenes products provide a user
13:49 interface and add capabilities and data
13:51 that aren't part of the model itself for
13:54 example the chat GPT product keeps track
13:56 of your message history while the GPT 4
13:58 model itself doesn't have any message
14:01 history as a developer you can use these
14:02 models to build your own AI powered
14:05 products and features for example let's
14:06 say you have an e-learning site you
14:08 could add a chat bot to answer questions
14:10 about the courses or as a recruitment
14:12 company you might build AI powered tools
14:14 to help evaluate candidates in both
14:16 these cases your users interact with
14:18 your product and then your product
14:19 interacts with the model this is done
14:21 via apis or application programming
14:23 interfaces which allow your code to talk
14:26 to the model so here's a simple example
14:29 of using open AI API to talk to GPT not
14:31 a lot of code needed and here's another
14:33 example of the automatic candidate
14:35 evaluation thing I talked about it takes
14:37 a job description and a bunch of CVS in
14:40 a folder and evaluates each candidate
14:42 automatically and incidentally the code
14:45 itself is mostly AI written as a product
14:48 developer you can use AI models kind of
14:50 like an external brain to insert
14:52 intelligence into your product very
14:55 powerful in order to use generative AI
14:57 effectively you need to get good at
14:59 prompt engineering or prompt design as I
15:01 prefer to call it this skill is needed
15:03 both as a user and as a product
15:05 developer because in both cases you need
15:07 to be able to craft effective prompts
15:09 that produce useful results from an AI
15:11 model here's an example let's say I want
15:14 help planning a workshop this prompt is
15:16 unlikely to give useful results because
15:18 no matter how smart the AI is if it
15:20 doesn't know the context of my workshop
15:22 it can only give fague high level
15:24 recommendations the second prompt is
15:26 better now I provided some context this
15:28 is normally done iteratively write a
15:30 prompt look at the result add a
15:31 follow-up prompt to provide more
15:34 information or edit the original prompt
15:35 and rinse and repeat until you get a
15:38 good result in this third approach I ask
15:40 it to interview me so instead of me
15:42 providing a bunch of context up front
15:43 I'm basically saying what do you need to
15:45 know in order order to help me and then
15:47 it will propose a workshop agenda after
15:49 I often combine these two I provide a
15:51 bit of context and then I tell it to ask
15:53 me if it needs any more information
15:54 these are just some examples of prompt
15:57 engineering techniques so overall the
15:59 better you get at prompt engineering the
16:00 faster and better results you will get
16:02 from AI there are plenty of courses
16:04 books videos articles to help you learn
16:06 this but the most important thing is is
16:08 to practice and Learn by doing a nice
16:09 side effect is that you will become
16:11 better at communicating in general since
16:13 prompt engineering is really all about
16:15 Clarity and effective
16:17 communication I think the next Frontier
16:19 for generative AI is autonomous agents
16:21 with tools these are AI powerered
16:23 software entities that run on their own
16:24 rather than just sitting around waiting
16:26 for you to prompt them all the time so
16:28 you go down to Einstein in your basement
16:29 and do what a good good leader would do
16:31 for a team you give him a high level
16:32 Mission and the tools needed to
16:34 accomplish it and then open the door and
16:36 let him out to run his own show without
16:38 micromanagement the tools could be
16:40 things like access to the internet
16:42 access to money ability to send and
16:45 receive messages order pizza or whatever
16:47 for this prompt engineering becomes even
16:49 more important because your autonomous
16:51 tool wielding agent can do a lot of good
16:54 or a lot of harm depending on how well
16:55 you craft that mission
16:58 statement all right let's wrap it up
17:00 here are the key things I hope you will
17:02 remember from this video generative AI
17:04 is a super useful tool that can help
17:06 both you your team and your company in a
17:08 big way the better you understand it the
17:10 more likely it is to be an opportunity
17:12 rather than a threat generative AI is
17:14 more powerful than you think the biggest
17:17 limitation is not the technology but
17:19 your imagination like what can I do and
17:21 your prompt engineering skills how do I
17:24 do it prompt engineeringdesign is a
17:27 crucial skill like all new skills just
17:29 accept that you will kind of suck at at
17:31 first but you'll improve over time with
17:34 deliberate practice so my best tip is
17:36 experiment make this part of your
17:38 day-to-day life and the Learning Happens
17:40 automatically hope this video was
