00:05ever since computers were invented
00:07they've really just been glorified
00:08calculators machines that execute the
00:11exact instructions given to them by the
00:13programmers but something incredible is
00:15happening now computers have started
00:16gaining the ability to learn and think
00:19and communicate just like we do they can
00:21do creative intellectual work that
00:23previously only humans could do we call
00:25this technology generative Ai and you
00:27may have encountered it already through
00:29products like GPT basically intelligence
00:32is now available as a service kind of
00:34like a giant brain floating in the sky
00:36that anyone can talk to it's not perfect
00:39but it is surprisingly capable and it is
00:40improving at an exponential rate this is
00:43a big deal it's going to affect just
00:45about every person and Company on the
00:47planet positively or negatively this
00:49video is here to help you understand
00:51what generative AI is all about in
00:53Practical terms beyond the hype the
00:54better you understand this technology as
00:56a person team or company the better
00:58equipped you will be to survive and
01:00thrive in the age of AI so here's a
01:03silly but useful mental model for this
01:05you have Einstein in your basement in
01:07fact everyone does and by Einstein I
01:10really mean the combination of every
01:12smart person who ever lived you can talk
01:14to Einstein whenever you want he has
01:16instant access to the sum of all human
01:18knowledge and will answer anything you
01:20want within seconds never running out of
01:21patience he can also take on any role
01:23you want a comedian poet doctor coach
01:27and will be an expert within that field
01:29he has some human-like limitations
01:31though he can make mistakes he can jump
01:33to conclusions he can misunderstand you
01:35but the biggest limitation is actually
01:37your imagination and your ability to
01:39communicate effectively with them this
01:41skill is known as prompt engineering and
01:43in the age of AI this is as essential as
01:46reading and writing most people vastly
01:49underestimate what this Einstein in your
01:51basement can do it's like going to the
01:53real Einstein and asking him to proof
01:55read a high school report or hiring a
01:56worldclass five-star chef and having him
01:59chop onion the more you interact with
02:01Einstein the more you will discover
02:02surprising and Powerful ways for him to
02:04help you or your company okay enough
02:07fluffy metaphors let's clarify some
02:08terms AI as you probably know stands for
02:11artificial intelligence AI is not new FS
02:14like machine learning and computer
02:15vision have been around for decades
02:17whenever you see a YouTube
02:18recommendation or a web search result or
02:21whenever you get a credit card
02:22transaction approved that's traditional
02:24AI in action generative AI is AI that
02:27generates new original content rather
02:29than just finding or classifying
02:30existing content that's the G in GPT for
02:33example large language models or llms
02:36are a type of generative AI that can
02:38communicate using normal human language
02:41chat GPT is a product by the company
02:43open AI it started as an llm essentially
02:46an advanced chatbot using a new
02:47architecture called the Transformer
02:49architecture which by the way is the T
02:51in GPT it is so fluent at human language
02:54that anyone can use it you don't need to
02:55be an AI expert or programmer and that's
02:57kind of what triggered the whole
02:58Revolution so how does it actually work
03:02well a large language model is an
03:03artificial neural network basically a
03:06bunch of numbers or or parameters
03:08connected to each other similar to how
03:09our brain is a bunch of neurons or brain
03:11cells connected to each other neural
03:12networks only deal with numbers you send
03:15in numbers and depending on how the
03:16parameters are set all the numbers come
03:18out but any kind of content such as text
03:20or images can be represented as numbers
03:22so let's say I write dogs are when I
03:25send that to a large language model that
03:27gets converted to numbers processed by
03:29the neural network and then the
03:30resulting numbers are converted back
03:31into text in this case the word animals
03:34dogs are animals so yeah this is
03:36basically a guest to next word machine
03:39the interesting part is if we take that
03:41output and combine it with the input and
03:43send it through the model again then it
03:45will continue adding new words that's
03:46what's going on behind the scenes when
03:48you type something in chat GPT in this
03:50case for example it generated a whole
03:51story and I can continue this
03:53indefinitely by adding more
03:55prompts a large language model may have
03:58billions or even trillions of of
03:59parameters that's why they're called
04:01large so how are all these numbers set
04:04well not through manual programming that
04:06would be impossible but through training
04:09just like babies learning to speak a
04:10baby isn't told how to speak she doesn't
04:13get an instruction manual instead she
04:15listens to people speaking around her
04:16and when she's heard enough she starts
04:18seeing the pattern she speaks a few
04:20words at first to the Delight of her
04:21parents and then later on full sentences
04:24similarly during a training period the
04:26language model is fed a mindboggling
04:28amount of text to learn from Mostly from
04:31internet sources it then plays guess the
04:33next word with all of this over and over
04:35again and the parameters are
04:37automatically tweaked until it starts
04:38getting really good at predicting the
04:40next word this is called back
04:41propagation which is a fancy term for oh
04:44I guessed wrong I better change
04:45something however to become truly useful
04:47a model also needs to undergo human
04:49training this is called reinforcement
04:51learning with human feedback and it
04:53involves thousands of hours of humans
04:55painstakingly testing and evaluating
04:57output from the model and giving
04:58feedback kind of like training a a dog
05:01with a clicker to reinforce good
05:02behavior that's why a model like GPT
05:04won't tell you how to rob a bank it
05:06knows very well how to rob a bank but
05:08through human training it has learned
05:09that it shouldn't help people commit
05:11crimes when training is done the model
05:13is mostly Frozen other than some fine
05:15tuning that can happen later that's what
05:17the P stands for in GPT pre-trained
05:19although in the future we will probably
05:20have models that can learn continuously
05:22rather than just uh during training and
05:24fine-tuning now although chat GPT kind
05:26of got the ball rolling GPT isn't the
05:29only model out there in fact new models
05:32mushrooms they vary a lot in terms of
05:34speed capability and cost some can be
05:37downloaded and run locally others are
05:38only online some are free or open source
05:41others are commercial products some are
05:43super easy to use While others require
05:45complicated technical setup some are
05:47specialized for certain use cases others
05:50are more General and can be used for
05:51almost anything and some are baked into
05:54products in the form of co-pilots or or
05:56chat windows it's it's the Wild West
06:00just keep in mind that you generally get
06:01what you pay for so with a free model
06:04you may just be getting a smart high
06:06school student in your basement rather
06:08than Einstein the difference between for
06:11example GPT 3.5 and GPT 4 is massive
06:15note that there are different types of
06:17generative AI models that generate
06:18different types of content textto text
06:20models like gpc4 take text as input and
06:23generate text as output the text can be
06:25natural language but it can also be
06:27structured information like code Json
06:29HTML I use this a lot myself to generate
06:32code when programming uh it saves an
06:34incredible amount of time and I also
06:35learn a lot from the code it generates
06:37text to image models will generate
06:39images describe what you want and an
06:41image gets generated for you you can
06:42even pick a style image to image models
06:45can do things like transforming or
06:47combining images and we have image to
06:50text models which describe the contents
06:52of a given image and speech to text
06:54models create voice transcriptions which
06:56is useful for things like uh meeting
06:58notes text Audio models they generate
07:00music or sounds from a prompt for
07:02example here is some sound generated
07:04from The Prompt people talking in a
07:08busy okay guys enough stop now thank you
07:13and there are even text to video models
07:15that generate videos from a prompt
07:17sooner or later we'll have infinite
07:18movie series that autog generate the
07:20next episode tailored to your tastes as
07:22you're watching kind of scary if you
07:24think about it one Trend now is
07:26multimodal AI products meaning they
07:28combine different models into one
07:30product so you can work with text images
07:32audio Etc without switching tools the
07:35chat GPT mobile app is a good example of
07:37this just for fun I took a photo of this
07:40room and I asked where I could hide
07:41stuff I kind of like that it mentioned
07:44the stove but warned that that it could
07:46get hot there when I have things to
07:48figure out such as the contents of this
07:50video I like to take walks using chat
07:52GPT as as a sounding board I start by
07:55saying always respond with the word okay
07:57unless I ask you for something that way
07:59it'll just listen and not interrupt
08:01after I finish dumping my thoughts I ask
08:03for feedback we have some discussion and
08:06then I ask it to summarize and text
08:07afterwards I really recommend trying
08:09this it's it's a really useful way to
08:11use tools like this turns out Einstein
08:13isn't stuck in the basement after all
08:15you can take him out for a walk
08:17initially language models were just word
08:19predictors statistical machines with
08:22limited practical use but as they became
08:24larger and were trained on more data
08:26they started gaining emergent
08:28capabilities expected capabilities that
08:30surprised even the developers of the
08:31technology they could roleplay write
08:34poetry write high quality code discuss
08:36company strategy provide legal and
08:38medical advice coach teach basically
08:41creative and intellectual things that
08:43only humans could do previously it turns
08:45out that when a model has seen enough
08:47text and images it starts to see
08:49patterns and understand higher level
08:51Concepts just like a baby learning to
08:53understand the world let's take a simple
08:55example I'll give gpc4 this little
08:57drawing that involves a string a pair of
09:00scissors an egg a pot and a fire what
09:03will happen if I use the scissors the
09:05model has most likely not been trained
09:07on this exact scenario yet it gave a
09:10pretty good answer which demonstrates a
09:11basic understanding of the nature of
09:13scissors eggs gravity and heat when GPT
09:164 was released I started using it as a
09:18coding assistant and I was blown away
09:20when prompted effectively it was a
09:22better programmer than anyone I've
09:23worked with same with article writing
09:25product design Workshop planning and
09:27just about anything I used it for for
09:29the main bottleneck was my prompt
09:32engineering skills so I decided to make
09:33a career shift and focus entirely on
09:35learning and teaching how to make this
09:37technology useful hence this video now
09:40let's take a step back and look at the
09:41implications for 300,000 years or so we
09:44homosapiens have been the most
09:46intelligent species on Earth depending
09:48of course on how you define intelligence
09:50but the thing is our intellectual
09:51capabilities aren't really improving
09:53that much our brains are about the same
09:55size same weight as they've been for
09:56thousands of years computers on the
09:58other hand have been around for only 80
10:00years or so and now with generative AI
10:02they are suddenly capable of speaking
10:04human languages fluently and carrying
10:06out an increasing number of intellectual
10:08creative tasks that previously only
10:10humans could do so we are right here at
10:12the Crossing Point where AI is better at
10:14some things and humans are better at
10:15some things but ai's capabilities are
10:17improving at an exponential rate while
10:19ours aren't we don't know how long that
10:22exponential Improvement will continue or
10:24if it will level off at some point but
10:25we're definitely entering a new world
10:27order now this isn't the first
10:29Revolution we've experienced we tamed
10:31fire we learned how to do agriculture we
10:33invented the printing press steam power
10:35Telegraph these were all revolutionary
10:37changes but they took decades or
10:39centuries to become widespread in the AI
10:42Revolution new technology spreads
10:44worldwide almost instantly dealing with
10:46this rate of change is a huge challenge
10:48for both individuals and
10:50companies I've noticed that people and
10:52companies tend to fall into different
10:54kind of mindset categories when it comes
10:56to AI on one side we have denial the
10:59belief that AI cannot do my job or we
11:02don't have time to look into this
11:03technology this is a dangerous place to
11:05be a common saying is AI might not take
11:08your job but people using AI will and
11:11this is true for both individuals and
11:13companies on the other side of the scale
11:15we have panic and despair the belief
11:16that AI is going to take my job no
11:18matter what AI is going to make my
11:19company go bankrupt neither of these
11:21mindsets are helpful so I propose a
11:24middle ground a balanced positive
11:26mindset AI is going to make me my team
11:28my company insanely productive
11:31personally with this mindset I feel like
11:33I've gained superpowers I can go from
11:35idea to result in so much shorter time I
11:38can focus more on what I want to achieve
11:40and less on the grunt workk of building
11:41things and I'm learning a lot faster too
11:43it's like having an awesome Mentor with
11:45me at all times this mindset not only
11:47feels good but it also equips you for
11:49the future makes you less likely to lose
11:51your job or your company and more likely
11:53to thrive in the age of AI despite all
11:56uncertainty so one important question is
11:59is human role X needed in the age of AI
12:02for example are doctors needed
12:03developers lawyers CEOs uh whatever so
12:06this question becomes more and more
12:08relevant as the AI capabilities improve
12:11well some jobs will disappear for sure
12:13but for most roles I think we humans are
12:15still needed someone with domain
12:17knowledge still needs to decide what to
12:19ask the AI how to formulate The Prompt
12:21what context needs to be provided and
12:23how to evaluate the result AI models
12:25aren't perfect they can be absolutely
12:27brilliant sometimes but sometimes also
12:30terribly stupid they can sometimes
12:32hallucinate and provide bogus
12:33information in a very convincing way so
12:36when should you trust the AI response
12:38when should you double check or do the
12:39work yourself what about legal
12:41compliance data security what
12:43information can we send to an AI model
12:45and where is that data stored a human
12:48expert is needed to make these judgment
12:50calls and compensate for the weaknesses
12:51of the AI model so I recommend thinking
12:54of AI as your colleague a genius but
12:56also an oddball with some personal
12:58quirks that you need to learn to work
12:59with you need to recognize when your
13:01Genius colleague is drunk as a doctor my
13:04AI colleague can help diagnose rare
13:06diseases that I didn't even know existed
13:08as a lawyer my AI colleague could do
13:10legal research and review contracts
13:12allowing me to spend more time with my
13:13client or as a teacher my AI colleague
13:16could grade tests help generate course
13:18content provide individual support to
13:20students Etc and if you're not sure how
13:22it can help you just ask it I work as X
13:26how can you help me overall I find that
13:29the combination of human plus AI That's
13:31where the magic lies it's important to
13:34distinguish between the models and the
13:36products that build on top of them as a
13:38user you don't normally interact with
13:39the model directly instead you interact
13:41with a product website or a mobile app
13:43which in turn talks to the model behind
13:45the scenes products provide a user
13:47interface and add capabilities and data
13:49that aren't part of the model itself for
13:51example the chat gbt product keeps track
13:54of your message history while the GPT 4
13:56model itself doesn't have any message
13:58history history as a developer you can
14:00use these models to build your own AI
14:02powered products and features for
14:04example let's say you have an e-learning
14:06site you could add a chat bot to answer
14:08questions about the courses or as a
14:10recruitment company you might build AI
14:12powered tools to help evaluate
14:13candidates in both these cases your
14:15users interact with your product and
14:17then your product interacts with the
14:18model this is done via apis or
14:20application programming interfaces which
14:22allow your code to talk to the model so
14:24here's a simple example of using open AI
14:29not a lot of code needed and here's
14:31another example of the automatic
14:33candidate evaluation thing I talked
14:34about it takes a job description and a
14:37bunch of CVS in a folder and evaluates
14:39each candidate automatically and
14:41incidentally the code itself is mostly
14:43AI written as a product developer you
14:46can use AI models kind of like an
14:48external brain to insert intelligence
14:50into your product very powerful in order
14:54to use generative AI effectively you
14:56need to get good at prompt engineering
14:58or prompt design as I prefer to call it
15:00this skill is needed both as a user and
15:02as a product developer because in both
15:04cases you need to be able to craft
15:06effective prompts that produce useful
15:08results from an AI model here's an
15:10example let's say I want help planning a
15:12workshop this prompt is unlikely to give
15:15useful results because no matter how
15:17smart the AI is if it doesn't know the
15:19context of my workshop it can only give
15:21fague high level recommendations the
15:23second prompt is better now I provided
15:25some context this is normally done
15:27iteratively write a prompt look at the
15:29result add a follow-up prompt to provide
15:31more information or edit the original
15:33prompt and rinse and repeat until you
15:35get a good result in this third approach
15:37I ask it to interview me so instead of
15:39me providing a bunch of context up front
15:42I'm basically saying what do you need to
15:43know in order order to help me and then
15:45it will propose a workshop agenda after
15:47I often combine these two I provide a
15:49bit of context and then I tell it to ask
15:51me if it needs any more information
15:53these are just some examples of prompt
15:54engineering techniques so overall the
15:57better you get at prompt engineering the
15:58faster and better results you will get
16:00from AI there are plenty of courses
16:02books videos articles to help you learn
16:04this but the most important thing is is
16:06to practice and Learn by doing a nice
16:08side effect is that you will become
16:09better at communicating in general since
16:11prompt engineering is really all about
16:13Clarity and effective
16:15communication I think the next Frontier
16:17for generative AI is autonomous agents
16:19with tools these are AI powerered
16:21software entities that run on their own
16:23rather than just sitting around waiting
16:24for you to prompt them all the time so
16:26you go down to Einstein in your basement
16:28and do what a good leader would do for a
16:29team you give him a high level Mission
16:31and the tools needed to accomplish it
16:33and then open the door and let him out
16:35to run his own show without
16:36micromanagement the tools could be
16:38things like access to the internet
16:40access to money ability to send and
16:42receive messages order pizza or whatever
16:45for this prompt engineering becomes even
16:47more important because your autonomous
16:49tool wielding agent can do a lot of good
16:51or a lot of harm depending on how well
16:54you craft that mission
16:55statement all right let's wrap it up
16:58here are the key things I hope you will
17:00remember from this video generative AI
17:02is a super useful tool that can help
17:04both you your team and your company in a
17:06big way the better you understand it the
17:08more likely it is to be an opportunity
17:10rather than a threat generative AI is
17:12more powerful than you think the biggest
17:14limitation is not the technology but
17:17your imagination like what can I do and
17:19your prompt engineering skills how do I
17:21do it prompt engineeringdesign is a
17:24crucial skill like all new skills just
17:27accept that you will kind of suck at it
17:29at first but you'll improve over time
17:31with deliberate practice so my best tip
17:33is experiment make this part of your
17:36day-to-day life and the Learning Happens
17:38automatically hope this video was