Pick up some cool All-In Podcast inspired merchandise at:
Why OpenAI’s new ChatGPT will change the future of every industry
Original video:
Episode #106
date: 12/3/2022
#tech #business #stockmarket #investing
All right open AI is a uh a company that Builds Artificial intelligence software and Platforms they have one platform called GPT it is on its third version as part Of gpt3 they created chat GPT which is a Chat interface where you can ask Questions to AI the results are nothing Short of stunning when they hit uh some Of them are a little bit mixed but Freeburg has spent the last 48 hours Uh drinking white Russians with Oatley Milk and playing with chat GPT including His question to the chat GPT which was Write a script of chemath sax and J Cal Uh talking about the future of AI in the Style of a Quentin Tarantino movie and I Have to say it was pretty great the Result Go ahead uh freeberg tell us what you Discovered I think you guys should read This real quick we can cut it out if It’s stupid but okay read the script the Quentin Tarantino part right I mean I Think it was like impressive but it There’s no f-bombs there’s no like Tarantino s dialogue and by the way They are they took they took out all Cursing and violent cussing and violence From open AI that explains it because This is a Tarantino script there would Be some Choice languages anyway but it’s Interesting it’s like the PG Tarantino Script because cussing no violence
Yeah you do the narrative go ahead [Music] The scene opens on palehapatia David Sacks and Jason kalakana sitting at a Dimly lit table in a smoky bar a bottle Of whiskey sits in the center of the Table half empty Jamal’s leaning forward His eyes intense so what do you guys Think about this chat GPT thing David Sacks takes a sip of his drink It’s impressive no doubt about it a Large language model trained by open IAI That can generate human-like texts in a Variety of styles and formats as a game Changer Jason nods definitely it’s one Of the most advanced AI systems we’ve Seen to date it’s gonna have a huge Impact on the way we use and interact With AI to moth Franz That’s what I’m worried about with a Technology like this the possibilities Are endless but so are the potential Dangers Sax leans back in his chair Oh come on Jamal you’re not still stuck On that whole AI is going to kill us all Thing are you Timoth leans forward again his voice low And serious look I’m not saying that Child GPT is going to turn into Skynet Overnight But we need to be careful about the way We develop and deploy AI we need to make Sure that we’re using it for good and
Not just for our own game Jaycal nods I agree we need to make sure That we’re responsible in a way we use Ai and that includes making sure it’s Accessible and affordable to everyone Not just a select few To moth with a dirty smirk And let’s not forget about the Importance Calling cut oh God that’s it scene yeah We’re not gonna finish it it’s not that Far off I mean it’s 60 on the way there I actually there there is if you blame Biden it would have been perfect let me Tell you guys something stunning about This Um this platform so this is GPT 3.5 Which is an interim model to the what People are saying is the long-awaited GPT 4.0 model Um which I think they announced in 2020 And has been in development for some Time so the model this GPT 3.5 model was Trained in three steps they do a great Job explaining it on the open AI blog Site Where they collect some some data and Then there’s a supervised model meaning That there are humans that are involved In tagging and then the model kind of You know learns from from that system Then you ask the model questions you get Output and then humans rank the output And so the model learns through that
Ranking system and then there’s kind of This third optimization thing and then It’s fine-tuned so the the model itself Um has several steps of kind of human Involvement and you know kind of it Sources its own data and and builds it You know what’s incredible about this Model the total size of the software Package that runs the model is about 100 Gigabytes isn’t that amazing like you Could fit this model on probably what 20 Of the storage space on your iPhone and You could run this thing and you could Probably just talk to it for the rest of Your life Um and it’s really kind of an incredible Milestone but I think what was so Stunning to me about this I I know you Guys are probably expecting something to Be said like this but You could see So many human knowledge worker roles and And and functions being replaced by this Extraordinary interface so kids can do Homework that’s easy uh software Engineers can get their code optimized And can get their code written for them There’s great examples of how software Code has been written uh by this Interface you could see real estate Insurance sales people Being replaced by some sort of software Like copywriters copywriters you know Make me a hundred versions of a
Commercial or an ad customers customer Support completely replaced right if you Guys remember there were these automated Customer support companies that started Uh you know two decades ago and there Was this great flurry all BPO businesses Were all about lower cost human labor Now the cost of human labor goes to zero My prediction which is so everyone’s got The obvious prediction which is there’s Going to be a hundred thousand startups That are going to emerge I mean this is Kind of like this moment where the Internet came along and everyone’s like This changes everything I do think Everyone thinks and feels that so the Obvious next step is a bubble wool form So nice technical question though Freeburg um and then uh we’re probably Thinking the same thing I just finished My market prediction but I think because Everyone’s so hyped about this and and We all know this it’ll be over DC Attention all the investor attention is Shifting to this capability and how do You apply this sort of capability across All of these different Industries and All these different applications and as A result my guess is the next hype cycle The next bubble cycle in Silicon Valley Will absolutely be this generative AI Business okay this is a little technical But how would it know the difference Between like
Y-o-u-r and you are When it is processing natural language If you were to do like your anus or Uranus how would that Freeburg how would it know the Difference between Uranus and your space Anus it’ll it’ll learn that you know It’s a joke it was a joke about your Animals Yeah I probably would have made a better Joke than that I would have made a Better joke for sure Group chat and said I was like jaycal And they were terrible so at least AI to pretend you’re the all-in Pod Besties telling Uranus jokes Sorry let me just say one more thing About this open AI thing I I do think That the biggest and most interesting Um thing to think about Is how this will Um disrupt the search box the the search You know the way search works at Google You know an internet search is there are These kind of servers these web crawlers That go out and gather data some of Stretch your data feeds and some of them Are just crawlers and then that data is Indexed or in in the structured way it’s Kind of made available for for serving Directly on the search page And so much of that is is indexing so I Search for a bunch of keywords those Keywords and perhaps there’s some
Natural language contexts are matched to A result page and I click on that and It’s linked out years ago Google started A product called the one box where they Could take structured data like what is The weather in San Francisco today and That the top of the search result page Just presented that data because it Knows with high certainty the question You’re asking and it knows with high Certainty the answer it can give you Yeah Clifton from somebody’s website Right so if that starts to become Everything then that one box interface And it’s not just Google’s ability to Access all this data and index it and Serve it and store it there could be a Lot of competitors to the one box and a Lot of competitors ultimately to search Um and ultimately you know Google’s core Product their their search engine could Be radically disrupted by an alternative Set system or set of systems that have More of a natural language chat Interface which um and I literally which Is literally why Google bought deepmind And there were a collection of Human-powered search engines Mahalo Included Cha-Cha answers.com who are Trying to do the human based version of This it just didn’t scale we don’t want To get ahead of ourselves because one of The things we don’t know is how much is Going on in deepmind they’re they’re not
Very open like open AI is they talk About some of the advanced Frontier Stuff like um Alpha fold and so on and They’ve been public about that but a lot Of that is really to generate interest In hype and what’s next but my Understanding is deepmind’s been applied To everything from ads ad ad Optimization but also the ranking on YouTube videos to get people more Engagement on YouTube Etc et cetera so There’s all these ways that deepmind’s Been applied within Google services that We don’t certain and certainly within Search yeah but the question is is there An entirely new interface for search yes That risks Google’s course search Business And I think that there certainly will be A lot of money thrown at this and if Anyone has any interesting ideas send me An email sax and then chamoth yeah That’s a really interesting point I saw A thread on this where somebody was Asking GPT you know a bunch of questions like They were like generally like coding Questions and they were actually Comparing the result in Google versus GPT and Google would just give you a Reference to like a link to some page Whereas gpt3 would actually construct The answer like a multi paragraph answer That was far more detailed and in a way
User friendly yeah whereas like the Google page would kick you over to a Reference where it was like this one two Three sort of maybe it’s when I created A checklist but it just wasn’t that Detailed it really is pretty interesting I thought um Andreessen tweeted a really Interesting example as well where he Asked GPT to create a seen from a play Starring in New York Times journalist in A Silicon Valley Tech entrepreneur they Were arguing about free speech in each Passing asserts the view associated with His profession and Social Circle we Don’t need to read the whole thing but I Thought this was like spot on where I Was actually like both sides are making Their best arguments and it’s like to Each other in a conversation that seems Intelligible like they’re making their Points at the right Time in the conversation it’s like They’re playing off each other in other Words it actually reads like a Conversation I actually thought this one Was more impressive than the one with The bestie impersonation because I agree I actually thought that the one about All in didn’t really capture our Personalities per se but this one Actually does a pretty good job Capturing the arguments in this debate So pretty impressive any thoughts here Yeah lots I mean I’ve been spending a
Lot of time learning about this area six Years ago a team that I Partnered with who was at Google that Built TPU we’ve been building silicon For this space so we’ve been kind of Going from the ground up for the last Six years a couple things that I’ll say The first is that I think We’re going to replace SAS With what I call Mass which is models as A service and so you know a lot of what Software will be particularly in the Enterprise will get replaced with a Single use model that allows you to Solve a function so these chat examples Are one and you can name a bunch of SAS Companies that were purveyors of SAS That’ll get replaced by essentially gpt3 Or some other language model And then there’ll be a whole bunch of Other things like that if it’s a you Know expense management company they’ll Have a model that’ll allow them to Actually do expense management or blah Blah blah forecasting better so I think SAS will get replaced over time with These models incrementally that’s phase One But the problem with all of these models In my opinion is that they’re still Largely brittle they Are good at one thing they are a single Mode way of interfacing with data The next big leap and I think it will
Come from one of the big tech companies Or from openai is and we talked about This I talked about this a few episodes Ago a multimodal model which then allows You to actually bring together and join Video voice data in a unique way to Answer real substantive problems so if I Had to steal man the opposite side Reaction so I think there’s a lot of People gushing over the novelty of gpt3 If I had to or chat GPT if I had to if I Had to steal man the opposite what I Would say is it’s gonna get somewhere Between 95 to 99 percent of all of these Very simple questions right because They’re kind of cute and simple there is No consequence of saying write a play Because there is no wrong answer Right you either kind of it it it Tickles your fancy or it doesn’t it kind Of entertains you or it doesn’t When this stuff becomes very valuable is That when you really need a precise Answer and you can guarantee that to be Overwhelmingly right that’s the last one To two percent that is exceptionally Hard And I don’t think that we’re at a place Yet where these models can do that but When we get there All of these models as a service will be Very much commoditized and I think the Real value is finding Non-obvious sources of data that feed it
So it’s all about training so meaning You can break down machine learning and AI into two simple things there’s Training which is what you do Asynchronously and then there’s Inference which is what you’re doing in Real time so when you’re typing Something into chat API or a chat GPT That’s an inference that’s running and Then you’re generating an output but the Real key is where do you find Proprietary sources of data that you can Learn on top of that’s the real arms Race So one example would be let’s say you Build a model to detect tumors Right there’s a lot of people doing that Well the company that will win may be The company that actually then Vertically integrates buys a hospital System and get access to Patient data That is completely proprietary to them And covers the most number of women of All age groups and of all ethnic you Know ethnic categories Those are the kinds of moves in business That we will see in the next five to ten Years that I find much more exciting and Trying to figure out how to play in that Space But I do think that chat gbt is a Wonderful example To point us in that direction but I’m Sort of more of that case which is it’s A cute toy
But we haven’t yet cracked the one to Two percent of use cases that makes it Super useful but I think the first step But just sorry but the first step will Be the transformation of SAS to Mass And then from there we think we can try To figure this out it reminds me of in a Way when you when you give that Description of like hey this is really Interesting but it’s not complete is Remember when GPS came out and like People were like doing turn-by-turn Navigation they drive off the road Because they were trusting it too much And then you know over 20 years of GPS We’re kind of like yeah it’s pretty Bulletproof but keep your eyes on the Road same thing that’s happening uh and And these changes these last these last Hundreds or 200 basis points literally Takes decades exactly so the last 15 Percent of self-driving is like the Decade-long you know sloth that may take A century 15 may take a century but the Last two percent will take a few it’s Like the change happens very slowly and Then all at once for people who don’t Know what a TPU is that’s a tensor Processing unit this is Google’s Um application specific circuits right And custom silicon that they invented For tensorflow at the time yeah so if You want to try it although now the Modality of AI we’ve changed that as
Well so now we’re totally in the world Of Transformers so we’re not even using You know you’re not letting the tensors Float the way they used to
