00:00SPEAKER- You mean to tell me they made a soldering iron smart?
00:03SPEAKER- Okay, when you say it that way, it sounds stupid.
00:07SPEAKERBut yeah, like I updated the firmware
00:09SPEAKERon my soldering iron last night.
00:11SPEAKER- I want out, I'm done.
00:14SPEAKERNo, I'm sorry, actually, I'm being extremely flip here.
00:18SPEAKERI could totally see how this would actually be very cool.
00:20SPEAKER- It's so freaking cool.
00:21SPEAKERWell, okay, so it's nice to have
00:23SPEAKERa temperature controlled soldering iron.
00:25SPEAKERLike for the first 40 years of my life,
00:28SPEAKERI basically had the $10 Radio Shack one that you plug in,
00:32SPEAKERit's always the same temperature,
00:33SPEAKERand it scorches some stuff,
00:34SPEAKERand it's not hot enough for other stuff,
00:35SPEAKERand there's no precision,
00:37SPEAKERand it has a big fat chunky point, which is usually fine.
00:40SPEAKERAnd then I got that Hakko,
00:42SPEAKERwhich we talked about on the soldering episode years ago,
00:46SPEAKERwhich is like a temperature controlled,
00:48SPEAKERlike it's a nice analog soldering station.
00:50SPEAKERMaybe it's a digital, I don't know, it's a soldering station.
00:52SPEAKERYou can control the temperature.
00:53SPEAKER- Yeah, if people haven't heard the patron spotlight
00:56SPEAKERwe did with Twinkle Twinkie on--
00:57SPEAKER- Yeah, that's it.
00:58SPEAKER- Largely soldering, not all soldering,
00:59SPEAKERbut a lot of soldering.
01:00SPEAKERThat's a great resource to go back to.
01:04SPEAKER- So then a few years ago, this company called Pine64
01:07SPEAKERthat makes like open hardware that's collaborative,
01:09SPEAKERand I think we should try to have them on the FOSPOD maybe,
01:12SPEAKERit seems like a good fit.
01:14SPEAKER- Are they the Pinebook people?
01:16SPEAKER- They make the Pinebook.
01:17SPEAKERThey make a bunch of other like kind of hardware widgets
01:20SPEAKERfor softwares, but a lot of their stuff lives in,
01:22SPEAKERbut the Pinebook is a modular laptop.
01:25SPEAKER- Okay, I'm sorry, I'm on their front page.
01:28SPEAKERThe image carousel just scrolled a meetup by.
01:31SPEAKERI counted at least a couple of gray beards in there.
01:33SPEAKER- There, look, I'm gonna go and tell you, it is, yeah.
01:37SPEAKEROkay, so this thing is like a pencil-sized soldering iron
01:41SPEAKERthat's powered off of USB-C.
01:43SPEAKER- Pine-soul-sized, yes, okay.
01:44SPEAKER- Yeah, oh God, that's probably why it's named that.
01:47SPEAKERI didn't realize it.
01:48SPEAKERAnyway, you plug a USB-C port in the back
01:50SPEAKERand then you plug that into a battery pack
01:53SPEAKERor your USB port on your computer or USB charger
01:55SPEAKERor whatever and it lets you do temperature control.
01:59SPEAKERIt has stuff like, like it goes to sleep
02:02SPEAKERif you lay it flat on the table and it doesn't move
02:04SPEAKER'cause it has an accelerometer in it.
02:06SPEAKERIt is so freaking cool.
02:07SPEAKER- Ooh, that alone, 'cause I still have the $10
02:11SPEAKERRadio Shack soldering iron from 10 years ago
02:13SPEAKERand boy, do you have to be careful putting that thing down.
02:15SPEAKERI can see something that just prevents careless accidents
02:19SPEAKERbeing pretty nice anyway.
02:20SPEAKER- Well, it's still gonna be really hot.
02:22SPEAKERSo if you put it down the wrong way,
02:23SPEAKERit's gonna do bad things to your desk,
02:25SPEAKERbut it won't burn forever is the nice thing.
02:28SPEAKERSo yeah, anyway, it's really freaking cool, man.
02:31SPEAKERAnd it was like $30.
02:32SPEAKERI was super stoked.
02:34SPEAKER- Dude, I'm on the product page.
02:37SPEAKERI don't know that I've ever read
02:39SPEAKERa more TechPod product tagline.
02:42SPEAKERA RISC-V soldering iron running an open firmware.
02:46SPEAKER- Yeah, right?
02:47SPEAKER- And then I gather, I think it just runs
02:50SPEAKERlike user software, like anybody can write software for it.
02:53SPEAKER- You can write plugins for it.
02:55SPEAKERIt has like a little screen and a couple of buttons.
02:58SPEAKERYou can change settings.
02:59SPEAKERIt has a little bit of flash memory inside.
03:03SPEAKERIt's not much, it's not big.
03:04SPEAKERIt's like 100, 200 kilobytes or something.
03:09SPEAKERBut it runs Iron OS, which I thought was kind of cool.
03:12SPEAKER- Okay.
03:13SPEAKER- And it fits standard tips, standard software.
03:20SPEAKERIt's standard B2 soldering iron tips.
03:23SPEAKERSo you can just snap them in there.
03:24SPEAKERIt's really cool.
03:25SPEAKERThere's a short tip and a long tip.
03:27SPEAKER- Apparently the PineSol is not even the only soldering iron
03:29SPEAKERthat Iron OS runs on.
03:32SPEAKER- I mean, look, it's part of the Iron OS ecosystem,
03:37SPEAKERwhich is the most metal of software.
03:39SPEAKER- I'm seeing like, yeah, seriously.
03:42SPEAKERI'm seeing like six different products on this GitHub repo
03:45SPEAKERthat this thing runs on.
03:47SPEAKER- Yeah, but PineSol also, the Pine64 people also do
03:50SPEAKERlike single board computers and a bunch of other stuff.
03:53SPEAKERIt's like, they're kind of,
03:55SPEAKERI feel like they maybe are one of the ones
03:57SPEAKERthat predated Raspberry Pi, but they do,
03:59SPEAKERthey build in batches rather than having continuous supply,
04:01SPEAKERwhich kind of continuous supply changes the dynamic
04:04SPEAKERa little bit.
04:05SPEAKER- Yeah, and honestly their product line
04:07SPEAKERis like amazingly diverse for how kind of grassroots
04:09SPEAKERand scrappy this whole thing looks.
04:10SPEAKERI mean, they make a phone, they make a smartwatch.
04:12SPEAKER- Yeah, I kind of want to try the smartwatch.
04:13SPEAKERThe smartwatch looks kind of cool.
04:15SPEAKER- This is some wild stuff.
04:16SPEAKERLike, I mean, okay, as we will see in the episode to come,
04:20SPEAKERlike maybe the future isn't always great,
04:22SPEAKERbut sometimes the future is kind of cool.
04:24SPEAKER(upbeat music)
04:46SPEAKER(upbeat music continues)
04:53SPEAKER- Welcome to Brad and Will Made a Tech Pod, I'm Will.
04:56SPEAKER- I'm Brad, hello.
04:57SPEAKER- Hello, hi, hello.
04:59SPEAKER- I kind of just want to keep talking
05:00SPEAKERabout your smart soldering iron,
05:01SPEAKERbut we have a topic here, but man, that's,
05:05SPEAKERI need to just close this pine64.org tab right now.
05:09SPEAKER- I haven't spent enough time.
05:10SPEAKERI just got the soldering iron like two days ago.
05:13SPEAKERSo I've used it.
05:14SPEAKERLike I just did a test, a couple of the wires together
05:16SPEAKERto test it out, but yeah, I'm super stoked about it.
05:21SPEAKERI can't wait.
05:21SPEAKERI need a project now.
05:23SPEAKER- Is it not widely available?
05:24SPEAKERI don't see a price on it on here.
05:27SPEAKERAnd when I go to, when I click go to store,
05:28SPEAKERI just see a bunch of accessories for it, but not it.
05:31SPEAKER- I got it on Amazon for 30 or $40.
05:33SPEAKER- Oh, that's it?
05:34SPEAKER- Yeah.
05:35SPEAKER- Dude, what?
05:36SPEAKERMaybe I might just order one right now.
05:37SPEAKER- I mean, I'm going to go and tell you,
05:39SPEAKERlike having a good soldering iron,
05:40SPEAKERif you have not ever had one before is a life-changing thing.
05:45SPEAKERSo I paid, it looks like I paid a $5 premium
05:49SPEAKERover their list price.
05:50SPEAKERThey listed it as $25 on the thing, but it's out of stock.
05:53SPEAKERSo they built, it's a batch thing.
05:55SPEAKER- Got it.
05:56SPEAKER- Like they, I probably almost certainly paid a premium
05:59SPEAKERfor it because they, somebody stocked some and held them.
06:02SPEAKERSo.
06:03SPEAKER- That's cool.
06:04SPEAKERLast question and then we will move on.
06:05SPEAKERDoes running it off USB-C limit the temperatures
06:07SPEAKERit can reach?
06:09SPEAKER- Yeah, but it works with USB PD so it can get high.
06:14SPEAKER- Okay, quite a bit.
06:14SPEAKER- Like you're not going to have soldering problems
06:17SPEAKERbecause of the lack of power.
06:18SPEAKER- Sure.
06:19SPEAKER- So if you, like I plugged it into my Anker,
06:22SPEAKERa USB, what is it?
06:25SPEAKERLike a, I don't know, PD-1 or whatever,
06:28SPEAKERthe one that does up to,
06:30SPEAKERI plugged into a brick that'll charge a switch
06:32SPEAKERor my laptop and it worked.
06:33SPEAKER- Okay.
06:34SPEAKERI've got this MacBook Pro charger, USB-C,
06:36SPEAKERso here on this desk that is like 90 something watts,
06:39SPEAKERI think it'll probably be fine.
06:40SPEAKER- You're going to be okay.
06:41SPEAKER- Yeah.
06:42SPEAKER- Yeah, no problems.
06:43SPEAKERBrent, what are we talking about this week?
06:44SPEAKERWhat do we, what do we, what do we got in the hopper?
06:46SPEAKER- I'm going to be okay running a soldering iron
06:47SPEAKERoff this power brick.
06:48SPEAKERI don't know if we as a society are going to be okay
06:51SPEAKERwith the rise of AI.
06:54SPEAKER- As long as the power bricks don't come to life.
06:58SPEAKER- I think anything is on the table.
07:00SPEAKERThat might be the theme of this episode is that it kind of
07:03SPEAKERall of a sudden feels like anything might be on the table.
07:06SPEAKER- Yeah, it's, it's funny because like we've lived through
07:09SPEAKERa couple of computing revolution, revolutions at this point,
07:12SPEAKERbut for whatever reason, this one seems.
07:15SPEAKER- I'm sorry, can I stop you?
07:16SPEAKER- Yeah.
07:17SPEAKER- Can I guess what you're talking about?
07:19SPEAKERIs, is the integrated circuit the microprocessor
07:21SPEAKERone of them or are we too young for that?
07:23SPEAKER- We're too young for that.
07:24SPEAKER- Okay, okay, the internet has to be one of them.
07:26SPEAKER- The rise of the internet and generalized computing, yeah.
07:29SPEAKER- Are those the two?
07:30SPEAKER- No, no, that's the same one.
07:31SPEAKERLike, like basically thinking, if you think about it,
07:34SPEAKERthe internet is the thing that made everybody
07:36SPEAKERwant to get a computer.
07:36SPEAKER- Oh yeah, of course.
07:37SPEAKER- Right, so we have that and then also the,
07:39SPEAKERthen that becoming mobile and sticking in our pockets
07:42SPEAKERis the other one.
07:42SPEAKER- Okay.
07:43SPEAKER- That's what I was, that was my thought at least.
07:45SPEAKER- Okay, yeah, those, those tracks, those check out.
07:48SPEAKERAnyway, yes.
07:51SPEAKER- So we, I don't know, I don't know if we're here to argue
07:53SPEAKERor on the verge of another sea change, but it kind of,
07:55SPEAKERI think we might be here to suggest that it's possible.
07:58SPEAKER- So if you had told me a year ago that we would be taking
08:00SPEAKERthis seriously, I would probably have said,
08:01SPEAKEReh, I don't think so.
08:02SPEAKER- Sure.
08:03SPEAKER- And, and the things that have happened in,
08:07SPEAKERwhat I'm going to say is the AI field, which is not like,
08:13SPEAKERokay, there's some language problems in this whole space
08:16SPEAKERthat are basically like AI doesn't mean anything anymore
08:19SPEAKER'cause people use it for 50 million different things.
08:21SPEAKERMachine learning is going that way.
08:24SPEAKERNeural networks are not what we're talking about.
08:26SPEAKERLike, like neural networks are a thing,
08:27SPEAKERare a thing that is used in this,
08:29SPEAKERbut it's not what we're talking about.
08:30SPEAKERWhat we're talking about is a combination of a bunch of
08:32SPEAKERdifferent machine learning techniques to,
08:35SPEAKERto create things that are indistinguishable
08:39SPEAKERfrom something that was created by a person.
08:41SPEAKER- Sure.
08:42SPEAKER- In a lot of cases.
08:43SPEAKER- Yes.
08:44SPEAKER- I would, I would add real quick here.
08:45SPEAKERTo me, I think the language problem is downstream
08:48SPEAKERof the money problem, which is to say that all the money
08:51SPEAKERin Silicon Valley is now being poured into this.
08:54SPEAKERAnd as soon as that starts happening,
08:55SPEAKERthen all the language becomes meaningless.
08:57SPEAKER- Probably we're on the back end of that trend now,
09:00SPEAKERwould be my guess.
09:01SPEAKERWhat happens is everybody makes their bets in an area.
09:04SPEAKERSo like, this happened with VR,
09:06SPEAKERthis happened with mobile stuff,
09:07SPEAKERthis happened with apps 20 years ago.
09:09SPEAKERPeople make their, like the big investors make their bets.
09:12SPEAKERSo your entries and Horowitz's and your, your, you know,
09:15SPEAKERwhatever, Sequoia's, Sequoia's, your, your, all of those,
09:19SPEAKERall of the big capital people will say, okay,
09:21SPEAKERwe're going to put a billion dollars into 10 companies
09:24SPEAKERthat do 10 different things.
09:25SPEAKEROr some of them do 10 companies that do the same things
09:27SPEAKER'cause they're fucking mental.
09:31SPEAKERBut yeah, and then they make them fight basically.
09:33SPEAKERJust, just to be clear, that's the mental part
09:34SPEAKERis that you invest in 10 competing products
09:36SPEAKERthat then all you need is one of them
09:39SPEAKERto succeed and you win.
09:40SPEAKER- Sure.
09:41SPEAKER- Anyway, the point is people put in a buttload of money,
09:46SPEAKERlike billions upon billions of dollars.
09:48SPEAKERThe results are starting to show from the first people
09:51SPEAKERthat were in that space.
09:53SPEAKERSo, and, and the interesting thing is,
09:56SPEAKERbecause one of those companies, Open AI, was founded
10:01SPEAKERto, to, to do good.
10:03SPEAKERLike their, their, their charter is interesting
10:06SPEAKERin that they say, hey, we're, we're, we're,
10:09SPEAKERwe are starting this company because we're concerned
10:11SPEAKERthat the proliferation of AI could be dangerous
10:14SPEAKERfor humanity.
10:16SPEAKERSo, you know, we're, they are profit limited.
10:21SPEAKERThey have a open, it's like a prop for profit,
10:25SPEAKERbut profit limited company tied to a,
10:28SPEAKERand also I don't know what a profit limited company means.
10:31SPEAKERLike that's a thing that I think they had to make up
10:33SPEAKERfor this, so who fucking knows?
10:36SPEAKERBut they're tied, and they're tied to a,
10:38SPEAKERlike a research/nonprofit entity as well.
10:43SPEAKERResearch, research entity, research lab.
10:47SPEAKERSo anyway, the upshot is it's really hard to know
10:51SPEAKERwhat's actually going on, but we're starting to see tools
10:54SPEAKERthat are coming out in this, in these spaces,
10:58SPEAKERlike well before we would have in previous times
11:01SPEAKERbecause Open AI is releasing a lot of their stuff
11:04SPEAKERfor people to just noodle with.
11:06SPEAKERSo we can, we can see the results,
11:10SPEAKERwhich also benefits them because it gives them more data
11:12SPEAKERto feed back into the machines.
11:14SPEAKERBut we can see the results in a time when normally
11:16SPEAKERat this part of the product cycle,
11:18SPEAKERnone of this stuff would be public, right?
11:20SPEAKER- Yeah, I'm looking at their list of products and definitely
11:24SPEAKERsome heavy hitters on here, like Dolly is on here,
11:26SPEAKERwhich helped me remember, like it's all a slurry
11:28SPEAKERin my memory at this point because stuff's moving so fast
11:30SPEAKERand there's so much of it.
11:31SPEAKERBut like everybody talks about mid journey
11:34SPEAKERand stable diffusion now,
11:35SPEAKERbut wasn't Dolly kind of the first one that made headlines
11:38SPEAKERthat became a household name in the image generation space?
11:41SPEAKER- Dolly was the first one that like produced images
11:45SPEAKERthat were, that looked, that could have been mistaken
11:47SPEAKERfor something done by a bad artist.
11:49SPEAKER- Right. - Yeah.
11:50SPEAKER- Okay, and then they are also chat GPT,
11:52SPEAKERwhich of course is everywhere at the moment.
11:55SPEAKER- Yep, mid journey is one of the other ones.
11:57SPEAKERIt's one of the other image generators.
11:59SPEAKER- Did they do that one too?
12:00SPEAKEROh no, you're just.
12:01SPEAKER- I'm just saying like, here are the things
12:03SPEAKERthat people are talking about now
12:04SPEAKERand people are using now.
12:05SPEAKER- Yeah, I feel compelled to mention that there are such
12:08SPEAKERpeople as Elon Musk and Peter Thiel also involved in OpenAI.
12:11SPEAKERSo.
12:12SPEAKER- Yeah, like, but okay, so yeah, Thiel and Musk
12:17SPEAKERand a handful of big value investors
12:19SPEAKERput in a couple billion dollars to start this in 2015,
12:22SPEAKERI think, 2016, and then Microsoft and Amazon
12:27SPEAKERhave dumped in a buttload of money.
12:29SPEAKERLike AWS specifically invested, which makes sense
12:32SPEAKER'cause a lot of the math for all,
12:33SPEAKERfor this kind of machine learning stuff runs in the cloud
12:36SPEAKER'cause you don't want, nobody wants to have,
12:38SPEAKERit doesn't make sense for people to buy a cluster
12:40SPEAKERof 85 cluster GPU boxes just to run this stuff
12:45SPEAKERwhen, you know, in two years it's,
12:48SPEAKERthe hardware's all gonna be obsolete,
12:49SPEAKERso you just let Amazon eat that cost.
12:52SPEAKER- I would like to have that in my backyard.
12:54SPEAKERI'm a weirdo that likes to look at the computer
12:57SPEAKERthat's doing the process.
12:58SPEAKER- I think your power bill is gonna disagree with that.
13:01SPEAKERI mean, but it's just the hard drive problem
13:04SPEAKERwith NASA's, right?
13:05SPEAKERLike unless you're really getting use out of the NASA,
13:07SPEAKERdoes it make sense to have a NASA with like five hard drives
13:09SPEAKERin it 'cause the cost of replacing the hard drives
13:11SPEAKERevery three years is a little, anyway.
13:12SPEAKER- Yes, it's just wish fulfillment anyway.
13:15SPEAKER- Yeah, so, okay, there's all these companies,
13:19SPEAKERthere's a ton of money in this space.
13:21SPEAKERThere's all these companies doing this work.
13:22SPEAKERWe're starting to see things like a chat, like GPT,
13:26SPEAKERwhich shows GPT-2 specifically, which is indistinguishable
13:30SPEAKERfrom like shitty blog posts written by people
13:33SPEAKERthat are bad at writing all the way up to,
13:37SPEAKERI mean, this came up because last night in the Discord,
13:41SPEAKERone of our bots has a chat GPT integration.
13:44SPEAKER- I saw that, did we know about that?
13:48SPEAKER- I turned it off, it was spammy.
13:50SPEAKER- Did we know about that?
13:51SPEAKER- We did not know about that, no.
13:52SPEAKERIt's a new feature, it's a new feature that's been added.
13:54SPEAKERThe bot does some handful of things for us.
13:57SPEAKERIt does like role assignments for people.
13:58SPEAKER- Yeah, just basic stuff and then all of a sudden
14:00SPEAKERthat's kind of the danger of running external software
14:02SPEAKERon your Discord server, right?
14:03SPEAKERIf they add a feature to that software,
14:04SPEAKERyou didn't anticipate and your users know about it
14:06SPEAKERand you don't.
14:08SPEAKER- Well, so most of the stuff you can turn off,
14:11SPEAKERlike most of the features of that particular bot
14:13SPEAKERare turned off.
14:16SPEAKERBut like I asked it to write a post that says,
14:19SPEAKERwrite a passive aggressive 1,500 word non-apology notice
14:22SPEAKERexplaining that you're removing a beloved feature
14:24SPEAKERnamed Ask Chat GPT from Me6, which is the bot name,
14:28SPEAKERwithout committing any concrete reasons
14:30SPEAKERwhy the real reason is spam.
14:31SPEAKERAnd it did a really nice job with it.
14:33SPEAKER- I have to find this, I'm scrolling frantically
14:36SPEAKERtrying to find this.
14:37SPEAKER- I'll send you the link.
14:38SPEAKER- That's amazing.
14:39SPEAKER- I'll read it to you actually if you want.
14:41SPEAKER- Okay, sure.
14:42SPEAKERDear Me6 community, we know how much you have come to love
14:45SPEAKERthe Ask Chat GPT feature on our platform.
14:48SPEAKERWe've enjoyed watching you use it
14:49SPEAKERto create amazing conversations
14:51SPEAKERand share your ideas with each other.
14:52SPEAKERUnfortunately, we have made the difficult decision
14:54SPEAKERto remove Ask Chat GPT from Me6.
14:56SPEAKERWe understand that this will be a disappointment
14:58SPEAKERto many of you and we apologize
14:59SPEAKERfor any inconvenience this may cause.
15:01SPEAKERWe understand that this is a beloved feature
15:03SPEAKERand we know that many of you have come to rely on it
15:05SPEAKERas part of your daily Me6 experience.
15:07SPEAKERWe are sorry to have made this decision,
15:09SPEAKERto have to make this decision,
15:11SPEAKERbut we feel that it's necessary
15:12SPEAKERin order to maintain the quality of our platform.
15:14SPEAKERWe understand that this decision may seem sudden
15:16SPEAKERand without explanation.
15:17SPEAKERWe assure you that we have given this decision
15:18SPEAKERa great deal of thought and consideration before making it.
15:22SPEAKERWe are confident that it is the right decision
15:23SPEAKERfor the future of Me6.
15:26SPEAKERWe understand that this decision may be difficult to accept
15:28SPEAKERand we are sorry for any disappointment or frustration
15:31SPEAKERthis may cause.
15:32SPEAKERWe hope that you can understand our decision
15:34SPEAKERand continue to enjoy Me6 as you have in the past.
15:37SPEAKERThank you for your understanding, the Me6 team.
15:40SPEAKER- A little verbose, could have used an editor
15:41SPEAKERor maybe redundant in spots,
15:43SPEAKERbut like kind of totally functional as a shitty,
15:45SPEAKERsoulless PR managed apology sort of note.
15:50SPEAKER- And I told it to do 1500 words.
15:52SPEAKERIt did write out 1500 words.
15:53SPEAKER- Okay.
15:54SPEAKER- Like I wanted it to be a long apology letter
15:56SPEAKERand not a short apology letter, right?
15:58SPEAKERMaybe 800 words would have been better, but anyway.
16:00SPEAKER- Like kind of almost indistinguishable
16:02SPEAKERfrom a human written version of this
16:05SPEAKERthat you might see in this context.
16:07SPEAKER- It didn't give any real reasons why.
16:09SPEAKERIt was like, we wanna maintain the quality of our platform.
16:11SPEAKERThat's what you always say when you wanna kill something
16:13SPEAKERthat's expensive or a pain in the ass.
16:14SPEAKER- That's exactly what I meant
16:15SPEAKERabout it being indistinguishable.
16:17SPEAKER- Yeah.
16:17SPEAKER- It's a lot of words that say basically nothing.
16:19SPEAKER- Yep.
16:21SPEAKERYeah. - Oh man.
16:22SPEAKER- So anyway, so we wanted to talk about it.
16:25SPEAKERLike we wanna talk about the quote unquote AI.
16:29SPEAKER- Just yeah, where things are at.
16:31SPEAKERThings feel weird right now in a way that is moving quickly.
16:35SPEAKERI think you would agree.
16:37SPEAKER- Yeah, I would say that the path from like Dolly
16:40SPEAKERto mid journey to stable diffusion,
16:41SPEAKERto stable diffusion two, to whatever's coming next
16:44SPEAKERis everything seems to be accelerating
16:47SPEAKERin a way that is like singularity-ish almost.
16:52SPEAKERIt's unsettling.
16:53SPEAKER- Yeah, it is for sure.
16:54SPEAKERI mean, that's a big word to throw around.
16:56SPEAKER- Well, maybe the point though,
16:58SPEAKERthe bog standard definition of the singularity
17:02SPEAKERfrom Kurzweil 30, 40 years ago whenever we wrote the book
17:05SPEAKERis, hey, it's the time at which you can't predict
17:08SPEAKERanything else, what comes beyond.
17:11SPEAKER- Yeah.
17:12SPEAKER- Yeah, the Kwitsatz Haderach thread.
17:14SPEAKER- Sure, yes.
17:15SPEAKERAnd I assume he was speaking in much broader terms
17:19SPEAKERthan AI, right?
17:20SPEAKEROr was he talking--
17:21SPEAKER- No, he was specifically talking about AI.
17:22SPEAKERHe was talking about AI as like a driver for change
17:26SPEAKERthat makes change accelerate to the point
17:28SPEAKERthat it's impossible to predict what comes next.
17:30SPEAKER- Yeah, I mean, but that concept obviously could apply
17:32SPEAKERto energy production or genetic manipulation
17:35SPEAKERor any number of other radical--
17:37SPEAKER- Specifically, it was about machines becoming smarter
17:41SPEAKERthan people and being able to make themselves smarter.
17:43SPEAKER- Sure, I mean--
17:44SPEAKER- So that at which point our brains become
17:46SPEAKERthe second tier brains on the planet
17:48SPEAKERand maybe the universe, who knows?
17:50SPEAKER- I tell you, I don't.
17:51SPEAKER- And then things go bad.
17:52SPEAKER- I saw the Terminator.
17:53SPEAKERI know how this goes.
17:54SPEAKER- Yeah, dun, dun, dun, dun, dun.
17:56SPEAKER- Yeah.
17:56SPEAKERSo anyway, yeah, things are unpredictable.
18:00SPEAKER- The robots aren't coming, Brad.
18:01SPEAKERThat's the thing to know.
18:02SPEAKERThe robots, look, we can still, like there was a great thread
18:07SPEAKERon Twitter or Mastodon or something a few months ago
18:09SPEAKERfrom a roboticist who was like, okay,
18:11SPEAKERyou don't have to worry about killer robot.
18:12SPEAKERLike killer robots aren't gonna be bipedal
18:14SPEAKERwalking around with machine guns.
18:16SPEAKERThere'll be cars and drones that depth charge you
18:18SPEAKERand stuff like that.
18:19SPEAKERIf you have a robot walking at you,
18:21SPEAKERall you do is get some dust, some Fuller's dust
18:23SPEAKERand just throw it at it.
18:24SPEAKERIt'll clog up the joints to the point that it seizes up.
18:27SPEAKER- Chaff, chaff is always the way.
18:29SPEAKER- Yeah, keep your dust on hand.
18:30SPEAKERAnyway, okay, so the point of this whole episode
18:32SPEAKERis we wanna talk about what stuff actually does
18:35SPEAKERas much as we can know.
18:36SPEAKER'Cause some of it happens internally.
18:37SPEAKERI mean, stable diffusion's open, so you can look at it.
18:39SPEAKERA lot of the open AI stuff is actually open,
18:42SPEAKERalthough I think they haven't,
18:43SPEAKERthey're a little weird about releasing source code
18:45SPEAKERfor stuff, it seems like.
18:49SPEAKERThis isn't generalized AI, right?
18:51SPEAKERLike these products, mid journey and dolly
18:56SPEAKERand stable diffusion and GPT and whatever comes next,
18:59SPEAKERthe, oh, we even talked about the audio stuff,
19:01SPEAKERthe ones that you feed your voice into
19:03SPEAKERand two minutes of voice and it records,
19:05SPEAKERit generates a voice generator that sounds kind of like you,
19:08SPEAKERbut not exactly.
19:09SPEAKER- Yeah.
19:13SPEAKER- They're not, these don't, they don't have general,
19:15SPEAKERthey don't have understanding.
19:16SPEAKER- I mean, I fucking hope not.
19:17SPEAKERLike if we have inadvertently created sentient programs
19:21SPEAKERand nobody is talking about it, we are in trouble.
19:24SPEAKER- Nope, we're not there yet.
19:25SPEAKER- No.
19:26SPEAKER- That guy at Google said that a couple of years ago
19:28SPEAKERand then was proudly debunked and also I think sacked.
19:30SPEAKER- I think he was fired.
19:31SPEAKERI mean, I think it was, he was like violating some.
19:34SPEAKER- He broke some of the contractor's clause.
19:38SPEAKER- Yeah.
19:38SPEAKER- Yeah, anyway, but.
19:39SPEAKER- But.
19:40SPEAKER- But.
19:41SPEAKER- And then like also, also like, you know,
19:42SPEAKERthere are all these intermediate steps on the way
19:44SPEAKERto a truly self-aware theoretical AI, you know,
19:48SPEAKERTuring test and stuff like that.
19:50SPEAKER- Well.
19:51SPEAKER- We might be there actually.
19:52SPEAKER- I think we're there.
19:53SPEAKER- We might be getting really close to things satisfying
19:55SPEAKERthe Turing test as a matter of course.
19:57SPEAKER- But I mean, the thing that's happening, okay.
20:00SPEAKERSo the thing that's happening, when you say,
20:03SPEAKERhey, I want to see pictures of pizza,
20:05SPEAKERdraw me a picture of pizza to mid journey or to.
20:08SPEAKER- I'm always saying that I want to, okay.
20:11SPEAKER- My understanding is that you're more bread than pizza,
20:13SPEAKERbut whatever.
20:15SPEAKERWhen you say, hey, I want to see pizza, it's not,
20:17SPEAKERthe machine doesn't know, oh, okay,
20:19SPEAKERBrad likes pizza here and it doesn't conceptually understand
20:24SPEAKERwhat a pizza is.
20:25SPEAKERIt knows that pizza is used with words like dough and sauce
20:29SPEAKERand cheese and pepperoni.
20:31SPEAKERAnd it knows that there's a specific color,
20:33SPEAKERlike there's a structure to images that are used
20:36SPEAKERin conjunction with the word pizza.
20:38SPEAKER- I have to, I assume it's just literally associating colors
20:41SPEAKERand shapes and that sort of thing.
20:43SPEAKER- I mean, it's more complicated than that.
20:45SPEAKER- I'm sure there's a lot more detail and category
20:48SPEAKERof information, but yes.
20:50SPEAKER- Yeah, it knows that it basically knows that certain things
20:53SPEAKERgo together and some things go together better than others.
20:57SPEAKERAnd there's some games that are like this,
21:00SPEAKERwhere you can put a word in and you have to,
21:04SPEAKERI can't remember what it's called.
21:06SPEAKERYou put a word in and your goal is to guess a word
21:09SPEAKERand it tries to use semantic analysis of the word
21:12SPEAKERthat you put in and the word that it's trying to get you to
21:14SPEAKERand tells you how close your word is.
21:17SPEAKERSo like the word might be water.
21:18SPEAKERAnd if you put ocean, that would be very close
21:20SPEAKER'cause an ocean is made of water.
21:21SPEAKERBut if you put a person, that would also be close
21:23SPEAKER'cause people are also made of water.
21:25SPEAKERAnd if you put a rock, that's gonna be probably
21:26SPEAKERa lot further away.
21:27SPEAKER- I saw there was a TikTok going around this morning
21:30SPEAKERof using your head to amplify the radio signal
21:33SPEAKERout of your old car key fob
21:34SPEAKERbecause your head is full of water
21:36SPEAKERand the water acts as an amplifier.
21:38SPEAKER- That doesn't sound good.
21:40SPEAKER- No, they did mention you could also use
21:42SPEAKERa full water bottle.
21:45SPEAKERPoint being, people have a lot of water in them anyway.
21:47SPEAKER- I didn't know you were on TikTok.
21:48SPEAKERAlso, we gotta have a conversation about this.
21:50SPEAKER- I view TikTok by way of Twitter.
21:54SPEAKER- Okay, okay.
21:56SPEAKER- You know what, oh boy, complete tangent here.
21:58SPEAKERIs that a minor sign of the impending death of Twitter
22:01SPEAKERthat a lot of Twitter content
22:04SPEAKERis just reposting TikToks these days?
22:06SPEAKERLook, a lot of it.
22:08SPEAKER- The Twitter situation, turning off the,
22:11SPEAKERon a related note, hey, PSA,
22:14SPEAKERif you have ever used Twitter to sign in for a website
22:17SPEAKERthat you find have information that is useful
22:19SPEAKERor valuable to you on it,
22:21SPEAKERgo to that website and put your email address into it
22:23SPEAKERand then hit the forgot password button
22:25SPEAKERso you can associate that Twitter account
22:28SPEAKERwith your email address
22:29SPEAKER'cause it's gonna stop working on February 8th
22:31SPEAKERat the end of the day
22:33SPEAKERbecause Twitter is full of numpties.
22:35SPEAKERAnyway. - By the time you hear this,
22:38SPEAKER- You have like four days. - Four days, yes.
22:40SPEAKER- Yeah.
22:42SPEAKER- Sorry, I did not mean to derail.
22:44SPEAKER- No, no, no, it's okay.
22:45SPEAKERBut the whole point is the machines don't understand things.
22:48SPEAKERThey aren't sentient, right?
22:49SPEAKERThey're just, they just know the relationship
22:51SPEAKERbetween everything to everything else,
22:55SPEAKERwhich I mean, maybe that's what sentience is.
22:57SPEAKERWho knows?
22:59SPEAKER- I mean, well, boy, we do not need to get
23:01SPEAKERinto like philosophy of consciousness stuff here,
23:03SPEAKERbut you know, like self-awareness and ambition
23:06SPEAKERand some kind of emotional state, I feel like,
23:10SPEAKERare some of the requisites anyway.
23:13SPEAKER- Yeah, there's no, there's no,
23:15SPEAKERas far as we know, the internal monologue
23:17SPEAKERis just about what's connected to everything else.
23:20SPEAKER- Yeah, can we just start here and define some terms
23:25SPEAKERor like, I mean, maybe try to like unpack
23:27SPEAKERa little bit of the investment-driven jargon.
23:30SPEAKER- Yeah.
23:31SPEAKER- Like AI is kind of the term
23:33SPEAKERthat is generally the buzzword these days.
23:34SPEAKERLike, is it pretty much 100% of the quote unquote AI
23:38SPEAKERin use now, machine learning driven?
23:40SPEAKER- Well, I mean, it's saying they're,
23:42SPEAKERboth of those words have been the same thing
23:44SPEAKERand both of them can mean anything
23:45SPEAKERthat people want them to mean.
23:46SPEAKER- So that's the problem
23:47SPEAKERand that's what I'm trying to get at here.
23:49SPEAKER- Machine, I would say that the people
23:51SPEAKERwho talk about machine learning generally are the people,
23:54SPEAKERwhen somebody says, hey, I have a new AI,
23:55SPEAKERthen you kind of go, oh, okay, whatever.
23:57SPEAKERThe machine, the people who know what they're talking about
23:59SPEAKERgenerally talk about machine learning
24:01SPEAKERor they talk about specific types of machine learning,
24:03SPEAKERlike GANs and stuff like that.
24:04SPEAKER- Okay, and I, like a kind of related question
24:07SPEAKERis I was gonna ask, is there another better term
24:10SPEAKERthan the kind of singular AI as in an AI,
24:14SPEAKERlike machine learning is an abstract concept,
24:19SPEAKERbut what do you call the result of machine learning,
24:21SPEAKERa model?
24:22SPEAKER- Yeah, models are built.
24:25SPEAKER- Is that the best term?
24:26SPEAKER- Machine learning models are built
24:28SPEAKERusing these different techniques
24:30SPEAKERand like, I don't, again, this is one of the situations
24:34SPEAKERwhere everybody calls stuff different things,
24:36SPEAKERso it's hard to know.
24:37SPEAKER- Right.
24:38SPEAKER- I mean, I've always used machine learning for this
24:40SPEAKERand most of these people I've known
24:42SPEAKERthat work in these spaces that are serious,
24:44SPEAKERyou describe things as machine learning
24:46SPEAKERwith AI as the thing that they talk about
24:48SPEAKERon the nightly news being not a term of art, I guess,
24:54SPEAKERbut I don't have a strong sense of that.
24:57SPEAKER- Yeah, it's just like, you know,
24:59SPEAKERwith AI both being an extreme business buzzword
25:02SPEAKERat the moment and also this decades-long legacy
25:04SPEAKERin science fiction, it's a little bit,
25:07SPEAKERit's got a lot of baggage.
25:09SPEAKER- I think we have to just move,
25:11SPEAKERI think we have to just like not think about that too much.
25:13SPEAKER- No, no, no, I agree, it's just an attempt to understand
25:16SPEAKERwhat is actually tangibly going on under the hood.
25:19SPEAKER- Yeah, it's, I mean, this always happens, right?
25:23SPEAKER'Cause like we didn't remember smartphones,
25:24SPEAKERwe didn't know what to call smartphones for a while,
25:26SPEAKERpeople were trying to call them other things
25:27SPEAKERand like the names will settle out in time.
25:30SPEAKEREventually the machines will just tell us
25:32SPEAKERwhat to call themselves.
25:33SPEAKERI am Steve.
25:34SPEAKER- It's true.
25:35SPEAKER- Yeah, so how does this stuff work?
25:39SPEAKERThat's the, like the upshot is they take a large data set,
25:46SPEAKERlike a massive amount of data called a corpus,
25:49SPEAKERwhich is a, it's structured data,
25:52SPEAKERso it's like pictures of pizza that are labeled pizza
25:55SPEAKERand they also have a bunch of other tags on them.
25:56SPEAKERLike they'll say cooked, you know, with sauce,
25:59SPEAKERall the descriptors that you could think of
26:01SPEAKERthat would go with that picture of pizza, right?
26:03SPEAKERThey'll have straight on shots, they'll have off-axis,
26:06SPEAKERthey'll have slices, they'll have whole pies,
26:08SPEAKERthey'll have the box, they'll have not in the box,
26:10SPEAKERthey'll have the little tables that are in there
26:11SPEAKERto keep the box from smashing into the pizza.
26:13SPEAKERThey'll have pizzerias, they'll have delivery guys,
26:16SPEAKERthey'll have the bags that the delivery comes in.
26:19SPEAKERThey'll all be labeled and tagged
26:20SPEAKERwith all the descriptors you could possibly imagine.
26:24SPEAKERA lot of the stuff was happening,
26:26SPEAKERit seems like with your captchas over the years
26:28SPEAKERas you were clicking pictures of bicycles and sunflowers
26:31SPEAKERand goats and all those things.
26:35SPEAKERBut the point is it's all structured
26:38SPEAKERand it's known good, right?
26:40SPEAKERSo you run these neural networks on these corpuses
26:45SPEAKERand they build connections between them
26:49SPEAKERthat are faster and more accessible
26:51SPEAKERthan a traditional database lookup
26:53SPEAKERand also kind of more arbitrary.
26:55SPEAKERSo they're not predictable, right?
26:57SPEAKERLike the stuff that it latches onto,
27:00SPEAKERbasically at this point they're running,
27:02SPEAKERwell, at least the old days,
27:03SPEAKERI don't know what the people inside OpenAI
27:05SPEAKERand Stable Diffusion and stuff are doing.
27:07SPEAKERBut they would run artificial neural networks
27:09SPEAKERthat mimic the way neurons create paths
27:12SPEAKERwhere you put in the data set on one side
27:14SPEAKERand you see what comes out the other side
27:16SPEAKERand you infer what's happening
27:17SPEAKERon the inside of the neural network
27:19SPEAKERfrom what goes in and comes out,
27:20SPEAKERwhat comes out when it goes in.
27:23SPEAKERAnd just to be clear, this isn't new.
27:26SPEAKERThis part isn't new.
27:27SPEAKERGoogle Photos has been doing this on photos
27:28SPEAKERyou upload to their servers for years.
27:30SPEAKERApple does the same thing with photos
27:32SPEAKERthat you keep in your Apple Photos account.
27:36SPEAKERI think Apple does it locally
27:38SPEAKERand Google does it in the cloud,
27:39SPEAKERbut the upshot is the same
27:42SPEAKERis that I can type dog into Google Photos
27:44SPEAKERand it gives me all the pictures of dogs
27:45SPEAKERthat I've uploaded to Google Photos in the last 10 years.
27:48SPEAKER- Yeah, honestly, I mean, maybe very assault on this,
27:51SPEAKERbut googling around, I had landed on, for example,
27:54SPEAKERlike an article that IBM posted saying like,
27:57SPEAKERhey, a lot of modern machine learning
27:59SPEAKERhas its roots in research that was done
28:01SPEAKERby IBM in the '60s, and of course,
28:03SPEAKERthey might want to lay claim to some of this,
28:05SPEAKERbut the point is they made it seem like the concept
28:08SPEAKERfor this stuff does go back quite a ways.
28:11SPEAKERI was gonna ask if stuff like Deep Blue, for example,
28:14SPEAKERin the '90s, did that work on a reduced
28:17SPEAKERor primitive version of the same concept?
28:19SPEAKER- Almost certainly not.
28:20SPEAKERI mean, it was just database lookups.
28:22SPEAKER- Okay, studying chess moves in that context.
28:26SPEAKER- And Deep Blue might have actually done it the other way
28:28SPEAKERwhere it just projected forward.
28:29SPEAKERI don't know if it had, I don't know how that stuff worked.
28:32SPEAKERSo I did my senior thesis when I was in 1998, '97
28:40SPEAKERin the biochemistry department.
28:41SPEAKERI did my senior thesis about neural networks,
28:43SPEAKER'cause at that point we were building biological model,
28:46SPEAKERbiological networks out of like actual neurons
28:49SPEAKERand modeling how they were extracted.
28:53SPEAKERSeeing if we could model them in computers,
28:55SPEAKERwe didn't have anywhere near the compute necessary
28:57SPEAKERto do that, but like you could see a future
28:59SPEAKERwhere it happened.
29:00SPEAKERI just thought it would take a lot longer
29:01SPEAKERthan it was going, than it did.
29:03SPEAKERI mean, I guess it took 20 years,
29:04SPEAKERwhich seems about right.
29:05SPEAKERSo, but like we used some basic machine learning stuff
29:10SPEAKERat FU to do the, to do the animation.
29:17SPEAKERSo like with, we were using a push-pull kind of setup
29:24SPEAKERfor the animation where the IK would push the limbs
29:27SPEAKERin one direction, and then we trained a corpus of data
29:30SPEAKERon actual movements that would use a GAN to,
29:36SPEAKERGAN is a generative adversarial network.
29:39SPEAKERAdversarial network, yeah.
29:40SPEAKERThat basically tried to pull the IK data back to the mean,
29:46SPEAKERso it prevented things like weird chicken arms
29:48SPEAKERand unpredictable behaviors from happening.
29:52SPEAKER- When you said compared to actual movements,
29:54SPEAKERlike what form did that take?
29:55SPEAKERWas it video of human beings,
29:57SPEAKERor- - We bought a giant set
29:59SPEAKERof data, of IK data that was captured.
30:02SPEAKER- Okay, so just pure- - Of people moving.
30:04SPEAKERYeah, it was just people- - Pure motion capture data?
30:06SPEAKER- Well, it was vector data, I think.
30:08SPEAKERI can't remember, but it was basically,
30:11SPEAKERit was basically like somebody in a motion capture suit
30:14SPEAKERdoing all of the limits of their human movement
30:17SPEAKERfor X period of time, and it was like 18 hours of that
30:20SPEAKERor something, it was some ridiculous,
30:22SPEAKERit was a giant amount of data.
30:25SPEAKERAnd that let us do something.
30:27SPEAKERSo the benefit of this stuff is that the neural networks,
30:31SPEAKEReven simulated and just running on TensorFlow cores
30:33SPEAKERon your GPU or in your CPU or whatever,
30:36SPEAKERare able to do this math, these lookups,
30:39SPEAKERway, way, way faster than you could in a normal,
30:42SPEAKERif you put all this stuff in a normal SQL database
30:44SPEAKERand search for pizza, you're not gonna get
30:47SPEAKERthe kind of weight and accuracy that you get
30:52SPEAKERby doing the trained neural network.
30:55SPEAKERWith the machine learning models.
30:57SPEAKERSo then, okay, so then how does that lookup
31:01SPEAKERturn into a drawing of a pizza in mid journey
31:04SPEAKERor something like that, right?
31:06SPEAKERYou put in your words, it parses the words
31:09SPEAKERto find what it thinks they mean, right?
31:11SPEAKERAnd then it's like, okay, here are all of the pictures
31:13SPEAKERthat we have that mean these words, right?
31:16SPEAKERSo you say, I wanna draw a picture of a pepperoni pizza,
31:20SPEAKERand I want it to look like a Looney Tunes cartoon,
31:23SPEAKERand I want Wile E. Coyote to be holding the pizza,
31:26SPEAKERand I want Big Bird to be taking a slice from the box.
31:31SPEAKERAnd it knows, these are what Big Bird looks like,
31:35SPEAKERhere's Big Bird taking something,
31:37SPEAKERhere's what a pepperoni pizza looks like,
31:39SPEAKERhere's what a pepperoni pizza looks like
31:40SPEAKERwhen it's in a box, here's the Looney Tunes,
31:42SPEAKERis this all flat shaded?
31:45SPEAKERThese pictures all have these same characteristics,
31:48SPEAKERthere's always edges, there's hard edges,
31:50SPEAKERthey're flat shaded, they're, you know,
31:53SPEAKERblah, blah, the limited palette, blah, blah, blah.
31:56SPEAKERAnd then it takes 15 different,
31:59SPEAKERall those different elements
32:00SPEAKERand kind of blurs the edges between them
32:02SPEAKERand smashes them together on one picture.
32:05SPEAKER- Okay, that might answer a question I was about to ask
32:07SPEAKERand have wondered for quite a while
32:08SPEAKERabout these generated images.
32:10SPEAKERLike, what is the most irreducible component
32:13SPEAKERof the image that comes out of these things?
32:15SPEAKERLike, is it building literally pixel by pixel,
32:17SPEAKERor is it, I mean, this is kind of undefinable
32:21SPEAKERbecause you don't know what images are being used
32:22SPEAKERin the result necessarily,
32:24SPEAKERbut is it like, is it actual like visible chunks
32:26SPEAKERof whole images that are being smashed together
32:29SPEAKERwith image processing, or is it literally going,
32:32SPEAKERis it generating images from whole cloth,
32:34SPEAKERor do you get what I'm asking?
32:36SPEAKER- It's generating,
32:37SPEAKERit understands the characteristics of images.
32:40SPEAKER- Okay.
32:41SPEAKER- So like when you say make an art, draw a goat
32:44SPEAKERthat looks like a Lisa Frank painting,
32:45SPEAKERit knows what a Lisa Frank,
32:46SPEAKERlike it knows what the characteristics
32:48SPEAKERof a Lisa Frank painting or Salvador Dali
32:50SPEAKERor whatever are, right?
32:52SPEAKERAnd it knows what a goat looks like,
32:54SPEAKERand it knows how to apply.
32:56SPEAKERIt can apply those, it's, it is,
32:58SPEAKERyou can create a goat picture out of whole cloth
33:01SPEAKERthat looks like a Lisa Frank painting,
33:02SPEAKEReven though Lisa Frank never painted a goat.
33:04SPEAKER- Got it, so there are not necessarily identifiable bits
33:08SPEAKERof existing images in the result.
33:10SPEAKER- Sometimes there are, people have definitely found them.
33:12SPEAKER- Yeah, absolutely.
33:13SPEAKERI mean, I saw something floating around last night,
33:14SPEAKERI should have clicked in and read more about it,
33:16SPEAKERthat there's an issue with some of the models,
33:18SPEAKERquote unquote, memorizing certain images
33:21SPEAKERthat it was trained on.
33:22SPEAKERLike the specific example given was,
33:25SPEAKERpart of the model was a photo of a specific living,
33:29SPEAKERyou know, not an AI,
33:31SPEAKERlike an actual woman who was alive with a name.
33:33SPEAKERAnd then they requested a picture
33:35SPEAKERof that woman from the model.
33:37SPEAKERAnd it spit out basically the exact same image
33:39SPEAKERthat was fed to it.
33:40SPEAKERAnd in that case, like that's not an original image at all.
33:43SPEAKER- Well, none of these are original images.
33:45SPEAKERThese are all like, so the thing that I've found,
33:48SPEAKERso I installed Stable,
33:49SPEAKERI've had Stable to Fusion on my computer
33:51SPEAKERjust to kind of noodle 'cause I like to poke at stuff
33:53SPEAKERand I needed to test a really overpowered GPU.
33:58SPEAKERAnd the thing I found is that stuff
34:01SPEAKERthat the more pictures there are of things,
34:03SPEAKERthe more likely you are to get an actionable,
34:06SPEAKERgood result out of them.
34:08SPEAKERFor example, if you type Mickey Mouse into Stable to Fusion,
34:12SPEAKERMickey Mouse, there's a lot of pictures of Mickey Mouse
34:14SPEAKERout in the world.
34:16SPEAKERYou get something that looks exactly like art created
34:19SPEAKERby Disney to promote Mickey Mouse,
34:24SPEAKERexcept it can't do hands.
34:26SPEAKERHands are really hard.
34:27SPEAKER- Hands and teeth seem like the things.
34:30SPEAKER- Yeah, wow, there's always way too many or not enough,
34:32SPEAKERor like sometimes there's just like one really wide tooth.
34:36SPEAKERThat's disturbing.
34:39SPEAKERYeah.
34:39SPEAKER- That hand is coming right out of Mickey Mouse's shoulder.
34:43SPEAKER- Yeah, and it also has the wrong number of fingers
34:44SPEAKERand the thumb is where the index finger should be.
34:47SPEAKER- And also there's something on the tongue
34:49SPEAKERthat could be shading or it could be a tooth
34:51SPEAKERcoming out of the tongue, I'm not sure.
34:52SPEAKER- But the Mickey Mouse head pretty,
34:55SPEAKERoh, that's a shine, I would say.
34:57SPEAKERThat's a shine on the tongue.
34:59SPEAKER- Yeah, I think so.
35:02SPEAKER- Just so you know, the prompt of this
35:03SPEAKERwas Mickey Mouse human photo reel.
35:06SPEAKERDidn't get the human or the photo reel part,
35:08SPEAKERbut the Mickey Mouse head is pretty solid.
35:12SPEAKER- So like when you do pictures of things,
35:16SPEAKERwhen you ask for like a celebrity,
35:18SPEAKERsomebody that has a lot of photos,
35:19SPEAKERor I asked for Will Smith slapping Chris Rock,
35:22SPEAKER'cause I thought that would be funny,
35:24SPEAKERand I got the amalgam of all the news blog pictures
35:28SPEAKERwhere they put like a separate picture of Will Smith
35:31SPEAKERand Chris Rock facing each other,
35:33SPEAKERlooking like they're both yelling.
35:34SPEAKER- Interesting.
35:35SPEAKERYeah, like it's literally divided down the middle,
35:37SPEAKERjust like it would be in one of those
35:40SPEAKERkind of faceless clickbait news sites.
35:42SPEAKER- Yeah, like when you ask for like Mars bases
35:46SPEAKERor moon bases, if you don't give specific constraints,
35:49SPEAKERlike if you want like a '70s pastel of a space station,
35:57SPEAKERlike the kind that NASA used to commission
35:59SPEAKERof the rotating space stations,
36:05SPEAKERyou have to really constrain the more specific things
36:08SPEAKERso it understands what you're asking for.
36:10SPEAKERAnd like anything else, the more specific you are
36:15SPEAKERwhen you're writing the query, the better you get,
36:19SPEAKERyou also can drill down in specific axes.
36:22SPEAKERLike if you like, I'm trying to think of an example here,
36:26SPEAKERbut like, so I was trying to get a Homer Simpson,
36:28SPEAKERbut with real human skin.
36:30SPEAKERAnd when you, it turned out to be really hard
36:34SPEAKERbecause there's a lot of pictures of Homer Simpson
36:35SPEAKERas a cartoon and no pictures of people
36:38SPEAKERthat look like Homer Simpson
36:39SPEAKERthat are tagged as Homer Simpson.
36:41SPEAKERSo it didn't know, like it knows what to do with line art.
36:44SPEAKERIt doesn't know what to do with a 3D.
36:47SPEAKERIt doesn't know how to put skin on a 2D representation
36:51SPEAKERof a 3D person, if that makes sense.
36:54SPEAKER- Everything you just said sounds horrifying.
36:56SPEAKER- It's not good, it's not good.
36:58SPEAKER- Like if you set out to make images
37:00SPEAKERthat explicitly are horrifying as a goal,
37:02SPEAKERyou can come up with some extremely disturbing stuff.
37:06SPEAKER- You should check out my Tumblr.
37:07SPEAKER- I might be good.
37:09SPEAKERI did want to ask real quick.
37:10SPEAKERSo you've got a pretty beefy PC.
37:12SPEAKERYou're like what, a 13900K and a 4090.
37:14SPEAKER- Yeah.
37:15SPEAKER- How long does it take on that beefy hardware
37:18SPEAKERto put in a request to give it a prompt
37:20SPEAKERand get something out?
37:22SPEAKER- Depends on the resolution you ask for.
37:24SPEAKERIf you're looking for like, these images,
37:26SPEAKERI think are 512 by 512.
37:28SPEAKERThat takes like 30, 45 seconds to generate a few.
37:31SPEAKER- Okay, sure.
37:33SPEAKER- It generally, for whatever reason,
37:35SPEAKERI don't understand why this works,
37:38SPEAKERbut for whatever reason, when you start it,
37:43SPEAKERif you do four images off the same query,
37:48SPEAKERit generates the last three much faster than the first one.
37:51SPEAKERSo I assume there's some magic happening there
37:55SPEAKERthat then is repeatable and it just pulls off.
37:58SPEAKERLike one of the ways to think about this
38:00SPEAKERis that you're building a three-dimensional map
38:02SPEAKERthrough this database or maybe it's a 10-dimensional map
38:06SPEAKERthrough this database, right?
38:07SPEAKERSo you're taking different branching paths at each node
38:11SPEAKERand it's not just a normal three-dimensional space.
38:14SPEAKERIt's like as many dimensions
38:16SPEAKERas there are tags off of each node.
38:19SPEAKERSo like you go in, like if you're looking at pizza,
38:24SPEAKERthere's a whole other dimensions of the data
38:28SPEAKERat dough, crust, sauce, cheese, pepperoni,
38:33SPEAKERItalian, Detroit, Sicilian,
38:36SPEAKERlike all the different things you get, recipe.
38:39SPEAKERLike there's different nodes on all of those things.
38:41SPEAKER- Checker tablecloth.
38:42SPEAKER- Yeah, and once it's found that spot in the database,
38:46SPEAKERit can, in the model, it can go back.
38:50SPEAKERSo the other thing to say about this
38:51SPEAKERis creating the models is like,
38:55SPEAKERcreating the images really computationally expensive.
38:58SPEAKERCreating the models that create the images
39:00SPEAKERis offensively computationally expensive.
39:03SPEAKERLike it is stupid, stupid expensive to make those models
39:07SPEAKERto the point that like Nvidia has a whole cluster
39:10SPEAKERof machines that they use to generate the DLSS models.
39:15SPEAKERAnd they run for weeks and weeks and weeks
39:18SPEAKERand are constantly updating them and things like that.
39:20SPEAKERSo there's a real, even like the total cost of the compute
39:25SPEAKERfor these machine learning things,
39:27SPEAKERisn't just like the energy that comes out of the wall
39:31SPEAKERand goes into your computer
39:32SPEAKERwhen you run against a stable diffusion.
39:35SPEAKERIt's also the cost of them building that model
39:38SPEAKERat stable diffusion, which is, you know,
39:42SPEAKERan enormous amount of energy and that's carbon,
39:44SPEAKERdepending on where they're buying their electricity from.
39:48SPEAKER- There's a lot.
39:50SPEAKER- It's a lot.
39:50SPEAKER- A lot of externalities, a lot of factors.
39:55SPEAKER- Yeah, I mean, and like, there's a lot,
39:58SPEAKEReven before you get into the, hey, this is,
40:03SPEAKERit gets, it's controversial for a lot of reasons,
40:06SPEAKERfor a lot of good reasons, just to be clear.
40:08SPEAKERLike they're taking art
40:11SPEAKERthat has questionable license to it at best.
40:15SPEAKER- Yeah, yeah, I mean, to be clear,
40:17SPEAKERlike we were just talking about,
40:18SPEAKERthere's a lot on just the pure technical end
40:20SPEAKERof implementation, but the ethical component
40:22SPEAKERof how this affects society is a broad topic, let's say.
40:27SPEAKER- Well, and I mean, yeah,
40:30SPEAKERso like artists understandably are pissed, right?
40:33SPEAKER'Cause if you're an artist and you type in,
40:35SPEAKERdraw a ladybug and then your name
40:38SPEAKERand you get a picture that looks like
40:39SPEAKERyou could have drawn it,
40:41SPEAKERthat's not great for a lot of reasons.
40:44SPEAKERI think the writers are less concerned,
40:46SPEAKERbut I mean, we've already seen like CNET,
40:48SPEAKERyour former employers, Red Ventures,
40:50SPEAKERreplaced the CNET writers with AI
40:53SPEAKERand then started posting a bunch
40:54SPEAKERof wildly inaccurate news stories as a result.
40:59SPEAKERBrad's being quiet now, sorry, Brad.
41:05SPEAKERYeah, like there's concerns about replacing people's jobs.
41:10SPEAKERWe as a society have done that multiple times in the past
41:14SPEAKERand it's worked out to date,
41:15SPEAKERbut it seems like the jobs are getting fewer and fewer.
41:19SPEAKER- Yeah, I mean, I'm repeating verbatim
41:21SPEAKERsomething I just said on the patron episode the other day,
41:23SPEAKERbut like, yes, it has worked out in a broad historical sense,
41:27SPEAKERbut I also worry that history has kind of forgotten
41:31SPEAKERa lot of the people who got left behind along the way.
41:33SPEAKERLike it's easy to look back at,
41:34SPEAKERI think you used the example of the cotton gin, right?
41:37SPEAKEROr I mean, the industrial revolution
41:39SPEAKERis rife with this stuff, right?
41:40SPEAKERBut like, it's easy now to look back and say like,
41:42SPEAKERoh, new jobs emerged and people found new lines of work.
41:45SPEAKERBut like in the moments when it happened,
41:48SPEAKERlike people who are not capable of finding new work-
41:52SPEAKER- Wasn't great for a lot of people.
41:52SPEAKER- Might disagree with you.
41:54SPEAKERBut like we have no visibility into that.
41:57SPEAKERI mean, I'm sure historical work has been done
41:59SPEAKERthat we could look up, but you know what I mean?
42:01SPEAKERLike it's easy to gloss over, I think,
42:02SPEAKERthe human cost at the time when we are decades
42:05SPEAKERor centuries hence and can say like,
42:08SPEAKERoh, everything's fine now.
42:10SPEAKER- Yeah, I mean, I think also,
42:15SPEAKERyou know, the industrial revolution gave rise
42:18SPEAKERto the modern labor movement, right?
42:19SPEAKERWhich is, I mean, it took a while.
42:22SPEAKER- Yeah.
42:24SPEAKER- But ultimately, ultimately the concern I think,
42:30SPEAKERthe concern I have is that these technologies
42:32SPEAKERconsolidate more of the power of labor in fewer hands, right?
42:39SPEAKERIt's like, it's the general critique
42:41SPEAKERof the tech industry in general, right?
42:42SPEAKERIs that when we're talking about car manufacturing,
42:47SPEAKERthat's consolidating, you know,
42:49SPEAKERthere's a lot of labor spread out over a lot of places
42:51SPEAKERto make one car.
42:53SPEAKERAnd when we talk about Google, it's a lot less labor
42:57SPEAKERthat generates a lot more capital,
42:59SPEAKERthus consolidating both the labor
43:01SPEAKERand the capital in smaller places.
43:03SPEAKERAnd like the outcomes from that have been often bad so far.
43:07SPEAKERSo anyway, I don't know, I'm not a labor theorist,
43:11SPEAKERso we should have Austin on to talk about this sometime.
43:13SPEAKER- Yeah.
43:16SPEAKER- But yeah, and then there's a whole bunch
43:20SPEAKERof other ethical considerations around,
43:22SPEAKERlike you can use this, you can train stable diffusion
43:26SPEAKERon a set of your own photos.
43:27SPEAKERSo if you wanna make photos of somebody that you know,
43:30SPEAKERor photos of a famous person against their,
43:35SPEAKERwithout their consent that put them doing things
43:37SPEAKERthat they wouldn't wanna be seen doing,
43:39SPEAKERit's really easy to do that.
43:40SPEAKERLike it's not, it is not like, hey,
43:43SPEAKERyou have to be really good at Photoshop now
43:45SPEAKERto make something that looks like a good Photoshop.
43:47SPEAKERIt's that, hey, you take 10 pictures of somebody
43:49SPEAKERand put them into stable diffusion
43:51SPEAKERand then put the picture you want them
43:53SPEAKERto look like they're in
43:54SPEAKERand you can make whatever you want out of that.
43:56SPEAKERAnd it's profoundly bad.
43:58SPEAKER- But that is another long-term concern.
43:59SPEAKERI mean, we're nowhere near there yet.
44:01SPEAKERThere's plenty of uncanny valley left to traverse here,
44:04SPEAKERbut like the potential to poison, to further poison
44:07SPEAKERthe information environment is also pretty alarming,
44:09SPEAKERsay 20 years down the line or something.
44:11SPEAKERIf you're able to, you know, right now, like you said,
44:14SPEAKERright now, it takes a ton of work and iteration
44:17SPEAKERand processing just to spit an image out that's usable.
44:20SPEAKERIt's going to be a long time before you can just request
44:22SPEAKERa realistic video of a politician doing something awful
44:25SPEAKERor whatever, but it doesn't seem like it will stay
44:29SPEAKERthat way forever.
44:30SPEAKER- As long as you don't show their hands,
44:31SPEAKERit's not gonna be that hard.
44:32SPEAKER- Yeah.
44:35SPEAKER- Well, the data cleanliness thing,
44:38SPEAKERthe sanitation of the data is another interesting place.
44:40SPEAKER'Cause like these tools are widespread.
44:43SPEAKERThe images aren't particularly well fingerprinted
44:45SPEAKERat this point.
44:47SPEAKERI mean, they're starting to be,
44:48SPEAKERbut there's a whole bunch of stuff out there
44:50SPEAKERthat wasn't up until this point.
44:52SPEAKERAnd like the analogy that people are using
44:56SPEAKERis the value of pre-World War II steel.
44:58SPEAKERSo if you want to do any work with a metal box,
45:02SPEAKERif you want to make a machine that tests radiation
45:06SPEAKERat a low level, you got to build a box
45:09SPEAKERthat's made out of steel that was smelted
45:13SPEAKERbefore the first nuclear bomb was blown up
45:15SPEAKERand here's Hiroshima.
45:17SPEAKER- Go on, this sounds fascinating, no?
45:20SPEAKER- Oh yeah, so like Geiger counters and research stuff,
45:25SPEAKERall is built with metal that was smelted
45:28SPEAKERbefore the Hiroshima bomb blew up.
45:31SPEAKER'Cause as soon as that happened,
45:32SPEAKERthere's radioactive active particulate in the air
45:35SPEAKERthat's just part of the increased background radiation
45:38SPEAKERof our world means that post-Hiroshima steel
45:41SPEAKERis more radioactive.
45:42SPEAKERSo there's going to be more signal
45:44SPEAKERunless you can't get as clean a signal out of that.
45:47SPEAKERSo people literally go around and salvage World War I,
45:53SPEAKERships sunk during World War I for the steel
45:57SPEAKERto then make scientific equipment and stuff out of.
46:00SPEAKER- Wow, geez, I have no idea.
46:02SPEAKERI mean, I don't know what the half-life is
46:03SPEAKERof this particular type of radiation or radioactive.
46:05SPEAKER- 200,000 years.
46:06SPEAKER- I was gonna say, it's probably not something we can just--
46:09SPEAKER- It's not long.
46:10SPEAKER- It's probably not something we can just wait out
46:11SPEAKERto get more usable steel back anyway.
46:14SPEAKER- But the point is, we're doing the same thing right now
46:17SPEAKERwith all the data on the internet.
46:18SPEAKERUp until this point, like up until five years ago,
46:21SPEAKERwe had clean data about what,
46:23SPEAKERwe could assume that everything that was on the internet
46:25SPEAKERwas probably made by people, right?
46:27SPEAKERWhich is what all these companies have done
46:29SPEAKERto make the models that, you know,
46:30SPEAKERthey built their corpuses out of this data.
46:34SPEAKERThe downside to the thing now
46:38SPEAKERis everybody's just posting the pictures
46:39SPEAKERthey're making in stable diffusion
46:40SPEAKERwhatever on their desktop computers.
46:42SPEAKERSome of that stuff's fingerprinted.
46:43SPEAKERSome of the tech stuff is fingerprinted.
46:45SPEAKERBut for the most part, eh, a lot of it,
46:48SPEAKERa lot of especially the early stuff
46:49SPEAKERis completely unlabeled in a way that like,
46:54SPEAKERwe're gonna start poisoning the corpuses
46:55SPEAKERof the future machine learnings
46:57SPEAKERand that's gonna have unpredictable effects.
46:58SPEAKER- Oh God.
46:59SPEAKER- 'Cause it's like running JPEGs,
47:00SPEAKERit's like running JPEGs
47:01SPEAKERthrough a bunch of compression loops, right?
47:03SPEAKER- Oh no.
47:04SPEAKER- So like the big question I have is,
47:07SPEAKERwith the way all this stuff works,
47:08SPEAKERare we just reverting to mean, right?
47:10SPEAKERIs everything that we're training these machines on,
47:15SPEAKERhow much work are they doing to not revert to mean,
47:18SPEAKERoh God, there's even more ethical stuff.
47:20SPEAKERLike there's an infinity of ethical stuff here.
47:22SPEAKERBut our, you know, the reversion to mean question is,
47:30SPEAKERare we, when you put in attractive woman
47:36SPEAKERinto some of, especially some dollies,
47:38SPEAKERsome of the early machine learning models,
47:41SPEAKERyou just get a white person who projects female
47:49SPEAKERwith North American,
47:51SPEAKERthat who adheres to North American beauty standards.
47:54SPEAKER- Okay, I was gonna ask you to define
47:57SPEAKERwhat mean in particular you were talking about.
47:59SPEAKER- Oh, average in this case.
48:00SPEAKER- Reverting to, but that makes it pretty clear.
48:02SPEAKER- Yeah, which is not, just to be clear,
48:05SPEAKERwhite woman who lives on the East Coast in New York,
48:08SPEAKERin the United States, is not the worldwide attractive mean
48:12SPEAKERfor attractiveness, is my understanding.
48:14SPEAKER- That's what I meant about defining the mean here, yes.
48:17SPEAKER- Yeah, but those are the people
48:19SPEAKERthat the most pictures have been taken of them
48:20SPEAKERand probably have posted the most pictures on Instagram
48:22SPEAKERor wherever else they stole the pictures from.
48:24SPEAKERSo here we have it.
48:26SPEAKER- I mean, what you're really talking about is the AI
48:29SPEAKERjust modeling existing human biases and proclivities.
48:33SPEAKER- Well, the AI specifically models the data that it's given.
48:36SPEAKER- Right. - Right.
48:37SPEAKERSo if you take your corpus from,
48:39SPEAKERif you build your corpus from Instagram,
48:41SPEAKERwhich is predominantly used in the West,
48:43SPEAKERyou're gonna get a different result
48:44SPEAKERthan if you take it from WeChat,
48:45SPEAKERwhich is used in the East.
48:47SPEAKERI mean, this is China specifically, I guess.
48:49SPEAKER- Craig, if I'm wrong, what was the Microsoft?
48:51SPEAKERWas it Tay?
48:53SPEAKER- No, the Nazi one?
48:54SPEAKER- Microsoft's Nazi bot.
48:56SPEAKER- Yeah, I mean, look, that's happening now with GPT, right?
49:00SPEAKER- That's the title for a podcast if I've ever heard one.
49:04SPEAKER- Well, the GPT stuff, people are using the Ask GPT feature
49:07SPEAKERthat we talked about earlier in this episode
49:09SPEAKERto get around, 'cause it doesn't,
49:10SPEAKERit's not subject to spam filter,
49:12SPEAKER'cause it comes through the bot that it's a moderator.
49:13SPEAKER- Right. - And as a result,
49:15SPEAKERthey were using that to post Nazi shit
49:17SPEAKERin discords that have that bot turned on.
49:19SPEAKERWe don't have that problem, just to be clear.
49:21SPEAKER- That's what I was getting at with Tay, right,
49:22SPEAKERwas there was no constraint or filtering
49:24SPEAKERon what they were allowing to enter into its model, right?
49:28SPEAKER- Yeah, and to be clear, the people that I followed
49:32SPEAKERon Twitter that have been talking about this stuff
49:33SPEAKERfor 20 years were like, yeah, you can't post
49:35SPEAKERan unrestrained chatbot on the internet,
49:36SPEAKER'cause it's gonna be a Nazi in about five minutes.
49:38SPEAKERAnd it turns out they were right.
49:39SPEAKERIt took 90 minutes, I think, if I remember right.
49:41SPEAKERBut anyway, the whole thing is really complicated.
49:51SPEAKERIs this gonna result in just mass homogenization
49:54SPEAKERof all of the, whatever you feed into it,
49:57SPEAKERis one of my big questions.
49:59SPEAKERAnd it seems like that's one of the reasons
50:01SPEAKERI'm not as worried about it, actually,
50:02SPEAKERis that if you can generate stuff, art, words, whatever,
50:11SPEAKERout of the corpus of human creativity up until this point,
50:16SPEAKERthen all of a sudden the people that can make stuff
50:17SPEAKERthat live outside of that space and that are novel
50:20SPEAKERare going to be even more valuable than they were before.
50:23SPEAKERRight? - Sure.
50:26SPEAKER- Now, is it possible to beat the brunt of,
50:31SPEAKERdoes anybody ever invent anything new?
50:33SPEAKERDoes anybody ever make new art?
50:34SPEAKERIs every drawing that's ever been done made before?
50:37SPEAKERProbably, I don't know.
50:38SPEAKER- Every story was told centuries ago.
50:40SPEAKER- Yeah, there's only seven stories anyway.
50:42SPEAKERSo it's weird, there's a bunch of companies.
50:46SPEAKERI'm profoundly uncomfortable with companies
50:49SPEAKERthat involve Peter Thiel and Elon Musk
50:52SPEAKERhaving say in this type of thing.
50:56SPEAKERMusk is no longer on the board of OpenAI,
50:58SPEAKERfor what it's worth. - Really? Interesting.
50:59SPEAKER- He went away, he's still an investor,
51:03SPEAKERbut he's no longer part of the decision-making process there.
51:10SPEAKERThe interesting thing about this whole thing
51:13SPEAKERis now that we know that he's kind of a dope,
51:16SPEAKERreading the quotes about the foundation of OpenAI
51:20SPEAKERthat were written by him,
51:22SPEAKERI'm reading them now in the context of,
51:24SPEAKERmaybe he watched "Terminator" the night before?
51:26SPEAKER- That's entirely possible.
51:27SPEAKER- So they asked him for this?
51:29SPEAKER- Look, glass of red wine, some Ambien magic.
51:32SPEAKER- Got real jacked up on mushrooms and watched "Terminator."
51:34SPEAKERI was like, fuck man, we gotta do something about the AI.
51:39SPEAKERThere's a bunch of other companies involved too.
51:40SPEAKERThere's a bunch of people
51:41SPEAKERthat are not posting all their stuff.
51:43SPEAKERI mean, OpenAI posting Dolly and chat GPT
51:46SPEAKERand all that stuff is good
51:49SPEAKERbecause it does let people see what's actually going on
51:51SPEAKERand let people understand
51:52SPEAKERand let this be a gradual transition
51:54SPEAKERrather than something terrifying
51:56SPEAKERwhere the machines just start telling us what to do.
52:00SPEAKERBut there's also a ton of companies
52:02SPEAKERthat have taken as much or more money
52:05SPEAKERthat are doing this stuff quietly
52:06SPEAKERwithout publicizing what they're working on.
52:09SPEAKER- Right, like, I mean, it's the thing, right?
52:11SPEAKERLike it's the thing you have to have now, right?
52:13SPEAKERLike I assume in the kind of like valley group think,
52:17SPEAKERit's like if you don't have an AI strategy,
52:18SPEAKERyou're nothing, right?
52:19SPEAKERI'm guessing, but I assume that is the notion
52:22SPEAKERin the investor class at this point.
52:24SPEAKER- But like I said,
52:25SPEAKERthey probably are already made their bets at this point
52:27SPEAKERfor most of them.
52:28SPEAKERThey try a little bit of money squirreled away,
52:30SPEAKERbut now we're in the Gartner hype cycle, right?
52:33SPEAKERWhere there's a huge round of investment
52:35SPEAKERand then nothing really comes of it.
52:37SPEAKERLike we'll see stuff out of this first round.
52:40SPEAKERNow, the real test of how screwed we are
52:42SPEAKERis does the hype cycle accelerate?
52:44SPEAKER'Cause usually the hype cycle is like a five-year thing,
52:46SPEAKERright?
52:47SPEAKERPeople get excited about it.
52:48SPEAKERThere's a huge round of investment.
52:49SPEAKERNothing happens for a long time.
52:51SPEAKERPeople get really down on it
52:52SPEAKERand everybody starts really talking smack about this.
52:54SPEAKERLike if you think about smartphones,
52:56SPEAKERthis period is like 2007, 2005 to 2010, right?
53:02SPEAKERAnd then the iPhone 4 comes out
53:04SPEAKERand all of a sudden you look around
53:06SPEAKERand everybody has an iPhone.
53:07SPEAKERBut at least in San Francisco,
53:08SPEAKERnormal places, it took a couple more years.
53:13SPEAKERIf they accelerate that process
53:15SPEAKERfrom the normal like two to five years,
53:18SPEAKERVR is still is 10 years in almost.
53:20SPEAKERAnd we're still not through the trough of despair.
53:24SPEAKERSo with the sad places
53:25SPEAKERwhere the investment isn't paying off yet.
53:31SPEAKERIf they accelerate that trough,
53:33SPEAKERthey accelerate through that trough, all bets are off.
53:37SPEAKER'Cause people will spend more money.
53:39SPEAKERAnd when that happens,
53:41SPEAKERit's like who fucking knows what's gonna happen next?
53:44SPEAKER- Yes, I think that is the gist
53:45SPEAKERor the theme of this whole thing
53:46SPEAKERis who fucking knows what's gonna happen next?
53:48SPEAKERThere was one thing, one distinction I wanted to make,
53:50SPEAKERlike when we talk about the whims of investors here,
53:52SPEAKERI think you're primarily talking about venture capitalists
53:55SPEAKERwho are funding new ventures, right?
53:57SPEAKERBut there's another component of that,
53:59SPEAKERwhich is the investors as a block
54:02SPEAKERof massive existing public companies
54:04SPEAKERthat influence the actions of those companies
54:07SPEAKERby I guess, by way of the board of directors,
54:09SPEAKERI assume is generally how that works.
54:10SPEAKERBut like point I'm making is like,
54:12SPEAKERinvestors are going to want to demand to know
54:15SPEAKERlike what is Microsoft's AI strategy?
54:16SPEAKERWhat is Google's like mass?
54:19SPEAKER- So Microsoft's put like 10 or $15 billion
54:21SPEAKERinto open AI and some other companies.
54:24SPEAKER- Like I'm asking this in the wake of like
54:25SPEAKERthese mass layoffs that have just occurred already,
54:28SPEAKERwhich make it very evident that like the people
54:31SPEAKERwho hold the money and the capital in this arrangements
54:34SPEAKERare quite willing to slash and burn
54:37SPEAKERthe second it is useful or beneficial
54:39SPEAKERor even feels beneficial.
54:41SPEAKER- Point I'm making is like, you know, again,
54:44SPEAKERlike mass disruption to the workforce
54:45SPEAKERis the thing that probably concerns me the most here.
54:48SPEAKERAnd like, that's what I'm getting at
54:50SPEAKERis just like there could be this groundswell of like,
54:53SPEAKERhey, what are you doing with AI?
54:55SPEAKERMore to the point, how are you going to start cutting jobs
54:58SPEAKERand replacing it with AI?
55:00SPEAKER- Well, yeah, and I think, I mean,
55:03SPEAKERI think that's the big question, I think.
55:07SPEAKERI think that's one of the big questions.
55:10SPEAKERI also, there's like big outstanding copyright questions.
55:13SPEAKERI think, I really do think that we're gonna see
55:15SPEAKERsome copyright lawsuits against these companies.
55:17SPEAKER- Those are brewing, I've seen talk about class actions
55:22SPEAKERcoalescing among artists who were able to prove
55:24SPEAKERthat their work was used to train existing models.
55:27SPEAKER- Yeah, and I think we'll see some rules
55:29SPEAKERaround disclosure, something like that.
55:30SPEAKERThe problem with this is it's all so early
55:32SPEAKERand the markets, like nobody's making money on it really
55:36SPEAKERto speak of yet, so there's not damages to be claimed,
55:40SPEAKERwhich means the copyright stuff will be kind of hold
55:42SPEAKERuntil that happens.
55:45SPEAKERThe legislation side of it's the other thing too,
55:47SPEAKER'cause like Congress is gonna fuck this up really bad
55:49SPEAKERfor a long time.
55:51SPEAKER- Even a normal Congress would probably not get it right,
55:55SPEAKERlet alone the force that is taking place now.
55:59SPEAKER- Well, the push pull on, hey, we don't want to fall behind
56:03SPEAKERother countries in this space, which is clearly important
56:06SPEAKERfor the future versus, hey, Hollywood wants us
56:09SPEAKERto protect their profits, the copyright arm wants us
56:14SPEAKERto protect the profits of Hollywood.
56:16SPEAKERLike the first time you could get a new Spider-Man,
56:18SPEAKERbut actually we're there, you can watch Watch Me Forever
56:21SPEAKERon Twitch right now and watch endless low poly Seinfeld
56:24SPEAKERon Twitch.
56:25SPEAKER- Yes, that's exactly what that is.
56:27SPEAKERRuns in Unity, if you haven't seen it,
56:28SPEAKERthe Twitch channel, Watch Me Forever.
56:31SPEAKER- Yeah, it's a show about nothing.
56:32SPEAKERI think we only know it runs in Unity
56:34SPEAKERbecause they had to restart it on stream at one point.
56:36SPEAKER- Oh, really?
56:37SPEAKER- Yes.
56:39SPEAKER- But it's a generated TV show.
56:41SPEAKERLike the characters talk, the characters create dialogue.
56:47SPEAKERThe dialogue is created by the machine.
56:49SPEAKERThe characters act it, perform it, the shots are made.
56:51SPEAKERThey change sets occasionally.
56:53SPEAKERThere doesn't seem to be any plot at all long-term,
56:55SPEAKERbut there's like a series of scenes with jokes
56:58SPEAKERand then normal TV show punchlines
57:03SPEAKERand laugh track at the end.
57:05SPEAKER- Yes, well, attempts at normal TV show punchlines.
57:08SPEAKERI wouldn't say many of them are successful
57:09SPEAKERor even comprehensible, but.
57:11SPEAKER- Well, they don't use the punchlines correctly,
57:13SPEAKERbut they often use the punchlines that you've heard.
57:15SPEAKERLike it wouldn't surprise me to hear somebody say bazinga
57:18SPEAKERon this at some point.
57:20SPEAKER- God, sure, yes.
57:22SPEAKER- Right, like it's a real weird thing.
57:26SPEAKERBut my point is when Congress,
57:32SPEAKERCongress is gonna have to decide
57:33SPEAKERwhether they want to invest in the future,
57:35SPEAKERwhether they wanna support the future,
57:37SPEAKERwhich is gonna involve like finally fucking up
57:40SPEAKERthe copyright rules that have frankly caused some problems
57:43SPEAKERfor us as a culture over the last hundred years,
57:46SPEAKERor giving up on AI and letting Europe and China lead instead,
57:53SPEAKERwhich I don't think they're gonna do.
57:54SPEAKER- Right.
57:56SPEAKER- In a more robust regulatory environment,
57:59SPEAKERI would have not as many concerns about this, but.
58:02SPEAKER- I mean, but at the same time, it's a scary thing.
58:06SPEAKER'Cause if like, like you don't wanna be the third one
58:09SPEAKERto the AI party.
58:10SPEAKER- Yeah, fair.
58:11SPEAKER- 'Cause if they started like, so, okay,
58:14SPEAKERhere's the other side of it.
58:15SPEAKERThis is, we're like into the 10th dimension here.
58:18SPEAKERI apologize, we're all over the place with this episode,
58:19SPEAKERbut there's a lot.
58:21SPEAKERThe GPT-2 stuff, you're not just, you don't,
58:24SPEAKERlike I used it to generate a blog post on my,
58:28SPEAKERon shitty.blog, shout out to shitty blog.
58:33SPEAKERSomebody, I posted a tomato soup recipe
58:35SPEAKERand somebody was like, hey, I wish,
58:36SPEAKERI don't read recipes on the internet.
58:37SPEAKERLet's say I have a really long meandering story at the start
58:40SPEAKERabout how important tomato soup is to you as a person.
58:43SPEAKERSo I put into the GPT-2 stuff,
58:49SPEAKERthe prompt was generate a long meandering blog post
58:54SPEAKERfull of, a long and meandering introduction
58:57SPEAKERfull of questionable history and moderately amusing
58:59SPEAKERanecdotes for people who like food blogs
59:01SPEAKERcourtesy of chat GPT.
59:03SPEAKERAnd it wrote a thousand word thing about like the history
59:08SPEAKERof tomato soup, told a story about an amusing anecdote
59:13SPEAKERabout the first time I had great tomato soup as a kid
59:16SPEAKERand how the other kids made fun of me for that.
59:18SPEAKER- Tomato soup has a rich history that dates back
59:20SPEAKERto the 18th century.
59:21SPEAKERIt was first created as a way to preserve tomatoes,
59:23SPEAKERwhich were not yet considered a food item
59:25SPEAKERand were mostly used for decoration.
59:27SPEAKER- No idea if that's true. - That's funny.
59:29SPEAKER- I have no idea if any of this is true.
59:32SPEAKERLike, yeah, anyway, so it talks about grilled cheese
59:36SPEAKERbeing an important pairing for tomato soup.
59:38SPEAKERIt talks about the use of pesticides
59:40SPEAKERand genetically modified seeds in the farming industry.
59:43SPEAKERLike, I honestly, I couldn't tell the difference
59:48SPEAKERbetween this and something that like a normal human wrote
59:50SPEAKERon some shitty recipe blog where they were ripping off
59:52SPEAKERrecipes from other sites.
59:53SPEAKER- Right, except for the factual basis of all of this
59:57SPEAKERis like pretty clearly not sound.
1:00:01SPEAKER- I don't know if it is or not.
1:00:02SPEAKERIt could be a hundred percent true.
1:00:04SPEAKER- I have to assume people considered tomatoes food
1:00:06SPEAKERearlier than the 18th century.
1:00:10SPEAKER- I mean, I could go to Wikipedia and find out
1:00:12SPEAKERit's an edible berry.
1:00:14SPEAKER- Yeah, I mean, sure, only kind of,
1:00:17SPEAKERis husbandry the right term for agricultural?
1:00:20SPEAKER- Botany.
1:00:20SPEAKERI don't know what kind of-
1:00:23SPEAKER- Horticulture.
1:00:24SPEAKER- Yeah, the manipulation of crops successively.
1:00:27SPEAKERI don't know what the term is there,
1:00:28SPEAKERbut I don't know when the,
1:00:30SPEAKERmaybe the tomato came about much later than I thought.
1:00:35SPEAKER- Yeah, I'm on the Wikipedia page right now.
1:00:40SPEAKER500 BC, see it's wrong.
1:00:41SPEAKER- Okay, yes, I kind of assumed.
1:00:45SPEAKER- Yeah, tomatoes were not grown in England until the 1590s.
1:00:48SPEAKERSee, it's complete horse shit, right?
1:00:50SPEAKERThis is the thing, you can't, but it's good enough.
1:00:52SPEAKERYou can't tell.
1:00:54SPEAKERSo yeah, anyway.
1:00:56SPEAKER- Yeah, I think I like another random example
1:00:59SPEAKERthat comes to mind.
1:00:59SPEAKERI don't know for a fact they used machine learning for this,
1:01:02SPEAKERbut I have to assume the middleware they used
1:01:05SPEAKERfor Darth Vader's voice in the Obi-Wan series
1:01:07SPEAKERwas very famously generated.
1:01:10SPEAKER- Oh yeah, that was absolutely machine learning.
1:01:12SPEAKERThey fed in a bunch of James Earl Jones,
1:01:14SPEAKERDarth Vader voices and a hundred percent.
1:01:16SPEAKERAnd- - There's so many companies
1:01:20SPEAKERworking on that, and we have friends who do that.
1:01:21SPEAKERIt's hard, but it's a short problem at this point, basically.
1:01:25SPEAKER- My understanding, and it's proprietary,
1:01:26SPEAKERso they didn't get a lot into the technology that I've seen.
1:01:28SPEAKERI would love to know more about it.
1:01:29SPEAKERI think that, I believe that was like kind of run
1:01:32SPEAKERagainst actual voiceover.
1:01:34SPEAKERIt might've been James Earl Jones doing acting
1:01:37SPEAKERthat they then applied a kind of younger sounding model to,
1:01:40SPEAKERor it may have been another actor,
1:01:41SPEAKERI don't think they've said.
1:01:42SPEAKERBut my sense was this was kind of a vocal marionette
1:01:45SPEAKERsituation of somebody was saying the lines
1:01:47SPEAKERand then they made the Darth Vader voice
1:01:49SPEAKERsay those same lines in the same way.
1:01:51SPEAKERAnd it was shockingly convincing.
1:01:53SPEAKERLike I don't think it was acted exactly the way
1:01:55SPEAKERthat it would have, intonation, dramatic flare,
1:02:00SPEAKERmaybe not quite the same as 1977 or whatever,
1:02:03SPEAKERbut like the tone and the pitch of the voice,
1:02:05SPEAKERthe resonance of it, shockingly convincing.
1:02:08SPEAKER- So there's two ways that people do that.
1:02:12SPEAKERThe one is that they record another actor
1:02:14SPEAKERand then they run the filter through it
1:02:15SPEAKERto make it sound like the new one.
1:02:18SPEAKERThe other is it uses a thing called,
1:02:19SPEAKERI think it's Voice ML or something like that,
1:02:22SPEAKERor Speech ML, which is a markup language
1:02:25SPEAKERthat basically you write a script
1:02:27SPEAKERand then you write the inflections
1:02:29SPEAKERthat you wanna have using markup language
1:02:30SPEAKERat different parts of the script.
1:02:33SPEAKERThe tools for that stuff are really bad right now.
1:02:35SPEAKERThere's opportunities there probably.
1:02:36SPEAKERBut anyway, yeah, I have no doubt
1:02:39SPEAKERthat that was a machine learning thing.
1:02:41SPEAKER'Cause there's enough James Earl Jones from 1977 dialogue
1:02:44SPEAKERthat they could probably generate something
1:02:45SPEAKERpretty useful off of it now.
1:02:51SPEAKERSo the point is, the point of the tomato soup example
1:02:55SPEAKERis that that's the idiot way to use GPT.
1:02:59SPEAKERYou can type Python.
1:03:00SPEAKERYou can ask for Python queries that do things.
1:03:02SPEAKERYou can ask for bash scripts.
1:03:04SPEAKERYou can ask for C++ that'll run in the Unreal engine.
1:03:08SPEAKERYou can ask for, the machines can write code.
1:03:13SPEAKERSo presumably at some point,
1:03:14SPEAKERsomebody's gonna be like, hey, what if we write code
1:03:15SPEAKERthat makes the code that is running here better?
1:03:18SPEAKERAnd then we just keep doing that
1:03:19SPEAKERuntil they get better and better and better.
1:03:22SPEAKEROr maybe we revert more toward the mean.
1:03:25SPEAKERI don't know.
1:03:26SPEAKERIt remains to be seen.
1:03:26SPEAKERIt's like, yeah, the whole thing's messy.
1:03:31SPEAKERGoogle is apparently freaked out
1:03:32SPEAKERbecause people are starting to use GPT-2 instead of Google
1:03:34SPEAKERto find things on the internet
1:03:35SPEAKERsince Google kind of sucks now.
1:03:37SPEAKER- Interesting. - And GPT-2
1:03:38SPEAKERis possibly apocryphal but generates more believable results.
1:03:44SPEAKER- I mean, we're edging toward more truthiness, Brad.
1:03:48SPEAKER- I do, seriously.
1:03:49SPEAKERIt's like, it's that whole thing of like,
1:03:50SPEAKERah, the things are this way.
1:03:52SPEAKERNo, they're not.
1:03:53SPEAKERWell, doesn't it mean something that I thought they were?
1:03:56SPEAKERLike that the whole thing is just like,
1:03:57SPEAKERhey, whatever feels good to believe is, yeah.
1:04:01SPEAKER- Yeah, and presumably all of the big tech companies
1:04:04SPEAKERare also doing stuff here just quietly in the background
1:04:06SPEAKERbecause either they're behind or like you look at Google
1:04:11SPEAKERand they should have a corpus of data
1:04:13SPEAKERthat's bigger than anybody else's in the world's probably.
1:04:19SPEAKERIt's all weird, man.
1:04:20SPEAKER- It sure is where things are at say five years from now,
1:04:23SPEAKER10 years from now, 20 years from now.
1:04:25SPEAKER- Well, so the moon room had asked in the Discord
1:04:28SPEAKERif we could do a predictions on AI episode.
1:04:30SPEAKER- I don't even like, that's really the kind of basis
1:04:35SPEAKERof my concern is that the number of different domains
1:04:38SPEAKERthis keeps spreading into and seemingly disrupting,
1:04:42SPEAKERwe're getting close to disrupting, just keeps growing.
1:04:44SPEAKER- Like singularity.
1:04:45SPEAKER- Like I don't, what other categories
1:04:48SPEAKERcould we see this deployed in like all of them?
1:04:51SPEAKER- Yeah, I think probably all of them.
1:04:52SPEAKER- Like does AI start, do they start relying on AI
1:04:55SPEAKERto make economic and political decisions and set policy?
1:04:58SPEAKER- I think we're probably already doing
1:04:59SPEAKERsome of the political, the economic decisions.
1:05:01SPEAKER- I'm sure there's a lot of like the stock market stuff
1:05:05SPEAKERis all based on machine learning stuff.
1:05:08SPEAKER- Yeah, and I assume there's a massive amount
1:05:10SPEAKERof machine learning and data science being used.
1:05:12SPEAKERSo like some of that stuff is happening already,
1:05:14SPEAKERbut I'm talking less like, hey, I'm gonna use this tool
1:05:16SPEAKERto profile data more efficiently
1:05:18SPEAKERand more I'm gonna ask this thing
1:05:19SPEAKERwhat the inflation rate should be and trust it.
1:05:23SPEAKERLike, I don't know.
1:05:25SPEAKER- Yeah, I mean, I think, I would hope,
1:05:30SPEAKERone would hope that you would use this stuff
1:05:32SPEAKERfor things that are like a good place for this
1:05:35SPEAKERis to use it for image recognition on factory lines
1:05:39SPEAKERto find flawed parts and to scan x-rays
1:05:41SPEAKERto find, you know, bubbles and stamped metal,
1:05:45SPEAKERyou know, cast metals and stuff like that, right?
1:05:48SPEAKERLike that's the thing that it should be able to do
1:05:49SPEAKERand probably is gonna be as reliable
1:05:51SPEAKERor like a supplement, a really great supplement to people.
1:05:56SPEAKERWe're seeing like the bad example is that the dumb ass
1:06:00SPEAKERwho hooked up a chat GPT to his person to person therapy app
1:06:07SPEAKERand did some human experimentation without consent
1:06:14SPEAKERbecause he, instead of connecting people to people
1:06:17SPEAKERfor therapy, which is an established valuable tool,
1:06:22SPEAKERhe connected people to this ambiguously good chat bot
1:06:29SPEAKERand didn't tell them they were talking to a machine.
1:06:35SPEAKERSo like, again, it's the concentration of the ability
1:06:39SPEAKERto do something stupid and bad
1:06:40SPEAKERin a relatively small number of hands.
1:06:43SPEAKERLike if that person had done that,
1:06:46SPEAKERif that person wanted to, you know, have dogs do therapy,
1:06:49SPEAKERthey would have had to first get a thousand dogs, right?
1:06:52SPEAKERAnd it's hard to get a thousand dogs.
1:06:54SPEAKER- Dude, we should get a thousand dogs.
1:06:56SPEAKERI don't know, that sounds pretty awesome.
1:06:58SPEAKER- It's a lot of poop, right?
1:06:59SPEAKER- It's true, there are downsides.
1:07:01SPEAKER- Yeah.
1:07:02SPEAKER- Since we're, real quick before we wrap,
1:07:04SPEAKERsince we're talking predictions, like, I mean,
1:07:05SPEAKERthis would require an extremely sound track record to trust,
1:07:09SPEAKERbut like I could see this being useful in medical science
1:07:12SPEAKERand healthcare at some point, specifically in the context
1:07:17SPEAKERof preventative medicine and kind of catching stuff early.
1:07:21SPEAKER- Like, yeah, but if you can use it for that,
1:07:23SPEAKERthen the insurance companies can use that
1:07:25SPEAKERto determine what your actuarial rate is, right?
1:07:27SPEAKER- Yes, everything has got a downside.
1:07:30SPEAKER- Yeah, it's like, I imagine that there's people
1:07:33SPEAKERdoing like radiology work right now
1:07:35SPEAKERthat are building software radiology assistance
1:07:38SPEAKERthat will help radiologists do their jobs better, right?
1:07:40SPEAKER- But that's the type of stuff I'm talking about.
1:07:42SPEAKER- Yeah, I would, like, as much as I don't like Elon Musk
1:07:49SPEAKERand Peter Thiel, I do tend to agree that they're,
1:07:51SPEAKERwith the initial assortation that Sam Altman
1:07:54SPEAKERand the people at OpenAI made, which is that this,
1:07:57SPEAKERthis is a technology that could be a powerful force for good
1:08:00SPEAKERor a powerful force for evil.
1:08:02SPEAKERAnd it's on us to make sure that we,
1:08:05SPEAKERwe make sure it's the former and not the latter.
1:08:08SPEAKER- I have bad news about all of history.
1:08:11SPEAKER- It's gonna get cheaper and easier to do it.
1:08:12SPEAKERI mean, the one thing that we have going for us
1:08:14SPEAKERis the corpuses are really, really, like they're expensive
1:08:17SPEAKERand time consuming to make and training the models
1:08:20SPEAKERis hard and time consuming.
1:08:22SPEAKERThe bad thing is I think the people who do it first
1:08:24SPEAKERwill probably maintain the lead forever
1:08:26SPEAKERbecause you'll always be a generation or two ahead
1:08:28SPEAKERof everybody else or so.
1:08:30SPEAKER- Anyway, until that company's AI takes over
1:08:33SPEAKERand becomes Skynet and then we're all done.
1:08:36SPEAKER- Yeah, I mean, look, some dipshit probably has a startup
1:08:39SPEAKERcalled Skynet Industries right now because like, yeah.
1:08:42SPEAKER- Yes, in a world where they named a surveillance company
1:08:45SPEAKERPalantir, yes, that absolutely is going to happen.
1:08:47SPEAKER- Yeah, and a food replacement Soylent.
1:08:49SPEAKERLike we're, the stupids, the stupids are everywhere, man.
1:08:52SPEAKER- You know what, we have to wrap this up.
1:08:54SPEAKERYou know what I always thought
1:08:55SPEAKERwas an interesting thought experiment about Terminator
1:08:58SPEAKERand Skynet was like Skynet was just a,
1:09:01SPEAKERessentially like a missile trajectory processing AI, right?
1:09:05SPEAKER- Yeah, it was a computer from War Games, whopper.
1:09:07SPEAKER- So like how it got from that to being able to manufacture
1:09:10SPEAKERbipedal robots and stuff like that is really interesting.
1:09:12SPEAKERLike how did it commandeer production lines?
1:09:14SPEAKERHow did it retool production lines?
1:09:16SPEAKERSkynet doesn't have arms and legs.
1:09:18SPEAKERThat's what I'm saying.
1:09:19SPEAKERIt's like Skynet isn't a thing that walks around
1:09:21SPEAKERwith opposable thumbs.
1:09:24SPEAKERIt can't manipulate things itself.
1:09:25SPEAKERLike it's just fun to think about like,
1:09:27SPEAKERwhat were the many baby steps required
1:09:30SPEAKERto commandeer and retool production lines
1:09:33SPEAKERto start generating, making actual robots?
1:09:35SPEAKER- Look here, you want the biggest argument
1:09:37SPEAKERagainst self-driving cars ever.
1:09:39SPEAKERIf you want to kill the most people as fast as possible,
1:09:42SPEAKERwhat you do, if you're Skynet, not if you're me, obviously.
1:09:45SPEAKERIf you're Skynet is you take over the self-driving cars.
1:09:49SPEAKER'Cause there'll be a bazillion of those around.
1:09:51SPEAKEREverybody will have them.
1:09:52SPEAKERAnd you just start smashing them into people,
1:09:54SPEAKERbuildings, whatever.
1:09:55SPEAKER- Stephen King. - It's really grim.
1:09:56SPEAKER- Stephen King wrote that story like 40 years ago.
1:09:59SPEAKER- Really? - Yeah.
1:10:00SPEAKER- Well, there's a guy named Daniel Suarez
1:10:02SPEAKERwho wrote another one of those about 15 years ago,
1:10:04SPEAKERprobably now.
1:10:05SPEAKER- This was not, this was more of a supernatural,
1:10:07SPEAKERoh God, all the vehicles have come to life
1:10:09SPEAKERand are killing us. - Oh.
1:10:10SPEAKER- Yeah, the Christine parable.
1:10:12SPEAKER- Yes, the trucks.
1:10:14SPEAKER- Yeah. - Yes.
1:10:14SPEAKER- No, 18-wheeler is terrifying.
1:10:16SPEAKERAnyway, you know what's not terrifying, Brad?
1:10:19SPEAKER- What's that?
1:10:20SPEAKER- Our wonderful patrons. - Oh, they're great.
1:10:22SPEAKER- As always, Brad Will Made a Tech Pod
1:10:24SPEAKERis a 100% listener-supported show,
1:10:26SPEAKERso we appreciate the generosity of all of our patrons.
1:10:30SPEAKERThank you, patrons.
1:10:31SPEAKER- Thank you, thank you, thank you.
1:10:33SPEAKERI said it three times, that's how you know it sticks.
1:10:35SPEAKER- I mean, yeah, that's, they'll remember.
1:10:38SPEAKERYou gotta say it three times
1:10:39SPEAKERfor people to remember anything.
1:10:40SPEAKERAs always, we appreciate all of our patrons,
1:10:42SPEAKERbut we especially appreciate our executive producer
1:10:44SPEAKERto your patrons.
1:10:45SPEAKERWe thank every episode, including Nick Johnston,
1:10:48SPEAKERPaddle Creek Games, Makers of Fractured Veil,
1:10:50SPEAKERAndrew Slosky, Octothorpe Bunny Crimes, Just Wed,
1:10:54SPEAKERJoel Krauska, Twinkle Twinkie, David Allen,
1:10:56SPEAKERand James Gammock.
1:10:57SPEAKERThank you all so much.
1:10:59SPEAKERWe will be back next week
1:11:01SPEAKERwith another edition of the Tech Pod.
1:11:04SPEAKER- Yeah, I hope people enjoyed this rambly episode.
1:11:06SPEAKERWe were all over the place,
1:11:07SPEAKERand we're certainly not subject matter experts,
1:11:09SPEAKERbut let us know.
1:11:10SPEAKER- I mean, it's a weird, it's hard,
1:11:14SPEAKERit seems like in order, you have to, it's complicated.
1:11:21SPEAKERThis one's really complicated.
1:11:23SPEAKER- It is. - Yeah.
1:11:24SPEAKER- It sure is.
1:11:25SPEAKER- There's a lot to it.
1:11:27SPEAKERThere's a lot of implications to the technology
1:11:31SPEAKERand how it works and what it's going to mean for us.
1:11:33SPEAKERMy prediction is that what is five years looks like
1:11:37SPEAKERis not gonna be anything like what we expect it is right now.
1:11:40SPEAKER- Yeah, yes, I think I would conclude by saying,
1:11:43SPEAKERare we in the singularity?
1:11:44SPEAKERMaybe not yet.
1:11:45SPEAKERAre we on our way there?
1:11:46SPEAKERMaybe?
1:11:46SPEAKER- I think my hunch is that we have a slow singularity coming
1:11:51SPEAKERand we're probably already there.
1:11:54SPEAKERAnd also, probably it's not gonna be as bad as we think
1:12:00SPEAKERif we're being pessimistic today,
1:12:02SPEAKERbut it's not gonna be as good as we think
1:12:03SPEAKERif we're being optimistic today.
1:12:05SPEAKER- So what you're saying is if as long as we assume
1:12:07SPEAKERthe absolute worst, then what we get
1:12:09SPEAKERwill invariably be better.
1:12:12SPEAKER- Yeah, you see, you say that, but then 2016 happened
1:12:14SPEAKERand it turns out that was way worse than I expected it to be.
1:12:16SPEAKERSo yeah, I don't know, see you all next week.
1:12:19SPEAKERBye everybody.
1:12:21SPEAKER(upbeat music)