Ep. 244: Cal Newport’s Thoughts on ChatGPT

00:00:00.000 | That is the deep question I want to address today.

00:00:05.280 | How does chat GPT work?

00:00:08.700 | And how worried should we be about it?

00:00:19.160 | I'm Cal Newport, and this is Deep Questions, the show about living and working deeply in

00:00:25.720 | an increasingly distracted world.

00:00:27.600 | I'm here in my Deep Work HQ, joined once again by my producer, Jesse.

00:00:37.440 | So Jesse, you may have noticed that we have been receiving a lot of emails in the last

00:00:43.360 | few months about chat GPT.

00:00:46.120 | Yeah.

00:00:47.240 | And related AI technologies.

00:00:49.080 | And our listeners want to know my thoughts on this, right?

00:00:51.880 | I'm a computer scientist.

00:00:52.880 | I've thought about the intersection of technology and society, and I've been silent about it.

00:00:57.360 | Well, I can reveal the reason why I've been silent about it is that I've been working

00:01:01.040 | on a big article for the New Yorker about exactly this technology, how it works and

00:01:06.480 | its implications for the world.

00:01:08.520 | And my general rule is when I'm writing an article, I don't talk about that subject publicly

00:01:13.160 | until the article is done.

00:01:15.760 | I mean, that's basic journalistic practice, but actually, Jesse, I've never told this

00:01:19.960 | story, but that rule was really ingrained in me when I was in college.

00:01:27.320 | So when I was coming up as a young writer, you know, I got started pretty early, wrote

00:01:30.120 | my first book in college.

00:01:31.960 | I was commissioned to write something for the New York Times.

00:01:34.720 | I don't remember exactly what it was, maybe an op-ed, something to do with college students

00:01:38.840 | or something like this.

00:01:39.840 | And I had an early blog at that time.

00:01:43.680 | And I wrote something on the blog like, "Hey, isn't this exciting?

00:01:47.640 | I'm going to write an article for the New York Times."

00:01:49.040 | And maybe like I put the short email on there and it was like, "Yeah, we'd love to have

00:01:52.080 | you write the piece or something."

00:01:53.760 | And that editor went ballistic.

00:01:55.400 | Really?

00:01:56.400 | Yeah.

00:01:57.400 | Cancelled the piece.

00:01:58.400 | Cancelled the piece?

00:01:59.400 | Chewed me out.

00:02:00.400 | Now, this was early internet, right?

00:02:02.320 | I mean, this was 2004 probably.

00:02:05.800 | So I don't know.

00:02:07.100 | Maybe it was more, it felt more like a breach then, but ever since then, if I'm writing

00:02:11.120 | an article.

00:02:12.120 | Did you ever talk to that editor again?

00:02:13.760 | No.

00:02:14.760 | I ended up writing a lot for the Times, but not really until 2012.

00:02:20.240 | Was that, you're going to be your first big splash?

00:02:22.040 | That would have been my first big splash.

00:02:24.080 | Were you like depressed for a couple of days?

00:02:26.040 | A little shook up.

00:02:27.520 | And then starting with So Good They Can't Ignore You and going forward, I had a really

00:02:30.720 | good relationship with the Times, especially through digital minimalism.

00:02:32.960 | I've written tons of articles for them, but there is a lesson learned.

00:02:37.040 | So I've written now, the day that we're recording this podcast, April 13th, my new article,

00:02:42.880 | The New Yorker has been published.

00:02:43.880 | So I am free.

00:02:45.000 | The gag order has been lifted and we can get into it when it comes to chat GPT.

00:02:51.320 | In fact, I'll even load the article up here on the screen.

00:02:54.600 | For those who are watching, you will see this on the screen.

00:02:58.280 | If you're not watching, you can watch at youtube.com/calnewportmedia.

00:03:03.520 | Look for episode 244.

00:03:05.360 | You can also find that at the deeplife.com episode 244.

00:03:08.520 | Here Jesse is the long awaited article.

00:03:11.200 | The title is What Kind of Mind Does Chat GPT Have?

00:03:15.240 | The subhead is large language model seems startlingly intelligent, but what's really

00:03:19.360 | happening under the hood?

00:03:22.720 | It's a big long article.

00:03:23.720 | So it's good.

00:03:24.720 | I'm excited to read it.

00:03:26.000 | Yes, we can talk chat GPT.

00:03:27.160 | I mean, you probably haven't been following it too closely just based on our conversation.

00:03:31.760 | Some people really in the weeds and some people don't want to know.

00:03:35.800 | I'm guessing you're not in the weeds on chat GPT, but I could be wrong.

00:03:39.400 | No, I'm not in the weeds at all.

00:03:40.880 | I listened to like a few of the hard fork episodes on it.

00:03:45.120 | That was about it.

00:03:46.360 | And what was the tone of those episodes?

00:03:49.040 | They were given some examples of what it was when it first came out.

00:03:52.200 | I probably listened to them like six weeks ago.

00:03:54.480 | Yeah.

00:03:55.480 | And then, yeah, that was kind of it.

00:03:58.320 | Well, so I'll give a quick primer then before we get into the guts of what I want to talk

00:04:02.240 | about today.

00:04:03.240 | So chat GPT is a chat bot.

00:04:05.640 | You can sign up for an account at open AI and it's a web interface.

00:04:10.320 | You type in questions and chat GPT or prompts or requests and chat GPT response types text

00:04:15.080 | back like you're talking to someone over Slack or instant messenger.

00:04:18.280 | So this was released in November, late November of last year, and almost immediately people

00:04:23.160 | began circulating online screenshots of particularly impressive interactions or particularly funny

00:04:30.280 | interactions that they had with chat GPT.

00:04:33.000 | Here's one of the first ones to go viral.

00:04:34.960 | I talk about this one in my article.

00:04:38.640 | So here's a tweet of a screenshot that went along.

00:04:41.700 | This was from a software developer named Thomas Pacek and he asked chat GPT the following,

00:04:48.880 | write a biblical verse in the style of the King James Bible explaining how to remove

00:04:54.060 | a peanut butter sandwich from a VCR.

00:04:57.360 | Chat GPT rose to the challenge and wrote a response that begins, and it came to pass

00:05:03.840 | that a man was troubled by a peanut butter sandwich for had been placed within his VCR

00:05:07.880 | and he knew not how to remove it.

00:05:09.640 | And he cried out to the Lord saying, Oh Lord, how can I remove this sandwich from my VCR

00:05:13.460 | for it is stuck fast and will not budge.

00:05:16.000 | And the response goes on.

00:05:17.860 | Here's another early viral example of chat GPT's prowess.

00:05:24.780 | This was a someone named Riley Goodside who asked chat GPT to write a sign failed scene

00:05:33.340 | in which Jerry needs to learn the bubble sort algorithm.

00:05:38.000 | And chat GPT once again rose to the occasion.

00:05:41.320 | A not a properly formatted script, but has some of the aspects of it.

00:05:46.760 | It opens in a monk's cafe.

00:05:50.160 | It says Jerry is sitting at the counter with George.

00:05:53.120 | Jerry sighs and says, I can't believe I have to learn to bubble sort algorithm for my computer

00:05:57.200 | science class.

00:05:58.640 | George laughs, bubble sort.

00:06:00.480 | That's the most basic sorting algorithm there is.

00:06:02.180 | Even a monkey could do it.

00:06:03.560 | Audience laughs.

00:06:04.560 | Jerry.

00:06:05.560 | Yeah, well, I'm not a monkey.

00:06:06.560 | I'm a comedian.

00:06:07.880 | And then the screen, the scene goes on from there.

00:06:10.360 | All right.

00:06:11.360 | So this is the type of thing chat GPT can do.

00:06:12.680 | These impressively perceptive answers to pretty esoteric requests.

00:06:18.240 | Now, if you go back and actually watch the media cycle around chat GPT, which I have

00:06:25.040 | to say is driven very strongly by Twitter.

00:06:27.160 | I think the fact that anyone can sign up for an account and that screenshots of your interactions

00:06:32.600 | can be easily embedded in the Twitter really helped get the hype cycle around this technology

00:06:38.560 | spinning much more furiously than it has for past artificial intelligence innovations.

00:06:43.160 | Anyways, if you go back and look at this media cycle, it took a week or two before the tone

00:06:48.120 | shifted from exuberance and humor.

00:06:53.280 | Like if you look at this example I just gave about Seinfeld, the tweet says, actually not

00:06:59.920 | that one I meant of the VCR.

00:07:01.840 | The tweet says, I'm sorry, I simply cannot be cynical about technology that can accomplish

00:07:06.400 | this.

00:07:07.400 | So it went from this sort of exuberance and happiness to something that was a little bit

00:07:10.320 | more distressing.

00:07:11.320 | Here's a couple of examples I want to bring up here.

00:07:14.440 | Here is an article from NBC news.

00:07:17.480 | The headline is chat GPT passes MBA exam given by a Wharton professor.

00:07:24.280 | Oh, that got people worried.

00:07:27.800 | Here is a another article from around this period from time magazine headline.

00:07:33.280 | He used AI to publish a children's book in a weekend.

00:07:36.920 | Artists are not happy about it.

00:07:38.900 | It details a product design manager who used chat GPT to write all the texts of a book,

00:07:43.680 | which he then self published on Amazon and started selling a bit of a stunt, but it implied

00:07:50.280 | certain types of future scenarios in which this technology was taking away creative work

00:07:55.320 | that really made people unsettled.

00:07:58.380 | As we get closer to the current period, I would say the tone shifted since the new year

00:08:01.720 | and in particular coming into March and April, the tone shifted towards one of alarm, not

00:08:08.040 | just about the focused economic impacts that are possible with this type of technology,

00:08:13.840 | but some of the bigger societal, if not civilization level impacts of these type of technologies.

00:08:20.680 | I would say one article that really helped set this tone was this now somewhat infamous

00:08:27.760 | Kevin Roos piece from the New York Times that is titled a conversation with Bing's chat

00:08:34.120 | bot left me deeply unsettled.

00:08:38.040 | Bing released a chat bot after chat GPT based on a very similar underlying technology.

00:08:43.640 | Kevin Roos was, I guess, beta testing or using this new tool and fell into this really sort

00:08:49.120 | of dark conversation with the chat bot where among other things, the chat bot tried to

00:08:53.600 | convince Kevin to divorce his wife.

00:08:56.000 | The chat bot revealed that she had a sort of hidden double identity.

00:09:01.320 | I think that identity was called venom, which was a very sort of dark personality.

00:09:05.680 | So Kevin set a tone of, Ooh, I'm a little bit worried.

00:09:08.540 | And it escalated from there.

00:09:10.840 | In late March, we get this op ed in the New York Times.

00:09:16.040 | This is March 24th, written by some prominent authors, Yuval Harari, Tristan Harris, and

00:09:22.240 | Azza Raskin.

00:09:24.520 | And they really in this article are starting to point out potential existential threats

00:09:30.000 | of these AIs.

00:09:31.520 | They are arguing strongly for, we need to take a break and step back from developing

00:09:36.960 | these AIs before it becomes too late.

00:09:39.760 | Here's their last paragraph.

00:09:41.760 | We have summoned an alien intelligence.

00:09:43.860 | We don't know much about it, except that it is extremely powerful and offers us bedazzling

00:09:48.120 | gifts but could also hack the foundations of our civilization.

00:09:52.680 | We call upon world leaders to respond to this moment at the level of challenge it presents.

00:10:00.880 | The first step is to buy us time to upgrade our 19th century institutions for an AI world

00:10:04.680 | and to learn to master AI before it masters us.

00:10:09.000 | A few days after this op ed, an open letter circulated signed by many prominent individuals

00:10:13.820 | demanding exactly this type of pause on AI research.

00:10:16.980 | Okay, so this is the setup.

00:10:19.860 | Chat GPT is released.

00:10:22.200 | Everyone's using it.

00:10:23.200 | Everyone's posting stuff on Twitter.

00:10:24.740 | Everyone's having fun.

00:10:25.740 | Then people start to get worried about, wait a second, what if we use it for X, what if

00:10:28.580 | we use it for Y?

00:10:29.580 | And then people got downright unsettled.

00:10:31.260 | Wait a second, what if we've unleashed an alien intelligence and we have to worry about

00:10:35.120 | it mastering us?

00:10:37.820 | We have to stop this before it's too late.

00:10:39.220 | So it really is a phenomenal arc and this all unfolded in about five months.

00:10:46.480 | So what I want to do is try to shed some clarity on the situation.

00:10:52.080 | The theme of my New Yorker piece, and I'm going to load it on the screen and actually

00:10:54.800 | read to you the main opening paragraph here.

00:10:59.140 | The theme of my New Yorker piece is we need to understand this technology.

00:11:04.740 | We cannot just keep treating it like a black box and then just imagining what these black

00:11:12.860 | boxes might do and then freak ourselves out about these stories we tell ourselves about

00:11:17.300 | things that maybe these black boxes could do.

00:11:19.540 | This is too important for us to just trust or imagine or make up or guess at how these

00:11:27.540 | things function.

00:11:28.540 | So here's my, the, the nut graph of my, my New Yorker piece.

00:11:32.660 | What kinds of new minds are being released into our world?

00:11:35.280 | The response to chat GPT and to the other chat bots that have followed in its wake has

00:11:38.900 | often suggested that they are powerful, sophisticated, imaginative, and possibly even dangerous.

00:11:43.760 | But is that really true?

00:11:45.800 | If we treat these new artificial intelligence tools as mysterious black boxes, it's impossible

00:11:50.180 | to say.

00:11:52.160 | Only by taking the time to investigate how this technology actually works from its high

00:11:56.180 | level concepts down to its basic digital wiring, can we understand what we're dealing with.

00:12:01.600 | We send messages into the electronic void and receive surprising replies, but what exactly

00:12:06.600 | is writing back?

00:12:09.360 | That is the deep question I want to address today.

00:12:13.240 | How does chat GPT work?

00:12:16.660 | And how worried should we be about it?

00:12:19.240 | And I don't think we can answer that second question until we answer the first.

00:12:21.720 | So that's what we're gonna do.

00:12:23.080 | We're gonna take a deep dive on the basic ideas behind how a chat bot like chat GPT

00:12:28.680 | does what it does.

00:12:30.160 | We'll then use that to draw some more confident conclusions about how worried we should be.

00:12:33.880 | I then have a group of questions from you about AI that I've been holding on to as I've

00:12:38.340 | been working on this article.

00:12:39.340 | So we'll do some AI questions and then the end of the show, we'll shift gears and focus

00:12:43.540 | on something interesting.

00:12:45.040 | So an unrelated interesting story that was arrived in my inbox.

00:12:47.720 | All right, so let's get into it.

00:12:49.360 | I want to get into how this actually works.

00:12:51.000 | I drew some pictures, Jesse, be warned.

00:12:53.960 | I am not a talented graphic designer.

00:12:57.800 | That's not true.

00:12:58.800 | I'm not much of an artist.

00:13:00.200 | So Jesse watched me hand drawing some of these earlier on the tablet.

00:13:03.600 | I got to say this ain't exactly Chiat Gay, the famous ad agency level work here, but

00:13:11.160 | you know what?

00:13:12.160 | It's going to get the job done.

00:13:13.160 | So I have five ideas here I'm going to go through.

00:13:15.360 | And my goal is to implant into the high level ideas that explain how a computer program

00:13:21.520 | can possibly answer with such sophisticated nuance.

00:13:26.160 | These weird questions we're asking it how it can do the Bible verse about the VCR, how

00:13:30.400 | it can do a Seinfeld scene with bubble sort.

00:13:33.640 | And we're going to do this at the high level.

00:13:34.800 | We're going to essentially create a hypothetical program from scratch that is able to solve

00:13:39.880 | this.

00:13:40.880 | And then at the very end, I'll talk about how these big ideas I'm going to these five ideas

00:13:43.360 | I'm going to present.

00:13:44.360 | We'll talk about how that's actually implemented on real computers, but we'll do that real

00:13:47.040 | fast.

00:13:48.360 | That's kind of a red herring.

00:13:50.440 | The neural networks and transformer blocks and multi headed attention.

00:13:53.560 | We'll get there, but we'll do that very fast.

00:13:55.120 | That's the big conceptual ideas that I care about.

00:13:57.680 | All right.

00:13:58.680 | I need idea number one about how these type of programs work is word guessing.

00:14:04.160 | Now, I got to warn you, this is very visual.

00:14:07.720 | Everything I'm talking about now is on the screen.

00:14:09.400 | So if you're a listener, I really would recommend going to YouTube dot com slash Cal Newport

00:14:13.520 | media and going to episode two forty four.

00:14:17.320 | And if you don't like YouTube, go to the deep life dot com and go to episode two forty four

00:14:20.840 | because it's very visual.

00:14:21.840 | What I'm going to do here.

00:14:22.840 | All right.

00:14:23.840 | So what we see on the screen here for idea number one is word guessing.

00:14:27.200 | And I have a green box on the screen that represents the LLM or large language model

00:14:32.180 | that would underpin a chat bot like chat GPT.

00:14:36.720 | So what happens is.

00:14:38.860 | If you put a incomplete bit of text into this box, an example here is I have the partial

00:14:45.480 | sentence fragment, the quick brown fox jumped.

00:14:50.340 | The whole goal of this large language model is to spit out what single next word should

00:14:56.560 | follow.

00:14:58.320 | So in this case, if we give it the quick brown fox jumped as input in our example, the language

00:15:03.560 | model has spit out over.

00:15:07.840 | This is the word I'm guessing should come next.

00:15:10.240 | All right.

00:15:11.500 | So then what we would do is add over to add the word to our sentence.

00:15:16.640 | So now our sentence reads the quick brown fox jumped over.

00:15:19.720 | So we've added the output.

00:15:20.960 | We've expanded our sentence by a single word.

00:15:24.140 | We run that into the large language model and it spits out a guess for what the next

00:15:27.080 | word should be.

00:15:28.080 | So in this case is the.

00:15:29.080 | And then we would now expand our sentence.

00:15:31.320 | The quick brown fox jumped over the.

00:15:33.720 | We put that as input into the model.

00:15:36.880 | It would spit out the next word.

00:15:39.800 | This approach, which is known as auto regressive text generation, is what the models underneath

00:15:47.320 | all of these new generation chatbots like chat GPT actually use.

00:15:51.520 | They guess one word at a time.

00:15:54.000 | That word is added to the text and the newly expanded text is put through the model to

00:15:58.040 | get the next word.

00:15:59.040 | So it just generates one word at a time.

00:16:01.360 | So if you type in a request to something like chat GPT, that request plus a special symbol

00:16:06.820 | that means, OK, this is the end of the request and where the answer begins, that is input.

00:16:12.120 | And it'll spit out the first word of its response.

00:16:14.960 | It'll then pass the request plus the first word of its response into the model to get

00:16:19.360 | the second word of the response.

00:16:21.040 | It'll then add that on and pass the request plus the first two words of its response into

00:16:24.400 | the model to get the third word.

00:16:25.480 | So this is generating text one word at a time.

00:16:30.360 | It just slowly grows what's generating.

00:16:33.060 | There's no recurrency in here.

00:16:34.960 | It doesn't remember anything about the last word it generated.

00:16:37.740 | Its definition doesn't change.

00:16:39.760 | The green box, my diagram never changes once it's trained.

00:16:43.680 | Text goes in, it spits out a guess for the next word to add to what's being generated.

00:16:47.920 | All right.

00:16:50.440 | Idea number two, relevant word matching.

00:16:53.600 | So how does it figure out how do these large language models figure out what word to spit

00:16:58.280 | out next?

00:16:59.280 | Well, at its core, what's really happening here, and I'm simplifying, but at its core,

00:17:04.640 | what's really happening here is the model is just looking at the most relevant words

00:17:11.320 | from the input.

00:17:13.200 | It is then going to match those relevant words to actual text that has been given.

00:17:20.360 | We call these source text in my article.

00:17:22.140 | So examples of real text, it can match the relevant words to where they show up in real

00:17:26.640 | text and say, what follows these relevant words in real text?

00:17:30.720 | And that's how it figures out what it wants to output.

00:17:33.840 | So in this example, the most relevant words are just the most recent words.

00:17:37.560 | So if the input into our box is the quick brown fox jumped over, perhaps the model is

00:17:43.960 | only going to look at the last three words, fox jumped over.

00:17:46.800 | Then it has this big collection over here to the side of real text that real humans

00:17:50.600 | wrote, all these examples.

00:17:53.080 | And it's going to look in there and it's going to say, okay, have we seen something like

00:17:57.040 | fox jumped over show up in one of our input texts?

00:17:59.600 | And okay, here's an input text that says, as the old saying goes, the quick brown fox

00:18:04.200 | jumped over the lazy brown dog.

00:18:05.720 | So it looks, fox jumped over.

00:18:07.240 | Here it is.

00:18:08.240 | We found it in one of the source texts.

00:18:09.240 | What came after the words fox jumped over?

00:18:11.360 | The great, let's make the what we guess.

00:18:14.640 | Now, of course, in a, in a real industrial strength, large language model, the relevant

00:18:20.760 | words aren't just necessarily the most recent words.

00:18:23.740 | There's a whole complicated system called self-attention in which this, the, the model

00:18:29.120 | will actually learn what type, which words to emphasize as the most relevant words.

00:18:33.200 | But that's too complicated for this discussion.

00:18:34.720 | The key thing is, is just looking at some words from the text, effectively finding similar

00:18:39.520 | words in real texts that it was provided and saying what happened in those real texts.

00:18:43.960 | And that's what it figures out, how it figures out what to produce next.

00:18:46.720 | All right.

00:18:47.720 | This brings us to idea number three, which is voting.

00:18:54.280 | So the way I just presented it before it was, you know, Hey, just start looking through

00:18:57.000 | your source text till you find the relevant words, see what follows output it.

00:19:00.560 | That's not actually what happens.

00:19:01.960 | We want to be a little bit more probabilistic.

00:19:04.880 | So what I would say a closer way of describing what happens is we can imagine that our large

00:19:10.520 | language model is going to look for every instance, every instance of the relevant words

00:19:19.240 | that we're looking for.

00:19:21.260 | And it's going to see what follows those instances and keep track of it.

00:19:24.400 | What are all the different words that follow in this example, fox jumped over.

00:19:30.280 | And every time it finds an example of fox jumped over, it says, what word follows next?

00:19:33.920 | Let's give a vote for that word.

00:19:36.200 | And so if the same word follows in most of the examples, it's going to get most of the

00:19:40.440 | votes.

00:19:41.440 | Now I'm using votes here sort of as a metaphor.

00:19:44.120 | What we're really doing here is trying to build a normalized probability distribution.

00:19:47.720 | So in the end, what we're going to get, what the, what the large language model is going

00:19:50.680 | to produce is for every possible next word, it is going to produce a probability.

00:19:59.180 | What is the probability that this should be the next word that follows?

00:20:01.880 | But again, you can just think about this as votes, which word received the most votes,

00:20:05.920 | which word received the second most votes, how many votes this word received compared

00:20:09.480 | to that word.

00:20:10.480 | And we're just going to normalize those is really what's happening.

00:20:13.240 | But you just think about it as votes.

00:20:15.820 | So in this example, we see the phrase, the quick brown fox jumped over the lazy brown

00:20:20.320 | dogs.

00:20:21.320 | I mean, that shows up in a bunch of different sources.

00:20:23.480 | So the word the gets a lot of votes.

00:20:25.000 | So it has sort of a high percentage here.

00:20:27.680 | But maybe there's similar phrases.

00:20:30.160 | Like look at this example here, the cat jumped over a surprised owner.

00:20:35.740 | Cat jumped over is not the same as fox jumped over because cat is different than fox.

00:20:41.200 | But in this voting scheme, we can say, you know what, cat jumped over is similar to what

00:20:44.600 | we're looking for, which is fox jumped over.

00:20:47.360 | So what word follows that?

00:20:48.920 | The word a, well, we'll give that a small vote.

00:20:52.880 | And so now what we're able to do is not only find every instance of the word, the relevant

00:20:56.960 | words that we're looking for and generate votes for what follows, we can also start

00:21:00.420 | generating weaker votes for similar phrases.

00:21:02.720 | And in the end, we just get this giant collection of every possible word and a pile of votes

00:21:08.520 | for each.

00:21:11.160 | And what the system will do is now randomly select a word.

00:21:15.120 | So that's why I have a picture.

00:21:17.640 | It's just, you would admit expertly drawn picture of a three dimensional dice.

00:21:22.600 | That's pretty good.

00:21:23.600 | Honest question.

00:21:24.600 | Did you know that was a dice before?

00:21:25.920 | Oh, yeah.

00:21:26.920 | Look at that guy's 3d rendering.

00:21:27.920 | Yeah.

00:21:28.920 | So for those who are listening, I have a picture of a dice to indicate randomness.

00:21:33.480 | It'll then randomly select which word to come next and it'll weigh that selection based

00:21:36.780 | on the votes.

00:21:38.640 | So if in this case, the has most of the votes, it almost certainly that's the, the word it's

00:21:42.800 | going to choose to output next, but look, a has some votes.

00:21:45.340 | So it's possible that it'll select a, it's just not as likely the word apple has zero

00:21:49.560 | votes, 0% probability because it never shows up after the phrase, anything similar to fox

00:21:54.280 | jumped over.

00:21:55.360 | No phrase similar to that is ever followed by apple.

00:21:56.920 | So there's no chance it'll select it and, and so on.

00:21:59.320 | And actually in these systems, the output is a vote or a percentage like this for every

00:22:04.360 | possible, they call them tokens, but word that could follow next.

00:22:08.200 | Here's a quiz, Jesse in the model on which chat GPT is based, how many different words

00:22:14.080 | do you think it knows?

00:22:15.080 | So in other words, when it, when it has to generate, okay, here's a pile of votes for

00:22:19.920 | every possible next word, how many words or punctuations are a billion?

00:22:25.000 | No, it's 50,000, 50,000.

00:22:28.440 | So it has a vocabulary of 50,000.

00:22:31.440 | It's not all words, but basically it knows like tens of thousands of words.

00:22:34.360 | And how big is like the biggest dictionary?

00:22:37.920 | That's a good question.

00:22:38.920 | Yeah.

00:22:39.920 | I don't think it's as, it probably, there's probably a lot of esoteric words.

00:22:43.040 | Yeah.

00:22:44.040 | Because the thing is it has some like vectors describing all these words, follow it along

00:22:49.440 | throughout the system.

00:22:50.440 | So it really does affect the size, how big, so it's like you want to have a big enough

00:22:53.200 | vocabulary to talk about a lot of things, but not so big that it really inflates the

00:22:57.480 | system.

00:22:58.480 | Right.

00:22:59.480 | So voting is just my euphemism for probability, but this is what's, what's happening.

00:23:02.280 | So now we have a bunch of source text and we imagine that for our relevant words, we're

00:23:08.720 | just finding all the places in these source texts where relevant words show up or similar

00:23:12.560 | relevant words show up and see what follows it.

00:23:15.440 | And in all cases, generate votes for what follows it, use those votes to select what

00:23:18.600 | comes next.

00:23:19.600 | All right.

00:23:20.600 | So this brings us to idea four, idea one through three can generate very believable text.

00:23:31.800 | This is well-known in natural language processing systems that do more or less what I just described.

00:23:37.740 | If you give it enough source text and have it look at a big enough window of relevant

00:23:44.380 | words and then just have it spit out word by word in the way we just described that

00:23:48.300 | auto regressive approach, this will spit out very believable text.

00:23:53.960 | It's actually not even that hard to implement.

00:23:55.560 | In my New Yorker article, I point towards a simple Python program I found online.

00:24:01.560 | It was a couple hundred lines of code that used Mary Shelley's Frankenstein as its input

00:24:08.200 | text.

00:24:09.200 | It looked at the last four words in the sentence being generated.

00:24:12.280 | That's what it uses to relevant words.

00:24:14.000 | And I showed in the article, this thing generated very good Gothic text.

00:24:18.700 | Right.

00:24:19.980 | So that's how you generate believable text with a program.

00:24:23.360 | And notice nothing we've done so far has anything to do with understanding the concepts the

00:24:28.900 | program is talking about.

00:24:31.720 | All of the intelligence, the grammar, the subtleties, all of that that we see so far

00:24:37.400 | is just being extracted from the human text that were pushed as input and then remixed

00:24:42.740 | and matched and copied and manipulated into the output.

00:24:45.140 | But the program is just looking for words, gathering votes, selecting, outputting blindly

00:24:51.340 | again and again and again.

00:24:52.340 | The program is actually simple.

00:24:54.180 | The intelligence you see in an answer is all coming at this point from the input text themselves.

00:24:58.180 | All right.

00:24:59.180 | But we've only solved half the problem.

00:25:01.300 | If we want a chat bot, we can't just have our program generate believable text.

00:25:06.300 | The text have to actually answer the question being asked by the user.

00:25:11.720 | So how do we aim this automatic text generation mechanism towards specific types of answers

00:25:17.520 | that match what the user is asking?

00:25:19.740 | Well, this brings in the notion of feature detection, which is the fourth out of the

00:25:25.420 | five total ideas I want to go over today.

00:25:28.300 | So what happens with feature detection is a response.

00:25:32.300 | We have a request and perhaps the answer that follows the request is being input into our

00:25:38.260 | large language model.

00:25:39.780 | So I've shown here a request that says write instructions for removing a peanut butter

00:25:45.100 | sandwich from a VCR.

00:25:46.100 | Then I have a bunch of colons and I have the beginning of a response.

00:25:50.280 | The first step is to write because everything gets pushed into the model.

00:25:54.620 | You get the whole original question and you get everything the model has said so far in

00:25:59.020 | its answer.

00:26:00.020 | Right.

00:26:01.020 | Word by word, we're going to grow the answer.

00:26:02.180 | But as we grow this answer, we want the full input, including the original question input

00:26:05.900 | into our models.

00:26:06.900 | That's what I'm showing here.

00:26:09.100 | Feature detection is going to look at this text and pattern match out features that it

00:26:15.420 | thinks are relevant for what text the model should be producing.

00:26:20.380 | So these yellow underlines here, instructions in VCR.

00:26:24.060 | So maybe that's one feature points out.

00:26:26.980 | It extracts from this text.

00:26:28.940 | These are supposed to be instructions about a VCR.

00:26:31.140 | And maybe this orange underline, another feature says the peanut butter sandwich is involved.

00:26:36.660 | And so now the model has extracted two features.

00:26:39.920 | These are VCR instructions we're supposed to be producing and they're supposed to involve

00:26:43.580 | a peanut butter sandwich.

00:26:46.360 | The way we then take advantage and by we, I mean the model, the way we take advantage

00:26:50.520 | of those features is that we have what I call in my article rules.

00:26:55.220 | I have to say, AI people don't like me using the word rules because it has another meaning

00:26:59.700 | in the context of expert decision systems.

00:27:01.880 | But just for our own colloquial purposes, we can call them rules that extract the each

00:27:07.420 | rule, think of it as an instructions for extracting features like a pattern matching instruction,

00:27:11.600 | and then a set of guidelines for how to change the voting strategy based on those particular

00:27:19.480 | features.

00:27:20.480 | So here's what I mean.

00:27:21.520 | Maybe there's a rule that looks for things like instructions in VCR and it figures out,

00:27:26.080 | okay, we're supposed to be doing instructions about a VCR.

00:27:29.600 | And its guidelines are then when looking to match the relevant words.

00:27:33.760 | And in this example, I have the, I'm saying the relevant words are step is to, so like

00:27:37.840 | just the end of the answer here.

00:27:40.060 | When looking to match step is to, when we find those relevant words, step is to showing

00:27:45.600 | up in a source text that is about VCR instructions, give extra strength to those votes.

00:27:54.080 | So here I have on the screen, maybe one of the input texts was VCR repair instructions.

00:27:58.160 | And it says when removing a jam tape, the first step is to open the tape slot.

00:28:03.700 | So we have step is to open.

00:28:06.000 | So open is a candidate for the next word to output here.

00:28:11.280 | Because this source document matches the feature of VCR instructions, our rule here that's

00:28:17.480 | triggered might say, hey, let's make our vote for open really strong.

00:28:22.400 | We know it's grammatically correct because it follows step is to.

00:28:26.340 | But we think it's also has a good chance of being semantically correct because it comes

00:28:30.720 | from a source that matches the type of things we're supposed to be writing about.

00:28:36.420 | So let's make ourselves more likely to do this.

00:28:40.260 | Now think about having now a huge number of these rules for many, many different types

00:28:46.240 | of things that people could ask about it.

00:28:47.880 | And for all of these different things, someone might ask your chat program about peanut butter

00:28:51.120 | sandwiches, VCR repair, Seinfeld scripts, the bubble sort algorithm for anything that

00:28:55.400 | someone might ask your chat program about.

00:28:58.000 | You have some rule that talks about what to look at in the source text to figure out it's

00:29:02.560 | relevant and very specific guidelines about how should we then change our votes for words

00:29:07.500 | that match source text, that match these properties, these complicated rules.

00:29:12.380 | If we have enough of these rules, then we can start to generate text that's not only

00:29:18.080 | natural sounding, but actually seems to reply to or match what is being requested by the

00:29:24.280 | user.

00:29:26.480 | Now I think the reason why people have a hard time grasping this step is they imagine how

00:29:32.220 | many rules them or them and a team of people could come up with.

00:29:36.880 | And they say, I could come up with a couple dozen.

00:29:39.980 | Maybe if I worked with a team for a couple of years, we could come up with like a thousand

00:29:42.560 | good rules.

00:29:43.560 | But these rules are complicated.

00:29:45.280 | Even a rule as simple as how do we know they're asking about VCR instructions and how do we

00:29:49.200 | figure out if a given text we're given is a VCR instruction text?

00:29:52.480 | I don't know.

00:29:53.480 | But I think about that and look at a lot of examples.

00:29:54.680 | And maybe if we worked really hard, we could produce a few hundred, maybe a thousand of

00:29:58.120 | these rules.

00:29:59.120 | And that's not going to be nearly enough.

00:30:01.080 | That's not going to cover nearly enough scenarios for all of the topics that the more than 1

00:30:06.280 | million users who've signed up for chat GPT, for example, all the topics they could ask

00:30:09.680 | about.

00:30:11.160 | It turns out that the number of rules you really need to be as adept as chat GPT just

00:30:18.240 | blows out of proportion, any scale, any human scale we can think of.

00:30:22.440 | I did a little bit of back of envelope math for my New Yorker article.

00:30:27.280 | If you took all of the parameters that define GPT-3, which is the large language model that

00:30:34.100 | chat GPT then refined and is based on.

00:30:36.480 | So the parameters we can think of as the things they actually change, actually train.

00:30:40.260 | So this is really like the description of all of its rules.

00:30:42.240 | If we just wrote out all of the numbers that define the GPT-3, we would fill over 1.5 million

00:30:52.240 | average length books.

00:30:55.040 | So the number of rules you would have to have if we were writing them out would fill a large

00:30:59.760 | university library full of rules.

00:31:03.260 | That scale is so big, we have a really hard time imagining it.

00:31:07.320 | And that's why when we start to see, oh my goodness, this thing can answer almost anything

00:31:12.560 | I send to it.

00:31:14.000 | It can answer almost any question I ask of it.

00:31:16.520 | We think there must be some adaptable intelligence in there that's just learning about things,

00:31:23.640 | trying to understand and interact with us because we couldn't imagine just having enough

00:31:27.160 | rote rules to handle every topic that we could ask.

00:31:30.080 | But there is a lot of rules.

00:31:31.080 | There's 1.5 million books full of rules inside this chat GPT.

00:31:34.400 | So you have to wrap your mind around that scale.

00:31:38.380 | And then you have to imagine that not only is that many rules, but we can apply them

00:31:41.200 | in all sorts of combinations.

00:31:43.000 | VCR instructions, but also about a peanut butter sandwich, also in the style of King

00:31:47.760 | James Bible, stack those three rules and we get that first example that we saw earlier

00:31:52.520 | on.

00:31:53.520 | All right.

00:31:54.520 | So then the final idea is how in the world are we going to come up with all those rules?

00:31:57.600 | 1.5 million books full of rules.

00:31:59.680 | How are we going to do that?

00:32:01.900 | And this is where self-training enters the picture.

00:32:06.580 | These language models train themselves.

00:32:09.600 | Here's the very basic way this works.

00:32:11.960 | Imagine we have this 1.5 million books full of rules and we start by just putting nonsense

00:32:16.960 | in every book.

00:32:19.160 | Nonsense rules, whatever they are.

00:32:20.160 | Right.

00:32:21.160 | So they don't do it.

00:32:22.160 | The system doesn't do anything useful right now, but at least we have a starting point.

00:32:26.000 | And now we tell the system, go train yourself and to help you train yourself, we're going

00:32:28.800 | to give you a lot of real text, text written by real humans.

00:32:31.480 | So when I say a lot, I mean a lot.

00:32:33.760 | The model on which chat GPT is based, for example, was given the results of crawling

00:32:39.420 | the public web for over 12 years.

00:32:41.960 | So a large percentage of anything ever written on the web over a decade was just part of

00:32:47.840 | the data that was given the chat GPT to train itself.

00:32:52.100 | And what the program does is it takes real text, little passages of real text out of

00:32:58.080 | this massive preposterously large data set.

00:33:02.480 | And it will use these passages one by one to make its rules better.

00:33:05.620 | So here's the example I have on the screen here.

00:33:08.160 | Let's say one of these many, many, many, many sample texts we gave chat GPT was Hamlet.

00:33:14.240 | And the program says, let's just grab some text from Hamlet.

00:33:18.280 | So let's say we're in Act 3 where we have the famous monologue to be or not to be, that

00:33:23.160 | is the question.

00:33:24.160 | What the program will do is just grab some of that text.

00:33:26.720 | So let's say it grabs to be or not to be.

00:33:29.720 | And then it's going to lop off the last word.

00:33:32.560 | So in this case, it lops off the word be.

00:33:35.260 | And it feeds what remains into the model.

00:33:37.640 | So when you lop off be here, you're left with to be or not to.

00:33:41.200 | It says, great, let's feed that into our model.

00:33:44.540 | We have 1.5 million books full of rules.

00:33:46.920 | They're all nonsense, because we're early in the training.

00:33:48.780 | But we'll go through each of those books and see which rules apply and let them modify our

00:33:51.900 | voting strategy.

00:33:52.900 | And we'll get this big vector of votes, then we'll randomly choose a word.

00:33:56.180 | And let's say in this case, the word is dog because it's not going to be a good word because

00:33:59.560 | the rules are really bad at first, but it'll spit out some words.

00:34:02.160 | Let's say it spits out dog.

00:34:04.920 | Now the good news is for the program, because it took this phrase from a real source, it

00:34:09.200 | knows what the next word is supposed to be.

00:34:12.040 | So on the screen here in orange, I'm showing it knows that be is what is supposed to follow

00:34:16.440 | to be or not to.

00:34:18.040 | So it can compare be to what it actually spit out.

00:34:20.120 | So the program spit out dog, it compares it to the right answer, the right answer is be.

00:34:25.360 | And here's the magic, it goes back and says, let me nudge my rules.

00:34:31.760 | There's a formal mathematical process it does to do this.

00:34:33.680 | But let me just go in there and just kind of tweak these rules.

00:34:36.640 | Not so the program accurately spits out be, but so it spits out something that is minutely

00:34:42.160 | more appropriate than dog, something that is just slightly better than the output it

00:34:48.240 | gave.

00:34:49.640 | So based on this one example, we've changed the rules a little bit, so that our output

00:34:53.680 | was just a teeny bit better.

00:34:55.800 | And it just repeats this again, and again, and again, hundreds of thousands of passages

00:35:01.040 | from Hamlet, and then from all the different Shakespeare works, and then on everything

00:35:04.080 | ever written in Wikipedia, and then on almost everything ever published on the web, bulletin

00:35:09.000 | board entries, sports websites, archived articles from old magazine websites, what just sentences

00:35:16.160 | of sentences, sentences, lop off a word, see what it spits out, compare it to the right

00:35:19.760 | answer, nudge the rules, take a new sentence, lop off the last word, stick in your model,

00:35:23.400 | see what it spits out, compare it to the real one, nudge the rules.

00:35:26.080 | And it does that again, and again, and again, hundreds of billions of times.

00:35:31.320 | There's one estimate I found online that said training chat GPT on a single processor would

00:35:37.320 | take over 350 years of compute time.

00:35:41.240 | And the only way that they could actually train on so much data so long was to have

00:35:46.040 | many, many processors working in parallel, spending well over a million dollars, I'm

00:35:50.320 | sure worth of compute time just to get this training done.

00:35:52.440 | And it still probably took weeks, if not months to actually complete that process.

00:35:55.840 | But here's the leap of faith I want you to make after this final idea.

00:35:59.440 | If you do this training, the simple training process on enough passages drawn from enough

00:36:05.000 | source text covering enough different types of topics from VCR instructions to Seinfeld

00:36:09.720 | scripts, these rules through all of these nudging, these 1.5 million books worth of

00:36:15.440 | rules will eventually become really, really smart.

00:36:18.840 | And it will eventually be way more comprehensive and nuanced than any one team of humans could

00:36:23.360 | ever produce.

00:36:24.920 | And they're going to recognize that this is a Bible verse.

00:36:26.840 | You want VCR instructions here and bubble sort is an algorithm.

00:36:29.920 | And this chapter from this textbook talks about bubble sort.

00:36:32.860 | And these are scripts and this is a script from Seinfeld.

00:36:35.500 | And actually, this part of the script for Seinfeld is a joke.

00:36:38.160 | So if we're in the middle of writing a joke in our output, then we want to really upvote

00:36:41.880 | words that are from jokes within Seinfeld scripts.

00:36:44.600 | All of these things we can imagine will be covered in these rule books.

00:36:47.320 | And I think the reason why we have a hard time imagining it being true is just because

00:36:49.960 | the scale is so preposterously large.

00:36:51.880 | We think about us filling up a book.

00:36:53.760 | We think about us coming up with two dozen rules.

00:36:55.880 | We have a hard time wrapping our mind around just the immensity of 1.5 million books worth

00:37:01.660 | of rules trained on 350 years worth of compute time.

00:37:06.260 | We just can't easily comprehend that scale.

00:37:10.020 | But it is so large that when you send what you think is this very clever request to chat

00:37:15.520 | GPT, it's like, oh, this rule, that rule, this rule, this rule.

00:37:18.800 | Boom, they apply.

00:37:20.680 | Modifier votes.

00:37:21.680 | Let's go.

00:37:22.680 | I think that amazes you.

00:37:24.880 | So those are the big ideas behind how chat GPT works.

00:37:27.040 | Now, I know all of the my fellow computer scientists out there with a background in

00:37:31.200 | artificial intelligence are probably yelling at your podcast headphones right now saying,

00:37:36.600 | well, that's not quite how it works, though.

00:37:37.880 | It's not it doesn't search for every word from the source text, and it doesn't have

00:37:43.440 | rules individually like that.

00:37:45.820 | It's instead in a much more complicated architecture.

00:37:48.160 | And this is all true.

00:37:49.160 | It's all true.

00:37:50.160 | I mean, the way that these models are actually architected or in something called a transformer

00:37:56.680 | block architecture, GPT three, for example, has ninety six transformer blocks arranged

00:38:01.440 | in layers one after another.

00:38:03.160 | Within each of these transformer blocks is a multi headed self attention layer that identifies

00:38:07.400 | what are the relevant words that this transformer block should care about.

00:38:10.880 | It then passes that on into a feed forward neural network.

00:38:14.360 | Is this neural networks that actually encode inside their weights, connecting their artificial

00:38:19.040 | neurons that actually encode in a sort of condensed, jumbled, mixed up manner, more

00:38:23.320 | or less the strategy I just described.

00:38:26.480 | So the feature detection that's built into the weights of these neural networks, the

00:38:32.400 | connection between certain features being identified, combined with certain relevant

00:38:36.680 | words, combined with vote strengths for what were should come next.

00:38:40.480 | All of that is trained into these networks during the training.

00:38:42.760 | So all the statistics and everything is trained into these as well.

00:38:46.840 | But in the end, what you get is a basically a jumbled, mixed up version of what I just

00:38:50.640 | explained.

00:38:52.600 | I sat down with some large language model experts when I was working on this article

00:38:55.960 | and said, let me just make sure I have this right.

00:38:58.180 | These high level five ideas.

00:39:00.760 | That's more or less what's being implemented in the artificial neural networks within the

00:39:04.040 | transformer block architecture.

00:39:05.040 | And they said, yeah, that's that's what's happening.

00:39:06.840 | It's again, it's mixed up, but that's what's happening.

00:39:10.300 | And so when you train the actual language model, you're not only training it to identify

00:39:15.280 | these features, you're baking in the statistics from the books and what happens with these

00:39:19.640 | for all that's getting baked into the big model itself.

00:39:21.760 | That's why these things are so large.

00:39:23.200 | That's why it takes 175 billion numbers to define all the rules for, let's say, GPT-3.

00:39:28.360 | But those five ideas I just gave you, that's more or less what's happening.

00:39:33.680 | And so this is what you have to believe is that with enough rules trained enough, what

00:39:37.960 | I just defined is going to generate really believable, impressive text.

00:39:41.300 | That's what's actually happening.

00:39:42.360 | Word guessing one word at a time.

00:39:45.140 | With enough rules to modify these votes and enough source text to draw from, you produce

00:39:50.260 | really believable text.

00:39:51.760 | All right, so if we know this, let us now briefly return to the second part of our deep

00:39:56.960 | question.

00:39:58.320 | How worried should we be?

00:40:01.540 | My opinion is once we have identified how these things actually work, our fear and concern

00:40:08.300 | is greatly tempered.

00:40:11.720 | So let's start with summarizing based on what I just said.

00:40:15.360 | What is it that these models like Chats GPT can actually do?

00:40:18.420 | Here's what they can actually do.

00:40:21.580 | They can respond to a question in arbitrary combination of known styles, talking about

00:40:31.000 | arbitrary combination of known subjects.

00:40:32.600 | So they can write about arbitrary numbers of known styles, talking about arbitrary combinations

00:40:38.040 | of known subjects.

00:40:39.040 | Known means it has seen enough of those things, enough of the style or enough writing about

00:40:43.200 | the topic in its training.

00:40:45.560 | That's what it can do.

00:40:47.080 | So say, write about this and this in this style.

00:40:51.920 | Bubble sort Seinfeld in a script.

00:40:55.320 | And it can do that and it can produce passable text if it's seen enough of those examples.

00:41:00.240 | And that's also all it can do.

00:41:01.720 | So let's start with the pragmatic question of is this going to take over our economy?

00:41:05.120 | And then we'll end with the bigger existential question.

00:41:07.200 | Is this an alien intelligence that's going to convert us into matrix batteries?

00:41:12.000 | So start with that.

00:41:13.000 | Is that capability I just described going to severely undermine the economy?

00:41:17.160 | And I don't think it is.

00:41:19.280 | I think where people get concerned about these Chats GPT type bots in the economy is they

00:41:25.240 | mistake the fluency with which it can combine styles and subjects with a adaptable fluid

00:41:30.280 | intelligence.

00:41:31.280 | Well, if it can do that, why can't it do other parts of my job?

00:41:33.900 | Why can't it handle my inbox for me?

00:41:36.740 | Why can't it build the computer program I need?

00:41:39.820 | You imagine that you need a human-like flexible intelligence to produce those type of texts

00:41:43.640 | that you see.

00:41:44.640 | And flexible human-like intelligence can do lots of things that are in our job.

00:41:47.480 | But it's not the case.

00:41:48.480 | There is no flexible human-like intelligence in there.

00:41:50.440 | There's just the ability to produce passable text with arbitrary combination of known styles

00:41:55.080 | on arbitrary combinations of known subjects.

00:41:57.160 | If we look at what most knowledge workers, for example, do in their job, that capability

00:42:01.240 | is not that useful.

00:42:03.120 | A lot of what knowledge workers do is not writing text.

00:42:07.100 | It is, for example, interacting with people or reading and synthesizing information.

00:42:11.600 | When knowledge workers do write, more often than not, the writing is incredibly narrow

00:42:16.520 | and bespoke.

00:42:17.520 | It is specific to the particular circumstances of who they work for, their job, and their

00:42:22.620 | history with their job, their history with the people they work for.

00:42:25.860 | I mentioned in my New Yorker piece that as I was writing the conclusion, that earlier

00:42:30.840 | that same day, I had to co-author an email to exactly the right person in our dean's

00:42:37.380 | office about a subtle request about how the hiring, faculty hiring process occurs at Georgetown,

00:42:44.520 | carefully couched, because I wasn't sure if this was the right person, carefully couched

00:42:47.820 | in language about, "I'm not sure if you're the right person for this, but here's why

00:42:50.360 | I think, and we talked about, this is why I'm asking about this.

00:42:53.920 | We had this conversation before."

00:42:56.440 | Writing in GPT, ChatGPT's broad training, could have helped it accomplish that narrow

00:43:01.440 | task on my behalf.

00:43:02.680 | And that's most of the writing that knowledge workers actually do.

00:43:07.000 | And even when we have relatively generic writing or coding or production of text that we need

00:43:12.360 | a system to do, we run into the problem that ChatGPT and similar chatbots are often wrong.

00:43:21.040 | Because again, they're just trying to make good guesses for words based on the styles

00:43:27.280 | and subjects you asked it about.

00:43:29.640 | These models have no actual model of the thing it's writing about.

00:43:33.720 | So they have no way of checking, does this make sense?

00:43:37.480 | Is this actually right?

00:43:39.820 | It just produces stuff in the style, like what is an answer supposed to more or less

00:43:43.000 | sound like if it's about this subject in this style?

00:43:46.220 | This is so pervasive that the developer bulletin board Stack Overflow had to put out a new

00:43:52.840 | rule that says no answers from ChatGPT can be used on this bulletin board.

00:43:57.840 | Because what was happening is ChatGPT would be happy to generate answers to your programmer

00:44:02.840 | questions.

00:44:03.840 | It sounded perfectly convincing.

00:44:05.300 | But as the moderator of the Stack Overflow board clarified, more often than not, they

00:44:09.640 | were also incorrect.

00:44:10.640 | Because ChatGPT doesn't know what a correct program is.

00:44:15.320 | It just knows I'm spitting out code.

00:44:18.720 | And based on other code I've seen and the features, this next command makes sense.

00:44:24.600 | And most of the commands make sense, but it doesn't know what sorting actually means.

00:44:29.560 | Or that there's a one off issue here.

00:44:33.560 | Or that equality isn't quite, you need equality, not just less than or whatever, right?

00:44:37.360 | Because it doesn't know sorting.

00:44:38.860 | It just says given the stuff I've spit out so far and the features I detected from what

00:44:41.720 | you asked me, and all this code I've looked at, here's a believable next thing to spit

00:44:44.880 | out.

00:44:45.880 | So it would spit out really believable programs that often didn't work.

00:44:48.480 | So we would assume most employers are not going to outsource jobs to an unrepentant

00:44:53.320 | fabulist.

00:44:54.320 | All right, so is it going to be not useful at all in the workplace?

00:44:57.800 | No, it will be useful.

00:44:58.800 | There'll be very bespoke things I think language models can do.

00:45:02.040 | It's particularly useful, what we've found in the last few months, where these technologies

00:45:06.600 | seems to be particularly useful is when you can give it text.

00:45:10.760 | They can do this too.

00:45:11.760 | You can give it text and say, rewrite this in this style or elaborate this.

00:45:15.960 | It's good at that.

00:45:16.960 | And that's useful.

00:45:18.200 | So if you're a doctor and you're typing in notes for electronic medical records, it might

00:45:21.320 | be nice that you can type them in sloppily.

00:45:23.760 | And a model like GPT-4, Chats GPT, might be able to take that and then transform those

00:45:29.980 | ideas into better English.

00:45:31.200 | It's the type of thing it can do.

00:45:33.600 | It can do other things like it can gather information for us and collate it in a way

00:45:37.680 | like a smart Google search.

00:45:39.160 | That's what Microsoft is doing when it's integrating this technology into its Bing search engine.

00:45:43.200 | It's a Google plus.

00:45:45.280 | So I mean, Google is already pretty smart, but you could have it do a little bit more

00:45:49.000 | actions.

00:45:50.000 | I mean, so there's going to be uses for this, but it is not going to come in and sweep away

00:45:54.740 | a whole swath of the economy.

00:45:57.040 | All right, let's get to the final deeper question here.

00:46:00.080 | Is this some sort of alien intelligence?

00:46:02.560 | Absolutely not.

00:46:03.720 | Once you understand the architecture, as I just defined it, there is no possible way

00:46:09.960 | that these large language model based programs can ever do anything that even approximates

00:46:14.480 | self-awareness consciousness or something we would have to be concerned about.

00:46:19.560 | There is a completely static definition for these programs once they're trained.

00:46:24.600 | The underlying parameters of GPT-3, once you train it up, for example, do not change as

00:46:30.660 | you start running requests through it.

00:46:32.420 | There is no malleable memory.

00:46:35.060 | It's the exact same rules.

00:46:36.320 | The only thing that changes is the input you give it.

00:46:38.360 | It goes all the way through these layers in a simple feed forward architecture and spits

00:46:42.600 | out a next word.

00:46:43.600 | And when you run it through again with a slightly longer request, it's the exact same layers,

00:46:47.920 | spits out another word.

00:46:49.400 | You cannot have anything that approaches consciousness or self-awareness without malleable memory.

00:46:54.160 | To be alive, by definition, you have to be able to have a ongoing updated model of yourself

00:47:00.080 | in the world around you.

00:47:02.100 | There's no such thing as a static entity where nothing can change.

00:47:08.800 | There's no memory that changes, nothing in it changes that you would consider to be alive.

00:47:13.520 | So no, this model is not the right type of AI technology that could ever become self-aware.

00:47:20.020 | There's other models in the AI universe that could be where you actually have notions of

00:47:24.320 | maintaining and updating models of learning, of thinking about yourself, interacting with

00:47:30.920 | the world, having incentives, having multiple actions you can take.

00:47:33.920 | You can build systems that in theory down the line could be self-aware.

00:47:38.240 | Large language model won't be it.

00:47:39.760 | Architecturally, it's impossible.

00:47:41.600 | All right, so that's where we are.

00:47:45.120 | We've created this really cool large language model.

00:47:49.160 | It's better than the ones that came before.

00:47:52.240 | It's really good at talking to people, so it's easy to use and you can share all these

00:47:55.280 | fun tweets about it.

00:47:57.240 | This general technology, one way or the other, will be integrated more and more into our

00:48:00.840 | working lives, but it's going to have the impact, in my opinion, more like Google had

00:48:04.740 | once that got really good, which was a big impact.

00:48:07.000 | You can ask Google all these questions, how to define words.

00:48:09.520 | It was very useful.

00:48:10.520 | It really helped people, but it didn't make whole industries disappear.

00:48:13.920 | I think that's where we're going to be with these large language models.

00:48:17.320 | They can produce text on arbitrary combinations of known subjects using arbitrary combination

00:48:21.760 | of known styles where known means they've seen it a sufficient number of times in their

00:48:25.520 | training.

00:48:26.520 | This is not a cow from 2001.

00:48:28.360 | This is not an alien intelligence that is going to, as was warned in that New York Times

00:48:33.480 | op-ed, deploy sophisticated propaganda to take over our political elections and create

00:48:39.580 | a one-world government.

00:48:41.860 | This is not going to get rid of programming as a profession and writing as a profession.

00:48:46.480 | It is cool, but it is not, in my opinion, an existential threat.

00:48:52.720 | What's transformative in the world of AI probably will not be in the immediate future transformative

00:48:57.440 | in your day-to-day life.

00:48:59.160 | All right?

00:49:00.160 | So, Jesse, there's my professor sermon for the day.

00:49:05.280 | >> JESSE: Solid.

00:49:06.280 | >> COREY: You don't want to get me started on computer science lectures because I could

00:49:09.320 | fall into my old habits, my professorial habits, and really bore you.

00:49:16.240 | >> JESSE: So how many rules will there be in five years?

00:49:18.640 | Will it double?

00:49:19.640 | >> COREY: I don't know how much bigger it can get.

00:49:21.520 | Yeah, it's a good question.

00:49:23.040 | So the jump from GPT-2 to GPT-3.

00:49:25.040 | So GPT-2 had some of the largest number of parameters before GPT-3 came out.

00:49:29.200 | It had 17 billion or something like this, and GPT-3 has 170 billion.

00:49:34.000 | I was talking to an expert at MIT about this.

00:49:36.920 | The issue about making this too much larger is they're already sort of giving it all of

00:49:43.040 | the text that exists.

00:49:45.880 | And so at some point, you're not going to get back bigger returns.

00:49:52.000 | So he said there's two issues with this.

00:49:53.360 | If you make your networks too small, they're not complicated enough to learn enough rules

00:49:57.060 | to be useful.

00:49:58.060 | But if you make them too large, you're wasting a lot of space.

00:50:00.000 | You're just going to have a lot of redundancy.

00:50:01.560 | I mean, it can only learn what it sees in its dataset.

00:50:06.140 | So if 175 billion parameters is well fit to this massive training data that we use for

00:50:12.400 | these chatbots, then just increasing the size of the network is not going to change much.

00:50:17.200 | You would have to have a correspondingly larger and richer training dataset to give it.

00:50:23.560 | And I don't know how much more, at least for this very particular problem of producing

00:50:28.200 | text, I don't know how much more richer or larger of a dataset we could give it.

00:50:31.880 | I actually think what the direction happening now is how do we make these things smaller

00:50:36.680 | again?

00:50:37.680 | GPT-3 is too big to be practical.

00:50:40.280 | 175 billion parameters can't fit in the memory of a single GPU.

00:50:43.840 | You probably need five different specialized pieces of hardware just to generate a single

00:50:47.280 | word.

00:50:48.280 | That's not practical.

00:50:49.280 | That means I can't do that on my computer.

00:50:51.440 | That means if everyone at my office is constantly making requests to GPT-3 as part of their

00:50:55.920 | work, we're going to have this huge computing bill.

00:50:58.340 | So actually a lot of the effort is in how do we make these things smaller?

00:51:01.600 | Just focus on the examples that are relevant to what these people actually need.

00:51:04.800 | We want them to be small.

00:51:05.840 | We want it eventually to have models that can fit in a phone and still do useful things.

00:51:10.880 | So GPT-3 I think was, and that's what all these other ones are based off of, that was

00:51:15.440 | open AI saying what happens if we make these things much bigger?

00:51:20.120 | And now we're going to go back to make them smaller.

00:51:21.400 | And if you actually read the original GPT-3 paper, their goal with making it 10 times

00:51:27.160 | bigger was not that it was going to have in a particular domain 10 times better answers.

00:51:32.480 | They wanted to have one model that could do well in many unrelated tasks.

00:51:37.060 | And if you read the paper, they say, look, here's a bunch of different tasks for which

00:51:41.660 | we already have these large language models that do well, but they're customized to these

00:51:46.980 | tasks.

00:51:47.980 | They can only do that one task.

00:51:49.720 | And what they were proud about with GPT-3, if you read the original paper, is this one

00:51:53.160 | model can do well on all 10 of these tasks.

00:51:56.840 | It's not that it was actually doing much better than the state of the art in any one of these

00:51:59.400 | things.

00:52:00.400 | You don't need necessarily to hand train a model for each task.

00:52:04.000 | If you make it big enough, it can handle all the different tasks.

00:52:06.120 | So it wasn't getting 10 times larger did not make GPT-3 10 times better at any particular

00:52:11.440 | task.

00:52:12.440 | In fact, in most tasks, it's as good as the best, but not much better.

00:52:15.560 | It was the flexibility and the broadness, but that's good to see.

00:52:20.760 | It's cool for these web demos.

00:52:23.720 | But going forward, the name of the game, I think is going to go back to, actually, we

00:52:26.800 | need to make these things smaller so that we can not have to use an absurd amount of

00:52:32.120 | computational power just to figure out that dog should follow the quick brown fox jumped

00:52:37.320 | over the lazy brown.

00:52:39.240 | We need to maybe be a little bit more efficient.

00:52:41.440 | But anyways, I'm not particularly...

00:52:45.080 | It's a cool technology, but I don't know.

00:52:47.960 | I think once you open this, it's just not as worrisome.

00:52:50.760 | When it's a black box, you can imagine anything.

00:52:54.480 | I definitely like that Yuval Harari op-ed was definitely influenced by Nick Bostrom's

00:52:58.840 | super intelligence, which we talked about on the show a few months ago, where he just

00:53:02.600 | starts speculating.

00:53:03.600 | He's a philosopher, not a computer scientist.

00:53:05.600 | Bostrom just starts speculating.

00:53:07.880 | What if it got this smart?

00:53:09.440 | What could it do?

00:53:10.440 | Well, what if it got this smart?

00:53:11.440 | What could it do?

00:53:12.440 | Just thinking through scenarios about...

00:53:13.440 | And he was like, "Well, if it got smart, it could make itself smarter, and then it

00:53:16.120 | can make itself even smarter, and it become a super intelligence."

00:53:18.360 | And then they have all these scenarios about, "Well, if we had a super intelligent thing,

00:53:20.960 | it could take over all world politics.

00:53:22.440 | Because it'd be so smart and understand us so well that it could have the perfect propaganda.

00:53:27.640 | And now the bot could get us all to do whatever it wanted us to do."

00:53:31.120 | It's all just philosophical speculation.

00:53:33.920 | You open up these boxes, and you see 175 billion numbers being multiplied by GPUs doing 1.5

00:53:41.560 | million books worth of pattern detection vote rules to generate a probability vector so

00:53:45.840 | it can select a word.

00:53:46.840 | All right.

00:53:47.840 | Well, there's my computer science sermon.

00:53:51.080 | I have a few questions I want to get to from you, my listeners, that are about artificial

00:53:55.400 | intelligence.

00:53:56.400 | First, I want to mention one of the sponsors that makes this nonsense possible.

00:53:59.040 | And we're talking about our friends at ZocDoc.

00:54:01.800 | ZocDoc is the only free app that lets you find and book doctors who are patient reviewed,

00:54:08.160 | take your insurance, and are available when you need them to treat almost every condition

00:54:13.000 | under the sun.

00:54:15.840 | So if you need a doctor, instead of just saying, "I'll ask a friend," or, "I'll look it up

00:54:20.160 | in the yellow pages," you can instead go to ZocDoc, your ZocDoc app, and it'll show you

00:54:26.600 | doctors who are nearby that take your insurance.

00:54:29.120 | You can read reviews.

00:54:30.640 | And then once you sign up with these doctors, they'll often use ZocDoc to make it easier

00:54:33.840 | for you to set up appointments, get reminders about appointments, send in paperwork.

00:54:39.400 | Both my dentist and my primary care physician use ZocDoc, and I find it really useful because

00:54:44.880 | all of my interactions with them happen to that interface.

00:54:47.800 | So go to ZocDoc.com/deep and download the ZocDoc app for free.

00:54:52.120 | Then find a book, a top rated doctor today, many who are available within 24 hours.

00:54:56.520 | That's Z-O-C-D-O-C.com/deep, ZocDoc.com/deep.

00:55:03.360 | The show is also sponsored by BetterHelp.

00:55:08.000 | As we often talk about when it comes to the different buckets relevant to cultivating

00:55:13.920 | a deep life, the bucket of contemplation is in there.

00:55:19.440 | Having an active and healthy life of the mind is critical to a life that is deep in many

00:55:25.960 | different ways.

00:55:26.960 | It is easy, however, due to various events in life to fall into an area where your relationship

00:55:31.940 | to your mind gets strained.

00:55:33.520 | Maybe you find yourself overwhelmed with anxious thoughts or ruminations or depressive moments

00:55:39.680 | where you feel lack of affect.

00:55:42.040 | Whatever it is, it is easy for your mind to get out of kilter.

00:55:45.560 | Just like if your knee started hurting, you would go to an orthopedist.

00:55:50.120 | If your mind started hurting, and by the way, there's really loud construction sound going

00:55:53.120 | on.

00:55:56.320 | This is the restaurant below us is being made.

00:55:58.520 | It's better be worth it.

00:56:00.120 | But returning to BetterHelp, let me add this example.

00:56:05.420 | If you find yourself becoming increasingly enraged because of restaurant construction

00:56:10.880 | noise that occurs during your podcast, among the other things that could happen to affect

00:56:15.040 | your mental health, you need a therapist.

00:56:18.040 | Orthopedists will fix your bum knee.

00:56:19.800 | Therapists helps make sure that your cognitive life gets healthy, gets resilient, gets back

00:56:23.880 | on track.

00:56:24.880 | The problem is it's hard to find therapists.

00:56:25.880 | If you live in a big city, all the ones near you might be booked or they might be really

00:56:30.040 | expensive.

00:56:31.040 | This is where BetterHelp enters the scene.

00:56:35.120 | If you're thinking about starting therapy, BetterHelp is a great way to give it a try

00:56:39.280 | because it's entirely online.

00:56:41.440 | It's designed to be convenient, flexible, and suited to your schedule.

00:56:44.640 | You just fill out a brief questionnaire to get matched with a licensed therapist, and

00:56:49.040 | you can switch therapists anytime for no additional charge.

00:56:53.920 | So discover your potential with BetterHelp.

00:56:57.120 | Visit BetterHelp.com/deepquestions, one word, today to get 10% off your first month.

00:57:03.680 | That's BetterHelp, H-E-L-P.com/deepquestions.

00:57:10.760 | I hope this restaurant's good, Jesse, after all the disruption.

00:57:14.680 | They put up the signage.

00:57:15.680 | I don't know if you saw that.

00:57:16.680 | I didn't see the signage.

00:57:17.680 | I've heard the music.

00:57:18.680 | Motocat.

00:57:19.680 | That's the name?

00:57:20.680 | Yeah.

00:57:21.680 | It's a Tacoma Park reference.

00:57:22.680 | Okay.

00:57:23.680 | Yeah.

00:57:24.680 | It's an old character from Tacoma Park history.

00:57:27.360 | Anyways, soon it will be open.

00:57:30.320 | I heard I was talking to the guy, he thought May.

00:57:33.000 | Yeah.

00:57:34.000 | So soon.

00:57:35.000 | Nice.

00:57:36.000 | All right, let's do some questions.

00:57:37.000 | What do we got?

00:57:38.000 | All right.

00:57:39.000 | First question is from Manav, a student at Yale.

00:57:43.200 | Looking at tools like ChatGPT makes me feel like there's nothing AI won't eventually do

00:57:47.000 | better than humans.

00:57:48.880 | This fear makes it hard to concentrate on learning since it makes me feel that there

00:57:52.360 | isn't certainty in my future.

00:57:55.560 | Are my fears unfounded?

00:57:56.960 | Well, Manav, hopefully my deep dive is helping dispel those fears.

00:58:03.560 | I want to include this question in part to emphasize the degree to which the hype cycle

00:58:08.200 | around these tools has really been unhelpful.

00:58:12.120 | Because you can so easily embed screenshots of interactions of ChatGPT, a lot of people

00:58:19.000 | started trying it because the attraction of virality is very strong for lots of people

00:58:23.720 | in our online age.

00:58:24.880 | So it brought ChatGPT to the awareness of a lot of people and generated a lot of attention.

00:58:30.160 | Now once we had a lot of attention, how do you one up that attention?

00:58:33.000 | Well, then you start thinking about worries about it.

00:58:35.120 | You start thinking about what if it could do this, what if it could do that?

00:58:38.440 | From what I understand, I'm not as plugged into these online worlds as others, but there's

00:58:42.240 | a whole tech bro push during the last few months that was increasingly trying to push.

00:58:46.560 | It can do this, it can do that, it can do this.

00:58:49.000 | Exactly the same tonality with which the same group talked about crypto two years ago.

00:58:53.440 | Nothing is going to be done by ChatGPT, just like currency will be gone in three years

00:58:56.880 | because of crypto.

00:58:57.880 | They turned all their attention onto that.

00:58:59.960 | And that got really furious.

00:59:02.240 | And everyone's trying to one up each other and do YouTube videos and podcasts about,

00:59:05.280 | "No, it could do this, no, it can do that."

00:59:06.840 | And then this created this weird counter reaction from the mainstream media, because the mainstream

00:59:12.760 | media has a, right now, an adversarial relationship with the Silicon Valley tech bro crowd.

00:59:18.080 | They don't like them.

00:59:19.520 | So then they started pushing back about, "No, it's going to be bad.

00:59:22.720 | No, no, these tech bros are leading us to a world where we're going to be able to cheat

00:59:26.920 | on tests.

00:59:27.920 | No, forget cheat on tests.

00:59:28.920 | It's going to take all of our jobs.

00:59:29.920 | No, forget take all of our jobs.

00:59:31.520 | It's going to take over the government and become super intelligent."

00:59:35.080 | So they started the counteraction to the overblown enthusiasm of the tech bros became an overblown

00:59:40.880 | grimness from the anti-tech bro mainstream media.

00:59:45.960 | All of it fed into Twitter, which like a blender was mixing this all together and swirling

00:59:50.240 | this spiral of craziness higher and higher until finally just the average person like

00:59:54.400 | Manav here at Yale is thinking, "How can I even study knowing that there will be no

00:59:59.400 | jobs and we'll be enslaved by computers within the next couple of years?"

01:00:03.400 | All right, so Manav, hopefully my deep dive helped you feel better about this.

01:00:08.280 | Chats GPT can write about combinations of known subjects and combinations of known styles.

01:00:13.360 | It does not have models of these objects.

01:00:16.920 | It has no state or understanding or incentives.

01:00:20.160 | You ask it the right about removing a peanut butter sandwich from a VCR.

01:00:23.160 | It does not have an internal model of a VCR in a sandwich on which it's experimenting

01:00:26.520 | with different strategies to figure out which strategy works best and then turns to its

01:00:29.800 | language facility to explain that to you.

01:00:32.760 | It just sees peanut butter as a possible next word to spit out and you ask about peanut

01:00:37.080 | butter and your response, so it puts more votes on it.

01:00:39.440 | Mixing, matching, copying, manipulating existing human text, the humor and the jokes it spits

01:00:44.800 | out, the accuracy and the styles it uses are all intelligence borrowed from the input that

01:00:50.680 | it was given.

01:00:51.680 | It does not have a broad adaptable intelligence that can in any significant sense impact the

01:00:57.960 | knowledge work sector.

01:01:00.960 | It's important to emphasize, Manav, it's not like we're one small step from making these

01:01:04.880 | models more flexible, more adaptable, able to do more things.

01:01:11.280 | The key to chat CPT being so good at the specific thing it does, which is producing text in

01:01:15.600 | known styles on known subjects, is that it had a truly massive amount of training data

01:01:19.560 | on which it could train itself.

01:01:20.640 | We could give it everything anyone had ever written on the internet for a decade and it

01:01:24.240 | could use all of that to train itself.

01:01:26.240 | This is the problem when you try to adapt these models to other types of activities

01:01:30.440 | that are not just producing text.

01:01:33.160 | You say, "What I really want is a model that can work with my databases.

01:01:38.560 | What I really want is a model that can send emails and attach files on my behalf."

01:01:44.320 | The problem is you don't have enough training data.

01:01:47.600 | You need training data where you have billions of examples of, "Here's the situation, here's

01:01:52.360 | the right answer."

01:01:53.360 | And then like most things that we do, we learn after a small number of examples.

01:01:57.920 | A model to do other activities other than produce text needs a ton of data.

01:02:01.240 | And in most other types of genres or activities, there's just not that much data.

01:02:04.640 | So one of the few examples where there is, is art production.

01:02:07.160 | This is how Dali works.

01:02:09.240 | You can give it a huge corpus of pictures that are annotated with text and it can learn

01:02:14.760 | these different styles and subject matters that show up in pictures then produce original

01:02:17.920 | artwork.

01:02:18.920 | But that's one of the few other areas where you have enough uniform data that it can actually

01:02:23.120 | train itself to be super adaptable.

01:02:24.880 | So I'm not worried that Manav, like all of our jobs will be gone.

01:02:28.040 | So your fears are unfounded.

01:02:29.120 | You can rest easy, study harder for your classes.

01:02:32.560 | All right, let's keep it rolling.

01:02:34.640 | What do we got, Jesse?

01:02:35.640 | - All right, next question's from Aiden.

01:02:37.800 | It seems almost inevitable that in 10 years, AI will be able to perform many knowledge

01:02:41.640 | workers jobs as well as a human.

01:02:44.400 | Should we be worried about the pace of automation and knowledge work and how can we prepare

01:02:48.440 | our careers now for increased power AI in the coming decades?

01:02:52.980 | - So as I just explained in the last question, this particular trajectory of AI technology

01:02:57.680 | is not about to take all of your jobs.

01:03:00.020 | There is however, and this is why I included this question.

01:03:03.260 | There is however, another potential intersection of artificial intelligence and knowledge work

01:03:07.640 | that I've been talking about for years that I think we should be more concerned about

01:03:14.080 | or at least keep a closer eye on.

01:03:17.480 | The place where I think AI is gonna have the big impact is less sexy than this notion of

01:03:22.880 | I just have this blinking chat cursor and I can ask this thing to do whatever I want.

01:03:26.520 | Now where it's really gonna intersect is shallow task automation.

01:03:32.680 | So the shallow work, the stuff we do, the overhead we do to help collaborate, organize,

01:03:37.800 | and gather the information need for the main deep work that we execute in our knowledge

01:03:41.360 | work jobs.

01:03:43.240 | More and more of that is gonna be taken over by less sexy, more bespoke, but increasingly

01:03:47.280 | more effective AI tools.

01:03:49.880 | And as these tools get better, I don't have to send 126 emails a day anymore because I

01:03:55.640 | can actually have a bespoke AI agent handle a lot of that work for me, not in a general

01:04:00.440 | it's intelligent sense, but in a much more specific like talking to Alexa type sense.

01:04:05.820 | Can you gather the data I need for writing this report?

01:04:10.680 | Can you set up a meeting next week for me with these three principles?

01:04:15.260 | And then that AI agent talks to the AI agents of the three people you need to set the meeting

01:04:19.840 | up with and they figure out together and put that meeting onto the calendar so that none

01:04:24.860 | of us, three of us have to ever exchange an email.

01:04:27.040 | The information it gathers from the people who have it by talking to their AI agents,

01:04:31.320 | and I never have to bother them.

01:04:32.600 | We never have to set up a meeting.

01:04:34.520 | It's able to do these rote tasks for us, right?

01:04:38.820 | This was actually a future that I was exposed to a decade earlier.

01:04:42.040 | I spoke at an event with the CEO of a automated meeting scheduling company called X.AI.

01:04:47.640 | I remember him telling me, this is the future.

01:04:51.240 | When you have a AI tool that can talk to another person's AI tool to figure out logistical

01:04:55.880 | things on your behalf so that you're never interrupted.

01:04:59.000 | I think that's where the big AI impact is going to come.

01:05:02.320 | Now, this does not automate your main work.

01:05:04.800 | What it does is it automates away the stuff that gets in the way of your main work.

01:05:08.040 | Why is that significant?

01:05:09.160 | Because it will immensely increase the amount of your main work you're able to get done.

01:05:14.520 | If you're not context switching once every five minutes, which is the average time the

01:05:18.200 | average knowledge worker spends between email or instant messenger chat checks, if you're

01:05:21.720 | not doing that anymore, you know how much you're going to get done?

01:05:26.160 | You know how much if you can just do something until you're done, and then the AI agents

01:05:29.760 | on your computer says, "Okay, we got the stuff for you for the next thing you need

01:05:32.280 | to work on.

01:05:33.280 | Here you go."

01:05:34.280 | And you have to have no overhead communicating or collaborating, trying to figure out what

01:05:36.680 | to do next.

01:05:37.680 | You can just execute.

01:05:38.680 | You know how much you're going to get done?

01:05:39.800 | I would say probably three to four X more of the meaningful output that you produce

01:05:43.680 | in your job will be produced three to four X more if these unsexy bespoke AI logistical

01:05:50.640 | automator tools get better.

01:05:53.640 | So this has this huge potential benefit and this huge potential downside.

01:05:57.200 | The benefit of course is your work is less exhausting.

01:05:59.760 | You can get a lot more done.

01:06:01.200 | Companies are going to generate a lot more value.

01:06:03.000 | The downside is that might greatly reduce the number of knowledge workers required to

01:06:09.060 | meet certain production outputs.

01:06:12.040 | If three knowledge workers now produce what it used to take 10, I could grow my company

01:06:18.040 | or I could fire seven of those knowledge workers.

01:06:20.440 | So I think this is going to create a disruption.

01:06:23.080 | We underestimate the degree to which shallow work and context shifting is completely hampering

01:06:28.280 | our ability to do work with our minds.

01:06:30.880 | But because it's like the pot that keeps getting hotter until the lobster is boiled, because

01:06:35.400 | it's inflecting everybody, we don't realize how much we're being held back.

01:06:39.160 | When computer tools aided by AI remove that, it's going to be a huge disruption.

01:06:43.480 | And I think ultimately the economy is going to adapt to it.

01:06:48.420 | The knowledge industry is going to explode in size and scope as we can unlock all this

01:06:52.320 | cognitive potential on new types of challenges or problems that we weren't thinking about

01:06:55.720 | before.

01:06:56.720 | Ultimately, it'll probably be good and lead to a huge economic growth.

01:06:59.780 | But I think there's going to be a disruption period because we really are at such, again,

01:07:05.840 | we just don't emphasize the degree to how inefficient we are and how much if we could

01:07:09.460 | remove that inefficiency, we don't need most of the people sitting here in this office

01:07:12.660 | to get the same work done.

01:07:14.360 | Getting over that, that's going to be the real disruption.

01:07:16.880 | And there's no scary how from 2001 type tool involved here.

01:07:19.840 | These things are going to be boring.

01:07:21.840 | Meeting, information scheduling, basically whatever you type in an email, it could do

01:07:26.360 | that for you.

01:07:28.400 | That's going to be the real disruption.

01:07:29.400 | I don't know when it's coming, but that's coming soon.

01:07:33.040 | A lot of money at stake.

01:07:34.360 | All right, this is good.

01:07:35.920 | I'm like debunking people's fears.

01:07:37.600 | I got a therapist today.

01:07:39.920 | All right, next question is from Ben, a Silicon Valley engineer.

01:07:44.640 | I've decided that web development freelancing will be the best possible career path to achieve

01:07:49.200 | my family's lifestyle vision.

01:07:50.840 | And I plan to freelance in addition to my full-time job until freelancing can support

01:07:55.020 | our life on its own.

01:07:57.200 | Over the last few weeks, however, I've been hearing about the breakthroughs of chat, GPT

01:08:01.560 | and other AI tools.

01:08:02.560 | Do you think I should stay on the path of learning the ropes of freelancing web development

01:08:06.600 | or should I focus more on the future of technology and try to stay ahead of the curve?

01:08:11.560 | Well, Ben, I'm using the inclusion of AI in your question to secretly get in an example

01:08:18.000 | of lifestyle-centric career planning, which you know I like to talk about.

01:08:22.080 | So I love your approach.

01:08:24.080 | You're working backwards from a vision of what you want you and your family's life to

01:08:27.760 | be like, a tangible lifestyle, not specific in what particular job or city, but the attributes

01:08:35.880 | of the lifestyle.

01:08:36.880 | So you're working backwards to say, what is a tractable path from where I am to accomplish

01:08:41.840 | that.

01:08:43.680 | And you're seeing here, web development could be there, freelance web development.

01:08:47.040 | And I don't know all the details of your plan, but I'm assuming you probably have a plan

01:08:50.200 | where you're living somewhere that is cheaper to live.

01:08:52.840 | Maybe it's more outside or country oriented where your expenses are lower.

01:08:57.200 | And because web development is a relatively high reward per hour spent type activity,

01:09:02.280 | that strategic freelancing could support your lifestyle there while giving you huge amounts

01:09:05.760 | of autonomy.

01:09:06.760 | And satisfying the various properties that I assume you figured out about what you want

01:09:10.520 | in your life, these sort of non-professional properties.

01:09:12.440 | So I really applaud this thinking.

01:09:13.920 | I also really applaud the fact that you're applying the principle of letting money be

01:09:18.760 | a neutral indicator of value.

01:09:21.640 | There's a strategy I talked about in my book, So Good They Can't Ignore You.

01:09:25.960 | This is a strategy in which instead of just jumping into something new, you try it on

01:09:30.800 | the side and say, can I actually make money at this?

01:09:34.300 | The idea here is that people paying you money is the most unbiased feedback you will get

01:09:39.400 | about how valuable or viable the thing you're doing actually is.

01:09:42.960 | And so I really like this idea.

01:09:44.640 | Let me try freelancing on the side.

01:09:46.680 | I mean, I want to see that people are hiring me and this money is coming in before I quit

01:09:50.320 | my job.

01:09:51.320 | It's a great way of actually assessing.

01:09:52.800 | Don't just ask people, hey, is this valuable?

01:09:54.440 | Do you think you would hire me?

01:09:56.080 | Look at dollars coming in.

01:09:57.080 | When the dollars coming in are enough to more or less support your new lifestyle, you then

01:10:02.560 | make that transition with confidence.

01:10:03.640 | So I appreciate that as well.

01:10:05.080 | Two Cal Newport ideas being deployed here.

01:10:08.560 | So let's get to your AI question.

01:10:10.640 | Should you stop this plan so you can focus more on the future of technology and try to

01:10:15.440 | stay ahead of the curve?

01:10:16.440 | I mean, I don't even really know what that means.

01:10:18.840 | My main advice would be whatever skill it is you're learning, make sure you're learning

01:10:22.100 | a course at the cutting edge of it.

01:10:24.500 | Get real feedback from real people in this field about what is actually valuable and

01:10:28.140 | how good you have to be to unlock that value.

01:10:30.280 | So I would say that don't invent your own story about how you want this field to work.

01:10:36.560 | Don't assume that if you know HTML and a little CSS, you're going to be living easy.

01:10:41.480 | What are the actual skills people care about?

01:10:43.160 | What web development technologies sell?

01:10:45.220 | How hard are they?

01:10:46.320 | How good you have to be at that to actually be desirable to the marketplace?

01:10:49.800 | Get hard, verified answers to those questions.

01:10:52.580 | That's what I would say when it comes to staying ahead of the curve.

01:10:54.720 | That's it.

01:10:55.800 | And as for some sort of fear that you becoming a web developer is quixotic because chat GPT

01:11:02.800 | is going to take that job away soon.

01:11:04.280 | Don't worry about that.

01:11:05.760 | So yes, make sure you're learning the right skills, not the skills you want to be valuable.

01:11:10.120 | But there's nothing wrong with this particular plan you have.

01:11:12.280 | I love the way you're executing it.

01:11:13.960 | I love the way you're thinking about it.

01:11:15.720 | I also appreciate, I cut it, Jesse, but Ben had a joke in the beginning of his response.

01:11:22.880 | He said, "I've been doing lifestyle-centric career planning.

01:11:26.360 | I've been thinking about it.

01:11:27.360 | I don't like my job.

01:11:28.360 | So what we're going to do is quit and I'm going to start a cattle ranch."

01:11:30.760 | And he was like, "Ah, just joking."

01:11:33.200 | I appreciate that.

01:11:34.560 | All right, let's do one more question before I kind of run out of the ability to talk about

01:11:41.760 | AI anymore.

01:11:42.760 | I'm getting, just, we're purging this.

01:11:44.040 | It's been months that people have been asking us about AI.

01:11:46.080 | I'm just purging it all out there.

01:11:48.720 | Next week, we're going to talk all about, I don't know, living in the country or minimalism,

01:11:56.000 | not using social media, but we're getting all the AI out of our system today.

01:11:58.840 | All right, here we go.

01:12:00.960 | Last question is from Anakin.

01:12:02.520 | "AI can solve high school level math word problems.

01:12:06.080 | AI can explain why a joke that has never seen before is funny.

01:12:09.920 | This one blows my mind.

01:12:11.400 | All this points to mass job loss within five years.

01:12:14.200 | What do you think?"

01:12:15.760 | Well, again, big thumbs down.

01:12:18.360 | The current trajectory of AI is not going to create mass job loss in the next five years.

01:12:24.000 | Chat GPT doesn't know why a joke is funny.

01:12:26.760 | It writes jokes that are funny because it knows what part of a script you can identify

01:12:31.340 | as a joke that's a pattern-mashing problem.

01:12:33.880 | And then it upvotes words from those parts of scripts when it's doing the word guessing

01:12:38.000 | game.

01:12:39.000 | And as a result, you get jokes that pull from existing structures of human and existing

01:12:43.680 | text.

01:12:44.680 | You don't actually know what humor is.

01:12:46.080 | You can see that in part if you look at that Seinfeld scene with bubble sort I talked about

01:12:50.480 | at the beginning of the program.

01:12:53.720 | There's non-sequitur jokes in there, things that are described as the audience laughing

01:12:57.800 | that aren't actually funny.

01:12:58.800 | And that's because it's not actually looking at its script and saying, "Is this funny?"

01:13:02.840 | It's guessing words, guessing words that things are accurate.

01:13:07.720 | But let's talk about...

01:13:08.720 | I want to use this as an excuse to talk about another trend in AI that I think is probably

01:13:13.980 | more important than any of these large language models that also is not getting talked about

01:13:18.520 | enough.

01:13:19.520 | So we talked about an earlier question, AI shallow work automation has been critical.

01:13:22.760 | The other major development that we're so used to now we forget, but I think is actually

01:13:27.840 | going to be the key to this AI shallow work automation, but also all sorts of other ways

01:13:33.160 | that AI interests our life is not these large language models, but it's what Google has

01:13:37.440 | in mind with Google Home.

01:13:39.620 | It's what Amazon has in mind with Alexa.

01:13:44.580 | These at-home appliances that you can talk to and ask to do things, they're ubiquitous

01:13:50.760 | and they're ubiquitous in part because these companies made them very cheap.

01:13:54.320 | They wanted people to use them.

01:13:55.320 | They're not trying to make a lot of money off them.

01:13:57.080 | Why?

01:13:58.080 | Why is Google or Amazon trying to get as many people as possible to use these agents at

01:14:01.680 | home that you could just talk to and it tries to understand you?

01:14:04.360 | It's data.

01:14:05.880 | The game they are playing is we want millions and millions of different people with different

01:14:09.760 | accents and different types of voices asking about all sorts of different things that we

01:14:13.960 | then try to understand.

01:14:16.320 | And we could then use this data set to train increasingly better interfaces that can understand

01:14:21.160 | natural language.

01:14:22.940 | That's the whole ballgame.

01:14:23.940 | Now, chat GPT is pretty good at this.

01:14:25.880 | They figured out a new model.

01:14:26.880 | I don't want to get into the weeds.

01:14:29.120 | They have a human semi-supervised, semi-human supervised reinforcement learning model that

01:14:33.760 | they inserted during the GPT-3 training to try to align its responses better with what's

01:14:38.760 | being asked.

01:14:40.180 | But this is the whole ballgame is just natural language understanding.

01:14:44.740 | And Google is working on this and Amazon is working on this and Apple is working on this

01:14:47.840 | with their Siri devices.

01:14:49.960 | And this is what matters, understanding people.

01:14:54.080 | The background activities, I think this is what we often get wrong.

01:14:58.240 | The actual activities that the disruptive AI in the future is going to do on our behalf

01:15:02.560 | are not that interesting.

01:15:04.360 | It's not, we're going to go write an ARIA.

01:15:06.640 | It's we're going to pull some data from an Excel table and email it to someone.

01:15:11.920 | It's we're going to turn off the lights in your house because you said you were gone.

01:15:17.080 | It's really kind of boring stuff.

01:15:19.080 | All of the interesting stuff is understanding what you're saying.

01:15:21.880 | And that's why Google and Amazon and Apple invested so much money into getting these

01:15:25.080 | devices everywhere is they want to train and generate the biggest possible data set of

01:15:29.280 | actually understanding what real people are saying and figuring out, did we get it right

01:15:32.280 | or did we get it wrong?

01:15:33.280 | And can we look at examples and let's hand annotate these examples and figure out how

01:15:37.680 | our models work.

01:15:38.680 | And I think this is really going to be the slow creep of AI disruption.

01:15:42.040 | It's not going to be this one entity that suddenly takes over everyone's job.

01:15:45.560 | It's going to be that more and more devices in our world are increasingly better at understanding

01:15:49.280 | natural language questions, whether it be typed or spoken, and can then act accordingly,

01:15:54.640 | even if what we're asking to do is simple.

01:15:56.360 | We don't really need these agents to do really complicated things.

01:15:58.720 | We just need them to understand what we mean.

01:16:02.040 | Most of what we do, that's a drag on our time, it's a drag on our energy, it's pretty simple.

01:16:06.640 | And it's something a machine could do if they just knew what it was we wanted.

01:16:10.480 | And so I think that's where we should be focusing is interfacing is what matters.

01:16:15.640 | These 175 billion parameter models that can generate all this text is really not that

01:16:21.080 | interesting.

01:16:22.080 | I don't need a Seinfeld script about bubble sort.

01:16:26.080 | I need you to understand me when I say, give me the email address of all of the students

01:16:31.280 | from my first section of my class.

01:16:33.560 | I need you just to understand what that means and be able to interface with the student

01:16:38.400 | database and get those emails out and format it properly so I can send the message to the

01:16:42.280 | class.

01:16:43.280 | That's what I want you to do.

01:16:44.840 | I don't want you to write an original poem in the style of Mary Oliver about a hockey

01:16:53.280 | game for my students.

01:16:55.080 | I need you to just go through when I say, look at the assignment pages for the problem

01:17:00.200 | sets I assigned this semester, pull out the grading statistics and put them all into one

01:17:05.800 | document just to kind of, okay, I know what that means.

01:17:08.080 | And now I'm doing a pretty rote automatic computer type thing.

01:17:11.640 | And I don't care if you come back and say, okay, I have three follow-up questions.

01:17:14.920 | So I understand what you mean.

01:17:15.920 | You mean this section, that section, this section, that's fine.

01:17:17.480 | They'll just take us another 10 or 15 seconds.

01:17:19.800 | I don't need, in other words, how from 2001, I just need my basic computer to understand

01:17:26.520 | what I'm saying.

01:17:27.520 | So I don't have to type it in or click on a lot of different buttons.

01:17:29.440 | So I mean, I think that's really where we're going to see the big innovations is the slow

01:17:32.800 | creep of better and better human understanding plugged into relatively non-interesting actions.

01:17:39.000 | That's really where the stuff's going to take a bigger and bigger picture.

01:17:41.200 | The disruption is going to be more subtle.

01:17:43.680 | This idea that it's now this all at once large language model represents, we have this new

01:17:49.280 | self-sufficient being that all at once we'll do everything.

01:17:52.800 | It's sexy, but I don't think that's the way this is going to happen.

01:17:55.840 | All right.

01:17:56.840 | So let's change gears here, Jesse.

01:17:57.840 | I want to do something interesting to wrap up the show.

01:18:01.240 | First, I want to mention one of the longtime sponsors that makes Deep Questions possible.

01:18:05.680 | That's our friends at Blinkist.

01:18:06.680 | When you subscribe to Blinkist, you get access to 15 minute summaries of over 5,500 nonfiction

01:18:14.320 | books and podcasts.

01:18:15.960 | These short summaries, which are called Blinks, you can either read them or you can listen

01:18:19.080 | to them while you do other activities.

01:18:22.520 | Jesse and I both use Blinkist as one of our primary tools for triaging which books we

01:18:28.800 | read.

01:18:29.800 | If there is a book we are interested in, we will read or listen to the Blink to see if

01:18:33.620 | it's worth buying.

01:18:36.760 | The other week, Jesse and I went through his own Blinkist list and we calculated that basically

01:18:41.360 | what it was roughly what 30%, I think you said 30% of the books for which you read Blinkist

01:18:45.680 | if you go on the buy.

01:18:47.000 | But there we go.

01:18:48.240 | That is a case in point of the value of Blinkist.

01:18:51.520 | It really helps you hone in on the right books to buy.

01:18:54.560 | The books you don't buy, you still know their main ideas.

01:18:56.520 | You can deploy them.

01:18:58.080 | Blinkist is really a tool that any serious reader needs to add to their tool kit.

01:19:04.680 | Right now, Blinkist has a special offer just for our audience.

01:19:06.880 | Go to Blinkist.com/deep to start your seven-day free trial and you will get 45% off a Blinkist

01:19:12.160 | premium membership.

01:19:14.080 | That's Blinkist spelled B-L-I-N-K-I-S-T.

01:19:17.040 | Blinkist.com/deep to get 45% off any seven-day free trial.

01:19:21.440 | That's Blinkist.com/deep.

01:19:23.400 | This offer is good through April 30th only.

01:19:27.480 | Even better news, for a limited time, you can take advantage of their Blinkist Connect

01:19:33.040 | program to share your premium account.

01:19:36.080 | You will, in other words, get two premium accounts for the price of one.

01:19:38.960 | So when you subscribe to Blinkist, you can give an account to a friend who you think

01:19:43.000 | would also appreciate it.

01:19:44.800 | So Blinkist.com/deep to find out more.

01:19:48.520 | I also want to talk about Latter.

01:19:51.480 | It's tax season.

01:19:53.360 | Tax season is when I often get stressed about putting things off because I don't even know

01:19:57.560 | where to get started.

01:19:59.040 | This got me thinking about the other classic thing that people put off, which is getting

01:20:04.360 | life insurance.

01:20:05.360 | This is one of those things that you know you need, but you don't know where to get

01:20:09.480 | started.

01:20:10.480 | Who do you talk to for life insurance?

01:20:11.960 | How much should life insurance cost?

01:20:14.000 | You're going to have to go to a doctor and get blood drawn and your eyeball scanned before

01:20:18.880 | you can actually get a policy.

01:20:20.920 | And we get paralyzed by all this complexity and say, forget it.

01:20:24.000 | Well, this is where Latter enters the scene.

01:20:29.160 | Latter makes it easy to get life insurance.

01:20:32.240 | It's 100% digital, no doctors, no needles, and no paperwork.

01:20:35.560 | When you're applying for $3 million in coverage or less, you just answer a few questions about

01:20:39.320 | your health on an application.

01:20:41.760 | You need just a few minutes in a phone or laptop to apply.

01:20:44.800 | Latter smart algorithms work in real time, so you'll find out instantly if you're approved.

01:20:50.600 | No hidden fees, cancel any time, get a full refund if you change your mind in the first

01:20:54.720 | 30 days.

01:20:56.720 | Life insurance is only going to cost more as you get older, so now is always the right

01:21:00.200 | time to get it.

01:21:02.320 | So go to ladderlife.com/deep today to see if you're instantly approved.

01:21:08.760 | That's LADDERLIFE.COM/DEEP.

01:21:14.400 | Ladderlife.com/deep.

01:21:15.400 | All right, Jesse, let's do something interesting.

01:21:20.500 | For those who are new to the show, this is the segment where I take something that you

01:21:24.400 | sent me to my interesting@calnewport.com email address that you thought I might find interesting.

01:21:31.360 | I take something that caught my attention and I share it.

01:21:35.460 | So here's something a lot of people sent me.

01:21:37.180 | This is actually from just a couple of days ago.

01:21:39.520 | I have it up on the screen now, so if you're listening, you can watch this at youtube.com/calnewportmedia

01:21:45.300 | episode 244.

01:21:46.540 | It is an article from NPR with the following music to my ears headline, NPR quits Twitter.

01:22:00.820 | There's a whole backstory to this.

01:22:02.860 | NPR is having a basically a feud with Twitter because Twitter started labeling the network

01:22:08.620 | as state affiliated media.

01:22:12.120 | The same description they use for propaganda outlets in Russia, China, and other autocratic

01:22:17.660 | countries that did not sit well with NPR.

01:22:20.880 | So then they changed it and said, okay, well, we'll call you government funded media.

01:22:24.520 | But NPR said only 1% of our annual budget comes from federal funding.

01:22:28.800 | That's not really accurate either.

01:22:30.540 | You know what?

01:22:31.540 | That's not enough.

01:22:32.540 | They're walking away.

01:22:33.540 | And they put NPR politics, for example, put out a tweet that said all of our 52 Twitter

01:22:39.140 | accounts, uh, we're not going to use them anymore.

01:22:42.380 | You want news for NPR, subscribe to our email newsletter, come to our website, listen to

01:22:47.540 | our radio program.

01:22:48.740 | We'll keep you up to date.

01:22:50.500 | You don't have to use this other guy's program.

01:22:53.700 | I really like to see that not because of the inter-scene political battles between Musk

01:22:59.780 | and the political, the different media outlets.

01:23:03.180 | I mean, I wish they would just make this move even without that, but whatever gets them

01:23:06.220 | there, I think is good.

01:23:08.580 | As I've talked about so many times on this show before, it is baked into the architecture

01:23:12.580 | of Twitter that it is going to generate outrage.

01:23:16.580 | It is going to manipulate your emotions.

01:23:19.300 | It is going to create tribalism and is going to really color your understanding of the

01:23:26.500 | world, your understanding of current events in ways that are highly inaccurate and often

01:23:30.980 | highly inflammatory.

01:23:32.260 | It's just built into the collaborative curation mechanism of all these retweets combined with

01:23:39.700 | the power law network.

01:23:40.700 | We've talked about this before.

01:23:42.620 | It's not a great way to consumer share information.

01:23:47.340 | Now more and more outlets are doing this.

01:23:49.020 | So a couple of weeks ago, we gave the example of the Washington post nationals baseball

01:23:54.580 | team coverage shifting away from live tweeting the games and instead having live updates

01:23:58.980 | on their Washington post.com website.

01:24:01.420 | And at the end of all those updates, then they write a recap article and it's all packaged

01:24:04.660 | together and you can see how it unfolded and they have more different people writing and

01:24:08.260 | it unfolds in real time.

01:24:09.260 | And I think the whole thing is great.

01:24:10.740 | There's no reason to be on someone else's platform, mixing tweets about the baseball

01:24:14.580 | game with tweets about, you know, the Martian that's going to come and destroy the earth

01:24:22.260 | because you didn't give it hydroxychloroquine or whatever the hell else is going on on Twitter.

01:24:27.540 | Twitter was a lot of fun.

01:24:28.540 | It's a lot of engaging.

01:24:29.540 | It's not very engaging.

01:24:30.740 | It's not the right way to consume news.

01:24:32.540 | It's not the right way to spread news.

01:24:34.120 | And I'm really happy about this trend.

01:24:35.780 | I think we would be in a much calmer place as individuals.

01:24:39.620 | I think we'd be in a much calmer place politically.

01:24:41.940 | I think we'd be in a much calmer place just from a civic perspective.

01:24:46.260 | If more and more sources of news, if more and more sources of expression, if more and

01:24:50.220 | more sources of commentary move to their own private gardens, here's my website, here's

01:24:55.700 | my podcast, here's my newsletter, not this giant mixing bowl where everyone and everything

01:25:02.180 | comes together in a homogenized interface.

01:25:05.940 | And we have this distributed curation mechanism rapidly amplifying some things versus others.

01:25:10.740 | That's not a healthy way for a society to learn about things and communicate.

01:25:17.620 | And I wrote about this in Wired Magazine early in the pandemic.

01:25:21.300 | I wrote an op-ed for Wired.

01:25:22.300 | I said one of the number one things we could do for public health right now at the beginning

01:25:26.860 | of the pandemic would be shut down Twitter.

01:25:29.380 | And I gave this argument in that article that, look, if professionals, if we retreated to

01:25:35.500 | websites, we could have considered long form articles with rich links in the other types

01:25:42.340 | of articles and sources, where the websites themselves, you could indicate authority by

01:25:50.020 | the fact that this website is hosted at a hospital or a known academic institution or

01:25:54.340 | a known news organization.

01:25:56.340 | We'd be able to curate on an individual sense of this website is at, the reference in old

01:26:03.500 | SNL skit, clownpenis.fart, and it has weird gifs of it of eagles that are flying away

01:26:09.900 | with Osama bin Laden.

01:26:10.900 | I'm not going to pay as much attention to that, to this long form article that's coming

01:26:15.340 | out of the Cleveland Clinic.

01:26:16.960 | And I said, if we went back, humans will be much better at consuming information.

01:26:20.620 | The information will be much better presented if we went back to more bespoke distributed

01:26:23.700 | communication, as opposed to having everyone, the cranks and the experts and the weirdos

01:26:29.320 | and everybody all mixed together in the exact same interface with the exact same, their

01:26:32.580 | tweets look exactly the same, and a completely dispassionate distributed curation mechanism,

01:26:37.660 | rapidly amplifying things that are catching people's attention.

01:26:41.420 | We need to get away from that automation.

01:26:42.760 | We need to get away from that distributed curation.

01:26:44.340 | And we get back to more bespoke things where we can learn from where you hosted, what does

01:26:48.760 | this look like?

01:26:49.760 | What are the texts?

01:26:50.760 | What are the things that have you written?

01:26:52.140 | We can really have context for information.

01:26:53.820 | Anyways, I don't want to go too far into this lecture, but basically this is a good trend.

01:26:57.820 | I think individually owned digital distribution of information, podcasts, websites, blogs,

01:27:06.940 | and news is going to be a much healthier trend than saying, why don't we all just have one

01:27:10.860 | or two services that everyone uses?

01:27:12.340 | So good job NPR.

01:27:15.160 | I hope more and more news sources follow your lead in the Washington Post's, Nationals,

01:27:19.100 | Reporters Leagues.

01:27:20.100 | I think this is the right direction.

01:27:21.100 | Do you know the clown penis dot fart reference?

01:27:23.820 | No.

01:27:24.820 | It was in the late nineties.

01:27:26.180 | It was a classic SNL skit and it was like an advertisement and they really had it.

01:27:32.500 | You know how they used to have those advertisements for brokerage firms?

01:27:35.180 | Where it's like super, I welcome, it's like Wilford Brimley, like welcome to such and

01:27:40.700 | such financial partners where trust is our number one, whatever.

01:27:44.540 | And so it's like this real serious ad and they're like, you can find out more at our

01:27:49.740 | website clownpenis.fart.

01:27:52.300 | And then they kind of reveal by the time we set up our website, it was the only address

01:27:57.860 | left.

01:27:58.860 | So it was like the premise of it, it was like this really serious financial brokerage firm,

01:28:02.100 | but in the late nineties, it felt like all the URLs were gone.

01:28:05.540 | And so it was like the only URL that was left.

01:28:07.140 | And so it was just a very serious commercial that they kept having to say clown penis dot

01:28:10.900 | fart, classic SNL.

01:28:13.740 | All right.

01:28:15.900 | I'm tired.

01:28:16.900 | Too much AI.

01:28:19.380 | I'm happy now to talk about AI whenever you want.

01:28:21.400 | My gag order has been lifted, but maybe it'll be a while until we get too much more deeply

01:28:25.460 | into that.

01:28:26.460 | But thank you all for listening.

01:28:27.460 | Thank you all for putting up with that.

01:28:28.460 | We'll be back next week with the next, hopefully much less computer science filled episode

01:28:32.660 | of the podcast.

01:28:33.660 | And until then, as always, stay deep.

01:28:36.300 | [MUSIC PLAYING]

01:28:39.960 | (techno music)

Ep. 244: Cal Newport’s Thoughts on ChatGPT

Chapters