AI YouTube Comments - Computerphile

Generating KZbin comments with a neural network trained on KZbin comments. What could possibly go wrong? Dr Mike Pound replied to our comment...
EXTRA BITS: • EXTRA BITS: More AI Ge...
Neural Networks & Deep Learning: • Neural Networks
Andrej Karpathy Blog post mentioned by Mike: bit.ly/C_Blog_RNN
Code that Mike used to create this: bit.ly/C_RNN_Code
/ computerphile
/ computer_phile
This video was filmed and edited by Sean Riley.
Computer Science at the University of Nottingham: bit.ly/nottscomputer
Computerphile is a sister project to Brady Haran's Numberphile. More at www.bradyharan.com

[@IceMetalPunk]

12:11 I'm in the video! :O This is amazing! The robots love me!

[@IceMetalPunk]

Also... this might be a sign that I need more to do with my life besides comment on KZbin videos...

[@Computerphile]

+IceMetalPunk hey, thanks for the comments! >Sean

[@user-tp2ly5ct3e]

Hi IceMetalPunk. ☺️

[@Klaevin]

Well, ask him for the code. That way, you'll never have to comment again!

[@danieledg94]

The code is in the description.

[@IceMetalPunk]

Anyway, excitement about my five minutes of robot fame aside, I did want to comment something related to the actual subject of this video :) My favorite method of generating plausible text is a simple multilayer Markov chain. Way back in my freshman year at uni, one of my assignments was to create a three-layer Markov chain trained on excerpts from the Wizard of Oz to generate a new page from the book. It was interesting. But then, we were allowed to train our Markov chain's dictionary on *any corpus* we wanted, so I chose the US Constitution. Needless to say, seeing brand-new laws come into existence at the (metaphorical) hands of a computer was extremely amusing.

[@cae289]

I love you, baby.

[@jennagentles1836]

Nice try computer. Your phony back story isn't fooling anyone. Well done on the improvements though. Your comments are getting better.

[@mickelodiansurname9578]

IceMetalPunk I am a knight and what I say to all this is Nee! Just nee, nee, nee!

[@DaVince21]

+IceMetalPunk A little late, but if you have any excerpts, please post them?

[@Flexy59]

After this video i will never be able to trust computerphile comments ever again.

[@watcherfox9698]

"I was able to want to be able to be happy."

[@veggiet2009]

me too... me toooo...

[@TiagoTiagoT]

Same

[@ThorkilKowalski]

That's an important first step.

[@daddyleon]

It has reached consciousness!

[@NL-tq1yr]

watcherFox they became self aware! 😱

[@SebSenseGreen]

Developing an AI based on KZbin's comment section might not be the brightest idea in the world. The "I" in AI stands for intelligence.

[@stensoft]

Maybe he was developing Artificial Idiocy

[@spacedoubt15]

I always preferred Artificial Incompetence.

[@dummyvariable3079]

Artificial ignorance.

[@wESTmanRO]

Artificial imitation

[@azmanabdula]

Space doubt wins....

[@xisumavoid]

IceMetalPunk comments on my videos too :-D Deffo real person!

[@noone-fe2pi]

xisumavoid xisuma here wow

[@brodern22]

Cool to see you here

[@SuperSand2000]

Hey Xisuma. Love your vids. good channel you watch here

[@CheekyBuilderGaming]

didnt expect to see you here :D

[@6infinity8]

xisumavoid Oh, hey xisuma

[@Computerphile]

First KZbin Comment

[@MAtukulis]

Mike is the best speaker IMO

[@QuoteVG]

Computerphile reply

[@Computerphile]

1:25 Don't right?

[@nitilpoddar]

do you support online learning from websites like "Free code amp" and 'Solo learn" 'Code academy". gosh! I am soo noob

[@Ezullof]

Great video!

[@Dsiluigi]

Train it on computerphile transcripts, then act out the output!

[@LimeGreenTeknii]

Yes! I second this entirely!!

[@LupusSolitarius590]

"This video was written by an AI."

[@pwhqngl0evzeg7z37]

even better, train it on classic literature, read its essays at a TED Talk

[@t400ml]

I find your own difference.

[@jeremyelliot4831]

This relayed having two first.

[@jeffirwin7862]

We all find your own difference on this blessed day.

[@Tesana]

Hmm, profound words; they art spilling from thine mouth.

[@jessehammer123]

I just wanted to let you know that I was just wondering if you were able to get the kids to school.

[@Biped]

Does it write "first"-comments?

[@Jan-vv1zk]

Ebumbaya ' first

[@ragnkja]

Probably.

[@Treddian]

Yes, but excessively so and at inappropriate times.

[@Djorgal]

@@Treddian So... like normal youtube comments then.

[@domninin]

In one of the screens you can see a comment which says "One!", which is kinda the same

[@sharkinahat]

Using comments as input on university computers? Would be a shame if someone would '); DROP TABLE Students;--

[@DKRCecer]

Hey! Leave Bobby out of this! :D

[@xjourner]

I wonder what lil Bobby has gotten himself into this time

[@ulilulable]

"I was able to want to be able to be happy" The network is trying to tell us something.

[@NVE]

I just discovered Numberphile and Computerphile recently and appreciate everything you guys are doing here. The content is definitely top notch. I'm lovin' it.

[@jamesgrimwood1285]

The output is nonsense, but it looks quite similar to badly translated Chinglish found in the manuals for cheap eBay electronics :) I am impressed with the way this isn't just randomly sticking words together, it's actually making the words themselves, letter by letter - without even really knowing what a "word" even is.

[@kanskejonasidag1]

I dunno why, but I really love this guy. He's very good at explaining things and he's funny. Thanks Mike!

[@poop121]

Agreed!

[@SamHizam]

I think it's his passion for explaining and kinda learning at the same time.

[@veggiet2009]

It would be interesting to create a neural network that would predict comment likes/dislikes based on the content of the comment.

[@entiretwix1480]

@@dejfcold that's an over simplification

[@abstractapproach634]

I predict your comment will receive 76 likes before it is forgotten.

[@samuelthecamel]

@@abstractapproach634 I don't think it's gonna make it :(

[@gpt-jcommentbot4759]

@@samuelthecamel Prob 42 likes

[@sharifsircar]

@IceMetalPunk Please drop a hello

[@IceMetalPunk]

Hello! :D

[@tomewyrmdraconus837]

You're internet famous now!

[@IceMetalPunk]

At least within the Computerphile-watching community, and for the next week or two. But I'll take that! :D

[@cobyambrose2915]

IceMetalPunk yey!

[@zss123456789]

That's amazing! how many comments have you posted lol!

[@minodragon]

"I find your own difference." It sounds so deep and profound! Now all we need to do in translate it into Latin, "Differentiae tuae invenio." That's one solid tattoo right there! Or it should be Chinese, if someone would like the chime in on that one?! :D

[@maxpeeters8688]

Hey Computerphile, I love your comment at 13:43, where you asked the question "How do you watch if you basically have one hardware?". Great video, quite fun to read through those generated comments to find some that almost make sense.

[@artemkovera5785]

Hi Max. If you're interested in AI, you can check out my introductory book "How to Create Machine Superintelligence" available for FRЕЕ in amazon kindle store till 6th October. In this book, I go over the following: - intelligence as a form of information processing - basics of classical computing - basics of quantum computing - some basics of machine learning and artificial neural networks - and also share some thoughts on building general AI and dealing with the control problem

[@Arthur-mj2vd]

Amazing stuff! And ironically, the fact that it outputs typos every now and then makes the comments much more realistic.

[@5cyndi]

I know, right? I found that to be simultaneously cool, and a bit thought-provoking. 💜

[@xplinux22]

Extremely cool demonstration! I find neural networks so fascinating.

[@kandy1249]

please keep on making videos. it really really really helps.

[@5cyndi]

11:47-12:01 makes me stop and think for a second about that for a little bit. Very interesting conclusion, as it's true, it is trained on data with typos, and it "learned" to make typos! That is both amazing and a little bit deep. Love this! 💜

[@natefuller1325]

He is so good at explaining this stuff

[@ZintomV1]

Mike's the best at Computerphile videos

[@gqh007]

IceMetalPunk is a KZbinr gamer who makes Minecraft let’s plays. This dude exists!

[@ragnkja]

I only know them as a KZbin viewer whose viewing habits often overlap with mine.

[@IceMetalPunk]

Well, I used to... I haven't had much time lately :( Having two jobs can really hinder a social (media) life XD

[@simbaonsteroids8836]

Isn't trying to generate an AI trained on KZbin comments a bit counterproductive?

[@daanwilmer]

Depending on the channel, it might indeed become an Artificial Stupidity. However, as this AI is trained on Computerphile comments it's probably above average.

[@simbaonsteroids8836]

Daan Wilmer tru nuff, it was a joke tho.

[@RitobanRoyChowdhury]

No, because even it the bot replicates natural stupidity, you've proved that you have an algorithm than can replicate human behaviour accurately.

[@DumpstaBoots]

first!'); DROP TABLE comments;--

[@IceMetalPunk]

Hello, little Bobby Tables :D

[@Snaperkid]

IceMetalPunk That should teach you to sanitize your inputs.

[@Bmac2112]

Clever, except that the machine learning server is probably quite remote from the administrative servers

[@oldbootz]

Yea because a Doctor of computer science would use SQL HAHAHAHA... Not.

[@maanvis81]

Cool :) I also did this a while ago when I learned about Andrew Karpathy's blog, to generate song lyrics with using a certain music style as input.

[@artemkovera5785]

Thanks, great vid. Also, would be great to hear from you about spiking neural networks. There isn't much free and quality information about spiking nets.

[@aybmnn]

how many times does 'Hitler' pops up in those generated comments

[@Biped]

Too often #Godwin’s law

[@FunIsGoingOn]

Watching this 5 years later now that we have ChatGPT and such...

[@gpt-jcommentbot4759]

and where he said that a Chatbot from this is "theoretically" possible and here we are

[@brianmcmullen95]

Great video!

[@LeonidasGrecos]

One more comment for your neural network :) I admire your work and your ambition.

[@drearyplane8259]

This was done by a KZbinr called CaryKH, he applies neural networks to various tasks.

[@Nerothe42]

Yea i love the stuff he does - like the language-recognision :)

[@charlesdorval394]

I was about to mention. I wonder what difference there is between their methods, if any

[@drearyplane8259]

Charles Thatisall There might be, if there is I'm voting for an epic NN battle.

[@BunnyOfThunder]

I imagine it would be possible to layer this with some sort of pre-processing? For example, I imagine it would be possible to parse the sentences into word objects pretty easily first, and then run the same kind of a network against words rather than letters?

[@cacheman]

I realize this is more powerful in a general sense, but to just generate text (what I know of as) a markov chain seems... easier. They're fun and extremely easy to get started with if nothing else, and I feel like playing with the generator order gives a sense of how (non)random language is.

[@superdau]

A recurrent neural network looks like a "recursive" markov chain to me.

[@IceMetalPunk]

Yeah, definitely. This isn't the best use of a recurrent neural network, but it does sort of help for demonstration purposes. But I love me some Markov chains! I always wanted to use Markov chains to generate songs based on a given artist's corpus of song lyrics, but the one time I tried to make one, I did it very crudely and naively and ended up accidentally DOS'ing a lyrics site... so I stopped XD I should get back to that one day, but do it more intelligently this time... if I ever have any free time for that anymore.

[@gpt-jcommentbot4759]

@@superdau A RNN is kinda like a Markov chain but using thousands of different states instead of one. Also each state is modulated by math/weighted connections instead of probability. It's much harder to generate with quality of RNN if the Markov Chain is working character by character.

[@v1sudo]

Nice to see you using Lua for this! Just recently began learning the language

[@TheiLame]

12:12 So did IceMetalPunk ever appear? Did the happy reunion ever happen? I so want to know!

[@IceMetalPunk]

I did appear! :D

[@Nerdnumberone]

Has there been much work on giving such systems initial information from which to build off of? For example, one might give the system a dictionary of English words, acronyms, etc (possibly letting the system expand it to some extent) with a list of the potential (or probable) types of word (nouns, verbs, adverbs, adjectives, etc.) and even, perhaps verb tenses. Besides going letter-to-letter and word-to-word, it could start building models of overall sentence structures. This would significantly increase the complexity of the system and prescribing rules, rather than letting the system learn rules, might limit it in certain ways and would require more work initially.

[@13lacle]

what would happen if you tried using this as a generator for a generative adversarial network and then make the classifier as real vs generated? is it too random to get to realistic comments?

[@himselfe]

It would be interesting to see what it comes up with if you train it to predict comments based on the video title (and perhaps description).

[@bobsters098]

could you use say a large neural network that was trained on high res numbers and see what is the "purest" image of each number by kind of going backwards?

[@olbluelips]

Is this a reupload? I feel like I remember this, down to the IceMetalPunk guy, but it could just be déjà vu

[@IceMetalPunk]

You probably just have seen my comments before... I comment way too much... hence why the AI noticed me XD

[@Dusk-MTG]

10:23 Neural network literally saying: "I don't think before doing" Mike Pound, you created a NN that can lie.

[@andrewjknott]

Great video. Phone and so, crunching (with the jump).

[@normannormiemates4844]

IceMetalPunk is a chat bot helping to train a chatbot to become a chatbot.

[@artemkovera5500]

:) Looks it is. If you're interested in AI, you can check out my introductory book "How to Create Machine Superintelligence" available for FRЕЕ in amazon kindle store till 6th October. In this book, I go over the following: - intelligence as a form of information processing - basics of classical computing - basics of quantum computing - some basics of machine learning and artificial neural networks - and also share some thoughts on building general AI and dealing with the control problem

[@normannormiemates4844]

I won't read it because if you're rationalising what intelligence is you're misrepresenting what intelligence is

[@johnfrancisdoe1563]

What technique is used by the somewhat similar Dissociated Press program, and how does it compare?

[@woopwoop769]

Can you post the code that generate the comment output? Would love to see it and potentially run it on my own data!

[@Sirlegna]

I am currently working on something very similar with reading books and writing small sub stories. I am happy I am not the only one getting the number loop problem

[@sygarth]

"I find your own difference" is amazingly deep!

[@eratzleretour1027]

So, the challlenge here is to give the NN a memory right? Is what you're doing approximately the same as giving a state the previous states as input ? I guess it's not totally equivalent since you said the hidden layers' weighs are shared. If it is different, why not do as I said ? Too many inputs ? Moreover, is the "memory" length fixed ? Like 30 characters for example ? I'm actually interested in this since I'm trying to figoure out how to design a NN for training on a racing game, and sometimes, I need to know that I've already been through some path, to not go there again, so I need memory. Thx for your answer!!!

[@gpt-jcommentbot4759]

LSTM is different from other neural nets because it incorporates timesteps. Every timestep, data is inserted into the model input. The one thousand neurons in Computerphile's LSTM decides the output based on the inputs and previous output. This allows the LSTM to remember, and unlike a Markov model, the memory capacity of an LSTM is theoretically infinite. Unfortunately in practice the memory is not actually infinite because Back-Prop through time (optimization algorithm) results in slowly vanishing gradients which limit it's memory to a few hundred characters.

[@Slada1]

Carykh

[@xanokothe]

I guess the network would output a lot of "first" as comments and a lot of "first" related replies

[@NicholasWong0112]

+ComputerPhile Are Markov chains related to the ideas in this video?

[@SomethingUnreal]

Since RNNs are ideal for predicting the next thing in a long sequence, it's really fun to train them on audio (predicting the next position of the waveform based on the audio leading up to now). It's much slower than dealing with text, though. In case anyone's curious, I've made a couple of videos showing results of that (using torch-rnn, the same software Mike uses here).

[@cptechno]

Please provide full references for scientific papers when used on the show. People like me might want to read it for ourselves.

[@OrangeC7]

What if there was another input into the network for some kind of "concept" the network is having, and you train it with sentences tagged with their concepts? Then, when generating sentences, you could theoretically generate them from concepts rather than thin air.

[@gpt-jcommentbot4759]

I dont think thats needed, the hidden units in the LSTM are responsible for remembering and generating the "concepts" and coordinating text outputs.

[@vincenzospaghetti]

what about using iconic authors from the past to get reasonable conversations or responses for the present time or maybe for commentary towards the future?

[@freeelectron8261]

Mike is one super enthusiastic computer scientist! Go Mike!

[@ProMoonLight]

If you would always choose the most likely character, what would be the average comment on Computerphile videos?

[@martinoandreascarpolini5128]

I love you, Mike Pound

[@TheUndeadLeech]

Great Video!

[@Houdini111]

Could you provide the code you used? I'd like to mess around and take a look.

[@Qmppu]

So how is this different what Markov chain does? Because to me this video sounded lot like only difference between Markov chain and neural network would be how it is made under the hood.

[@IceMetalPunk]

Yeah, this demonstration would work better (or at least be easier) with a Markov chain. But recurrent neural networks allow machines to learn more than just sequences, which Markov chains can't do.

[@antoncid5044]

Where's IceMetalPunk? Someone let them know.

[@IceMetalPunk]

Well, aside from the many comments my videos are now getting to let me know about this, I'm also subscribed to Computerphile. Hence why I leave enough comments for the AI to notice me, senpai :D So I saw it! :D

[@MJay_]

+IceMetalPunk .. we are waiting!

[@IceMetalPunk]

Wait no longer! :D

[@APSStevens]

@Computerphile Suggestion: Does it makes sense to use service like "Grammarly" to auto-correct sentences before we feed them to make this idea more sensible?.

[@sujitmaharjan565]

Is it character level RNN? And how much time did you train the network?

[@PaulPaulPaulson]

We welcome our new AI overlord! 👑

[@enricoalessio5434]

I was No pretty puzzling the code gaming so they would have found out the livelutoo tred s+so the larger stop information. I don't think before doing "is when it can predict it as computing and A/P?" R it wasn't a high-fact a bit stamp idea

[@irazorx4768]

Those over the sholder shots have quite a large desync between audio and video

[@kingpopaul]

It would have been interesting to have a 3 way chat with this LSTM network, a Markov model trained on the same data and a GAN. Has anyone made a text generating GAN yet, it should be trivial but the focus is mostly on image generation these days.

[@gpt-jcommentbot4759]

In 2023 this should be a 4 way chat with progressively more complex chatbots: Preprogrammed response chatbot, Markov bot, LSTM, ChatGPT.

[@jony7779]

This video was spectacular. - A real person

[@shinoobie1549]

How far we have come

[@4.0.4]

Can this read and generate MIDI? I wonder which styles it fails/works on.

[@Benimation]

"I said yesterday I walked to the park 2 days ago."

[@DustinRodriguez1_0]

Hahaha, it generated my name in the random username part! (Yeah I've commented on multiple Computerphile videos in the past) I did not expect to see that.

[@alexitosworld]

Nice one. Looks like me using keyboard suggestions to generate test text when developing an app ^^

[@artemkovera5785]

Hello. If you're interested in AI, you can check out my introductory book "How to Create Machine Superintelligence" available for FRЕЕ in amazon kindle store till 6th October. In this book, I go over the following: - intelligence as a form of information processing - basics of classical computing - basics of quantum computing - some basics of machine learning and artificial neural networks - and also share some thoughts on building general AI and dealing with the control problem

[@IceMetalPunk]

There are several different ways that predictive text systems can be built, but it seems the most common way is actually a trie-based Markov chain system, not a neural network. The result is basically the same in the case of predicting text, though :)

[@zakeryclarke2482]

That ASCII art is awesome

[@Interpause]

Wow its better than mine the chatbot my team made only could successfully achieve contextual awareness. we used LSTMs (a type of node for RNNs)

[@code-dredd]

IceMetalPunk: You've been targeted for termination.

[@IceMetalPunk]

D: NO, I welcome my robot overlords! I'm honored they picked me and I will work with them however they see fit! Don't kill me!

[@StefanReich]

"I have no ideas so I use a neural network"

[@antopolskiy]

why are you so mean?

[@StefanReich]

Because there is better AI. I am not mean, just angry for not being heard.

[@MrUtak]

Why Lua and nor other languages? Make a video about that!

[@tommykarrick9130]

Kind of old video now, but let’s say you trained it separately on comments on videos about computer science history, then AI safety videos, etc Would it start to generate more plausible sentences? Or is there too much variety even in that specific set that it’d devolve to nonsense

[@3Nightguy]

This is the same thing that powers the subreddit simulator! that sub actually gives nice plausible posts and stories! ( sometimes )

[@frognik79]

So what is the most likely sentence taking the highest percentage letters every time? What do all our replies combined give?

[@Biped]

frognik79 "First!"

[@MichaelMoore99]

I'd like to see your opinion on RoboRosewater using one of these networks.

[@steef7142]

It would be cool if you train this network on the whole(or part of) the wikipedia archive.

[@gpt-jcommentbot4759]

GPT-2:

[@mrr0r508]

This is the greatest solution of life changing information and the best of /.

[@jpphoton]

I think going below word level is like SHA1: You can't get it back. In this case, the semantics of the language itself. you could have a word level neural net run in parallel. that'd be neat.

[@TotteringGuy]

10:24 "I was able to want to be able to be happy."

[@neilisbored2177]

"I find your own difference." -Hessil200, 2017

[@SleeveBlade]

surprised it didn't say "WE LOVE YOU MIKE" all the time

[@Gooberslot]

Could you analyze a single users comments and generate comments that that user might be likely to say?

[@valshaped]

Giths transistor a benefit of cracking something reciever, not match obviously available?

[@Arnaz87]

Immediatly came to see IceMetalPunk's comment.

[@IceMetalPunk]

I commented a few comments just now... actually, I'm sure my habit of commenting multiple times throughout the course of watching a video helped the AI notice me xD

[@jonathantanner6701]

Gratz IceMetalPunk You win the comments...! xD

[@IceMetalPunk]

Extremely belated thank you!

[@TheCaerbanog]

Automatic writing to a whole new level! André Breton would be proud!

[@ponysopher]

This guy needs his own channel.

[@Seegalgalguntijak]

Does this take into account which user posted which type of comment? If not, the comments posted by IceMetalPunk may just be some random thing. How could you even assume to analyze whether a comment is positive or negative, which is several meta-layers above that?