0:02 AI is simpler than you think let me show
0:10 so chapter one the types of AI before AI
0:11 machines were built to follow our
0:13 Specific Instructions they did things
0:16 based on conditions we programmed into
0:18 them they were rules-based but AI is
0:19 different instead of giving machines
0:22 instructions on what to do and how to do
0:24 it we trained them to think and do
0:26 things on their own like raising a child
0:29 this is why people call AI a black box
0:31 it turns your input into an output based
0:33 on similar things you've shown it in the
0:35 past but you don't know what's happening
0:37 in the Box exactly AI basically combines
0:39 your input in the moment with the data
0:41 it's been trained on to generate an
0:43 output and there's three types of AI
0:45 boxes you should know about and you can
0:47 differentiate them based on their output
0:50 first we have predictive AI which labels
0:52 Things based on prior data like marking
0:55 an email as spam identifying someone in
0:57 a picture or recommending what you
1:00 should watch generative AI creates new
1:02 content like text images and videos a
1:05 gentic AI outputs actions based on a
1:07 given task like self-driving cars they
1:10 take a destination as an input then they
1:12 plan the trip drive the car while
1:14 stopping at traffic lights following
1:16 speed limits and not running over
1:19 pedestrians another example is AI agents
1:21 like Cloud computer use if you give them
1:23 a task they can access your computer
1:26 apps or accounts to execute actions on
1:28 your behalf these are the types of AI
1:31 but you can also Define AI by its level
1:33 of intelligence narrow AI is built for
1:36 specific tasks which applies to all AI
1:38 systems we have today artificial general
1:41 intelligence or AGI is theoretical for
1:43 now but would match human intelligence
1:46 across all tasks and scenarios but
1:48 nobody agrees on a formal definition of
1:50 what AGI actually is and finally
1:52 artificial super intelligence is also
1:54 theoretical but would surpass human
1:58 intelligence by far chapter 2 how AI is
2:01 created imagine adopted a pet robot and
2:03 you wanted to teach him how to think and
2:05 do things on his own this is called
2:06 machine learning and there's three ways
2:08 you can do this you could take your
2:10 robot around the world and point at
2:12 every object to teach him what it is
2:14 over and over again you're spoon feeding
2:16 him information until he recognizes what
2:18 each thing is this is supervised
2:20 learning and you're feeding a machine a
2:23 sandwich of labeled data like reviews
2:26 emails and even x-rays you could also
2:27 take a hands-off approach by letting him
2:30 learn on his own imagine you were living
2:31 the house and you told your robot to
2:33 sort the dishes it would start grouping
2:36 things based on similarities in shape
2:38 size and color this is UN supervised
2:40 learning and it Powers systems that
2:42 group similar items when you're shopping
2:44 browsing pictures or getting song
2:46 recommendations the final way you could
2:48 teach him is with positive and negative
2:51 reinforcement like a coach imagine you
2:52 asked your robot to make you a smoothie
2:55 it has no idea how to do it at first so
2:56 it starts experimenting with random
2:58 recipes after you taste each smoothie
3:00 you give it a thumbs up or a thumbs down
3:02 over time the robot learns to make
3:04 smoothies just the way you like them
3:06 this is reinforcement learning and it's
3:08 one of the ways chat GPT was trained to
3:10 improve its answers and the reason why
3:12 your algorithm knows you too well
3:13 creating intelligent machines isn't just
3:15 about the Training Method it's also
3:18 about the quality of data that you use
3:19 if your robot learns from low quality
3:21 information he'll only give you low
3:24 quality results garbage in garbage out
3:26 high quality data is more valuable than
3:28 ever but we might run out of it soon one
3:30 research group predicts that will run
3:31 out of publicly available data between
3:35 2026 and 2032 at the current pace of AI
3:37 development this is why AI companies are
3:39 signing expensive licensing deals to
3:41 find new training data the other
3:42 solution being explored is using
3:46 synthetic data generated by AI to create
3:49 new training data chapter three how AI
3:51 becomes biased training an AI system is
3:53 like raising a child the parents play a
3:55 big role in deciding the values they
3:57 teach them what they can watch on TV and
3:59 where they live and these factors shape
4:01 a child's worldview imagine a child who
4:03 grew up in Antarctica and the only
4:05 animal they ever saw was a penguin they
4:07 would think that penguins are the only
4:08 animals in the world but they have an
4:10 incomplete worldview because they
4:12 haven't seen all the other animals that
4:14 exist AI has the same problem which can
4:16 limit its abilities but also create
4:18 unfavorable outcomes for some people and
4:19 this bias can come from different
4:21 sources first the beliefs and values of
4:23 people can influence how AI systems are
4:25 designed and trained if the child's
4:27 father was an avid penguin Enthusiast he
4:29 might constantly talk about how penguins
4:31 are the best animals on Earth the child
4:33 would adopt the same belief because it's
4:34 all they've been exposed to and this is
4:36 the same reason why some chat Bots will
4:38 respond differently if you ask them
4:40 controversial questions bias can also
4:42 exist within the training data if it has
4:44 any imbalances that ignore the nuances
4:46 of the world if the parents only give
4:48 their child books about penguins there's
4:49 no way for them to learn about new and
4:51 different species this is why language
4:53 models might give you lower quality
4:55 responses in languages other than
4:57 English bias can also emerge from
4:59 Incorrect and overs simplistic rules
5:01 that machine identifies during training
5:03 the child who has only seen Penguins
5:04 would think that animals are only
5:07 creatures that are black white and have
5:08 flippers and this is why image
5:11 generators associate certain jobs with
5:13 specific genders and races unfortunately
5:14 there's no reliable solution to
5:17 eliminate bias from AI after all bias is
5:19 part of being human and if we're trying
5:21 to replicate the way humans think and
5:24 act bias is inevitable AI is only a
5:26 mirror image of us after all chapter 4
5:29 how AI generates text imagine you wanted
5:30 to become a DJ you would start by
5:32 listening to as many songs as possible
5:34 you would break down each song to
5:35 analyze the sequence of Beats and
5:37 instruments that were used and after
5:38 lots of listening you become comfortable
5:40 creating unique beats and mixes that
5:42 sound good to the ear large language
5:43 models are built in a similar way
5:45 they're trained on millions of pages of
5:47 text each piece of text is broken down
5:49 into tokens which are numerical
5:51 representations of words with billions
5:53 of examples machines can remember the
5:55 sequences of tokens to create language
5:57 that sounds good to the ear asking AI to
5:59 generate text for you is like requesting
6:00 a song from from a DJ based on your
6:02 request the songs learned during
6:05 training and the mixers the DJ creates a
6:07 unique music mix for you with language
6:08 models your text prompt is combined with
6:10 the training data and parameters to
6:12 create a new mix of text text generation
6:14 is like a math formula it takes your
6:16 inputs multiplies them by specific
6:18 parameters to give you a token of text
6:20 except that language models have
6:22 billions of parameters AI generates this
6:24 new mix of text by predicting the
6:26 sequence of tokens one at a time based
6:28 on all previous tokens and each token is
6:30 the result of hundreds of billions of
6:33 calculations like a DJ anxiously trying
6:35 to keep the crowd going after every beat
6:37 switch but sometimes things go wrong
6:39 since every token is a best guess these
6:41 guesses can be wrong when asking for
6:43 factual information these are called
6:44 hallucinations remember that language
6:47 models are text generators not truth
6:50 generators chapter 5 how AI generates
6:52 images imagine you were a sculptor and
6:53 someone asks you to make a statue you
6:55 would get a stone block and start
6:56 chisling away until you get your
6:58 Masterpiece AI generates images in a
7:00 similar way it starts with an image full
7:03 of random pixels and gradually adjusts
7:05 them until you get the final result but
7:07 how does it do that the process is
7:08 called diffusion these models are
7:10 trained on billions of images with text
7:12 descriptions but machines don't see
7:14 images like we do they see them as grids
7:17 of pixels and each pixel is represented
7:19 by three numbers for red green and blue
7:21 when you train an AI model on billions
7:23 of images it starts recognizing what
7:25 words are associated with which pixel
7:27 patterns and values this creates a
7:29 multi-dimensional map called the latent
7:31 space and it Maps specific features like
7:34 a shape artistic style or color to their
7:36 pixel values and patterns to capture all
7:38 these fine details these models are
7:40 trained by taking each image and slowly
7:42 randomizing all the pixels until it
7:44 becomes Pure Noise then the model is
7:46 trained to reverse the process to
7:48 reconstruct the original image so when
7:50 you ask AI to make an image for you it's
7:52 replicating that reversal process by
7:54 taking an image with random pixels
7:56 adjusting the pixel values based on the
7:58 latent space until it constructs the
8:01 image you want Che chapter 6 the energy
8:03 cost of AI every time you ask AI a
8:05 question that request is sent to a Data
8:07 Center and processed by thousands of AI
8:09 chips because the hardware in our phones
8:11 and laptops isn't powerful enough to do
8:14 it these data centers need two things to
8:16 operate first they need electricity to
8:19 run in 2024 data centers use 2% of the
8:20 world's electricity but this number is
8:22 for all data centers that keep the
8:24 internet running we don't know what
8:25 share of that can be attributed to AI
8:28 alone it's estimated that a single AI
8:30 request uses 10 times more energy than a
8:32 standard search but this estimate is
8:34 based on AI models from 2023 which are
8:36 much smaller than the ones we have today
8:38 second they need cooling systems to
8:41 prevent overheating most cooling systems
8:43 use fresh water and the average data
8:45 center consumes 300,000 gallons of water
8:48 every day which is equivalent to 19,000
8:50 showers and depending on the cooling
8:52 method used data centers may lose some
8:55 or all of this water and need constant
8:57 replenishment unfortunately we don't
8:59 have precise or reliable numbers for how
9:02 much energy or water AI uses alone
9:03 because tech companies aren't revealing
9:06 the size of their newest AI models what
9:08 we do know is that AI developments and
9:10 usage are likely to keep increasing over
9:12 the next few years which will require
9:14 more data centers AKA more electricity
9:17 and water than before thankfully some
9:19 solutions are starting to emerge to make
9:21 AI more energy efficient and here are
9:22 three you should know about first we
9:24 could run AI on our personal devices
9:26 instead of data centers because they've
9:28 been getting more powerful Hardware this
9:30 is called Edge AI and we're already
9:31 seeing signs of it for example the new
9:34 iPhone can run basic AI features on your
9:36 device directly without sending those
9:37 requests to a data center another
9:40 solution is called Model distillation
9:41 the idea is to create smaller models
9:44 based on large models to make them more
9:46 energy efficient and cost efficient for
9:48 specific tasks because we don't always
9:50 need the biggest and most advanced model
9:51 if a task is simple the same way we
9:53 don't need a Formula 1 car to deliver a
9:55 pizza finally the AI chips and data
9:57 centers are getting more efficient which
9:59 means they need less energy to produce
10:01 the same output over the past 8 years
10:03 the energy cost to generate one token on
10:07 an Nvidia GPU went down from 177,000
10:11 Jews to 0.4 chapter 7 a brief history of
10:13 AI the story of intelligent machines
10:16 begins in 1950 When Alan Turing famously
10:19 asked can machines think and proposed
10:20 the Turing test as a measure of
10:22 humanlike intelligence but the term
10:24 artificial intelligence wasn't born
10:27 until 1956 during a summer research
10:29 project at Dartmouth College from the '
10:32 50s to the 70s the AI field received
10:34 generous Government funding from DARPA
10:36 and the field witnessed its first
10:38 breakthroughs like Eliza a chatbot that
10:41 could talk to you like a psychotherapist
10:43 the hype and expectations around AI at
10:44 the time were Skyhigh High some
10:47 researchers even made bold predictions
10:48 like saying that machines would be as
10:51 smart as we are within 8 years but the
10:53 industry faced a massive blow once it
10:55 couldn't deliver on its promises two
10:57 government reports from the US and the
10:59 UK concluded that AI was more expensive
11:02 Ive less reliable slower and not solving
11:04 any useful problems yet this sent the
11:06 industry into its first winter where
11:09 funding dried up until the80s the field
11:11 went through a slight Resurgence After
11:13 Japan invested heavily in its Computing
11:15 industry which pressured the UK and the
11:17 us to do the same but they failed to
11:19 deliver on their promises again and
11:22 entered a second winter at this point 40
11:24 years had gone by with no fruitful
11:26 applications of AI and the industry was
11:29 at Rock Bottom some computer scientists
11:30 were even ashamed to say that they
11:33 worked on AI and avoided the term
11:35 completely but things were about to take
11:37 a dramatic turn in the late 9s because
11:39 the perfect storm was brewing first
11:41 computers were becoming way more
11:43 powerful quickly computer chips today
11:46 are 40 million times more powerful than
11:48 they were 50 years ago second there was
11:50 way more Digital Data available as we
11:53 moved all of our lives and work online
11:54 which could be collected as training
11:57 data and both of these things enabled
11:58 machine learning algorithms that can
12:01 replic ate and simulate intelligence
12:04 since the 2010s the AI field has slowly
12:06 recaptured The public's attention and
12:07 became an overnight sensation with the
12:09 release of chat GPT big tech companies
12:12 are racing to the top everyone has an AI
12:14 startup and people are Torn Between the
12:16 benefits AI can provide the harm it
12:18 creates and most importantly the
12:21 existential question if AI can do it all
12:23 what am I good for the future is hard to
12:25 predict but it feels like the real story
12:27 is just beginning I hope all of this
12:29 gives you a good understanding of what
12:32 AI actually is and thank you for
12:33 watching my first YouTube video if you
12:36 have any questions or feedback leave
12:37 them in the comments I'll take a look at
12:39 them so I can keep making better videos
Chrome Extension