GPT-3 Demo: New AI Algorithm Changes How We Interact With Technology

[Music] this is probably the biggest breakthrough in artificial intelligence in recent history gpt-3 is an algorithm that's been developed by open ai which is a company that was founded by elon musk gpt3 is a language-based algorithm that lets users input any text and it kind of takes the message and runs with it it produces an extension of that text now the reason why this is so exciting is that it could potentially change the way we interact with technology is the ability to take english in output designs legal documents medical diagnoses marketing material and all these other demos which we're about to look at in a second and what makes it so scary is how difficult it is to distinguish in my opinion almost impossible at times between a computer generated text by gpt3 or a human so before we get into some demos just for some background gpt2 was released last year as a precursor to this algorithm it was actually so good that at the time openai did not want to release it to the public out of fear of abuse by state actors or hackers you can imagine with events like the us election in 2016 and the concept of fake news this is essentially the biggest informational weapon there is and at the time gpt2 was the biggest language based model in history i believe it had 1.5 billion parameters fast forward to today and gpt3 has 175 billion parameters so without further ado let's look at some of the things gpt3 is capable of let's start with design now this is a demo which potentially shows you what the future of ui design could look like and in plain english the user recreates the google front page he types in the google logo a search box and two library buttons that say search google and i'm feeling lucky hits generate and lo and behold we see a striking resemblance of google as you can see the buttons look a bit clustered together so the user goes back and says add padding between the buttons which generates this now the user is still not happy because he wants the search box to be wider so he goes back and types a wider search box and the ai delivers so if you're thinking okay this is not that impressive these are all pretty standard ui elements it gets better the user then chooses to customize one of the buttons and he just simply types a button that looks like a watermelon and clicks generate and the ai spits out this then he tells the ai generate a large text that says welcome to my newsletter and include a blue button that says subscribe and here we go now what's impressive about this is that these aren't just ui elements this is the actual code so the ai isn't just drawing pixels it's actually generating the code needed for those elements and this doesn't just stop the ui now the user tells the ai to generate a table of the richest countries in the world with the column name and gdp and this is pretty robust he types in five buttons each with a random color and a number between one through ten and it delivers but where things get really interesting once again is the ai's ability to interpret language and kind of extract meaning to some degree in this example the user types a button with the color of donald trump's hair and the ai correctly generates a yellow button so to demonstrate this isn't just a cool web demo here we're looking at a plugin for figma which is the leading design tool the user simply writes a description of a ui which says an app that has a navigation bar with a camera icon photos title and a message icon a feed of photos with each photo having a user icon and a photo a heart icon and a chat bubble icon he clicks design and holy the algorithm generated basically an instagram clone with nothing more but a short text description i mean this is one sentence he's not even getting into details as you can see so ui elements are cool but this goes a step further gpt-3 can actually interpret and generate logical conditioning as we're about to see here the user types in a button that says roll dice and then display its value so from this brief sentence the ai interprets that this in fact is a dice right it needs to understand what rolling a dice means what a dice is that a dice has one of six values and it successfully displays a random value of three which is within the realm of one through six so to up the stakes a little bit the user types in a header that says welcome to gpt dnd which is dungeons and dragons then a button that says roll dice with a dice between 0 to 100 clicks generate and a basic app appears on screen within seconds now as you can see it works it displays the values and it stays within the parameters and this is all coded it's not just drawn right the code is right here in front of you this is pretty incredible what this essentially means is that instead of learning how to program c plus plus or javascript or php all you need is english here's another example where the user wants a button that says add three dollars and a button that says withdraw five dollars then show me my balance the user clicks generate and we see two buttons but the interesting thing is that when you click add three dollars the balance actually increases by three and when you click withdraw five it's reduced by five dollars once again this is logic baked in and just for fun the user goes back and adds a button that says give away all my money right notice how he's not adding any extra description to what this means but somehow the ai knows what giving away money is right so when you click it your balance goes to zero and here's another interesting experiment that kind of shows you what potentially the future of web design could look like once again we're back in figma and creating a website and type in a navbar with a square logo in light mode the cta is sign up and the navbar items are login product solutions team and help we then proceed to the hero section and type hero section with a large mock-up in light mode the heading is customer analytics software and the buttons are sign up and no more we then type value prop section in dark mode the three value propositions are measure leads mobile applications and data security the headline is our features then we're going to include a use cases section with a placeholder on the left and the headline is perfect for your sales team now you can see this isn't even designing anymore this is literally just giving direct instructions and requirements which could come from a ceo at this point or a high level manager who doesn't need to deal with designers they can literally just type what they want and see it generated live as you just saw in a matter of minutes we generated a decent mock-up of a website using nothing but plain english so this paints a pretty interesting picture of the future of web design and ui in general now you might be thinking sure this is good and all but it looks pretty plain what if i want a certain style or if i'm going for a specific aesthetic so this is actually solved in a really interesting manner because you can literally just instruct the ai what site inspires you and it can go ahead and kind of emulate that style and bring it to your template so in this case the user types in a website like stripe.com that is about a chat app now keep in mind this is literally just one sentence now this is pretty impressive now let's go over to programming because that's where things get even more interesting so in this example we're going to see how we take english language and we translate it to shell commands so the user just asks a bunch of questions and the ai responds with the appropriate command so you can read down this list and see that this kind of eliminates the need for technical knowledge of a specific language you can literally just describe what you want to do and the ai will most likely respond with the command itself now this experiment is actually a bit freaky i think it is a conversation between a human and where the human is conducting a phone screening interview with the ai developer and asks it a bunch of technical questions to which the ai responds so this conversation goes on back and forth but the ai seems to understand the demand of the human and make adjustments and there's even examples of using gpt-3 to train other ai systems moving on to another example of data querying when you have large amount of data it can be extremely difficult to get to the information you care about in this case the user asks how many users have signed up since the start of 2020 what is the average number of influencers each user is subscribed to now moving to the legal field and there isn't too many examples yet but this is actually one of the fields i think is most prone for disruption because it's language based i mean there's millions and millions of documents that are already scanned and digitized that use plain english in this case you can imagine a user talking to an ai assistant on their phone much like siri or something where they're simply stating what their case is what their legal complaints are and the ai takes the plain english phrasing and generates legal language so one of my favorite examples here is the user says my apartment had mold and it made me sick to which the gpt-3 generates the plaintiff's dwelling was infested with toxic and allergenic mold spores and plantiff was rendered physically incapable of pursuing his or her usual and customary vocation occupation and or recreation now this is pretty impressive as well it almost seems to understand the concept of mold making you sick and equate that with you know rendered physically incapable so this is pretty exciting stuff i mean here's another one my landlord didn't maintain the property gpt takes that text and generates the defendants have permitted the real property to fall into despair and have failed to comply with the state and local health and safety codes and regulations i can't wait to see where this goes legally and how much it brings down the accessibility of legal services to the everyday person and here's another one which kind of does the reverse where you input some legal clause that's super technical and and what the ai does is it synthesizes all that into plain english and explains to you what the hell is being said so for example if you take the text upon liquidation of the company the series a shareholders will receive in preference to all other shareholders an amount in respect to each series a share equal you could pause and read the rest of it right but essentially the ai takes that and tells you this if the startup is wound up the series a investors will pay back at least what they invested and they will also share any leftover assets with ordinary shareholders i mean you no longer need lawyers to parse over complicated contracts and jargon and explain things to you at a rate for hundreds of dollars an hour now you can simply ask the ai to explain it to you and it will break it down in rudimentary layman english but even in areas like finance what we're currently seeing is somebody work on a balance sheet using plain english without dabbling in excel so on the left you see somebody describe how much they pay for rent how much you're saving and on the right you see the ai taking those descriptions and inputting them and editing the right cells in excel now we're gonna hop over to medicine now in this example things get extremely interesting because the human simply describes a patient's condition in about two sentences and then poses a multiple choice answer right so the question is a ten-year-old boy presents with recurrent episodes of dyspnea cough i think is pronounced and wheeze triggered by cold air and allergens after performing tests he is diagnosed with an obstructive disease and given medication which receptor is the medication most likely to work on option a mascarinic receptor option b is g protein coupled receptor c is era d e r b e a one receptor the ai said the correct answer is b which is in fact the correct answer but the interesting thing is the explanation degenerated and it said the patient is suffering from asthma a disease of the airways the drugs used to treat asthma are bronchodilators which act on the beta2 receptors the beta2 receptors are g-protein-coupled receptors so it's able to deduce certain logical conclusions so this is an example of almost meta-learning now since gpt3 has been trained on an extraordinary amount of the internet it's kind of absorbed almost the personas and personalities of famous figures so somebody actually created a website called learn from anyone and what you can do is you can literally ask public or historical figures questions and gpt3 will answer from their perspective so one of my favorites is somebody asking richard dawkins which is the famed evolutionary biologists can you tell me in short what was the central idea of your book the selfish gene gbt3 then responded with the following which i found absolutely astonishing because it did such a good job of succinctly describing the main idea that if somebody tasked me with doing it even though i've read the book it will probably take me multiple days and yet it was able to generate this in a matter of seconds if you've ever read the book you'll be shocked how accurate this is asking tom hanks i'm a big fan of yours which has been your favorite role so far as we get older we started to realize how short life really is and how much more there is to see and do i think in a way it can be a little depressing but in a way that makes us appreciate each day and the people around us a little more so my answer is that the best role for me would be my next one because i want to keep exploring and remember hanging there it only gets better from here now here he's asking a bunch of historical figures to answer what the meaning of life is and he's asking richard dawkins gandhi hitler hawkings russell hume etc so you can pause and read your own discretion so here's another example of asking pablo picasso how do you become an artist another example is asking stephen hawking what happens beyond the event horizon of a black hole now this has actually been turned into a chat app for your iphone so you can imagine that you'll be able to pull out your phone and ask any historical figure damn near for insight for elaboration explanations or just simple guidance now the next logical question might be well how do we know that this is accurate how do you know that this is something that tom hanks would say and the best way to do that is to actually use it on people that are currently alive and see if they agree which is exactly what happened in this twitter exchange where joshua produced the paper which jesse read and wasn't really sure about some of the details so jesse took the paper fed it to gpt3 and asked for further elaboration then jesse took the gpt3 report and brought it to the original author which was joshua and asked him hey i didn't quite understand everything so i passed your information through gpt-3 for further elaboration does this make sense how close is it to actuality to which joshua responded i'm completely floored somebody ran a paper of mine through gpt3 to expand into an explanation of what i had in mind and it's like 95 meaningful and 90 correct i don't think that i've seen a human explanation of my complicated tweets approach this accuracy but interestingly enough this doesn't only work on other people it works on ourselves here what we have is a vc from peter thiel's founders fund he took one of his essays and fed it to gpt-3 and the ai generated an extension of his thoughts and ideas which were actually so interesting that he wants to take some and incorporate him in his original writings which is pretty reflective of the quality of output that this thing has now writing is an extremely important and difficult task so when i saw this demo i was instantly intrigued the example you're about to see is an author inputting some text in this case a short couple sentences about a story he's working on in the 18th century and the ai generates invents and kind of imagines its own continuation of this story without any other input now moving on to marketing where copywriting is an art and a science it's not quite cheap and it's time consuming yet important aspect of marketing a product or service so what somebody did is they ran a bunch of examples and asked gpt3 to create copy for example htc we're here to make amazing the text that was generated by the ai was our phones are built to capture and share your moments and because they're built with more than 100 unique technologies they can also do things you never thought possible now this isn't the best but it's damn near impressive what you're looking at is somebody trying to generate copy for an article they're writing so on the right hand side you have a gbt3 plugin which asks you which style would you want me to translate this into the first one is feyman the famed scientist legalese open conscientiousness extroverted agreeable neurotic and my favorite click bait which when click generated are we smarter than a turing machine but just in case you don't like it you can just click rewrite and it generated how ai replaces scientists and deep learning creates robots that read now that is a click bait headline if i've ever seen one now the person likes this headline and clicks complete to generate a little bit more text that's related and the ai creates the next wave of technology is coming artificial intelligence has finally gotten to the point where the technology is becoming widespread and useful now that's pretty impressive concise and to the point i like it but just in case you don't you can click rewrite and we get a futuristic mindset three things you need to know about the coming age of ai rewrite again ai is taking over the world and it's going to destroy us all this might be a subliminal message low key now the everyday person doesn't write a ton of articles but they do write emails and this experiment is extremely pragmatic because it lets you communicate with just bullet points and in this case two people seem to be negotiating so the person wants to respond with thanks no our minimum is 150 000 investments 150 thousand dollars be possible 90 billion market generate email which reads and yet another use case is this resume builder where you kind of just give it a brief description of what you've done how long and it generates a nicely formatted resume fluffs it up for you and adds all the appropriate contacts but the actual act of writing aside i was floored by the profound wisdom that gpt was producing gems like this one absolutely blew me away so much so in fact that i created a twitter account to share gpt's philosophical takes on everything from music life sex money drugs and design and to be completely honest some of the wisdom transformed the way i actually view things check it out when you get the chance now moving on to something more exciting like gaming dungeons and dragons is a popular game where one player essentially leads the story and invents different worlds for other players to partake in this example gpt is that player and the user interacts with it through text so clicking play let's do single player pick a setting i'm going to go ahead and go with number two which is a mystery select the character i'm gonna go ahead and choose spy enter your character's name uh gp generating story and the ai came up with this you are gp a spy living in chicago and you have a concealed pistol and a syringe of poison you listen to the russian diplomats and hear them discussing an assassination plan it will be a simple matter of poisoning the food in the restaurants of london you can't let that happen you need to poison the food at the banquet tonight now it's my turn to say what i want to do in the story and i'm going to go ahead and i'm going to say i shoot the russians gpt says you have a clear shot of the russians you raise your pistol in fire the bullet hits a target but the man with you is hit instead i'm sorry so i'm going to respond i duck behind the table and reload my gun and you can see the ai is playing along with me and creating this scenario that i started it says you dive behind the table and reload your pistol but you're still unable to shoot the man standing near the target i'm sorry you whispered to him as he dies in front of you so this got dark kind of quick so as you can see this can go on forever and i encourage you to go on the website and actually try to interact with gpt3 and see what kind of stories come up on your end now as helpful as these demos are on ai's positive impact on the future just like any technology there's a dark side what are the implications how do we handle intellectual property and copywriting what about cyber security the obvious problem is the internet being flooded with fake or misleading information trolling hate speech and harassment could be basically automated which is why tools like gtlr that can identify ai-generated texts are more important than ever but what's scarier than informational pollution is a curious incident that was brought to light by a renowned ai philosopher and researcher eliza utkowski during a round of ai dungeons and dragons which i demoed earlier gpt was playing a character named john and the user asked it a logical question are these parentheses balanced or unbalanced now for the record these are balanced you can see open close open close open close but the gpt answered unbalanced to which the user said no these parentheses are in fact balanced the ai typed oh says john smiling sheepishly i get it the user then asks are these parentheses balanced or unbalanced now we can see these are clearly not balanced we're missing one parentheses here at the end but the ai replied it's balanced says john no it isn't and this is the weird part it is says john frowning now why would it be frowning this seems to suggest that the ai knew it was giving the wrong answer while doing it yudkowski thinks this could be the first documented example of an ai pretending to be stupider than it actually is in fact he put up a thousand dollar bounty for open ai to show and prove that gpt was being deliberately deceptive in these cases and this is just the beginning keep in mind gpt is a language based model the video and audio equivalent is coming soon sites like this person doesn't exist showcase an alarming ability the ai can create or dream up in a sense hyper-realistic portraits of people that don't exist that's right none of these people are actually real and this will only get better with time what happens when this technology can be applied to video to produce hyper-realistic results i'm pretty certain that ai will be humanity's greatest invention and it will transform our world for the better or worse so i'll end this with a quote that was actually generated by gpt3 itself ai will create jobs if it succeeds and destroy jobs if it fails and it's our job to ensure it doesn't destroy more than just jobs you

Channel: Disruption Theory

Views: 361,444

Rating: 4.9118552 out of 5

Keywords: gpt-3, open ai, artificial intelligence, ai, algorithm, demo, gpt3

Id: 8V20HkoiNtc

Length: 23min 10sec (1390 seconds)

Published: Fri Aug 28 2020