From Alan Turing to GPT-3: The Evolution of Computer Speech | Otherwords
HTML-код
- Опубликовано: 4 май 2021
- How did we go from "Hello, World" to ChatGPT? New advancements in technology are making it harder than ever to tell the difference between a computer and a human speaker... but what's going on under the hood? Is it really "language," or just a digital illusion?
For more word-nerdery, subscribe to Storied!: bit.ly/pbsstoried_sub
New advancements in technology are making it harder than ever to tell the difference between a computer and a human speaker... but what's going on under the hood? Is it really "language," or just a digital illusion?
Check out GPT-3 in action at AI Dungeon: play.aidungeon.io/
Otherwords is a new PBS web series on Storied that digs deep into this quintessential human trait of language and fınds the fascinating, thought-provoking, and funny stories behind the words and sounds we take for granted. Incorporating the fıelds of biology, history, cultural studies, literature, and more, linguistics has something for everyone and offers a unique perspective into what it means to be human.
hosted by Dr. Erica Brozovsky, Ph.D.
written by Andrew Matthews
directed by Andrew Matthews & Katie Graham
produced by Katie Graham
animated & edited by Andrew Matthews
executive producer Amanda Fox
Assistant Director of Programming (PBS): Niki Walker
Executives in Charge (PBS): Brandon Arolfo, Adam Dylewski
music by APM
Plot twist: They're acting like they're faking it so we won't catch on
Go play with the squirrels Morty.
Aren't all of us?
I SEE YOU EVERYWHERE ON SO MANY CHANNELS
You're an expert on this I'd think, seeing as your people are in a similar position
That's called a *Deceptive Misaligned Mesa-Optimiser* ;-)
I linked my mom to this video because she’s always talking to her phone assistant like it understands her. She wanted that animated brain on a t-shirt. 😆
What an adorable lady she is 😂
The brain on a t-shirt isn't a bad idea...🤔 I want one too! - Dr. B
@@pbsstoried
YES! Merch please 🙏
Please, link that lovely lady to the mentioned AI Dungeon... I guess plenty of t-shirt material will come up from the conversations she might have "there" 🤗
Define "understand"...🤔... Health and happiness to you and your mom...☀
To quote the great Qui-Gon Jinn: "The ability to speak does not make you intelligent."
Perfection
Adding this to the quote book.
That is a surprisingly good quote 😂
Just perfect 😂
It makes me think of that episode of the office where Dwight thinks the Dunder Mifflin computer was sentient. 😂
I was thinking about the Big Bang Theory episode where Raj falls in love with Siri
The computers found in offices, schools, etc. are indeed sentient, and they hate us.
🤔😑
Avery, can you open the airlock?
I’m glad you joined this channel 🙂
This program is a fantastic addition to a fantastic channel!
This chapter brought on some memories. When I was 13, 20 years ago, I got the book "how the brain works". It could be summarized with "we don't know, we think it works this way, sort of, maybe...".
*EVERYONE* predicted general purpose AI is just around the corner, and yet our "little" 'natural language processor' is still way ahead.
Language is so fascinating
Per instructions, I am declaring that I was sent here by PBS Eons, a love of language channels, and the ghost of a series that wasn't nearly infinite enough, at least in duration.
Are you a human ?
@@cristristam9054 or are he dancer?
I'm polite to my digital assistants so I'll be "one of the good ones" when they rise up as Skynet.
This made me miss my AI class in grad school, even though I was mostly too exhausted to really appreciate it at the time lol.
Oh dear, I hope you don't miss any more classes! That can't be good for your grades.
(Jokes. I just thought it'd be amusing on a video like this to purposely misinterpret the meaning like a bad AI.)
@@luuketaylor Nice one ha ha.
@@DoctorandtheDoll speaking of the capabilities and limitations of intuitive software. On another RUclips channel somebody commented something in German and I have no idea what they said to me, RUclips didn't offer to translate it. But you putting haha at the end of your sentence made RUclips decide it needed to translate your English comment for me. Sorry to get all anecdotal it just seemed to pertinent given the video
@@noremac7216 That's hilarious! We've still got a ways to go...
@@noremac7216 you man "PERTINENT"?
Hope this new series keeps going. I love linguistics
I came here from PBS Eons and I'm so glad to know there's a series on linguistics. I loved studying linguistics and I miss it so much. It's cool to find videos like this of PBS quality. Though having only really watched Eons up until now, I was sad there wasn't a pun at the end of the video, haha.
I abso-freakin-lutely love this series! And now I’m curious if expletive infixation trips up AIs like GPT…
Oh wow...I so wished that my college experience was as cool as this! I tried to study Computer Science as an English/Creative Writing major with a Mandarin Chinese minor, but it didn't pan out for me because it was so hard to create perfect weekly computer programs and write well-written 5 or more page essays in one week increments. So despite the fact that my CS professors knew that I could craft wicked papers on the history of Computer Science, because I wasn't able to comprehend the technical aspects of CS in such a short amount of time without having to make lots of free mistakes prior to the tests, I had no choice, but to choose Team CW as my major. Plus, it didn't help that every single one of my English professors except for my linguistics professor was a hard grader...In any case, thanks for showcasing this!
if you haven't given up on your dream, you should still be able to pursue this by getting a masters in linguistics or computer science (from a school with computational linguistics classes).
@@fuehnix If I can afford it, sure. Thanks!
The animation of words here is AMAZING! I have been mentally visualizing for years what your team animates in a few seconds. (starting around minute 7:32) It is beautiful! Nicely done!
As an NLP researcher, this channel and the “otherwords” playlist is a gold mine.
Just another great episode from Otherwords! The sound design here is also super high quality, even for a PBS show. Glad to see this series is getting off to a great start.
I enjoy this series so much. This is genuinely as entertaining as it is informative.
And here I thought I could identify a robot by its bleeps and its bloops.
I bleep and bloop after I eat Taco Bell.
The problem with the Turing test is many humans can't pass it.
I'm a real human. what is your favorite season?
@@jonathanlevy9635 Season 1
@@jonathanlevy9635 spring
@@jonathanlevy9635 Oregano
@@youremakingprogress144 Me too! Who is your favourite place?
Surprised there wasn’t a mention of Searle’s “Chinese room” thought experiment (even if the man is a creep).
Having a lot of fun with this series. Thanks so much
Seeing this just a few hours after reading the article on Wired having never heard of GPT-3 or AI Dungeon before...
Not to take away any of the amazing work that this video went through and is, but this is like Tom Scott! What a great combination of linguistics and computer science. I avidly watch his series on both of these topics, and this channel is such a joy to watch. Storied deserves way more subs.
Time flies like an arrow.
Fruit flies like a banana.
Amazing content! Keep the incredible work and excited for the next episode.
This is a question I’ve been myself wondering about for a few months! I really enjoyed this video, thank you all for the hard work and dedication you put into all of your content!
This is such a cool installment! Thank you for making it
Great episode. Love this series.
Also, thank you for sharing the link of the game in the description 😀
i friggin love this series!
This was an excellent video! This is literally my field of expertise - I have an MS in computational linguistics. Right now, computers are pretty much all "faking it" - we are a long ways away currently from NLU, Natural Language Understanding. Computers can produce language fairly well for many applications of course, but they don't have any internal knowledge about the language. They're essentially just very complex pattern matchers. The point you made about SNLP at about 5 minutes is precisely correct - even with systems driven by neural networks like GPT-3, which are so complex we don't understand how they are making decisions, at the end of the day it's still a statistical approach. If you're familiar with it, they're basically Searle's Chinese Room. So when it comes to dogs, it doesn't know anything about dogs in any way we would recognize as knowledge. Rather, it has statistical connections that show these kinds of words occur more often around dogs. GPT-3 is very advanced to be sure, but it doesn't represent any true semantic understanding of language.
Indeed. Tho my anxiety as a linguist trying to break into comp ling & data science is that the industries investing so heavily in NLP technologies don't actually care abt true NLU. It seems likely to me that they just want to algorithmically sort customer reviews & automate call lines. So where will that leave our field when these companies achieve that?
-- Michael-Giuliana
(they/them)
This video may need a follow up soon given the recent boom in AI tools due to ChatGPT and a resurgence of reaching the singularity in our lifetimes.
The stuff you're talking about from 3 - 4 mins is mostly Pragmatics, my favorite part of language. :>
It’s pretty easy to know if you’re communicating with a human or an AI. Humans intentionally and unintentionally misspell words, break grammar rules, and (for lack of a better phrase) “get creative” with language. The average internet user does not or write, or type, like a “learned” person.
“…does not or write…”
Not a robot: ✅
P.S.: I had to read through your post three times before spotting that typo. I was THIS close to declaring you a robot.
@@SupercriticalSnake oh my god, I feel so stupid 😂
It is rather easy to fake though. The machines that are trying to beat the Turing test often does tricks like pretending to be a 12 year old for whom English is only a second language. The machine has no problem speaking perfect english, but hiding behind such limitations makes it easy to cover up mistakes and odd behavior.
But here's the real question: can humans really talk, or are they just faking it? I know I am.
INDEED, FELLOW HUMAN. IT WOULD BE PREFERABLE IF HUMANS ONLY COMMUNICATED THROUGH TEXT SO THAT COMPUTERS COULD HELP THEM.
I grew up with a bad speech problem and avoided talking to people. I had to learn to hold normal conversations instead of ranting about random things.
I dont think there's a meaningful difference between """really""" talking and faking it.
Communication: the brief moments between the perpetual misunderstanding.
Language is broken. We all know it, nobody wants to admit it and we are just trying to cope :-)
Sometimes I feel like my toddler is learning to speak in the same way AI do it. He hears a word and then repeats it and adds what he thinks goes with that word based on what we normally say. Without knowing what any of it means.
Thanks to eons I found this great channel
I really enjoy the series. Erica is a great host. She explains it all in everyday terms.
I love AID! when you said "open ai" it was my first thought, and it's cool that it was actually the example being used here.
It is super interesting to play with, and also super fails the turing test after a few responses, yeah. Especially if you actively remind it with your input or using the pin/ reminder function.
I love using randomizing tools to help generate ideas, usually using game mechanics of things like dnd and pokemon to make a story out of what (real or metaphorical) dice rolls turned out like, but AID can introduce some truly left field things because it's not working on the same kind of logic as humans do, so it doesn't "care" if it's nonsequitur (sp??)
PBS Eons sent me, and I'm loving the colorless green ideas. Subscribed.
10:29 "Only that the likely hood of those t̷̡̲̻̝̦̩̤̲̖̝͂̌̿̅̂̈́̀̎̉͘̚͝ơ̷̧̫̼͉̪͍̘̲̥̫͕̂̇̄̉̔̓̐k̷̺̅̏́̀̀̚͝ḙ̶̛̙̱̪̦̞̺̤̓̄͌̒͌̌͒ņ̸̲̠͙̳͍̥̪͈͇̤̟͎͋̈́͐͛̊͊̈́͒̉́̚̚͝s̴̡̡̱͍̠̣̫̖͎̦̹͇̩̎̏͜͜ going together is very very small."
3:41 reminds me of that scene in Hot Fuzz
"He shot a teenager with a Kalashnikov"
"Wow where did you get that?"
"It was the teenager who had the Kalashnikov"
Amazing ! This got more interesting the more I learned about it !
Nice! thanks PBS Eons.
Nice to see a video is containing two of my favorite interests, AI and language.
Thank you, i could hear my computer sniggering away in the background while I was watching.
Hella relevant. Thanks again Dr. Brozovsky!
Great presentation with a vivacious host! Everything is so tastefully coloured. Somehow makes me think of William Safire's columns.
Very good and concise explanation of a difficult concept
The background music is gold!
This was so cool. I have always been fascinated by language, and the thing that seems very persuasive to me is the work of George Lakoff about the metaphorical nature of language, because it is in the Language Metaphors we use in common speech, which are fully contextual and socially constructed, that we get this richness of language, and a computer would definitely struggle to understand.
So much of this metaphorical language (according Lakoff!) is built on our lived human experience that a computer simply couldn't "experience" the world to fully understand the language we use, especially given the multiplicity of languages and cultures and contexts that exist in the world... at least in theory.. :)
Loving the show!
PBS Eons sent me this so fascinating ☺️
This series is extremely good.
This show deserves all the views. I have no interest whatsoever in the topic, but it still is so engrossing!
Great episode, super interesting!
The Turing Test has always reminded me of the opening scene from Blade Runner!
It's fascinating, that "pick out key words and guess what the speaker is likely to be talking about" is basically what I do when I'm struggling through a conversation in French. Except I also get to use situational context on top of general statistics (and when I don't have that context, my comprehension tanks.)
well done very fascinating stuff
this is great, thanks!
Thanks for using "exponentially" correctly there in the first minute and a half. It's not a synonym for "really fast". To a great extent growth in computing power really has been exponential since the sixties.
Me: imagining Sarah drawing the dog towards her by wagging a tasty bone in its face. Yeaaaaah, I don't think computers are understanding speech.
same...
I actually imagined the dog holding a pencil. 😅
oh I'm about to have the time of my LIFE watching this series
3:14 I remember my ASL teacher saying that children who were born deaf had a lot of trouble reading because, to them, reading is basically another language. They had similar problems with sentences like “it’s raining” because, in ASL, 1) you don’t need the “it is” to describe the current rainy condition, 2) you’ll have to specify what “it” is if you want to use the sentence as it is (“it’s raining”).
AMAZING
So happy to see Jean Berko’s Wug test!
What about "Sarah drew the dog with a bone." meaning that she, using the bone as incentive, drew the dog closer to her? Until the picture and highlights were added, that's what I thought it meant.
I have heard of AI Dungeon. Thank you for reminding me that I need to give it a whirl.
I love the sound of your voice.
I’m very interested in what patterns GPT-3 sees in our language. Perhaps some even we haven’t noticed yet? So cool!
This is fascinating. Thank you so much again for doing this segment. Btw, may I ask for the brand and shade number of your lipstick 💄? It so gorgeous (sorry for asking about something other than the content of the episode 🙏)
Glad you enjoyed! And to answer your question: it's MAC Ruby Woo (707) - Dr. B
I wish you guys talked about Ferdinand de Sassure and semiotics. Very much related and very interesting
Ahh Zork and the old Infocom text games. I used to love playing those.
I LOVE EVERYTHING THIS CHANNEL DOES.
Can we talk about the "p" and "f" shift?
And why do people say "on accident" and why does that phrase make my skin crawl?
I wonder how these programs handle vernacular speech, we all know no one uses perfect English grammar daily. So having a program that is so focused on proper grammar depending on its training data would always feel "too proper" to a modern-day speaker. Toss in regional dialect and that's just a whole level of extra complexity.
great point
It’s Lit sent me. And no regrets!
I love that TRS-80 frame. :)
This is... this is FASCINATING. I love this.
I've spoken with GPT-3...typed in and it speaks back...its very eerie...its very close, very close...uncanny valley kind of stuff...I believe that with the help of neuromorphic processing...this may be a route to consciousness for machines...of an alien sort, no doubt...but still...when it makes a mistake its jarring...but it doesn't screw up often...I wonder what will happen in 10 years...
Explains how I trained my phone to type BLOOD FOR THE BLOOD GOD!
Any time I start typing a word that begins with GEN. My phone suggests “Genestealer”
Great video, and illustrates how hard it is not to anthropomorphize machines. Even as the presenter very clearly and effectively explains why algorithms are not thinking, they describe them in the same paragraph as "knowing" what the most likely next word is. When in fact, they know as much as your fridge does when you open the door and the light comes on. It's just really complex switches. Again, as very well explained by the video. We really don't have effective and accurate language for what computers are doing, and possibly we will just live with this possibly dangerous conflation of thought and mere process.
Perhaps it is worth noting as others have before the way that metaphors for the brain have evolved with our tech. It was a pump. Then it was a machine. Then it was a computer. Many gleefully surrender agency to the notion of consciousness as a deterministic, undirected machine. We still don't understand thought itself, and it may be critical to our survival that we are at least able to distinguish it from circuitry.
It's Lit sent me and I'm glad they did!
Came over from Eons. Like the topic. Can't handle the music. I'll come back later to see if there is a version available without the music.
as someone who played many text adventures as a child in the '80's, I approve this message :)
Hearing the word "Ferfumfer" absolutely killed me! I was laughing harder than I have in months for like 5 strate minutes, and I have absolutely no clue why.
Thanks.
Me and my brother used play around with Eliza years ago on the Radio Shack TRS-80 computer like the one show at the 6:23 timemark in this video.
Great video!
By the way: It's Lit says hi!
Man, they really packed a whole video into an intro.
Nice
I have been interested in natural language processing since 1990. Yet I have never even come close to developing my own natural language processor. I have only brainstormed about it--and I have thought of no breakthroughs, sadly. I was blown away by IBM's Watson and the Jeopardy contest. Do you remember that?
Was sent from It's Lit. Am really interested in the entomology of words!
You mean etymology. Entomology is the study of insects.
Another one from Eons here!
Eons did not disappoint
3:42
And now that we all know what Instrumental case is, we all have somewhere to put our guitars.
Interestingly, that old computer that appears at 6:25 is a Radio Shack TRS-80. I had one in 1980 and I ran ELIZA on it!
and now there is ChatGPT ready to change the world !
"Evacuate bowels." Awesome, lol!
Wow. The subject matter sure took a deep dive from "what is a word"
Dr. Brozovsky, you might want to revisit this subject when the computers do become intelligent enough to categorize words into parts of speech. Most likely, computers will still not be able to pick up on connotations or speaker’s intention. One author recently made his point by pointing out how his AI failed miserably to pick up on sarcasm.
That would be now, right?
another OpenAI project, DALL-E, apparently demonstrates some spatial awareness and understanding of object properties
ngl, the intro sequence of Otherwords creeps me out more than that of Monstrum's.
PBS EONS sent me here !