AI transcript
0:00:05 In the end of the day, when you think of the future, the first route is a utopia.
0:00:07 The second one is a dystopia.
0:00:10 The third one is something that I kind of envisioned.
0:00:14 Hey, welcome to the Next Wave Podcast.
0:00:15 My name is Matt Wolfen.
0:00:20 I’m here with my co-host, Nathan Lanz, and today we’ve got an amazing guest.
0:00:25 She has an amazing YouTube channel all about Python and how to code.
0:00:26 Her tutorials are amazing.
0:00:30 So many people learn Python from her, and we’re excited to have her on the show to
0:00:36 talk about the future of coding and how it overlaps with AIs.
0:00:40 When all your marketing team does is put out fires, they burn out.
0:00:44 But with HubSpot, they can achieve their best results without the stress.
0:00:50 Tap into HubSpot’s collection of AI tools, Breeze, to pinpoint leads, capture attention,
0:00:53 and access all your data in one place.
0:00:57 Keep your marketers cool and your campaign results hotter than ever.
0:01:03 Visit hubspot.com/marketers to learn more.
0:01:05 So welcome to the show, Maria Shaw.
0:01:06 Thank you for being on.
0:01:07 Yeah.
0:01:08 Absolutely.
0:01:09 Thank you for inviting me.
0:01:10 Great to meet you.
0:01:14 So let’s talk a little bit about the AI world, because I know, you know, when we were out
0:01:18 at GTC, the talk of the whole event was AI, right?
0:01:23 Like everything is AI, and obviously coding is one of those areas where AI has completely
0:01:24 taken over.
0:01:30 I know Nathan was one of the sort of early adopters of GitHub co-pilot, and yeah, I guess
0:01:33 I just want to know, what are your sort of like overall thoughts?
0:01:36 Let’s just kind of start from like the 30,000 foot like general overview.
0:01:39 Like what are your thoughts of like that overlap of AI and coding?
0:01:43 Are you excited that AI is making coding easier?
0:01:45 Are you worried about coders losing jobs?
0:01:48 Like where do you kind of stand on the whole thing right now?
0:01:53 I’m slightly confused because I’ve tried a few of the AI models to kind of see what
0:01:54 they’re all about.
0:01:59 And I don’t see a lot of difference between just kind of copying the prompt and pasting
0:02:00 it in a search engine.
0:02:04 Like I think the biggest difference I see is the fact that when you do so with a search
0:02:08 engine, you can see a whole bunch of sources and then you decide which one of these is
0:02:09 your favorite.
0:02:10 And you kind of go from there.
0:02:17 But I find these co-pilots as a middleman between the developer and the documentation.
0:02:20 And for me, being so nerdy, I really like documentation.
0:02:22 I really appreciate it.
0:02:23 It’s my favorite place to be.
0:02:26 And that’s how I write most of my tutorials.
0:02:31 So getting it from the source is more important to me than kind of saving a minute or so
0:02:32 or even less.
0:02:38 So that’s where I’m a bit confused because I hear a lot of my viewers really exciting,
0:02:44 really excited about ChedGPT, about co-pilot and Devin probably, even though I don’t even
0:02:49 know if it’s open to the wide public yet, but maybe Nathan can share his opinion about
0:02:50 it.
0:02:53 So that’s something I was thinking about too, for your YouTube channel.
0:02:58 It’s got to be– I’ve seen other YouTubers as well who are like– their channel is primarily
0:03:00 about coding that they’re very skeptical of it.
0:03:04 And it feels like there’s a slight conflict there, too, though, because if AI does get
0:03:09 so good that you don’t need to learn coding anymore, it’s like a lot of those channels
0:03:14 that have to pivot to some other kind of content, which I’m sure you’re coming from a genuine
0:03:15 place.
0:03:17 And yeah, there’s definitely major limitations right now.
0:03:22 And yeah, I just have a– like you guys were talking about the NVIDIA conference, Jensen
0:03:25 said that in the future, you won’t need coders.
0:03:29 When you think of the future, you need to think of all the possible routes that we can
0:03:30 go through.
0:03:37 And the whole purpose of artificial intelligence is to be competitive or better than us in
0:03:40 all the cognitive tasks that humans are engaged in.
0:03:42 And this was from the inception of AI.
0:03:44 There is nothing new about it.
0:03:52 This is in– people’s loss of jobs has been a very major ethical concern of AI.
0:03:55 And people were reasoning about it since the 1950s.
0:03:57 I noticed a very weird trend.
0:04:01 I’m kind of– I’m taking you guys off topic because when– chat GPT–
0:04:02 Let’s go.
0:04:06 Matt loves wrapped holes, so.
0:04:12 When chat GPT emerged, there was a wave of folks who were saying that it will never replace
0:04:13 us.
0:04:16 There were people saying that this will never happen.
0:04:18 And I was kind of warning people about it.
0:04:23 I was saying that, hey, if you’re filming videos and you’re saying on record that chat
0:04:29 GPT is better than you encoding, what are the chances that your employer will hear it?
0:04:30 Why would you do that?
0:04:33 Like, even if it’s true, why would you admit something like that?
0:04:37 And I found that a lot of people were telling me, Maria, you’re crazy.
0:04:38 It’s not going to happen.
0:04:39 You’re just paranoid.
0:04:42 You’re just speaking doomsday stuff.
0:04:48 Like now that folks like Jensen are saying very similar things, maybe they take me a
0:04:49 bit more seriously.
0:04:50 I’m not sure.
0:04:54 Yeah, there was a recent tweet from– I think his name is Ethan Malik, the professor from
0:04:58 Wharton, where he was sharing some kind of recent study that showed– I think he showed
0:05:03 that 85% of employees in polls are saying that they’re using chat GPT at work.
0:05:07 And then also, I think it was something like 77% of them don’t tell their employer.
0:05:12 So it’s like, a lot of people are using chat GPT at work for emails and all kinds of different
0:05:14 tasks, and they’re just not telling their bosses.
0:05:15 So yeah.
0:05:19 Well, it’s funny because I look at the whole scenario from somebody who doesn’t really
0:05:20 know code.
0:05:21 Like, I don’t know anything about Python.
0:05:23 I don’t know how to write JavaScript.
0:05:27 I know a little HTML and CSS, and that’s about it, right?
0:05:32 And so when I look at it from that perspective, I was actually able to get into chat GPT.
0:05:35 This was back when it was 3.5, I think.
0:05:40 And I was actually able to develop a game using JavaScript that was playable.
0:05:41 It had graphics.
0:05:46 I actually used mid-journey to generate some images, and it was this little side-scroller
0:05:50 game where you jumped and collect coins, and it looked really good because I used mid-journey
0:05:52 and all the code worked.
0:05:58 So from somebody who doesn’t know JavaScript at all, I was able to go from– I have an idea
0:06:02 for a real simple game I want to build to– I actually have a playable game, and I never
0:06:04 actually touched a line of code myself.
0:06:09 So I think, from the perspective of somebody who doesn’t code, I think it’s really exciting
0:06:12 to see, oh, maybe I’ll actually be able to code now.
0:06:16 From people that actually know how to code, it’s a different story, right?
0:06:20 So I’m at this Cisco conference now out in Vegas, so anybody who’s watching the video
0:06:23 and sees a different background, that’s why.
0:06:28 But somebody on stage was talking about how right now, if you know how to code, it’s still
0:06:34 actually faster to just code something than to use chat GPT, because if you use chat GPT,
0:06:38 it’ll write the code, and then you spend just as much time double-checking and debugging
0:06:40 the code to get it to work.
0:06:45 But from somebody who’s never been able to code JavaScript, all I did was go, “That didn’t
0:06:46 work.
0:06:47 What do I try next?
0:06:48 That didn’t work.
0:06:49 What do I try next?”
0:06:53 It took me hours to code the game, but I actually eventually got there without ever touching
0:06:54 a line of code myself.
0:06:59 So it’s kind of like– I can see why there– whereas there’s excitement around it from
0:07:03 non-coders, but I can also see why coders would be like, “Ah, right now it’s more of a nuisance
0:07:04 than it is helpful.”
0:07:06 For sure, for sure.
0:07:12 There are a few reasons why folks who are coding for living, they wouldn’t be using it, but
0:07:17 I think it’s cool that it gave you these superpowers that you weren’t able to use JavaScript.
0:07:21 You never learned it, and suddenly you have your own JavaScript game, which is amazing.
0:07:26 I find it to be an incredible way to take your creativity and kind of manifest it in
0:07:28 a way that you never thought you could.
0:07:29 So it’s pretty cool.
0:07:35 So in terms of the coding, the world of coding, there are a few reasons why.
0:07:39 It’s probably not a good idea to use these type of models for that, because our industry
0:07:41 is very dynamic.
0:07:43 Things change on a daily basis.
0:07:46 Just because something works now, it doesn’t mean that it’s going to work tomorrow or in
0:07:48 a few days.
0:07:53 Python has new versions almost every couple of weeks, and it’s something that keeps changing.
0:08:00 So you need to be on top of things, and it’s an occupational hazard, you can say, that
0:08:04 you learned something in university, and by the time you’re done, nobody’s using it anymore.
0:08:06 So that’s number one.
0:08:11 If you’re using a model like ChatGPT, how often is it being updated?
0:08:17 If there’s an update happening in one of the libraries, how long before ChatGPT is aware
0:08:18 of it?
0:08:21 There’s a bit of a problem there.
0:08:26 Another issue is vulnerabilities, cybersecurity, computer security.
0:08:30 Because these models, they tend to produce very similar code.
0:08:34 Whenever you have a similar prompt, it will show you a similar code.
0:08:40 Basically it creates a way for very malicious people to exploit a lot of software all at
0:08:41 once.
0:08:47 So if you write a prompt of, help me make this game, this car game.
0:08:54 When everybody gets the same piece of code for their car game, if somebody wants to write
0:08:59 a malicious software to target this specific bit of code, it will apply on all the software
0:09:00 at once.
0:09:04 So whenever you’re using this code, make sure you do a bit of variation.
0:09:06 That’s what I would do, at least.
0:09:11 Yeah, it feels like both of those are just current limitations, at least to me.
0:09:17 Because you’re talking about the complexity and things changing fast, and I think AI could
0:09:19 be way better at dealing with that than humans.
0:09:24 Humans are not very good at dealing with fast change and massive amounts of data and processing
0:09:25 that data.
0:09:33 So yeah, chat GPT, GPT4 kind of suck at this, but I think GPT5, GPT6, I think these problems
0:09:35 are going to be solved.
0:09:40 Perplexity also, they pull in more real-time data and feed it into the LLM.
0:09:44 So I think in terms of packages changing and things like that, you probably already could
0:09:45 get past that.
0:09:49 Maybe just no one’s done it in a good way yet, but it feels like that’s something that’s
0:09:50 going to be solvable.
0:09:51 No, you’re absolutely right.
0:09:57 The things that humans find easy, usually that’s what AI finds very complex, and the
0:10:03 things that humans find complex, such as mathematics, analytics, and reading large bodies of text,
0:10:05 making conclusions out of them.
0:10:07 It’s something that models are doing better than us.
0:10:13 So basically, it’s all about a symbiotic relationship, I think, in the end of the day.
0:10:19 We need to use these models to make our lives better, but it’s a challenge.
0:10:22 Yeah, engineering is all about problem solving, right?
0:10:25 I think ultimately, coding is just a way that you solve problems.
0:10:34 So long-term, I kind of imagine it will have a new class of Uber engineers who are probably,
0:10:39 they know how to code, but they’re also manning an army of AI bots that are helping them do
0:10:41 things that they don’t want to spend time doing.
0:10:43 The menial kind of coding task, right?
0:10:48 But then they have the ability to go in, possibly with the help of a co-pilot as well, to kind
0:10:51 of go in and where they need to to modify things.
0:10:55 I think that’s going to be a really interesting world where you have these brilliant engineers
0:10:59 who can go off and have a swarm of AI helping enhance what they do.
0:11:00 For sure.
0:11:06 Now, this is more of a philosophical question, but I’m definitely curious to hear your answer.
0:11:12 If Jensen is right, and in five years, we don’t need coders, do you think, do you still
0:11:15 think it’s important that people learn to code right now?
0:11:20 I think this is a problem that is not unique to folks who code.
0:11:22 I can say that about many industries.
0:11:23 I can say it about accountants.
0:11:25 I can say it about truck drivers.
0:11:28 I can say it about many, many professions.
0:11:29 So…
0:11:30 100%.
0:11:31 Is it worth doing anything?
0:11:32 Right?
0:11:37 Definitely a fair question, given the current state of AI.
0:11:43 I think that there’s a few routes that we can go through in the future.
0:11:45 The first route is a utopia.
0:11:47 The second one is a dystopia.
0:11:50 The third one is something that I kind of envisioned.
0:11:55 So I imagined, instead of everyone using the same type of models, instead of millions of
0:12:02 people using chat GPT, let’s say that everyone will have their own type of AI that is customized
0:12:03 to ourselves.
0:12:09 So, for example, if I have Maria GPT, I will train it on the books I read, on the movies
0:12:17 I watched, on the values that my parents taught me, on the collection of knowledge that I experienced
0:12:18 through the years.
0:12:20 I will train it on the countries I visited.
0:12:24 I will train it on anything that will make it closer to myself.
0:12:26 I will train it on my political affiliations.
0:12:29 I will give it my biases, because everyone has biases.
0:12:32 My model would love Python.
0:12:37 And if somebody else uses this model, who loves C++, they’ll be very upset, right?
0:12:39 But I’m the only one who’ll be using my model.
0:12:43 And needless to say, these models will be not public.
0:12:45 Everyone will have it stored.
0:12:50 And I can see how this would be an enhancement of yourself.
0:12:55 And it’s going to be proprietary to you, and it’s going to be used in a way to make you
0:13:00 indispensable, rather than just part of the herd that is using the exact same model to
0:13:02 accomplish the same task.
0:13:07 Yeah, that’s, I agree, it’s kind of like, for some reason, I’m thinking of like Star Trek
0:13:08 versus Star Wars.
0:13:12 I’m thinking like the Borg and Star Trek versus like the more like space pirate kind
0:13:16 of vibe of Star Wars, you know, where like people are more independent and doing their
0:13:17 own thing.
0:13:21 Like, so I’m a big proponent of open source, but I’m also, you know, we’ve taught that
0:13:24 this on the podcast several times now, like, I’m a big proponent of open source, and I’m
0:13:30 also very worried that, and excited, I’m very mixed feelings about open AI, I think they’re
0:13:34 probably very far ahead of other people, based on like rumors I’ve heard about GPT-5 from
0:13:37 friends in San Francisco, I think they’re very far ahead.
0:13:42 And so yeah, that could lead to a Borg type of scenario where, yeah, everyone’s outsourcing
0:13:46 all of their thinking to this genius AI brain.
0:13:48 And so everyone acts very, very similar.
0:13:51 Now Sam Altman has said that like they plan to make it a little bit more personalized
0:13:52 over time.
0:13:55 And everyone, he said the same kind of thing, everyone has biases.
0:13:58 And so they want to make it, they don’t want to make it left wing or right wing, whatever.
0:14:02 It’s like, you know, whatever your biases are, like it learns that and it kind of adapts
0:14:03 to you.
0:14:07 So I hope they actually take that seriously long term so we don’t end up in a Borg type
0:14:08 scenario.
0:14:09 I mean, I’m on the same page.
0:14:11 I think that’s where all of this is headed.
0:14:14 I think that’s kind of like Sam Altman’s vision right now.
0:14:19 I think that’s what, you know, Satya Nadella over at Microsoft, they’re trying to do that
0:14:23 with everything at Microsoft, Sundar over at Google, starting to do that with Google.
0:14:26 I think that’s sort of the vision all of these big companies have is turning this into
0:14:33 like this personal AI assistant that is totally trained on you and what you like.
0:14:37 And it can see your calendar, it can see your emails, it can, you know, go back through
0:14:41 some of the conversations you’ve had and it just kind of knows everything about you
0:14:45 and what you need to do next and can direct you through the day, hey, don’t forget, you’ve
0:14:50 got this meeting at two o’clock and just sort of like be there with you all the time to
0:14:51 assist you.
0:14:53 I do think that’s sort of where that’s, where it’s all headed.
0:14:57 And there was that interview with the CEO of Bumble as well, I don’t know if you saw
0:15:01 that clip where she talked about like, that’s going to be the future of dating too, where
0:15:07 you create like this AI avatar of yourself and your AI avatar that’s trained on you goes
0:15:10 and dates other people’s avatars.
0:15:16 And then when the two avatars find a compatibility, it comes back and says, here’s a match we’ve
0:15:17 found for you.
0:15:19 To me, it sounds awful, actually.
0:15:23 That’s a dystopian one, yeah.
0:15:27 We’ll be right back, but first I want to tell you about another great podcast you’re going
0:15:28 to want to listen to.
0:15:33 It’s called Science of Scaling hosted by Mark Roberge and it’s brought to you by the HubSpot
0:15:38 Podcast Network, the audio destination for business professionals.
0:15:43 Each week, host Mark Roberge, founding chief revenue officer at HubSpot, senior lecturer
0:15:47 at Harvard Business School and co-founder of Stage 2 Capital, sits down with the most
0:15:53 successful sales leaders in tech to learn the secrets, strategies, and tactics to scaling
0:15:54 your company’s growth.
0:16:00 He recently did a great episode called How Do You Solve for a Siloed Marketing in Sales
0:16:02 and I personally learned a lot from it.
0:16:06 You’re going to want to check out the podcast, listen to Science of Scaling wherever you
0:16:10 get your podcasts.
0:16:17 So for me, a future where Sam Altman controls our personal AIs that knows everything about
0:16:20 us, it’s also kind of dystopian.
0:16:25 You mentioned open source and for many, many years, the type of researchers that were in
0:16:32 the field of AI, they all were accompanied by a paper that tells you with full transparency
0:16:36 what kind of data the model was trained on, the architecture of the model, even what kind
0:16:42 of equipment it used so that you can recreate the same conditions on your end and so you
0:16:47 can provide what is called a peer review because it’s not truly science unless somebody else
0:16:53 can recreate it and for many, many years I’ve been reading those papers and they were kind
0:16:59 of the standard of publishing new architecture, so for example, AlexNet, which I’m sure you
0:17:05 guys are familiar with, there’s ResNet, there’s VGG, they all are accompanied by papers that
0:17:12 are fully transparent so when we are entering this realm of proprietary models where people
0:17:18 are hiding the type of data that they’re using and it worries me because it hasn’t been the
0:17:21 case for many, many years so I wonder what changed.
0:17:25 It did kind of seem to change with open AI, obviously, I don’t know if you’ve seen the
0:17:33 whole argument on X that Elon Musk and Jan Lacuna have been having, where Jan’s tried
0:17:38 to claim that it’s not science unless there’s a paper and Elon, yeah, there’s a whole drama
0:17:39 going on.
0:17:42 Both of them are scientists, both of them.
0:17:46 The fact that they are passionately arguing with one another, it’s just an example of
0:17:48 how scientific both of them are.
0:17:54 Yeah, it is interesting though, even Anthropic, who seems to be pretty ethical, pretty above
0:17:59 board, pretty safety-minded, still has a closed model that they don’t share what’s going on
0:18:03 underneath the model, which is interesting to me, but yeah, I wonder why that is.
0:18:10 It does seem like it kind of started from open AI because isn’t GPT2 fully documented?
0:18:14 Can’t you go and isn’t that one openly available to run off of?
0:18:17 You can download it, you can use it in your software, you can get it right now if you
0:18:18 want, from Huggingface.
0:18:21 Yeah, so it must have been GPT3 that started it all.
0:18:25 Yeah, they’re saying it’s from the advancement of the capability and what that enables is
0:18:30 what they’re saying, Bellagy, I don’t know if you know Bellagy, but he tweeted yesterday
0:18:35 something about the people who are against open source AI, they started with left wing
0:18:40 talking points and then now they’ve migrated to right wing talking points and they’re
0:18:44 trying to get both sides on the same page because they’re like, “Okay, we need this
0:18:46 because otherwise it’s going to be biased against certain people,” and so they started
0:18:50 on that angle and then now they went to more the national security thing of like, “You
0:18:54 don’t want China to get this,” and that’s starting to get some traction even in Silicon
0:18:55 Valley.
0:18:59 So I have a lot of connections of people, I used to mentor for Peter Till’s 2020, so
0:19:03 I have some connections at Founders Fund and they’ve been tweeting stuff about this recently
0:19:10 where they seem to be more supporting closed source AI for this national security reason,
0:19:15 whereas obviously A16Z and most other VCs are really supporting open source because, I mean
0:19:19 even if it’s for their own personal reasons, you can’t really have many AI startups if
0:19:26 everything’s closed source and you have to rely on open AI, but it is interesting.
0:19:31 I would say the national security one is the one that I have mixed feelings about because
0:19:33 I totally get what they’re saying.
0:19:36 They’re like, “Yeah, if we open all of this up and then China or Russia or whoever can
0:19:40 just copy it,” that is the one argument that I have really mixed feelings about.
0:19:44 Yeah, I’m going to say something kind of controversial here, I think, but I think that’s kind of underselling
0:19:45 China too.
0:19:48 I think China is pretty far ahead.
0:19:52 So many of the papers that I’ve seen lately have been coming out of China, not the US.
0:19:56 But their LLMs are really far behind, and my understanding is their reason is because
0:19:57 of their censorship.
0:20:01 So they actually have a lot of issues there where their censorship is actually hurting
0:20:04 how they can train their models because they don’t even want the people who are training
0:20:08 the models to be able to talk about sensitive topics.
0:20:11 And so it’s actually like really, apparently that’s one of the reasons they’re behind.
0:20:12 That’s what I’ve heard from friends from China.
0:20:18 I’m sure they’re trying to really bake their sort of governmental bias into all the AI.
0:20:20 Yeah, which slows everything down, right?
0:20:24 I think that when it comes to China, they’re more focused on facial recognition and kind
0:20:27 of preventing you from using certain services.
0:20:32 Because the whole idea about China is that they have an app that is like the everything
0:20:34 app.
0:20:35 They call it WeChat.
0:20:37 They use it to pay for transactions.
0:20:40 They use it to communicate to one another, send messages.
0:20:42 They get coupons for food through this app.
0:20:45 So everything you want to do, they do on this app.
0:20:51 And obviously all this information goes to the source that trains those artificial intelligence
0:20:52 models.
0:20:57 So even though their LLM capabilities, it’s reasonable that they’re not there, obviously,
0:20:59 because there’s a lot of censorship there.
0:21:05 But there are other models that, I bet you, they’re way more advanced than what we have.
0:21:11 And we wouldn’t know because they haven’t been used against us yet.
0:21:17 Also wonder if TikTok is directly connected to China, because there’s a lot of information
0:21:20 that people voluntarily post on TikTok.
0:21:22 I don’t have a TikTok, so I don’t know.
0:21:27 But I know that a lot of people are posting three videos a day about all the things they
0:21:28 do.
0:21:32 I wonder if they already know us really, really well.
0:21:40 And they don’t really need LLMs because they just use the video input to train a model
0:21:44 that is also looking at the facial expressions you have.
0:21:47 I think you’re probably right there because TikTok is known for having the best algorithm
0:21:48 in social media.
0:21:50 So I think it’s a fair point.
0:21:51 The best or the most addictive?
0:21:53 I think it’s probably the most addictive.
0:21:54 Both.
0:21:58 It’s the most addictive because they know people the best.
0:22:04 They understand people and I do believe, like I’m a proponent of TikTok being divested
0:22:09 because when COVID went down, I had a lot of friends, like I actually studied Mandarin,
0:22:10 I lived in Taiwan.
0:22:15 So I’m probably slightly biased on the Taiwan side, but because I have a lot of Silicon
0:22:21 Valley friends who are all like Taiwan founders or Taiwanese Americans.
0:22:25 And like during COVID, there was like a thing that kind of went down that was not really
0:22:27 talked about because everyone was talking about COVID.
0:22:32 But during the same time period, China just like took over the boards of like every single
0:22:33 major tech company in China.
0:22:38 And people don’t really talk about that, but it’s kind of why it was like a wild transition
0:22:39 to happen.
0:22:42 Because before it was like, yeah, maybe they kind of control it, but it’s not like directly
0:22:43 controlling it.
0:22:45 It’s like, yeah, they can threaten them or whatever.
0:22:47 Maybe the US government even does that kind of stuff.
0:22:51 But like during the COVID time period, they literally inserted board members into all the
0:22:52 major tech companies.
0:22:56 And so they basically control every single major tech company in China.
0:23:00 And so yeah, I would not be surprised if they also have the same, you know, China is involved
0:23:03 in TikTok and using that data to get way ahead in AI.
0:23:05 And like you said, probably we don’t even know.
0:23:08 Maybe it’s an architecture that is way beyond, you know, LLMs.
0:23:10 They just don’t share it with us.
0:23:16 So I don’t know if this arguments of national security is entirely applicable, because yeah,
0:23:19 they can copy it, but you still need to train it, right?
0:23:24 Because even if you show people how you made your model, you still need to have the equipment
0:23:28 and the time, you know, the processing power to train it.
0:23:32 And we’re talking about models like chat GPT that are being trained actively for years
0:23:34 on supercomputers.
0:23:39 And it’s very hard to catch up to them, even if they come up like, and that’s the reason
0:23:44 why I don’t understand why they don’t post, you know, the data, because no matter how
0:23:45 …
0:23:48 But they want Taiwan and all the computing is happening in Taiwan.
0:23:49 Of course.
0:23:50 Well, it’s an important center.
0:23:51 Yeah.
0:23:56 I think every piece of electronic in the world depends on this manufacturing facility.
0:23:58 Even cars, even cell phones, things like that.
0:24:05 So if they’re being, you know, taken out of the equation, we’re facing a very weird future.
0:24:08 At least those of us who haven’t invested in a very fancy computer.
0:24:10 There’s not going to be a lot of parts.
0:24:11 Yeah.
0:24:16 I know, like, I know Jan LeCun, he’s been a big … he’s been very outspoken about
0:24:21 his belief that AGI isn’t going to come from large language models, right?
0:24:26 It’s going to come from some other sort of AI model, like some sort of world simulator
0:24:27 model or something like that.
0:24:28 Who knows?
0:24:31 You know, maybe China’s already got something like that and they’re like, yeah, let them
0:24:33 have fun with their little LLMs.
0:24:35 This is what we’re working on, you know?
0:24:36 You never know.
0:24:42 They probably wouldn’t make it public and put it on GitHub if they were working on that.
0:24:45 You will know about it when it’s too late.
0:24:46 Right.
0:24:47 Exactly.
0:24:48 Exactly.
0:24:52 I’m curious about your thoughts on the sort of ethics of how the code is trained, right?
0:24:59 Because recently we heard about this partnership between Stack Overflow and OpenAI, allowing
0:25:03 OpenAI to basically train on Stack Overflow’s data and then people started going and trying
0:25:10 to sabotage the code that was on Stack Overflow to try to poison the data.
0:25:17 How do you feel about coder’s work just being trained into the data without their permission?
0:25:22 I think if it was a different organization, if it was an open AI, if it was like some
0:25:27 open source type of model, Mr. Al, or like something along these lines that were entering
0:25:32 this partnership with Stack Overflow, I don’t think those moderators will go to the length
0:25:36 of like deleting their contribution, right?
0:25:37 Because it’s been there for years.
0:25:42 It is general knowledge, like it’s something that I’ve been using for many years and since
0:25:47 this website existed, it was helping many people solve their errors.
0:25:53 I think that people mistrust OpenAI in particular, first of all, because it’s not really open.
0:25:57 How can you call your company OpenAI if it’s not open?
0:25:58 That’s the number one.
0:25:59 Closed AI.
0:26:00 Right?
0:26:01 It’s like the number one warning sign.
0:26:02 That’s number one.
0:26:03 Mm-hmm.
0:26:10 Second of all, I think that there’s just a major level of mistrust to something that
0:26:12 is not open source in the field of programming.
0:26:17 We support open source from the very get go, you know, in every, even in the computer science
0:26:21 program that I’m taking right now, I’m doing a computer science BSE.
0:26:24 My lecturer was teaching us about open source.
0:26:30 He was basically explaining why Windows is not as good as Linux, you know, these type
0:26:31 of stuff.
0:26:34 We have it for years and we’ve been teaching people for years about these concepts.
0:26:37 And suddenly there’s this big company that everyone is using.
0:26:39 It is far ahead of everyone else.
0:26:45 They got a supercomputer for free from Jensen, you know, back in the time because they were
0:26:51 open and free and they suddenly decided to kind of close all the doors and take all this
0:26:54 transparency and just throw it to the garbage.
0:26:56 Why would they do it?
0:26:57 You know, is it because of money?
0:26:58 I think they have enough money.
0:27:00 I don’t think it’s because of that.
0:27:02 So what are you hiding?
0:27:04 And a lot of people are thinking that.
0:27:08 So these moderators, I don’t think that you just have a problem with some AI learning
0:27:11 because you can web scrape any piece of information from any website.
0:27:15 You can make a bot that copies all the data you want.
0:27:16 It’s not a problem.
0:27:20 Plus all the data and stack overflow, as far as I’m concerned, it’s public.
0:27:22 You don’t need to log in to be able to access it.
0:27:24 So from a legal perspective, you can already use it.
0:27:26 You don’t need to go into this partnership.
0:27:32 So I think it’s a it’s a very weird, it’s very weird to me that Stack Overflow agreed
0:27:35 because I don’t know what they are earning in this equation.
0:27:36 I think there’s worried.
0:27:37 Yeah.
0:27:41 I think because like apparently their traffic’s way down because of people not using Stack
0:27:42 Overflow.
0:27:43 They just asked chat to be the same question.
0:27:47 I don’t think they had a plan B. I think I think chat GPD came up and ate a bunch of
0:27:49 Stack Overflow’s lunch.
0:27:53 Stack Overflow tried to create their own stack AI or overflow AI or whatever they were calling
0:27:54 it.
0:27:58 They created their own model and nobody cared because they were already indoctrinated into
0:28:02 chat GPD and they went, well, okay, if we can’t beat them, join them.
0:28:08 I think that’s kind of what happened, honestly, but I also think, you know, the whole like
0:28:14 ethos of like open source is like, I’m putting this code out there, but if you build something
0:28:16 with this code, also make it open source, right?
0:28:20 Like that’s sort of the whole like open source code, whatever you build with this open source
0:28:22 stuff, make that available.
0:28:28 And so it’s weird to me that companies like chat GPD or open AI can go train on all of
0:28:34 this code and then close it off when all of this stuff was put publicly available and was
0:28:39 designed for whatever you build with this, also make that publicly available.
0:28:40 You know?
0:28:41 Yeah.
0:28:42 I agree.
0:28:43 It’s kind of shady.
0:28:46 You know, I don’t know what’s the purpose of it, but it makes me mistrust.
0:28:50 And when you mistrust, you know, you had a lot of trust towards Stack Overflow as a developer,
0:28:54 you know, from the get go, like it’s the best source to solve your errors.
0:28:58 And, you know, you can find multiple solutions to the same problem and choose what kind of
0:29:01 solution you prefer, basically.
0:29:06 So it’s a very trusted name that has a lot of integrity, right?
0:29:11 So when it goes into partnership with something that folks like ourselves, like the majority
0:29:15 of the audience that is still using Stack Overflow, we never left.
0:29:17 We kept using it, right?
0:29:23 It’s kind of a slap in the face, and I don’t know, I don’t think that Stack Overflow should
0:29:26 have been chasing them, you know, preventing them from deleting their comments because that’s
0:29:28 their decisions, you know, they enter this partnership.
0:29:31 They should have known there’s going to be a backlash.
0:29:36 You know, they have marketing experts that, you know, chat with users, they understand
0:29:41 what’s going on, you know, it was clear to me, and if it was clear to me, I assume it
0:29:42 was clear to Stack Overflow.
0:29:48 So don’t, like, antagonizing them, like, further, basically, just kind of, you can’t
0:29:49 touch your account.
0:29:50 It’s suspended.
0:29:55 We are checking if you really meant to delete your messages or not, like, come on, you guys,
0:29:56 that’s even worse.
0:30:01 You just build, burning the creators that create on your platform even more.
0:30:03 They’re training on all the GitHub data too, right?
0:30:06 I think that’s even a bigger deal, but that’s the actual code itself.
0:30:11 GitHub is owned by Microsoft and, you know, Microsoft and OpenAI are, they’re like this.
0:30:15 You know, I would imagine all the GitHub stuff is, you know, freely available to OpenAI.
0:30:16 Yeah.
0:30:17 And why not?
0:30:21 You know, it’s public data, like the, maybe the private repositories, you cannot really
0:30:27 access them, but as long as something is public data, and I know it because there was a company,
0:30:31 there was a company that was scraping data from platforms, and they were, they were
0:30:34 accused, bright data, their name.
0:30:38 They were accused of, like, stealing content, they were sued, but when they came to the
0:30:43 court, when it was time to kind of actually talk about the laws behind it, it’s like,
0:30:49 hey, we never logged in to, you know, whatever platform we were, we were taking information
0:30:50 from.
0:30:51 Anyone can copy it.
0:30:56 So what is the difference if a bought, if an automated entity is copying it or somebody
0:30:58 that physically sits with a keyboard does it?
0:31:00 And I think they won.
0:31:01 Yeah.
0:31:02 Yeah.
0:31:03 I mean, we’ve been seeing a lot of the same sort of arguments over on the art side as
0:31:04 well, right?
0:31:10 Um, you know, a lot of these AI art generators have trained on other people’s arts and, and
0:31:14 a lot of that has gone to court and been fought in court and, and same with the writing too,
0:31:18 we had a lot of authors take chat GPT to court saying they trained on their writing and almost
0:31:23 all of these cases end up getting thrown out because well, it’s publicly available data.
0:31:25 They’re, I guess they’re allowed to train on it.
0:31:30 So yeah, it’s, it’s definitely interesting, but I personally have very mixed feelings
0:31:31 about it all, right?
0:31:36 Like I was actually for a little while, um, doing photography and uploading it to sites
0:31:41 like shutter stock to earn, you know, the, the income from selling the stock photography.
0:31:46 Well, a couple of years go by and now we find out that, okay, shutter stock trained all
0:31:49 of the images that were uploaded into like their new AI model.
0:31:52 And Adobe did the same thing with Adobe stock, right?
0:31:56 Everything that’s been uploaded to Adobe stock over, you know, however long Adobe stocks
0:32:01 been around now, all of a sudden that became training data for our AI art generator.
0:32:02 Well, cool.
0:32:06 When I was uploading photos years ago to try to make money, I didn’t really say you had
0:32:09 permission to use that photo as part of the training data.
0:32:13 I don’t really care that much because I’m, I’m like in AI and I’m an AI optimist and
0:32:17 I, I’m kind of fine with it, but I don’t, I definitely see both sides of that token,
0:32:18 you know?
0:32:19 Yeah, absolutely.
0:32:20 Yeah.
0:32:21 I think it’s a, I think it’s a gray area.
0:32:26 I think, I think it’s a gray area where like, you know, they, there’s a reason open AI doesn’t
0:32:27 want to talk about it, right?
0:32:30 Like the CTO, there was that famous clip of her stumbling where they were asked if she
0:32:36 was training on, if they were training on YouTube data and she was like, I don’t know
0:32:38 exactly, you know, kind of, kind of thing.
0:32:41 And then obviously she knows the answer, like obviously she knows.
0:32:45 And I think it’s because probably when they even started the company or when they started
0:32:48 really thinking about this stuff, they would have talked with lawyers and they would have,
0:32:52 it’s kind of like when Uber launched where technically Uber could be argued that it was
0:32:56 illegal when it started, but it was a gray area where they knew they had a legal argument
0:32:58 because there was no precedent exactly.
0:33:03 And so there’s no exact precedent here because, you know, you have public data, but you can’t
0:33:04 just copy public data.
0:33:06 So they’re going to argue, well, the AI is learning from it.
0:33:10 It’s just like if you go to a museum and you see some art or whatever, you’re learning
0:33:11 from that.
0:33:12 You’re not copying it.
0:33:15 And so I’m pretty sure that’s like the legal case they will make long term, but I think
0:33:20 they’re really trying to delay having that huge legal battle, but I think it’ll happen
0:33:21 at some point.
0:33:27 Because right now I decide to train my artificial intelligence model on a subset of data.
0:33:29 You can only see the final product.
0:33:31 You will not see what kind of data I trained on.
0:33:32 I don’t have to share it.
0:33:35 I don’t have to say anything about it, right?
0:33:43 It’s unethical, but you would never know because all you do is you take a bunch of images,
0:33:50 you feed it into a neural network, and then it finds patterns in those images and it basically
0:33:53 creates something new out of this data.
0:33:55 So you can say it’s like a musician.
0:34:02 So for example, the Beatles were inspired by Elvis and Pink Floyd were inspired by the
0:34:03 Beatles.
0:34:04 So who copied from who?
0:34:06 It’s the chicken and the egg.
0:34:12 And you’re right when you’re saying that it’s inspiration because what the model creates
0:34:16 is not an exact match to the training data.
0:34:18 It has to be something else.
0:34:22 They pull it out of late in space when it comes to a generative AI models.
0:34:28 And this late in space, I don’t think anybody owns it, and those type of things, we should
0:34:32 have thought of it way before.
0:34:37 Right now when we’re talking about the legal copyright implications, it’s something we
0:34:41 could have talked about years ago before it became a problem, before people’s livelihoods
0:34:44 were affected from Shatterstock and places like that.
0:34:49 By the way, Firefly, Adobe’s Firefly, is amazing.
0:34:51 I might actually use it in my final project.
0:34:56 I’m doing a final project for university, and I think that that’s my way to go.
0:34:58 I’m going to make a database.
0:35:02 But even in Firefly though, they were pitching that it was the responsible way to use AI
0:35:06 images, and then apparently they trained on mid-journey images.
0:35:09 They basically just went to another layer down and were like, “Yeah, we didn’t do it.
0:35:10 They did it.”
0:35:14 Yeah, well, I mean, they didn’t specifically go out of their way to train on mid-journey,
0:35:20 but they allowed people to upload mid-journey images to be sold as stock images.
0:35:24 So the mid-journey images that were uploaded to Adobe stock, and then when Adobe went to
0:35:29 go and train their Firefly model, well, there was a ton of mid-journey and stable diffusion
0:35:34 generated images in it already, because they were allowing people to sell those images as
0:35:35 stock.
0:35:39 So they didn’t go and train specifically on mid-journey, but mid-journey and stable
0:35:43 diffusion and dolly images were in the data set because people weren’t allowed to sell
0:35:46 that stuff on Adobe stock.
0:35:47 That makes more sense.
0:35:48 It’s crazy.
0:35:55 If you start nitpicking what pieces of data went into the model, it’s an ever-ending story.
0:35:59 You can always find somebody who contributed to create this image.
0:36:00 So what?
0:36:02 They deserve to get paid too.
0:36:03 It’s a rabbit hole that I don’t…
0:36:04 And how much?
0:36:05 Right?
0:36:06 I’m a very simple person.
0:36:12 I like the internet how it was when I was 12, which was basically wild, wild west.
0:36:14 You could copy anything.
0:36:15 You could get anything.
0:36:16 You can download anything.
0:36:20 And as a content creator, I know that it’s a terrible thing to say because people will
0:36:21 copy my content.
0:36:22 You know what?
0:36:26 If they want to, please, it only helps Python being thought.
0:36:28 I make enough money from what I do.
0:36:35 I don’t care if somebody else is basically using some of it to kind of grow this Python
0:36:38 world and kind of teach others.
0:36:42 If a teacher is using it in his lecture, I’m not going to chase him, asking him for some
0:36:43 royalties.
0:36:44 It doesn’t make any sense.
0:36:45 Why would I do it?
0:36:54 Same goes for a musician where some 12-year-old kid decides to use his piece of art, his music,
0:36:55 in his video.
0:36:58 Why would this musician chase the kid?
0:36:59 He is sharing his music with the world.
0:37:04 So yeah, you don’t get paid for it, but you get paid enough.
0:37:10 It should cover other fan art if you want to call it.
0:37:15 So I find it upsetting that people are really nitpicky with those copyright laws.
0:37:18 And if we weren’t, maybe the internet would have been a nicer place.
0:37:19 Yeah.
0:37:20 Yeah.
0:37:23 I mean, I really think we’re getting to a point where copyright in general just needs
0:37:24 to be rethought.
0:37:28 I don’t know the solution, but I think it just needs to be rethought because we’re going
0:37:32 to be able to generate images that look like other people’s images but weren’t created
0:37:33 by that person.
0:37:37 We can already create songs that sound like other people’s songs that wasn’t created
0:37:38 by them.
0:37:42 When we get into Sora and Veo and some of these new video generators, we’re going to
0:37:46 be able to generate videos that look like the style of other people’s videos but weren’t
0:37:48 created by them.
0:37:53 It’s just going to get so muddy that I don’t feel like the way copyright law was originally
0:37:55 written is still going to be relevant.
0:37:57 It just needs to be rethought.
0:37:58 I agree.
0:37:59 Absolutely.
0:38:00 But I don’t know how to rethink it.
0:38:02 I’m not the one to figure that out.
0:38:06 Actually, I had meetings in the Library of Congress with people in the Copyright Office.
0:38:10 I don’t know if you know my last startup, Binded.
0:38:14 We started off not being involved in copyright and then unfortunately our startup, we kind
0:38:18 of pivoted towards copyright, but we were not trying to be doing enforcement and stuff
0:38:21 like that with more just attribution and things like this.
0:38:24 I spent a lot of time thinking about how can you change this and I met with people in
0:38:28 the government and now I’m convinced that I can’t.
0:38:31 They have no interest in changing things.
0:38:34 I don’t think there’s going to be a fundamental like, “Oh, the government’s going to decide
0:38:36 that copyright should change” or something.
0:38:37 Who knows?
0:38:38 Who knows?
0:38:39 I mean, it could be a generational thing.
0:38:40 Right?
0:38:47 Maybe once there’s a little bit more turnover from the more elderly people in the governments.
0:38:51 Maybe some of the younger generations will see that, “Okay, technology has changed a
0:38:55 lot since a couple hundred years ago when so many of these laws were written.
0:38:59 Maybe we should update some of this stuff for the way the world is now instead of the
0:39:01 way the world was a hundred, two hundred years ago.”
0:39:02 Yeah.
0:39:04 Maybe GPT-6 will help us figure it out, you know?
0:39:05 Absolutely.
0:39:07 Well, this has been an absolutely amazing conversation.
0:39:12 We’ve had a blast talking to you, Maria, and I want to make sure that people can go check
0:39:13 out your stuff.
0:39:16 Where should people go and follow you after tuning into this episode?
0:39:19 Where’s the best place to learn tutorials from?
0:39:20 Yeah, the best place is always YouTube.
0:39:21 That’s my main platform.
0:39:24 This is where you can find shorts.
0:39:26 You don’t like TikTok style shorts.
0:39:29 You can find tutorials that are quite long.
0:39:33 This is where you find me, explain very complex concepts in a simple language that even a
0:39:36 six-year-old can understand, hopefully.
0:39:37 It’s called Python Simplified.
0:39:38 Python Simplified.
0:39:39 Yeah.
0:39:40 Python Simplified.
0:39:41 You can find me on YouTube.
0:39:44 You can find me on X as well, even though I’m not there very often.
0:39:46 Maria Simplified.
0:39:51 The best place is to find me is YouTube, Maria Shah, Python Simplified.
0:39:52 Well, very cool.
0:39:55 Thank you so much for hanging out and talking code with us.
0:39:59 This was a conversation that we wanted to have, but we wanted to bring on somebody that
0:40:02 knows a little bit more about the coding world than us.
0:40:05 I’m so thankful that you were able to join us and actually have this conversation with
0:40:06 us.
0:40:07 I really appreciate it.
0:40:08 Awesome.
0:40:25 Thank you.
0:00:07 The second one is a dystopia.
0:00:10 The third one is something that I kind of envisioned.
0:00:14 Hey, welcome to the Next Wave Podcast.
0:00:15 My name is Matt Wolfen.
0:00:20 I’m here with my co-host, Nathan Lanz, and today we’ve got an amazing guest.
0:00:25 She has an amazing YouTube channel all about Python and how to code.
0:00:26 Her tutorials are amazing.
0:00:30 So many people learn Python from her, and we’re excited to have her on the show to
0:00:36 talk about the future of coding and how it overlaps with AIs.
0:00:40 When all your marketing team does is put out fires, they burn out.
0:00:44 But with HubSpot, they can achieve their best results without the stress.
0:00:50 Tap into HubSpot’s collection of AI tools, Breeze, to pinpoint leads, capture attention,
0:00:53 and access all your data in one place.
0:00:57 Keep your marketers cool and your campaign results hotter than ever.
0:01:03 Visit hubspot.com/marketers to learn more.
0:01:05 So welcome to the show, Maria Shaw.
0:01:06 Thank you for being on.
0:01:07 Yeah.
0:01:08 Absolutely.
0:01:09 Thank you for inviting me.
0:01:10 Great to meet you.
0:01:14 So let’s talk a little bit about the AI world, because I know, you know, when we were out
0:01:18 at GTC, the talk of the whole event was AI, right?
0:01:23 Like everything is AI, and obviously coding is one of those areas where AI has completely
0:01:24 taken over.
0:01:30 I know Nathan was one of the sort of early adopters of GitHub co-pilot, and yeah, I guess
0:01:33 I just want to know, what are your sort of like overall thoughts?
0:01:36 Let’s just kind of start from like the 30,000 foot like general overview.
0:01:39 Like what are your thoughts of like that overlap of AI and coding?
0:01:43 Are you excited that AI is making coding easier?
0:01:45 Are you worried about coders losing jobs?
0:01:48 Like where do you kind of stand on the whole thing right now?
0:01:53 I’m slightly confused because I’ve tried a few of the AI models to kind of see what
0:01:54 they’re all about.
0:01:59 And I don’t see a lot of difference between just kind of copying the prompt and pasting
0:02:00 it in a search engine.
0:02:04 Like I think the biggest difference I see is the fact that when you do so with a search
0:02:08 engine, you can see a whole bunch of sources and then you decide which one of these is
0:02:09 your favorite.
0:02:10 And you kind of go from there.
0:02:17 But I find these co-pilots as a middleman between the developer and the documentation.
0:02:20 And for me, being so nerdy, I really like documentation.
0:02:22 I really appreciate it.
0:02:23 It’s my favorite place to be.
0:02:26 And that’s how I write most of my tutorials.
0:02:31 So getting it from the source is more important to me than kind of saving a minute or so
0:02:32 or even less.
0:02:38 So that’s where I’m a bit confused because I hear a lot of my viewers really exciting,
0:02:44 really excited about ChedGPT, about co-pilot and Devin probably, even though I don’t even
0:02:49 know if it’s open to the wide public yet, but maybe Nathan can share his opinion about
0:02:50 it.
0:02:53 So that’s something I was thinking about too, for your YouTube channel.
0:02:58 It’s got to be– I’ve seen other YouTubers as well who are like– their channel is primarily
0:03:00 about coding that they’re very skeptical of it.
0:03:04 And it feels like there’s a slight conflict there, too, though, because if AI does get
0:03:09 so good that you don’t need to learn coding anymore, it’s like a lot of those channels
0:03:14 that have to pivot to some other kind of content, which I’m sure you’re coming from a genuine
0:03:15 place.
0:03:17 And yeah, there’s definitely major limitations right now.
0:03:22 And yeah, I just have a– like you guys were talking about the NVIDIA conference, Jensen
0:03:25 said that in the future, you won’t need coders.
0:03:29 When you think of the future, you need to think of all the possible routes that we can
0:03:30 go through.
0:03:37 And the whole purpose of artificial intelligence is to be competitive or better than us in
0:03:40 all the cognitive tasks that humans are engaged in.
0:03:42 And this was from the inception of AI.
0:03:44 There is nothing new about it.
0:03:52 This is in– people’s loss of jobs has been a very major ethical concern of AI.
0:03:55 And people were reasoning about it since the 1950s.
0:03:57 I noticed a very weird trend.
0:04:01 I’m kind of– I’m taking you guys off topic because when– chat GPT–
0:04:02 Let’s go.
0:04:06 Matt loves wrapped holes, so.
0:04:12 When chat GPT emerged, there was a wave of folks who were saying that it will never replace
0:04:13 us.
0:04:16 There were people saying that this will never happen.
0:04:18 And I was kind of warning people about it.
0:04:23 I was saying that, hey, if you’re filming videos and you’re saying on record that chat
0:04:29 GPT is better than you encoding, what are the chances that your employer will hear it?
0:04:30 Why would you do that?
0:04:33 Like, even if it’s true, why would you admit something like that?
0:04:37 And I found that a lot of people were telling me, Maria, you’re crazy.
0:04:38 It’s not going to happen.
0:04:39 You’re just paranoid.
0:04:42 You’re just speaking doomsday stuff.
0:04:48 Like now that folks like Jensen are saying very similar things, maybe they take me a
0:04:49 bit more seriously.
0:04:50 I’m not sure.
0:04:54 Yeah, there was a recent tweet from– I think his name is Ethan Malik, the professor from
0:04:58 Wharton, where he was sharing some kind of recent study that showed– I think he showed
0:05:03 that 85% of employees in polls are saying that they’re using chat GPT at work.
0:05:07 And then also, I think it was something like 77% of them don’t tell their employer.
0:05:12 So it’s like, a lot of people are using chat GPT at work for emails and all kinds of different
0:05:14 tasks, and they’re just not telling their bosses.
0:05:15 So yeah.
0:05:19 Well, it’s funny because I look at the whole scenario from somebody who doesn’t really
0:05:20 know code.
0:05:21 Like, I don’t know anything about Python.
0:05:23 I don’t know how to write JavaScript.
0:05:27 I know a little HTML and CSS, and that’s about it, right?
0:05:32 And so when I look at it from that perspective, I was actually able to get into chat GPT.
0:05:35 This was back when it was 3.5, I think.
0:05:40 And I was actually able to develop a game using JavaScript that was playable.
0:05:41 It had graphics.
0:05:46 I actually used mid-journey to generate some images, and it was this little side-scroller
0:05:50 game where you jumped and collect coins, and it looked really good because I used mid-journey
0:05:52 and all the code worked.
0:05:58 So from somebody who doesn’t know JavaScript at all, I was able to go from– I have an idea
0:06:02 for a real simple game I want to build to– I actually have a playable game, and I never
0:06:04 actually touched a line of code myself.
0:06:09 So I think, from the perspective of somebody who doesn’t code, I think it’s really exciting
0:06:12 to see, oh, maybe I’ll actually be able to code now.
0:06:16 From people that actually know how to code, it’s a different story, right?
0:06:20 So I’m at this Cisco conference now out in Vegas, so anybody who’s watching the video
0:06:23 and sees a different background, that’s why.
0:06:28 But somebody on stage was talking about how right now, if you know how to code, it’s still
0:06:34 actually faster to just code something than to use chat GPT, because if you use chat GPT,
0:06:38 it’ll write the code, and then you spend just as much time double-checking and debugging
0:06:40 the code to get it to work.
0:06:45 But from somebody who’s never been able to code JavaScript, all I did was go, “That didn’t
0:06:46 work.
0:06:47 What do I try next?
0:06:48 That didn’t work.
0:06:49 What do I try next?”
0:06:53 It took me hours to code the game, but I actually eventually got there without ever touching
0:06:54 a line of code myself.
0:06:59 So it’s kind of like– I can see why there– whereas there’s excitement around it from
0:07:03 non-coders, but I can also see why coders would be like, “Ah, right now it’s more of a nuisance
0:07:04 than it is helpful.”
0:07:06 For sure, for sure.
0:07:12 There are a few reasons why folks who are coding for living, they wouldn’t be using it, but
0:07:17 I think it’s cool that it gave you these superpowers that you weren’t able to use JavaScript.
0:07:21 You never learned it, and suddenly you have your own JavaScript game, which is amazing.
0:07:26 I find it to be an incredible way to take your creativity and kind of manifest it in
0:07:28 a way that you never thought you could.
0:07:29 So it’s pretty cool.
0:07:35 So in terms of the coding, the world of coding, there are a few reasons why.
0:07:39 It’s probably not a good idea to use these type of models for that, because our industry
0:07:41 is very dynamic.
0:07:43 Things change on a daily basis.
0:07:46 Just because something works now, it doesn’t mean that it’s going to work tomorrow or in
0:07:48 a few days.
0:07:53 Python has new versions almost every couple of weeks, and it’s something that keeps changing.
0:08:00 So you need to be on top of things, and it’s an occupational hazard, you can say, that
0:08:04 you learned something in university, and by the time you’re done, nobody’s using it anymore.
0:08:06 So that’s number one.
0:08:11 If you’re using a model like ChatGPT, how often is it being updated?
0:08:17 If there’s an update happening in one of the libraries, how long before ChatGPT is aware
0:08:18 of it?
0:08:21 There’s a bit of a problem there.
0:08:26 Another issue is vulnerabilities, cybersecurity, computer security.
0:08:30 Because these models, they tend to produce very similar code.
0:08:34 Whenever you have a similar prompt, it will show you a similar code.
0:08:40 Basically it creates a way for very malicious people to exploit a lot of software all at
0:08:41 once.
0:08:47 So if you write a prompt of, help me make this game, this car game.
0:08:54 When everybody gets the same piece of code for their car game, if somebody wants to write
0:08:59 a malicious software to target this specific bit of code, it will apply on all the software
0:09:00 at once.
0:09:04 So whenever you’re using this code, make sure you do a bit of variation.
0:09:06 That’s what I would do, at least.
0:09:11 Yeah, it feels like both of those are just current limitations, at least to me.
0:09:17 Because you’re talking about the complexity and things changing fast, and I think AI could
0:09:19 be way better at dealing with that than humans.
0:09:24 Humans are not very good at dealing with fast change and massive amounts of data and processing
0:09:25 that data.
0:09:33 So yeah, chat GPT, GPT4 kind of suck at this, but I think GPT5, GPT6, I think these problems
0:09:35 are going to be solved.
0:09:40 Perplexity also, they pull in more real-time data and feed it into the LLM.
0:09:44 So I think in terms of packages changing and things like that, you probably already could
0:09:45 get past that.
0:09:49 Maybe just no one’s done it in a good way yet, but it feels like that’s something that’s
0:09:50 going to be solvable.
0:09:51 No, you’re absolutely right.
0:09:57 The things that humans find easy, usually that’s what AI finds very complex, and the
0:10:03 things that humans find complex, such as mathematics, analytics, and reading large bodies of text,
0:10:05 making conclusions out of them.
0:10:07 It’s something that models are doing better than us.
0:10:13 So basically, it’s all about a symbiotic relationship, I think, in the end of the day.
0:10:19 We need to use these models to make our lives better, but it’s a challenge.
0:10:22 Yeah, engineering is all about problem solving, right?
0:10:25 I think ultimately, coding is just a way that you solve problems.
0:10:34 So long-term, I kind of imagine it will have a new class of Uber engineers who are probably,
0:10:39 they know how to code, but they’re also manning an army of AI bots that are helping them do
0:10:41 things that they don’t want to spend time doing.
0:10:43 The menial kind of coding task, right?
0:10:48 But then they have the ability to go in, possibly with the help of a co-pilot as well, to kind
0:10:51 of go in and where they need to to modify things.
0:10:55 I think that’s going to be a really interesting world where you have these brilliant engineers
0:10:59 who can go off and have a swarm of AI helping enhance what they do.
0:11:00 For sure.
0:11:06 Now, this is more of a philosophical question, but I’m definitely curious to hear your answer.
0:11:12 If Jensen is right, and in five years, we don’t need coders, do you think, do you still
0:11:15 think it’s important that people learn to code right now?
0:11:20 I think this is a problem that is not unique to folks who code.
0:11:22 I can say that about many industries.
0:11:23 I can say it about accountants.
0:11:25 I can say it about truck drivers.
0:11:28 I can say it about many, many professions.
0:11:29 So…
0:11:30 100%.
0:11:31 Is it worth doing anything?
0:11:32 Right?
0:11:37 Definitely a fair question, given the current state of AI.
0:11:43 I think that there’s a few routes that we can go through in the future.
0:11:45 The first route is a utopia.
0:11:47 The second one is a dystopia.
0:11:50 The third one is something that I kind of envisioned.
0:11:55 So I imagined, instead of everyone using the same type of models, instead of millions of
0:12:02 people using chat GPT, let’s say that everyone will have their own type of AI that is customized
0:12:03 to ourselves.
0:12:09 So, for example, if I have Maria GPT, I will train it on the books I read, on the movies
0:12:17 I watched, on the values that my parents taught me, on the collection of knowledge that I experienced
0:12:18 through the years.
0:12:20 I will train it on the countries I visited.
0:12:24 I will train it on anything that will make it closer to myself.
0:12:26 I will train it on my political affiliations.
0:12:29 I will give it my biases, because everyone has biases.
0:12:32 My model would love Python.
0:12:37 And if somebody else uses this model, who loves C++, they’ll be very upset, right?
0:12:39 But I’m the only one who’ll be using my model.
0:12:43 And needless to say, these models will be not public.
0:12:45 Everyone will have it stored.
0:12:50 And I can see how this would be an enhancement of yourself.
0:12:55 And it’s going to be proprietary to you, and it’s going to be used in a way to make you
0:13:00 indispensable, rather than just part of the herd that is using the exact same model to
0:13:02 accomplish the same task.
0:13:07 Yeah, that’s, I agree, it’s kind of like, for some reason, I’m thinking of like Star Trek
0:13:08 versus Star Wars.
0:13:12 I’m thinking like the Borg and Star Trek versus like the more like space pirate kind
0:13:16 of vibe of Star Wars, you know, where like people are more independent and doing their
0:13:17 own thing.
0:13:21 Like, so I’m a big proponent of open source, but I’m also, you know, we’ve taught that
0:13:24 this on the podcast several times now, like, I’m a big proponent of open source, and I’m
0:13:30 also very worried that, and excited, I’m very mixed feelings about open AI, I think they’re
0:13:34 probably very far ahead of other people, based on like rumors I’ve heard about GPT-5 from
0:13:37 friends in San Francisco, I think they’re very far ahead.
0:13:42 And so yeah, that could lead to a Borg type of scenario where, yeah, everyone’s outsourcing
0:13:46 all of their thinking to this genius AI brain.
0:13:48 And so everyone acts very, very similar.
0:13:51 Now Sam Altman has said that like they plan to make it a little bit more personalized
0:13:52 over time.
0:13:55 And everyone, he said the same kind of thing, everyone has biases.
0:13:58 And so they want to make it, they don’t want to make it left wing or right wing, whatever.
0:14:02 It’s like, you know, whatever your biases are, like it learns that and it kind of adapts
0:14:03 to you.
0:14:07 So I hope they actually take that seriously long term so we don’t end up in a Borg type
0:14:08 scenario.
0:14:09 I mean, I’m on the same page.
0:14:11 I think that’s where all of this is headed.
0:14:14 I think that’s kind of like Sam Altman’s vision right now.
0:14:19 I think that’s what, you know, Satya Nadella over at Microsoft, they’re trying to do that
0:14:23 with everything at Microsoft, Sundar over at Google, starting to do that with Google.
0:14:26 I think that’s sort of the vision all of these big companies have is turning this into
0:14:33 like this personal AI assistant that is totally trained on you and what you like.
0:14:37 And it can see your calendar, it can see your emails, it can, you know, go back through
0:14:41 some of the conversations you’ve had and it just kind of knows everything about you
0:14:45 and what you need to do next and can direct you through the day, hey, don’t forget, you’ve
0:14:50 got this meeting at two o’clock and just sort of like be there with you all the time to
0:14:51 assist you.
0:14:53 I do think that’s sort of where that’s, where it’s all headed.
0:14:57 And there was that interview with the CEO of Bumble as well, I don’t know if you saw
0:15:01 that clip where she talked about like, that’s going to be the future of dating too, where
0:15:07 you create like this AI avatar of yourself and your AI avatar that’s trained on you goes
0:15:10 and dates other people’s avatars.
0:15:16 And then when the two avatars find a compatibility, it comes back and says, here’s a match we’ve
0:15:17 found for you.
0:15:19 To me, it sounds awful, actually.
0:15:23 That’s a dystopian one, yeah.
0:15:27 We’ll be right back, but first I want to tell you about another great podcast you’re going
0:15:28 to want to listen to.
0:15:33 It’s called Science of Scaling hosted by Mark Roberge and it’s brought to you by the HubSpot
0:15:38 Podcast Network, the audio destination for business professionals.
0:15:43 Each week, host Mark Roberge, founding chief revenue officer at HubSpot, senior lecturer
0:15:47 at Harvard Business School and co-founder of Stage 2 Capital, sits down with the most
0:15:53 successful sales leaders in tech to learn the secrets, strategies, and tactics to scaling
0:15:54 your company’s growth.
0:16:00 He recently did a great episode called How Do You Solve for a Siloed Marketing in Sales
0:16:02 and I personally learned a lot from it.
0:16:06 You’re going to want to check out the podcast, listen to Science of Scaling wherever you
0:16:10 get your podcasts.
0:16:17 So for me, a future where Sam Altman controls our personal AIs that knows everything about
0:16:20 us, it’s also kind of dystopian.
0:16:25 You mentioned open source and for many, many years, the type of researchers that were in
0:16:32 the field of AI, they all were accompanied by a paper that tells you with full transparency
0:16:36 what kind of data the model was trained on, the architecture of the model, even what kind
0:16:42 of equipment it used so that you can recreate the same conditions on your end and so you
0:16:47 can provide what is called a peer review because it’s not truly science unless somebody else
0:16:53 can recreate it and for many, many years I’ve been reading those papers and they were kind
0:16:59 of the standard of publishing new architecture, so for example, AlexNet, which I’m sure you
0:17:05 guys are familiar with, there’s ResNet, there’s VGG, they all are accompanied by papers that
0:17:12 are fully transparent so when we are entering this realm of proprietary models where people
0:17:18 are hiding the type of data that they’re using and it worries me because it hasn’t been the
0:17:21 case for many, many years so I wonder what changed.
0:17:25 It did kind of seem to change with open AI, obviously, I don’t know if you’ve seen the
0:17:33 whole argument on X that Elon Musk and Jan Lacuna have been having, where Jan’s tried
0:17:38 to claim that it’s not science unless there’s a paper and Elon, yeah, there’s a whole drama
0:17:39 going on.
0:17:42 Both of them are scientists, both of them.
0:17:46 The fact that they are passionately arguing with one another, it’s just an example of
0:17:48 how scientific both of them are.
0:17:54 Yeah, it is interesting though, even Anthropic, who seems to be pretty ethical, pretty above
0:17:59 board, pretty safety-minded, still has a closed model that they don’t share what’s going on
0:18:03 underneath the model, which is interesting to me, but yeah, I wonder why that is.
0:18:10 It does seem like it kind of started from open AI because isn’t GPT2 fully documented?
0:18:14 Can’t you go and isn’t that one openly available to run off of?
0:18:17 You can download it, you can use it in your software, you can get it right now if you
0:18:18 want, from Huggingface.
0:18:21 Yeah, so it must have been GPT3 that started it all.
0:18:25 Yeah, they’re saying it’s from the advancement of the capability and what that enables is
0:18:30 what they’re saying, Bellagy, I don’t know if you know Bellagy, but he tweeted yesterday
0:18:35 something about the people who are against open source AI, they started with left wing
0:18:40 talking points and then now they’ve migrated to right wing talking points and they’re
0:18:44 trying to get both sides on the same page because they’re like, “Okay, we need this
0:18:46 because otherwise it’s going to be biased against certain people,” and so they started
0:18:50 on that angle and then now they went to more the national security thing of like, “You
0:18:54 don’t want China to get this,” and that’s starting to get some traction even in Silicon
0:18:55 Valley.
0:18:59 So I have a lot of connections of people, I used to mentor for Peter Till’s 2020, so
0:19:03 I have some connections at Founders Fund and they’ve been tweeting stuff about this recently
0:19:10 where they seem to be more supporting closed source AI for this national security reason,
0:19:15 whereas obviously A16Z and most other VCs are really supporting open source because, I mean
0:19:19 even if it’s for their own personal reasons, you can’t really have many AI startups if
0:19:26 everything’s closed source and you have to rely on open AI, but it is interesting.
0:19:31 I would say the national security one is the one that I have mixed feelings about because
0:19:33 I totally get what they’re saying.
0:19:36 They’re like, “Yeah, if we open all of this up and then China or Russia or whoever can
0:19:40 just copy it,” that is the one argument that I have really mixed feelings about.
0:19:44 Yeah, I’m going to say something kind of controversial here, I think, but I think that’s kind of underselling
0:19:45 China too.
0:19:48 I think China is pretty far ahead.
0:19:52 So many of the papers that I’ve seen lately have been coming out of China, not the US.
0:19:56 But their LLMs are really far behind, and my understanding is their reason is because
0:19:57 of their censorship.
0:20:01 So they actually have a lot of issues there where their censorship is actually hurting
0:20:04 how they can train their models because they don’t even want the people who are training
0:20:08 the models to be able to talk about sensitive topics.
0:20:11 And so it’s actually like really, apparently that’s one of the reasons they’re behind.
0:20:12 That’s what I’ve heard from friends from China.
0:20:18 I’m sure they’re trying to really bake their sort of governmental bias into all the AI.
0:20:20 Yeah, which slows everything down, right?
0:20:24 I think that when it comes to China, they’re more focused on facial recognition and kind
0:20:27 of preventing you from using certain services.
0:20:32 Because the whole idea about China is that they have an app that is like the everything
0:20:34 app.
0:20:35 They call it WeChat.
0:20:37 They use it to pay for transactions.
0:20:40 They use it to communicate to one another, send messages.
0:20:42 They get coupons for food through this app.
0:20:45 So everything you want to do, they do on this app.
0:20:51 And obviously all this information goes to the source that trains those artificial intelligence
0:20:52 models.
0:20:57 So even though their LLM capabilities, it’s reasonable that they’re not there, obviously,
0:20:59 because there’s a lot of censorship there.
0:21:05 But there are other models that, I bet you, they’re way more advanced than what we have.
0:21:11 And we wouldn’t know because they haven’t been used against us yet.
0:21:17 Also wonder if TikTok is directly connected to China, because there’s a lot of information
0:21:20 that people voluntarily post on TikTok.
0:21:22 I don’t have a TikTok, so I don’t know.
0:21:27 But I know that a lot of people are posting three videos a day about all the things they
0:21:28 do.
0:21:32 I wonder if they already know us really, really well.
0:21:40 And they don’t really need LLMs because they just use the video input to train a model
0:21:44 that is also looking at the facial expressions you have.
0:21:47 I think you’re probably right there because TikTok is known for having the best algorithm
0:21:48 in social media.
0:21:50 So I think it’s a fair point.
0:21:51 The best or the most addictive?
0:21:53 I think it’s probably the most addictive.
0:21:54 Both.
0:21:58 It’s the most addictive because they know people the best.
0:22:04 They understand people and I do believe, like I’m a proponent of TikTok being divested
0:22:09 because when COVID went down, I had a lot of friends, like I actually studied Mandarin,
0:22:10 I lived in Taiwan.
0:22:15 So I’m probably slightly biased on the Taiwan side, but because I have a lot of Silicon
0:22:21 Valley friends who are all like Taiwan founders or Taiwanese Americans.
0:22:25 And like during COVID, there was like a thing that kind of went down that was not really
0:22:27 talked about because everyone was talking about COVID.
0:22:32 But during the same time period, China just like took over the boards of like every single
0:22:33 major tech company in China.
0:22:38 And people don’t really talk about that, but it’s kind of why it was like a wild transition
0:22:39 to happen.
0:22:42 Because before it was like, yeah, maybe they kind of control it, but it’s not like directly
0:22:43 controlling it.
0:22:45 It’s like, yeah, they can threaten them or whatever.
0:22:47 Maybe the US government even does that kind of stuff.
0:22:51 But like during the COVID time period, they literally inserted board members into all the
0:22:52 major tech companies.
0:22:56 And so they basically control every single major tech company in China.
0:23:00 And so yeah, I would not be surprised if they also have the same, you know, China is involved
0:23:03 in TikTok and using that data to get way ahead in AI.
0:23:05 And like you said, probably we don’t even know.
0:23:08 Maybe it’s an architecture that is way beyond, you know, LLMs.
0:23:10 They just don’t share it with us.
0:23:16 So I don’t know if this arguments of national security is entirely applicable, because yeah,
0:23:19 they can copy it, but you still need to train it, right?
0:23:24 Because even if you show people how you made your model, you still need to have the equipment
0:23:28 and the time, you know, the processing power to train it.
0:23:32 And we’re talking about models like chat GPT that are being trained actively for years
0:23:34 on supercomputers.
0:23:39 And it’s very hard to catch up to them, even if they come up like, and that’s the reason
0:23:44 why I don’t understand why they don’t post, you know, the data, because no matter how
0:23:45 …
0:23:48 But they want Taiwan and all the computing is happening in Taiwan.
0:23:49 Of course.
0:23:50 Well, it’s an important center.
0:23:51 Yeah.
0:23:56 I think every piece of electronic in the world depends on this manufacturing facility.
0:23:58 Even cars, even cell phones, things like that.
0:24:05 So if they’re being, you know, taken out of the equation, we’re facing a very weird future.
0:24:08 At least those of us who haven’t invested in a very fancy computer.
0:24:10 There’s not going to be a lot of parts.
0:24:11 Yeah.
0:24:16 I know, like, I know Jan LeCun, he’s been a big … he’s been very outspoken about
0:24:21 his belief that AGI isn’t going to come from large language models, right?
0:24:26 It’s going to come from some other sort of AI model, like some sort of world simulator
0:24:27 model or something like that.
0:24:28 Who knows?
0:24:31 You know, maybe China’s already got something like that and they’re like, yeah, let them
0:24:33 have fun with their little LLMs.
0:24:35 This is what we’re working on, you know?
0:24:36 You never know.
0:24:42 They probably wouldn’t make it public and put it on GitHub if they were working on that.
0:24:45 You will know about it when it’s too late.
0:24:46 Right.
0:24:47 Exactly.
0:24:48 Exactly.
0:24:52 I’m curious about your thoughts on the sort of ethics of how the code is trained, right?
0:24:59 Because recently we heard about this partnership between Stack Overflow and OpenAI, allowing
0:25:03 OpenAI to basically train on Stack Overflow’s data and then people started going and trying
0:25:10 to sabotage the code that was on Stack Overflow to try to poison the data.
0:25:17 How do you feel about coder’s work just being trained into the data without their permission?
0:25:22 I think if it was a different organization, if it was an open AI, if it was like some
0:25:27 open source type of model, Mr. Al, or like something along these lines that were entering
0:25:32 this partnership with Stack Overflow, I don’t think those moderators will go to the length
0:25:36 of like deleting their contribution, right?
0:25:37 Because it’s been there for years.
0:25:42 It is general knowledge, like it’s something that I’ve been using for many years and since
0:25:47 this website existed, it was helping many people solve their errors.
0:25:53 I think that people mistrust OpenAI in particular, first of all, because it’s not really open.
0:25:57 How can you call your company OpenAI if it’s not open?
0:25:58 That’s the number one.
0:25:59 Closed AI.
0:26:00 Right?
0:26:01 It’s like the number one warning sign.
0:26:02 That’s number one.
0:26:03 Mm-hmm.
0:26:10 Second of all, I think that there’s just a major level of mistrust to something that
0:26:12 is not open source in the field of programming.
0:26:17 We support open source from the very get go, you know, in every, even in the computer science
0:26:21 program that I’m taking right now, I’m doing a computer science BSE.
0:26:24 My lecturer was teaching us about open source.
0:26:30 He was basically explaining why Windows is not as good as Linux, you know, these type
0:26:31 of stuff.
0:26:34 We have it for years and we’ve been teaching people for years about these concepts.
0:26:37 And suddenly there’s this big company that everyone is using.
0:26:39 It is far ahead of everyone else.
0:26:45 They got a supercomputer for free from Jensen, you know, back in the time because they were
0:26:51 open and free and they suddenly decided to kind of close all the doors and take all this
0:26:54 transparency and just throw it to the garbage.
0:26:56 Why would they do it?
0:26:57 You know, is it because of money?
0:26:58 I think they have enough money.
0:27:00 I don’t think it’s because of that.
0:27:02 So what are you hiding?
0:27:04 And a lot of people are thinking that.
0:27:08 So these moderators, I don’t think that you just have a problem with some AI learning
0:27:11 because you can web scrape any piece of information from any website.
0:27:15 You can make a bot that copies all the data you want.
0:27:16 It’s not a problem.
0:27:20 Plus all the data and stack overflow, as far as I’m concerned, it’s public.
0:27:22 You don’t need to log in to be able to access it.
0:27:24 So from a legal perspective, you can already use it.
0:27:26 You don’t need to go into this partnership.
0:27:32 So I think it’s a it’s a very weird, it’s very weird to me that Stack Overflow agreed
0:27:35 because I don’t know what they are earning in this equation.
0:27:36 I think there’s worried.
0:27:37 Yeah.
0:27:41 I think because like apparently their traffic’s way down because of people not using Stack
0:27:42 Overflow.
0:27:43 They just asked chat to be the same question.
0:27:47 I don’t think they had a plan B. I think I think chat GPD came up and ate a bunch of
0:27:49 Stack Overflow’s lunch.
0:27:53 Stack Overflow tried to create their own stack AI or overflow AI or whatever they were calling
0:27:54 it.
0:27:58 They created their own model and nobody cared because they were already indoctrinated into
0:28:02 chat GPD and they went, well, okay, if we can’t beat them, join them.
0:28:08 I think that’s kind of what happened, honestly, but I also think, you know, the whole like
0:28:14 ethos of like open source is like, I’m putting this code out there, but if you build something
0:28:16 with this code, also make it open source, right?
0:28:20 Like that’s sort of the whole like open source code, whatever you build with this open source
0:28:22 stuff, make that available.
0:28:28 And so it’s weird to me that companies like chat GPD or open AI can go train on all of
0:28:34 this code and then close it off when all of this stuff was put publicly available and was
0:28:39 designed for whatever you build with this, also make that publicly available.
0:28:40 You know?
0:28:41 Yeah.
0:28:42 I agree.
0:28:43 It’s kind of shady.
0:28:46 You know, I don’t know what’s the purpose of it, but it makes me mistrust.
0:28:50 And when you mistrust, you know, you had a lot of trust towards Stack Overflow as a developer,
0:28:54 you know, from the get go, like it’s the best source to solve your errors.
0:28:58 And, you know, you can find multiple solutions to the same problem and choose what kind of
0:29:01 solution you prefer, basically.
0:29:06 So it’s a very trusted name that has a lot of integrity, right?
0:29:11 So when it goes into partnership with something that folks like ourselves, like the majority
0:29:15 of the audience that is still using Stack Overflow, we never left.
0:29:17 We kept using it, right?
0:29:23 It’s kind of a slap in the face, and I don’t know, I don’t think that Stack Overflow should
0:29:26 have been chasing them, you know, preventing them from deleting their comments because that’s
0:29:28 their decisions, you know, they enter this partnership.
0:29:31 They should have known there’s going to be a backlash.
0:29:36 You know, they have marketing experts that, you know, chat with users, they understand
0:29:41 what’s going on, you know, it was clear to me, and if it was clear to me, I assume it
0:29:42 was clear to Stack Overflow.
0:29:48 So don’t, like, antagonizing them, like, further, basically, just kind of, you can’t
0:29:49 touch your account.
0:29:50 It’s suspended.
0:29:55 We are checking if you really meant to delete your messages or not, like, come on, you guys,
0:29:56 that’s even worse.
0:30:01 You just build, burning the creators that create on your platform even more.
0:30:03 They’re training on all the GitHub data too, right?
0:30:06 I think that’s even a bigger deal, but that’s the actual code itself.
0:30:11 GitHub is owned by Microsoft and, you know, Microsoft and OpenAI are, they’re like this.
0:30:15 You know, I would imagine all the GitHub stuff is, you know, freely available to OpenAI.
0:30:16 Yeah.
0:30:17 And why not?
0:30:21 You know, it’s public data, like the, maybe the private repositories, you cannot really
0:30:27 access them, but as long as something is public data, and I know it because there was a company,
0:30:31 there was a company that was scraping data from platforms, and they were, they were
0:30:34 accused, bright data, their name.
0:30:38 They were accused of, like, stealing content, they were sued, but when they came to the
0:30:43 court, when it was time to kind of actually talk about the laws behind it, it’s like,
0:30:49 hey, we never logged in to, you know, whatever platform we were, we were taking information
0:30:50 from.
0:30:51 Anyone can copy it.
0:30:56 So what is the difference if a bought, if an automated entity is copying it or somebody
0:30:58 that physically sits with a keyboard does it?
0:31:00 And I think they won.
0:31:01 Yeah.
0:31:02 Yeah.
0:31:03 I mean, we’ve been seeing a lot of the same sort of arguments over on the art side as
0:31:04 well, right?
0:31:10 Um, you know, a lot of these AI art generators have trained on other people’s arts and, and
0:31:14 a lot of that has gone to court and been fought in court and, and same with the writing too,
0:31:18 we had a lot of authors take chat GPT to court saying they trained on their writing and almost
0:31:23 all of these cases end up getting thrown out because well, it’s publicly available data.
0:31:25 They’re, I guess they’re allowed to train on it.
0:31:30 So yeah, it’s, it’s definitely interesting, but I personally have very mixed feelings
0:31:31 about it all, right?
0:31:36 Like I was actually for a little while, um, doing photography and uploading it to sites
0:31:41 like shutter stock to earn, you know, the, the income from selling the stock photography.
0:31:46 Well, a couple of years go by and now we find out that, okay, shutter stock trained all
0:31:49 of the images that were uploaded into like their new AI model.
0:31:52 And Adobe did the same thing with Adobe stock, right?
0:31:56 Everything that’s been uploaded to Adobe stock over, you know, however long Adobe stocks
0:32:01 been around now, all of a sudden that became training data for our AI art generator.
0:32:02 Well, cool.
0:32:06 When I was uploading photos years ago to try to make money, I didn’t really say you had
0:32:09 permission to use that photo as part of the training data.
0:32:13 I don’t really care that much because I’m, I’m like in AI and I’m an AI optimist and
0:32:17 I, I’m kind of fine with it, but I don’t, I definitely see both sides of that token,
0:32:18 you know?
0:32:19 Yeah, absolutely.
0:32:20 Yeah.
0:32:21 I think it’s a, I think it’s a gray area.
0:32:26 I think, I think it’s a gray area where like, you know, they, there’s a reason open AI doesn’t
0:32:27 want to talk about it, right?
0:32:30 Like the CTO, there was that famous clip of her stumbling where they were asked if she
0:32:36 was training on, if they were training on YouTube data and she was like, I don’t know
0:32:38 exactly, you know, kind of, kind of thing.
0:32:41 And then obviously she knows the answer, like obviously she knows.
0:32:45 And I think it’s because probably when they even started the company or when they started
0:32:48 really thinking about this stuff, they would have talked with lawyers and they would have,
0:32:52 it’s kind of like when Uber launched where technically Uber could be argued that it was
0:32:56 illegal when it started, but it was a gray area where they knew they had a legal argument
0:32:58 because there was no precedent exactly.
0:33:03 And so there’s no exact precedent here because, you know, you have public data, but you can’t
0:33:04 just copy public data.
0:33:06 So they’re going to argue, well, the AI is learning from it.
0:33:10 It’s just like if you go to a museum and you see some art or whatever, you’re learning
0:33:11 from that.
0:33:12 You’re not copying it.
0:33:15 And so I’m pretty sure that’s like the legal case they will make long term, but I think
0:33:20 they’re really trying to delay having that huge legal battle, but I think it’ll happen
0:33:21 at some point.
0:33:27 Because right now I decide to train my artificial intelligence model on a subset of data.
0:33:29 You can only see the final product.
0:33:31 You will not see what kind of data I trained on.
0:33:32 I don’t have to share it.
0:33:35 I don’t have to say anything about it, right?
0:33:43 It’s unethical, but you would never know because all you do is you take a bunch of images,
0:33:50 you feed it into a neural network, and then it finds patterns in those images and it basically
0:33:53 creates something new out of this data.
0:33:55 So you can say it’s like a musician.
0:34:02 So for example, the Beatles were inspired by Elvis and Pink Floyd were inspired by the
0:34:03 Beatles.
0:34:04 So who copied from who?
0:34:06 It’s the chicken and the egg.
0:34:12 And you’re right when you’re saying that it’s inspiration because what the model creates
0:34:16 is not an exact match to the training data.
0:34:18 It has to be something else.
0:34:22 They pull it out of late in space when it comes to a generative AI models.
0:34:28 And this late in space, I don’t think anybody owns it, and those type of things, we should
0:34:32 have thought of it way before.
0:34:37 Right now when we’re talking about the legal copyright implications, it’s something we
0:34:41 could have talked about years ago before it became a problem, before people’s livelihoods
0:34:44 were affected from Shatterstock and places like that.
0:34:49 By the way, Firefly, Adobe’s Firefly, is amazing.
0:34:51 I might actually use it in my final project.
0:34:56 I’m doing a final project for university, and I think that that’s my way to go.
0:34:58 I’m going to make a database.
0:35:02 But even in Firefly though, they were pitching that it was the responsible way to use AI
0:35:06 images, and then apparently they trained on mid-journey images.
0:35:09 They basically just went to another layer down and were like, “Yeah, we didn’t do it.
0:35:10 They did it.”
0:35:14 Yeah, well, I mean, they didn’t specifically go out of their way to train on mid-journey,
0:35:20 but they allowed people to upload mid-journey images to be sold as stock images.
0:35:24 So the mid-journey images that were uploaded to Adobe stock, and then when Adobe went to
0:35:29 go and train their Firefly model, well, there was a ton of mid-journey and stable diffusion
0:35:34 generated images in it already, because they were allowing people to sell those images as
0:35:35 stock.
0:35:39 So they didn’t go and train specifically on mid-journey, but mid-journey and stable
0:35:43 diffusion and dolly images were in the data set because people weren’t allowed to sell
0:35:46 that stuff on Adobe stock.
0:35:47 That makes more sense.
0:35:48 It’s crazy.
0:35:55 If you start nitpicking what pieces of data went into the model, it’s an ever-ending story.
0:35:59 You can always find somebody who contributed to create this image.
0:36:00 So what?
0:36:02 They deserve to get paid too.
0:36:03 It’s a rabbit hole that I don’t…
0:36:04 And how much?
0:36:05 Right?
0:36:06 I’m a very simple person.
0:36:12 I like the internet how it was when I was 12, which was basically wild, wild west.
0:36:14 You could copy anything.
0:36:15 You could get anything.
0:36:16 You can download anything.
0:36:20 And as a content creator, I know that it’s a terrible thing to say because people will
0:36:21 copy my content.
0:36:22 You know what?
0:36:26 If they want to, please, it only helps Python being thought.
0:36:28 I make enough money from what I do.
0:36:35 I don’t care if somebody else is basically using some of it to kind of grow this Python
0:36:38 world and kind of teach others.
0:36:42 If a teacher is using it in his lecture, I’m not going to chase him, asking him for some
0:36:43 royalties.
0:36:44 It doesn’t make any sense.
0:36:45 Why would I do it?
0:36:54 Same goes for a musician where some 12-year-old kid decides to use his piece of art, his music,
0:36:55 in his video.
0:36:58 Why would this musician chase the kid?
0:36:59 He is sharing his music with the world.
0:37:04 So yeah, you don’t get paid for it, but you get paid enough.
0:37:10 It should cover other fan art if you want to call it.
0:37:15 So I find it upsetting that people are really nitpicky with those copyright laws.
0:37:18 And if we weren’t, maybe the internet would have been a nicer place.
0:37:19 Yeah.
0:37:20 Yeah.
0:37:23 I mean, I really think we’re getting to a point where copyright in general just needs
0:37:24 to be rethought.
0:37:28 I don’t know the solution, but I think it just needs to be rethought because we’re going
0:37:32 to be able to generate images that look like other people’s images but weren’t created
0:37:33 by that person.
0:37:37 We can already create songs that sound like other people’s songs that wasn’t created
0:37:38 by them.
0:37:42 When we get into Sora and Veo and some of these new video generators, we’re going to
0:37:46 be able to generate videos that look like the style of other people’s videos but weren’t
0:37:48 created by them.
0:37:53 It’s just going to get so muddy that I don’t feel like the way copyright law was originally
0:37:55 written is still going to be relevant.
0:37:57 It just needs to be rethought.
0:37:58 I agree.
0:37:59 Absolutely.
0:38:00 But I don’t know how to rethink it.
0:38:02 I’m not the one to figure that out.
0:38:06 Actually, I had meetings in the Library of Congress with people in the Copyright Office.
0:38:10 I don’t know if you know my last startup, Binded.
0:38:14 We started off not being involved in copyright and then unfortunately our startup, we kind
0:38:18 of pivoted towards copyright, but we were not trying to be doing enforcement and stuff
0:38:21 like that with more just attribution and things like this.
0:38:24 I spent a lot of time thinking about how can you change this and I met with people in
0:38:28 the government and now I’m convinced that I can’t.
0:38:31 They have no interest in changing things.
0:38:34 I don’t think there’s going to be a fundamental like, “Oh, the government’s going to decide
0:38:36 that copyright should change” or something.
0:38:37 Who knows?
0:38:38 Who knows?
0:38:39 I mean, it could be a generational thing.
0:38:40 Right?
0:38:47 Maybe once there’s a little bit more turnover from the more elderly people in the governments.
0:38:51 Maybe some of the younger generations will see that, “Okay, technology has changed a
0:38:55 lot since a couple hundred years ago when so many of these laws were written.
0:38:59 Maybe we should update some of this stuff for the way the world is now instead of the
0:39:01 way the world was a hundred, two hundred years ago.”
0:39:02 Yeah.
0:39:04 Maybe GPT-6 will help us figure it out, you know?
0:39:05 Absolutely.
0:39:07 Well, this has been an absolutely amazing conversation.
0:39:12 We’ve had a blast talking to you, Maria, and I want to make sure that people can go check
0:39:13 out your stuff.
0:39:16 Where should people go and follow you after tuning into this episode?
0:39:19 Where’s the best place to learn tutorials from?
0:39:20 Yeah, the best place is always YouTube.
0:39:21 That’s my main platform.
0:39:24 This is where you can find shorts.
0:39:26 You don’t like TikTok style shorts.
0:39:29 You can find tutorials that are quite long.
0:39:33 This is where you find me, explain very complex concepts in a simple language that even a
0:39:36 six-year-old can understand, hopefully.
0:39:37 It’s called Python Simplified.
0:39:38 Python Simplified.
0:39:39 Yeah.
0:39:40 Python Simplified.
0:39:41 You can find me on YouTube.
0:39:44 You can find me on X as well, even though I’m not there very often.
0:39:46 Maria Simplified.
0:39:51 The best place is to find me is YouTube, Maria Shah, Python Simplified.
0:39:52 Well, very cool.
0:39:55 Thank you so much for hanging out and talking code with us.
0:39:59 This was a conversation that we wanted to have, but we wanted to bring on somebody that
0:40:02 knows a little bit more about the coding world than us.
0:40:05 I’m so thankful that you were able to join us and actually have this conversation with
0:40:06 us.
0:40:07 I really appreciate it.
0:40:08 Awesome.
0:40:25 Thank you.
Episode 12: Are coding jobs at risk with the rise of AI? Matt Wolfe (https://x.com/mreflow) and Nathan Lands (https://x.com/NathanLands) dive into this compelling topic with guest Mariya Sha (https://x.com/mariyasha888), a seasoned coder and the creator of the popular YouTube channel Python Simplified.
This episode delves into the contradictions and synergies between artificial intelligence and coding, featuring Mariya Sha, who started coding at a young age and later found success with her YouTube channel that simplifies Python programming. Together, they explore the changing landscape of coding due to AI advancements, ethical concerns, and the future of AI-integrated coding environments. Mariya shares her skepticism and hopes for the future, particularly AI’s potential impact on coding jobs and the importance of a personalized touch in YouTube content creation.
Check out The Next Wave YouTube Channel if you want to see Matt and Nathan on screen: https://lnk.to/thenextwavepd
—
Show Notes:
- (00:00) Confusion about AI models and documentation use.
- (05:35) Exciting potential for non-coders to code.
- (08:36) AI is better at handling fast change.
- (09:49) Engineering and coding solve problems, with AI help.
- (14:51) Future AI control raises transparency and ethical concerns.
- (17:03) Debate over open source AI vs national security.
- (19:33) Concerns about LLM capabilities and potential surveillance.
- (25:20) Janssen’s free supercomputer and transparency questioned.
- (26:24) Lack of plan B led to GPT domination.
- (32:02) AI model training ethics and inspiration discussion.
- (35:04) Sharing is important, copyright laws are nitpicky.
- (36:43) Startup pivoted towards copyright, government unwilling to change.
—
Mentions:
- Grab HubSpot’s free AI-Powered Customer Platform and watch your business grow https://clickhubspot.com/tcp
- Mariya Sha: https://www.linkedin.com/in/mariyasha888/
- Python Symplified: https://www.youtube.com/PythonSimplified
- Stack Overflow: https://stackoverflow.com/
—
Check Out Matt’s Stuff:
• Future Tools – https://futuretools.beehiiv.com/
• Blog – https://www.mattwolfe.com/
• YouTube- https://www.youtube.com/@mreflow
—
Check Out Nathan’s Stuff:
- Newsletter: https://news.lore.com/
- Blog – https://lore.com/
The Next Wave is a HubSpot Original Podcast // Brought to you by The HubSpot Podcast Network // Production by Darren Clarke // Editing by Ezra Bakker Trupiano