YT YouTube2Text

YouTube2Text → Transcripts → Harvard Professor Explains Algorithms in 5 Levels of Difficulty | WIRED

Harvard Professor Explains Algorithms in 5 Levels of Difficulty | WIRED — Transcript

by WIRED · 5,390 words · 757 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00hello world my name is David J Ma and
0:02I'm a professor of computer science at
0:04Harvard University today I've been asked
0:05to explain algorithms in five levels of
0:08increasing difficulty algorithms are
0:11important because they really are
0:12everywhere not only in the physical
0:14world but certainly in the virtual world
0:16as well and in fact what excites me
0:17about algorithms is that they really
0:19represent an opportunity to solve
0:21problems and I dare say no matter what
0:23you do in life all of us have problems
0:25to
0:27solve so I'm a computer science
0:29professor so I spend a lot of time with
0:31computerss how would you define a
0:32computer for them well a computer is
0:36electronic like a phone but it's um a
0:39rectangle and you like can type like
0:43tick tick tick and you work on it nice
0:46do you know any of the parts that are
0:48inside of a computer um no can I explain
0:52a couple of them to you yeah so like
0:53inside of every computer is some kind of
0:56brain and the technical term for that is
0:58CPU or Central processing unit and those
1:01are the pieces of Hardware that know how
1:04to respond to those instructions like
1:06moving up or down or left or right knows
1:09how to do math like addition and
1:11subtraction and then there's at least
1:12one other type of Hardware inside of a
1:15computer called memory or Ram if you've
1:17heard of this I know memory because you
1:19have to memorize stuff yeah exactly and
1:21computers have even different types of
1:23memory they have what's called Ram
1:25random access memory which is where your
1:27games where your programs are stored
1:29while they being used but then it also
1:31has a a hard drive or a solid state
1:33drive which is where your data your high
1:36scores your documents once you start
1:38writing essays and and stories in the
1:40future stays there stays permanently so
1:42even if the power goes out the computer
1:44can still remember that information it's
1:46still there because the computer can't
1:49just like delete all the words itself
1:53because your fingers can only do that
1:55like you have to use your finger to
1:58delete all the stuff exactly have you
2:00heard of an algorithm before um yes
2:02algorithm is a list of instructions to
2:06tell people what to do or like a robot
2:09what to do yeah exactly it's so it's
2:11just stepbystep instructions for doing
2:14something for solving a problem for yeah
2:16so like if you have a bedtime routine
2:18then first you say I get dressed I brush
2:22my teeth I read a little story and then
2:25I go to bed all right well how about
2:27another algorithm like um what do you
2:28tend to eat for for lunch any types of
2:30sandwiches you like uh I eat peanut
2:32butter and let me get some supplies from
2:34the cupboard here so should we make an
2:36algorithm together for why don't we do
2:39it this way why don't we pretend like
2:40I'm a computer or maybe I'm a robot so I
2:42only understand your instructions and so
2:45I want you to feed me no pun intended an
2:47algorithm so step-by-step instructions
2:49for solving this problem but remember
2:52algorithms you have to be precise you
2:54have to give the right instructions the
2:57right instructions just do it for me
2:59soep step one was what open the bag okay
3:02opening the bag of bread stop now grab
3:06the bread and put it on the plate grab
3:07the bread and put it on the
3:10plate take all the bread back and put it
3:13back in there so that's like an undo
3:15command little control Z okay take one
3:19bread and put it on the plate take the
3:21lid off the peanut butter okay take the
3:23lid off the peanut butter put the lid
3:26down okay take the knife take the knife
3:29put the blade inside the peanut butter
3:32and spread the peanut butter on the
3:34bread I'm going to take out some peanut
3:36butter and I'm going to spread the
3:38peanut butter on the bread I put a lot
3:41of peanut butter on because I love
3:42peanut butter apparently I thought I was
3:44messing with you here but I think you're
3:46happy with this Put The Knife down and
3:49then grab one bread and put it on top of
3:52the second bread
3:53sideways
3:56sideways like put it flat on oh flat
3:59ways okay and now done you're done with
4:01your sandwich should we take a delicious
4:03bite yep let's take a bite okay here we
4:07go what would be the next step be you
4:09here clean all this mess up clean all
4:12this mess up right we made an algorithm
4:15step-by-step instructions for solving
4:17some problem and if you think about now
4:18how we made peanut butter and jelly
4:20sandwiches sometimes we were imprecise
4:22you didn't give me quite enough
4:24information to do the algorithm
4:25correctly and that's why I took out so
4:26much bread Precision being very very
4:30correct with your instructions is so
4:32important in the real world because for
4:33instance when you're using the worldwide
4:35web and you're searching for something
4:36on Google or being you want to do the
4:39right thing so like if you type in just
4:42Google then you won't find the answer to
4:45your question pretty much everything we
4:47do in life is an algorithm even if we
4:49don't use that fancy word to describe it
4:51because you and I are sort of following
4:53instructions either that we came up with
4:55ourselves or maybe our parents told us
4:57how to do these things and so those are
4:59just algorithms but when you start using
5:01algorithms in uh computers that's when
5:03you start writing
5:08code what do you know about algorithms
5:11nothing really um at all honestly I
5:13think it's just probably a way to store
5:15information um in computers and I dare
5:17say even though you might not have put
5:19this word on it odds are you executed as
5:21a human multiple algorithms today even
5:25before you came here today like what
5:26were a few things that you did I got
5:28ready okay and get ready what does that
5:30mean brushing my teeth brushing my hair
5:32okay put getting dressed okay so all of
5:34those frankly if we really um Dove more
5:37deeply could be broken down into
5:39stepbystep instructions and presumably
5:41your mom your dad someone in the past
5:44sort of programmed you as a human to
5:46know what to do and then after that as a
5:48smart human you can sort of take it from
5:49there and you don't need their help
5:50anymore but that's kind of what we're
5:51doing when we program computer something
5:54maybe even more familiar nowadays like
5:55odds are you have a cell phone your
5:57contacts or your dress book but let me
5:59ask you why that is like why does Apple
6:02or Google or anyone else bother
6:03alphabetizing your contacts I just
6:05assumed it would be easier to navigate
6:07what if your friend happened to be at
6:10the very bottom of this randomly
6:11organized list like why is that a
6:13problem like he or she is still there I
6:15guess it would take a while to get to
6:16while you're scrolling that in of itself
6:17is kind of a problem or it's an
6:19inefficient solution to the problem so
6:21it turns out that back in my day before
6:22there were cell phones like everyone's
6:24numbers from high schools like were
6:26literally printed in a book and everyone
6:27in my town and my city my state was
6:29printed in an actual phone book even if
6:31you've never seen this technology before
6:33how would you propose verbally to find
6:35John in this phone book or I would just
6:37flip through and just look for the J I
6:39guess yeah so let me propose that we
6:40start that way I could just start at the
6:42beginning and step by step I could just
6:44look at each page looking for John
6:47looking for John now even if you've
6:49never seen this here technology before
6:51it turns out this is exactly what your
6:52phone could be doing in software like
6:55someone from Google or Apple or the like
6:57they could write software that uses a
6:59technique in programming known as a loop
7:01and a loop as the word implies is just
7:02sort of do something again and again
7:04what if instead of starting from the
7:05beginning and going one page at a time
7:07what if I or what if your phone goes
7:09like two pages or two names at a time
7:11would this be correct do you think well
7:13you could skip over John I think in what
7:15sense if he's in one of the middle pages
7:17that you skipped over yeah so sort of
7:19accidentally and frankly with like 50/50
7:21probability John could get sandwiched in
7:22between two pages but does that mean I
7:24have to throw that algorithm out alt
7:27together maybe you could use that
7:28strategy until you get close to the
7:30section and then switch to going one by
7:31one okay that's nice so you could kind
7:33of like go twice as fast but then kind
7:35of pump the brakes as you near your exit
7:37on the highway or in this case near the
7:38J section of the book exactly and maybe
7:41alternatively if I get to like a b c d e
7:43f g h i j k if I get to the K section
7:46then I could just double back like one
7:48page just to make sure John didn't get
7:50sandwiched between those pages so the
7:51nice thing about that second algorithm
7:53is that I'm flying through the phone
7:54book like two pages at a time so 2 4 6 8
7:5710 12 it's not perfect it's not
7:59necessarily correct but it is if I just
8:00take like one extra step so I think it's
8:02fixable but what your phone is probably
8:04doing and frankly what I and like my
8:06parents and grandparents used to do back
8:08in the day is we'd probably go roughly
8:09to the middle of the phone book here and
8:11just intuitively if this is an
8:13alphabetized phone book in English what
8:15section am I probably going to find
8:16myself in roughly K okay so I'm in the K
8:19section is John going to be to the left
8:21or to the right to the left yeah so John
8:23is going to be to the left to the right
8:24and what we can do here though your
8:25phone does something smarter is tear the
8:27problem in half throw half of the
8:29problem away being left with just 500
8:32pages now but what might I next do I
8:34could sort of naively just start at the
8:36beginning again but we've learned to do
8:37better I can go roughly to the middle
8:39here do it again yeah exactly so now
8:41maybe I'm in the E section which is a
8:43little to the left so John is clearly
8:46going to be to the right so I can again
8:48tear the problem portly in half throw
8:51this half of the problem away and I
8:53claim now that if we started with a
8:55th000 Pages now we've gone to 500 250
8:57now we're really moving quick
9:07not on that page and I can call him
9:09roughly how many steps might this third
9:11algorithm take if I started with a th000
9:13Pages then went to 500 250 125 like how
9:17many times can you divide 1,000 and half
9:20maybe 10 that's roughly 10 because in
9:23the first algorithm looking again for
9:24someone like Zoe in the worst case might
9:26have to go all the way through thousand
9:28pages but the second algorithm you said
9:29was 500 maybe 500 in1 essentially the
9:32same thing so twice as fast but this
9:34third and final algorithm is sort of
9:36fundamentally faster because you're
9:38you're sort of dividing and conquering
9:40it in half and half and half not just
9:42taking one or two bites out of it at of
9:44a time so this of course is not how we
9:46used to use phone books back in the day
9:47since otherwise they'd be single use
9:49only but it is how your phone is
9:51actually searching for zoy for John for
9:53anyone else but it's doing it in
9:55software oh that's cool so here we've
9:57happened to focus on searching
9:58algorithms looking for John in the phone
10:00book but the technique we just used can
10:02indeed be called divide and conquer
10:03where you take a big problem and you
10:05divide and conquer that is you try to
10:07chop it up into smaller smaller smaller
10:09pieces a more sophisticated type of
10:11algorithm at least depending on how you
10:12implement it something known as a
10:14recursive algorithm recursive algorithm
10:16is essentially an algorithm that uses
10:19itself to solve the exact same problem
10:21again and again but chops it smaller and
10:24smaller and smaller
10:26[Music]
10:28ultimately hi my name is Patricia
10:30Patricia nice to meet you where are you
10:31a student at I'm starting my senior year
10:33now at NYU oh nice and what have you
10:35been studying the past few years I study
10:37computer science and data science if you
10:38were chatting with a non-cs nonata
10:40science friend of yours like how would
10:41you explain to them what an algorithm is
10:44some kind of like systematic way of like
10:46solving a problem or like a set of like
10:49steps to kind of solve a certain like
10:51problem you have so you probably recall
10:53learning topics like binary search
10:55versus linear search and the like so
10:57I've come here uh complete with a actual
10:59chalkboard with some magnetic numbers on
11:01it here like how would you tell a friend
11:03to sort these I think one of the first
11:06things we learned was something called
11:07bubble sort it was kind of like focusing
11:10on like smaller like bubbles I guess I
11:12would say it like of the problem like
11:14looking at like smaller segments rather
11:16than like the whole thing at once what
11:17is I think very true about what you're
11:19hinting at is that bubble sort really
11:21focuses on like local small problems
11:25rather than taking a step back trying to
11:26fix the whole thing let's just fix the
11:28obvious problem in front of us so for
11:29instance when we're trying to get from
11:31smallest to largest and the first two
11:32things we see are eight followed by one
11:35this looks like a problem cuz it's out
11:36of order so what would be the simplest
11:38fix the least amount of work we can do
11:40to at least fix one problem just like
11:41switch those two numbers cuz one is
11:43obviously smaller than eight perfect so
11:45we just swap those two then you would
11:47switch those again yeah so that further
11:50improves the situation and you can kind
11:51of see it that the one and the two are
11:53now in place how about 8 and six switch
11:55it again switch those again 8 and three
11:57switch it again
12:00and conversely now the one and the two
12:02are closer to and coincidentally are
12:04exactly where we want them to be so are
12:07we done no okay so obviously not but
12:10what could we do now to further improve
12:14the situation go through it again but
12:16like you don't need to check the last
12:18one anymore because we know like that
12:20number is bubbled up to the top Yeah
12:22because it has indeed bubbled all the
12:23way to the top so one and two yeah keep
12:26it as is okay two and six keep it as is
12:28okay six and three then you switch it
12:30okay we switch or swap those six and
12:31four swap it again okay so four and uh
12:34six and seven uh keep it okay seven and
12:36five swap it okay and then I think per
12:39your point we're pretty darn close let's
12:41go through once more one and two Keep It
12:452 three keep it 3 four keep it 4 six
12:48keep it 6 five and then switch it all
12:49right we'll switch this and now to your
12:51point we don't need to bother with the
12:52ones that already bubbled their way up
12:54now we're 100% sure it's sorted yeah and
12:57certainly the search engines of the
12:58world Google and Bing and so forth they
13:00probably don't keep web pages in sorted
13:02order because that would be a crazy long
13:04list when you're just trying to search
13:05the data but there's probably some
13:07algorithm underlying what they do and
13:08they probably similarly just like we do
13:11a bit of work upfront to get things
13:13organized even if it's not strictly
13:15sorted in the same way so that people
13:16like you and me and others can find that
13:19same information so how about social
13:21media can you Invision where the
13:23algorithms are in that world like maybe
13:25for example like Tik Tok like the for
13:27you page it's kind of like
13:29cuz those are like Rec like
13:30recommendations right it's like sort of
13:32like Netflix recommendations except more
13:34constant because it's just like every
13:36video you scroll it's like that's a new
13:38recommendation basically and it's like
13:39based on like what you've liked
13:40previously what you've like saved
13:42previously what you search up so I would
13:44assume there's some kind of algorithm
13:45there kind of figuring out like what to
13:47put on your foru page absolutely just
13:49trying to keep you presumably more
13:50engaged so the better the algorithm is
13:52the better your engagement is maybe the
13:54more money the company then makes on the
13:56platform and so forth so it all sort of
13:58feeds together together but what you're
13:59describing really is more artificially
14:01intelligent if I may because presumably
14:04there's not someone at Tik Tok or any of
14:05these social media companies saying if
14:07Patricia likes this post then show her
14:10this post if she likes this post then
14:12show her this other post because the
14:13code would sort of grow infinitely long
14:15and there's just way too much content
14:17for a programmer to be having those
14:19kinds of conditionals those those
14:21decisions being made behind the scenes
14:24so it's probably a little more
14:25artificially intelligent and in that
14:27sense You have topics like neuron
14:28networks and uh machine learning which
14:30really describe taking as input things
14:33like what you watch what you click on
14:34what your friends watch what they click
14:35on and sort of trying to infer from that
14:38instead what should we show Patricia or
14:40her friends next okay yeah yeah that
14:42makes like the distinction more makes
14:44more sense now
14:48yeah I am currently a fourth year PhD
14:51student at NYU I do robot learning so
14:54that's half and half Robotics and
14:55Mission learning sounds like you've
14:57dabbled with quite a few algorith so how
14:59does one actually research algorithms or
15:01invent algorithms the most important was
15:03just trying to think about
15:04inefficiencies and also think about
15:06Connecting Threads the way I think about
15:08it is that algorithm for me is not just
15:11about the way of doing something but
15:12it's about doing something efficiently
15:14learning algorithms are practically
15:16everywhere now CU Google I would say for
15:18example is learning every day about like
15:21oh what what articles with links might
15:23be better than others and reranking them
15:26um there are recommender systems all
15:28around us
15:29right like content feeds and social
15:31media or you know like YouTube or
15:33Netflix what we see is in a large part
15:36determined by this kind of learning
15:38algorithms nowadays there's a lot of
15:40concerns around some applications of
15:42machine learning and like deep fakes
15:44where it can kind of learn how I talk
15:46and learn how you talk and even how we
15:48look and generate videos of us we're
15:50doing this for real but you could
15:51imagine a computer synthesizing this
15:53conversation eventually but how does it
15:55even know what I sound like and what I
15:57look like and how to replicate that all
15:59of this learning algorithms that we talk
16:01about right uh a lot like what goes in
16:04there is just lots and lots of data so
16:06data goes in something else comes out
16:08what comes out is whatever objective
16:10function that you optimize for like
16:11where is the line between algorithms
16:13that like play games with and without AI
16:17I think when I started off my undergrad
16:19the current AI machine learning was not
16:23very much synonymous okay and even in my
16:25undergraduate in the AI class they
16:27learned a lot of classical gorithms for
16:29game plays like for example the AAR
16:31search right that's a very simple
16:33example of how you can play a game
16:35without having anything learned this is
16:38very much oh you are at a game State you
16:40just search down see what are the
16:42possibilities and then you pick the best
16:45possibility that it can see versus what
16:47you think about when you think about I
16:49gameplay like the alpha zero for example
16:53or Alpha star or there are a lot of you
16:55know like fancy new machine learning
16:57agents that are you know even like
16:59learning very difficult games like go
17:01and those are learned agents as in they
17:04are getting better as they play more and
17:06more games and as they get more games
17:09they kind of refine their strategy based
17:11on the data that they seen and once
17:13again this high level abstraction is
17:15still the same you see a lot of data and
17:17you learn from that right but the
17:19question is what is objective function
17:21that you're optimizing for is it winning
17:23this game is it forcing a tie or is it
17:25you know like opening a door in a
17:27kitchen so if the world is very much
17:28focused on supervised unsupervised
17:31reinforcement learning now like what
17:32comes next 5 10 years where's the world
17:34going I think that this is just U going
17:37to be more and more I don't want to use
17:40the word encroachment but that's what it
17:42feels like of algorithms into our
17:43everyday life like even when I was
17:45taking the train here right the trains
17:47are being routed with algorithms but
17:48this has existed for you know like 50
17:51years probably but as I was coming here
17:53as I was checking my phone those are
17:55different algorithms and you know
17:57they're they're kind of getting all
17:59around us getting they're with us all
18:01the time they're making our life better
18:03most places most cases and I think
18:06that's just going to be continuation of
18:07all of those and it feels like they're
18:08even in places you wouldn't expect and
18:10there's just so much data about you and
18:12me and everyone else online and this
18:13data is being mind and analyzed and
18:15influencing things we see and here it
18:17would seem so there is sort of a
18:19Counterpoint which might be good for the
18:20marketers but not necessarily good for
18:22you and me as individuals you know like
18:24we're human beings but for someone we
18:26might be just a pair of eyes who are you
18:29know carrying a wallet and are there to
18:31buy things but there is so much more
18:33potential for this algorithms to just
18:35make our life better without you know
18:38like changing much about our
18:42life I'm Chris Wiggins from associate
18:44professor of Applied Mathematics at
18:45Columbia I'm also the chief data
18:47scientist of the New York Times the data
18:48science team at the New York Times
18:50develops and deploys machine learning
18:51for Newsroom and business problems but I
18:53would say the things that we do mostly
18:55you don't see but it might be things
18:56like personalization algorithm are
18:58recommending different content and do
19:00data scientists which is rather distinct
19:02from the phrase computer scientist do
19:04data scientists still think in terms of
19:06algorithms as driving a lot of it oh
19:08absolutely yeah in fact so in data
19:10science and Academia often the role of
19:12the algorithm is the optimization
19:14algorithm that helps you find the best
19:16model or the best description of a data
19:17set okay in data science and industry
19:20the goal often it's centered around an
19:22algorithm which becomes a data product
19:24right so a data scientist in Industry
19:26might be developing and deploying the
19:28algorithm which means not only
19:30understanding the algorithm and its
19:31statistical performance but also all of
19:33the software engineering around systems
19:35integration making sure that that
19:37algorithm receives input that's reliable
19:40and has output that's useful as well as
19:42I would say the organizational
19:43integration which is how does a
19:45community of people like the set of
19:46people working at the New York Times
19:48integrate that algorithm into their
19:49process interesting and I feel like AI
19:51based startups are all their R and
19:53certainly within Academia are there
19:54connections between Ai and the world of
19:56data science absolutely the algorithms
19:58that there in can you connect those dots
19:59for you're right that AI as a field has
20:01really exploded I would say particularly
20:03many people experienced a chatbot that
20:05was really really good today when people
20:06say I AI they're often thinking about
20:09large language models or they're
20:11thinking about generative AI or they
20:12might be thinking about a chatbot one
20:14thing to keep in mind is a chatbot is a
20:16special case of generative AI which is a
20:18special case of using large language
20:20models which is a special case of using
20:22machine learning generally which is what
20:24most people mean by AI you may have
20:26moments that are um what John kthy
20:28called look M No Hands results where you
20:30do some fantastic trick and you're not
20:32quite sure how it worked I think it's
20:33still very much early days large
20:35language models is still in the point of
20:37what might be called alchemy that people
20:39are building large language models
20:40without a real clear a priori sense of
20:42what the right design is for a right
20:44problem many people are trying different
20:46things out often in large companies
20:47where they can afford to have many
20:48people trying things out seeing what
20:50works publishing that instantiating it
20:52as a product and that itself is part of
20:54the scientific process I would think too
20:56yeah very much well science and
20:57engineering because often you're
20:59building a thing and the thing does
21:01something amazing to large extent we are
21:03still looking for basic theoretical
21:05results around why deep neural networks
21:08generally work why are they able to
21:10learn so well they're huge billions of
21:12parameter models and it's difficult for
21:14us to interpret how they are able to do
21:16what they do and is this a good thing do
21:18you think or an inevitable thing that we
21:20the programmers we the computer
21:21scientist the data science who are
21:23inventing these things can't actually
21:25explain how they work because I feel
21:27like friends of mine industry even when
21:29it's something simple and relatively
21:30familiar like autocomplete they can't
21:32actually tell me like why that name is
21:34appearing at the top of the list whereas
21:36years ago when these algorithms were
21:38more deterministic and more procedural
21:40you could even point to the line that
21:42made that name bubble up to the top so
21:44is this a good thing a bad thing that
21:45we're sort of losing control perhaps in
21:47some sense of the algorithm it has risks
21:49I don't know that I would say that it's
21:50good or bad but I would say there's lots
21:52of scientific precedent there are times
21:53when an algorithm works really well and
21:55we have finite understanding of why it
21:57works or a model works really well and
22:00sometimes we have very little
22:01understanding of why it works the way it
22:02does in classes I teach certainly spend
22:04a lot of time on fundamentals algorithms
22:06that have been taught in classes for
22:08decades now whether it's binary search
22:09linear search bubble swort selection
22:11sort or the like but are if we're
22:14already at the point where I can pull up
22:15chat GPT copy paste a whole bunch of
22:17numbers or words and say sort these for
22:20me does it really matter how chat GPT is
22:23sorting it does it really matter to me
22:25as the user how the software is sorting
22:27it like do the fundamentals become more
22:29dated and less important do you think
22:31now you're talking about the ways in
22:32which code and computation is a special
22:35case of Technology right so for driving
22:38a car you may not necessarily need to
22:39know much about organic chemistry even
22:41if if the organic chemistry is how the
22:44car works right so you can drive the car
22:46and use it in different ways without
22:48understanding much about the
22:49fundamentals so similarly with
22:50computation we're at a point where the
22:52computation is so high level right as
22:54you you know you can import pyit learn
22:56and you can go from zero to machine
22:57learning and 30 seconds it's depending
22:58on what level you want to understand the
23:00technology where in the stack so to
23:02speak um it's possible to understand it
23:05and make wonderful things and Advance
23:06the world without understanding it at
23:08the particular level of somebody who
23:10actually might have originally designed
23:11the actual optimization algorithm I
23:13should say though for many of the
23:14optimization algorithms there are cases
23:16where an algorithm works really well and
23:18we publish a paper and there's a proof
23:20in the paper and then years later people
23:22realize actually that prove was wrong
23:23and we're really still not sure why that
23:24optimization works but it works really
23:26well or it inspires people to make new
23:28optimization algorithms so I I do think
23:32that the the goal of understanding
23:34algorithms is Loosely coupled to our
23:36progress in advancing grade algorithms
23:38but they don't always necessarily have
23:39to require each other and for those
23:41students especially or even adults who
23:43are thinking of now steering into
23:45computer science into programming who
23:47were really jazzed about heading in that
23:49direction up until for instance November
23:50of 2022 when all of a sudden for many
23:53people it looked like the world was now
23:54changing and now maybe this isn't such a
23:57promising path this isn't such a
23:58lucrative path anymore are llms are
24:01tools like chat GPT reason not to
24:03perhaps steer into the field large
24:05language models are a particular
24:06architecture for predicting let's say
24:08the next word or a set of tokens more
24:10generally the algorithm comes in when
24:12you think about how is that llm to be
24:15trained or also how to be fine-tuned so
24:19the P of GPT is a pre-trained algorithm
24:22the idea is that you train a large
24:24language model on some Corpus of text
24:27could be encyclopedias or textbooks or
24:30what have you and then you might want to
24:32fine-tune that model around some
24:35particular task or some particular
24:37subset of texts so both of those are
24:39examples of training algorithms so I
24:42would say people's perception of
24:43artificial intelligence has really
24:45changed a lot in the last 6 months
24:47particularly around November of 2022
24:51when people experienced a really good
24:52chatbot the technology though had been
24:54around already before academics had
24:56already been working with chat gpt3
24:58before that and GPT 2 and gpt1 and and
25:01for many people it sort of opened up
25:03this conversation about what is
25:04artificial intelligence and what could
25:05we do with this and what are the
25:07possible good and bad right like any
25:08other piece of technology cransberg
25:10first law of technology technology is
25:12neither good nor bad nor is it neutral
25:14every time we have some new technology
25:15we should think about its capabilities
25:17and the good and the possible bad as
25:20with any area of study algorithms offer
25:22a spectrum from the most basic to the
25:24most advanced and even if right now the
25:26most advanced of the those algorithms
25:28feels Out Of Reach because you just
25:29don't have that background with each
25:31lesson you learn with each algorithm you
25:33study that endgame becomes closer and
25:36closer such that it will before long be
25:38accessible to you and you will be at the
25:40end of that most advanced
25:42[Music]
25:45Spectrum

About this transcript

This page contains the full transcript of Harvard Professor Explains Algorithms in 5 Levels of Difficulty | WIRED by WIRED, generated from the public captions YouTube serves with the video. The transcript has 5,390 words across 757 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to translate the transcript into 100+ languages.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, translation to 100+ languages, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.