OpenAI tools and the future of programmers

From CitconWiki
Jump to navigationJump to search


Transcript: All right. The Evernote recording is now live. I'm happy to send it to anyone if they like in the future and I'll figure out a way to Put it on the wiki at some point. I don't know Yeah, okay, I'm sorry to interrupt you Mache(sp). What were you saying?

Let's go to PJ.

Oh my well my point was gonna be that for me the all of this AI stuff is Is doing what temp, you know, like Archetypes were doing before it's giving me the template right and the template is just getting better and better and better so now my job is about editing testing You know doing the real value add for my business rather than writing all this boilerplate crap The you know although I enjoy boilerplate crap It's because I'm a weird person right, so yeah, you just said testing it seems to me that that's a realm that AI could really excel in Because you know from the beginning we've always seen testing as like this mathematical proof, right? You know feed it a and you should get B back and We all know I mean you've written many books on this we've been talking about testing for years And it's still a topic of conversation to get developers to write tests like so I think there's an opportunity there for like AI just Throw AI at this and that's a problem. We're talking about designing tests. So what I mean all of it He's just saying you know Take all testing You know like why doesn't I mean it's doing it's the it's the task that I at least you know, it's like hate Yeah, we're still talking about it today, right? I mean it's still like an asking developers right test I mean you were talking about like being around TV people I mean like I'm like 15 plus years old like that we've been talking about this right? It's still like what so yeah There you go. Sorry such bad news for all of you I am gonna burst your bubble because it ultimately testing is about getting feedback About the extent to which your actions met your intentions And at the very lowest levels that's unit tests and at the much higher levels It's did we make revenue, but it's always do a thing see what happens and in the world of AI Especially where we are today like I love your perfect world But we're not there yet and right now chat GPT and their ill are far more on the end of it is plausibly Correct, then it is accurate and that means that the work shifts to testing We got a thing didn't meet our intention. So I have such bad news Now I also have a good news because it turns out that if you don't think testing is fun, you're doing it wrong I've been hearing that for 20 years I'm still talking about the same thing So the question I have is you brought up capitalism earlier, which No, it's we can talk about whatever we want but Masha you were talking earlier about with the example of Circuit boards right that there can be done better than a human can more efficiently and probably at some point cheaper Although I don't know what let's assume enterprise scale AI looks like right now So because we do live in capitalism the big concern I would think for people who design circuit boards is Why would the company pay you if they can just have this AI built and then you're out of the picture Yeah, I think especially when we get to this level two that you're talking about where Eventually the AI is strong enough that you don't even need to program All you need to be able to do is ask it a question and then see if it's right with a test. So Yeah, I mean, what do you guys think is that like every layer, you know technology advances through like layers of abstraction. I think it was kind of You know saying something along these lines as well that our lives are going to get better as a result of this, right? But you know, we all use compilers underneath all the code that we write. How many of us work on compilers today? You know, how many of us know someone who works on compilers today? You know, how many people are there that we work on compilers today, right? You know like back when I started programming Yeah, I knew people who worked on compilers, you know Well everyone has to take compilers. I don't know whether that's still the case You know, I think like if you don't take a compiler course and you get a CS degree What kind of Mickey Mouse degree did you get? You know But yeah, you know writing a low-level compiler is just kind of not something You know that a programmer today And they're would ever use in their day-to-day life are like 99 point some percent programmers You know, so I mean I think we're just going you know select the next level of abstraction makes the layer below kind of a niche at that point. There are still people who write compilers, and they're great at it, and they're experts, but we just never think about them. And there will be people who will write this software just like the AI software. Someone is going to write Torch or whatever framework it's going to be running on. We're just not going to think about them. They're going to be off somewhere and it's just magically going to be produced. Because it's harder and harder to predict the future. So that charge-adp is exactly as IBM 360 that your father worked on 50 years ago. So we cannot even think what's going to happen in 50 years. I mean, charge-adp will be ancient, archaic thing. So we have no way to imagine what's going to happen in 50 years. You guys probably read that book, Singularity. So lots of thoughts there. But we have no idea how is this, because time is accelerating. So inventions are accelerating. We have no idea what's going to happen, maybe not even in 20 years. So I'd like to continue Elizabeth's train of thought. So I'm part of a geek cabal. We've been testing charge-adp and it is really horrible at facts. If you talk to it about history, it's very Trumpian. It will sound very, very convincing, but it's just wrong. Oh, you don't mean embracing the political philosophy. It's just the biases that probably were in the data that it was trained on. In the facts, it's a probabilistic. So basically, yeah, if you talk about historical events and you say, no, but really this isn't, that's the fact. It's like, oh, I'm so sorry. And it will continue to, it will take your input, put it into the buffer or whatever, right? And in your session, it will get a little bit more accurate, because now you've told it and it will keep going. And also, it's really bad at math, like it doesn't understand math at all. And there was an interesting article by Stephen Wolfram on this topic where he was saying, chagivity has some good points, and then what you should really do is marry it with the alpha, or the alpha Wolfram. Yeah, alpha. And because we actually have a lot of friends. No, the guy who came up with Wolfram, who thinks you should use it, is like, crazy. Right, but that's a good point. The point is, if you are right now, if you don't understand how it works and you submit your paper somewhere, you can be very embarrassed. So just to award a warning, it is very creative, very convincing, but really bad at facts and doesn't know anything about math. Well, I think it's really good for creating a framework of your paper. And then, remember, David, you always have a chance to press the button, make another one. Right? And then it may just vary dramatically from version A to B, right, to C and D and F. Right? It's just like, I tried it. I wasn't happy with the results that were presented to me. Then I pressed it again. And it actually came up with a different set of variables and facts that spin it out at the end of the session. So I think Andy made a really good point about early searching. And I actually was just talking to someone about this, maybe last night even, that when I was a kid and learning how to search, my stepmom, Leanne, most of you met during check-in, was trying to think of a song. She couldn't remember what it was called because she didn't really know at the time how to use search effectively. And I was able to find it in five seconds. And that was magic to her and to a lot of my less technical friends. So I think you're right that the skill is going to be how to ask the specific AI to get the output that you want. But speaking of chat GPT, I know you said it's not good with facts and it's not good with dates or math or whatever. But if you're in an organization, and let's say instead of having a team of software engineers, I have two software engineers that then manage this AI, that AI will specifically be trained on doing what the team used to do. So you're not going to have those issues. Yeah, you probably can't replace your engineering teams with chat GPT today. But in five years, if you as engineers are told, hey, build this AI and then you won't have a job, I know a couple of people here are optimistic. But is anyone worried about that? Because I would be if my current job was technical. You're not worried at all. I think one thing that I think about, you mentioned people who write compilers. I would not be surprised there were more people writing compilers today than there was 50 years ago just because they... industry has grown so much, it's going to be a smaller percentage of the total value that's created by the software ecosystem. So I do think that, like, I have a friend who's in a different session and they say, like, they think that writing more code using robots just means there's going to be a lot more code, which means there's going to be a lot more bugs that we can define, which means that there's going to be, there's still going to be demand for humans within that system. You just basically said that the humans are going to be doing the shit jobs, right? We're going to give the fun jobs to the robots and then we're going to have to clean up the frigging toilets and the, you know, truck stops. No, finding the bugs and fixing the bugs. That is literally cleaning out the toilet and the truck stops. Well, it's finding the thing that needs to happen. That's fun. No, no, no, but I think, one second, let's look at another one. What if we all need to get really good at specification, right? And actually asking the system exactly what it is that we want, right? Maybe that's the job. And to me, that's how I'm, like, to me, specification has a lot to do with being good at abstract thinking and, you know, it has a lot to do with, like, searching the space of possibilities, you know, edge conditions and all these things, well, these things that some testers do. And to me, that's actually fun. Like, I moved from being a, you know, software developer into being a tester and I'm having fun and there's some interesting tools coming into our field and, no, I think it's all good. If you're not having fun testing, you're doing it wrong. That's the rumor. If testing is checking that the software satisfies the intent, the other big challenge is specifying the intent. Exactly. And knowing the intent. I mean, that's fun. Hello. We've been talking about Agile for 20 years and what is that about? We don't know what we want. If it was so easy to specify intent, we wouldn't be talking about Agile because it would have been specified and the specs for the software would be perfect and we know exactly what to go off and build, right? But that problem is inherently, you know, what, NP complete or whatever, NP hard problem, right, specifying intent. So I don't think that Agile says don't specify. I think it says don't put the phone all the front. I'm just saying that Agile deals with the challenges of being able to specify intent. And if it was so easy to do, we wouldn't need Agile. So maybe that just means we have jobs for the foreseeable future because, yeah, our bet is that, well, maybe it's both. If we've been struggling for 50 years to figure out how to specify intent correctly, so either that's just a hard problem. Or computers will figure it out very quickly. Well, I mean, I think if PJ said, you know, move up high value add, right, you know, I don't need to write the boilerplate code. I can move to higher value add stuff, you know, and I mean, yeah, we've always been moving in that direction. But a lot of, you know, like for me, the code that I write, you know, it's like, yeah, I kind of, I have an idea of what I want, you know, I have an idea of what I wanted to do, but, you know, I work, so for me, I work a lot with data. And, you know, a lot of it is just exploring, you know, so I can't tell you ahead of time, you know, well, I can tell you ahead of time, this is what I want to try, you know, but I can't tell you ahead of time, the final piece that I'm going to want to end up with, it's going to be this, this, this, and this, right, because there's no way for me to know that because I don't have enough experience with it. I don't know enough about the domain yet, right, so I'm learning as I'm trying things and, you know, in order to try things, I need to write code to analyze the data, you know, so that I can learn from it and decide what next thing to do. It's kind of an iterative process, as I think kind of most software, you know, development, like the business owner, you know, they have a rough idea, you know, but I think it's kind of very similar, it's like you try something, you know, and it either works or it kind of doesn't work as well as you were hoping, but it gives you another idea, so, you know, you kind of then go to the next thing. But I think you just described gender to the AI. Yeah. It just happens. That's what I was thinking. And it takes us weeks or years of lifetime. Okay, all right, all right, all right. Yeah, that's the way I use ChatchiBT now, right, I give it the first prompt and it will be off by a little bit, right, and then I say, update this to include these other cases. Yeah. And then it gives me a revised version. And then I look at that and I think about it and I say, oh yeah, and don't you know and also add in this Give give it to me again So the cool thing is it iterates really quickly. Yeah, right You know within five minutes I get to like I was writing personas the other day for specific software cases, right and within Ten minutes I had you know six fully written personas With all of the different characteristics that I that I think are important Added in you know and if I had had to sit down and write that myself it would have taken me a couple hours, right? Yeah, just writing text So for me, that's the way I use the chat GPT thing is it's this iterative me giving it feedback I love the conversational style of the way that the chat GPT thing works This is really really exciting for me now. I haven't done Enough like code stuff with it yet to see but I I can imagine say asking you know give me You know give me a Java class that will connect to a database of this URL and And you know select all of the user records, right? Yeah, and it would give it to me and then and then I would iterate Right, and it's gonna give me Java code And I can get to the point where I could then take that and then I'm doing the testing You know what else is interesting. It's really good at getting a job like at a fan company, right? Like you give it any lead code problem. It will ace it. Yeah, that was clear. That's because there's a bunch of answers on the internet right, right, but but the point is then You know when we talking about jobs of the future versus jobs of today, right? Like to get into a fan company at least I have to my age spend like months and months and months practicing for these Interviews that yeah, as soon as they get higher that will like never do that work again Right, what they will ask me to do is to specify the system right right because that's my that's my real job But I have to you know so so that like the question is like are we really having like people? You know doing the coding these days are they really having you know dream jobs? Do they really really enjoy it or they fooling themselves, right? Because I think like a lot of the coding jobs today are still way way too low level and not that much fun and like a lot of grind So that's that's my experience. I think that's a really interesting point Like I think the like even the v2 that Andy was talking about some of these chat GPT things Like I think it really comes down like as a software engineer Like I need to be really good at using these tools to like effectively do my job and I'm not scared of these tools replacing my job but Is like a member of like I work on a platform team and like we have a really wide array of services I can only focus on so many things at one time like I need to take these two wings and like Be more effective across the board like all my services are actually leveling up because I know there's things that I'm neglecting There's things that I'm not doing right like Level one level two of this stuff is really like how do we bring this in and how do we start using it effectively? And that's like what I'm trying to understand and even don't understand and even like you're saying with some of the like Iterating on some of this stuff like how do you really get something from these like open source tools that you can use internally? because internally you have all these libraries you have all these things right that like you have to go in and fix and go in and do it so like Is there a way for us to take this and just like easily fill in those puzzle pieces without having to like Constantly that I think maybe the answer is like you have like an internal team that's like feeding this data, right? Yeah, but even if you have that team feeding this data It's only gonna be as effective as the person asking those questions Sure putting those pieces in place. So yeah, it kind of goes back to what Andy said. I actually have a question Does anybody here hire people? right now You hire people so what is your solution to? This problem where any problem you give to an engineer they can immediately solve it with an AI or Is that a problem? Do you say well? They have you know, like you're saying they have a tool that they can use to solve the problem That's all I need It's an interesting question for what it's worth. I am not hiring or what when I interview I'm interviewing leaders Okay, it's a different skill set gotcha, but even when I was hiring let's say I see I I've always been less interested in like can you whatever solve the big end problem or whatever it is big Oh sure and more on like what's your passion? And what do you like what are you interested in and can you solve any problem that's thrown at you? That's right. I'm not so much worried about that paradigm shift the technical replacement I Would kind of take objection with your premise. Okay, right? I mean I think You know a lot of like hard engineering problems I don't you know There's many approaches like take scalability Right? You know, there's, you know, asking an AI, I have not tried this, you know, but I think a lot of times, you know, I'd be very surprised if AI could effectively solve, you know, like hardcore scalability challenges, right? Even if it could propose, you know, I mean, I'm just thinking back to, like, what Brian Delasanti, you know, used to the work he did, you know, approaches that sound, I mean, like when we did our product, you know, there were generations of evolution of kind of scalability, you know, you kind of do the, you know, easy, the easier thing first, and get to the max scalability on that, and then you have to go down lower level, you know, and you max out the scalability of that, and then you go down even lower, and you kind of, you know, so I mean, I mean, maybe AI could give you that solution, but, you know, isn't part of the solution kind of like this whole path? You know what I'm saying? Yeah. And maybe that's the specification thing. I don't know. And there's probably 10 potential bottlenecks the system like yours could have. Yeah. And it could give you 10 solutions for each of the 10 bottlenecks, but the question is what bottleneck are you actually going to have? Right. Which of those 10 is actually going to work, and which is achievable and within your skillset, and, and, and. That makes it really tricky. So interesting. So this is old tech, and I'm a firm believer of like, whenever we say the tech doesn't do that, that like we will be proven wrong in some point in the future. However, at Netflix, we did some early, I'll say AI stuff with, you know, AI stuff. Or around looking for errors on individual instances. So you deploy an app and we watch it. And then the idea was to alert developers before the actual thing got triggered. So the system would learn, okay, like when this starts, you know, when whatever CPU starts spiking, like alert now, not later. And what we discovered, and this is a project like maybe six or seven years ago, is that I was gonna say nine times out of 10, but more than half the time when the alert would, when you get the alert, Eric, you'd be like, I already know what that is, and it's not, it's not a problem. And so then we'd have to go back to the system and say, okay, if it's this, don't do that. And it became a rules engine with lots of ifs, like, right. And so we abandoned it because it was like, there was no value in this. And to your point, they're like the humans have at least then, and I'm sure still today, like we have this ability to kind of infer like, that's not really like the book says that's a problem, but in this system, it is not a problem. We've seen this before, it's not gonna break. So I do wonder about that kind of paradigm. Will this, you know, the computers ever get to a point where they could reason kind of that. And today, the answer's no, but I'm sure eventually they will. Yeah. And on that one, the whole ops space is working really hard. Yeah. Paul, to go back to Anton's question about the interviewing, right? So I recently had to take home tests for an interview and I used chat GVT. Right. And then during the interview review session, I told them, I use chat GVT to do this. Here's the problems I ran into. Here's how I corrected those. Right. And then I published it. Right. So I was just very straightforward with it. And I thought to myself, if I was interviewing people, one of the questions I would ask them is why didn't you use chat GVT to generate the base code for you now? Right. You know, why did you spend eight hours writing base code when you could have gotten it in a minute? I do have a blocker around why I'm sometimes reluctant to type stuff into chat GVT. And that is sometimes if I'm working on something professionally, I'm not sure where that data is gonna go. Right. So, yeah, am I allowed to do, like, you know, so that's one thing. There's a solution to that. Maybe chat GVT has like a paid version where they'll throw away your data in the future, perhaps. Right. It's coming up. There's a $20 subscription model that they're going to introduce at the end of this month. Yeah. More professional. But it's more interesting. There was nothing about throwing away your data. Well, I'm pretty sure that there is configurability of some sort. Like, you probably get some kind of a gain over the tool, but my, like, being here in the valley for a long time, I'm just wondering, you know, like this whole relationship between the chat GPT, the parent company and Microsoft, what that's gonna produce and how is that going to affect the alphabet in general? Like whether they're going to be in a world of hurt or they're gonna come up with their own set of tools. You're in the search business right now. You are in a world of hurt. It's gonna. Yeah. You can figure out how to handle it. Yeah. It's just literally from Facebook methods. Apple and everybody there, they're slashing their throats and there's nothing but blood. And then all of the digital marketers that were solely reliant on very accurate data, that data isn't there anymore. So the actual accuracy is just going through. You lost me on this accuracy of the data, digital marketers? Well, because there was a superfluous amount of data on the users until Apple introduced the version 14 of their operating system, iOS 14, which makes the user opt-in to let the application know about certain things about themselves. So the business model for Facebook and Instagram just went directly into the drain. Now the same goes with AdSense for the Google right now because of the... On iOS. For certain things, I'd much rather ask chatGPD and pay 20 bucks and then be done with it, right? Yeah, but I mean, Google, okay, my perspective, and I'm not in the valley, so I don't get the talk and all of that or not part of the grapevine. But I mean, Google has been more, in my mind, more on the leadership side of AI research and development than open AI for a very long time. So the fact that they don't have their chatGPD right now, in my mind at least, is not a capabilities issue. They have more of a business, other issues that they've kind of alluded to, like safety. It's okay for open AI to release this because it's open AI and if it starts spewing wrong information, as we all know it does, it's okay because it's open AI. If it's a Google product and Google released it and it's now spouting wrong information, that's not okay because we trust Google to give us correct information, right? So there's a far greater potential detriment to Google's business model for releasing this, which I think is probably the reason why they haven't released something like that yet because the downside is far greater, the risk is far greater for them, not because they don't have that technology. So in my mind, I don't know for a fact, but from my understanding they are way ahead in AI research. I don't know if they are or they aren't because essentially when you grow and then the companies are getting bigger and bigger, you kind of underestimate what other folks are doing and you're not stepping into the market and that's what happened with Yahoo. They were like the dominant force and then MySpace was swiped out by Facebook. I'm just saying that the relationship with open AI and the Microsoft and talking about Bing being part of it, I'm not worried about it. I mean, I love the open AI tool and then whether that's with Bing or with whoever else, I'm going to be definitely using it, right? Yeah. I think too how to point it about that. But I'm just worried about the Google. I wanted to respond directly to the Microsoft open AI. Yesterday, there was an announcement that Google invested a major stake in Anthropic, which is another large language model company. So I'm actually not sure that if it becomes a big thing, I think that Google is going to try to be as fast a follower as possible. But they're followers. Yeah. What do you mean by that? Yeah. Like, if it actually looks like Bing's eating the market share of Google, I would expect that Google, even if they have safety concerns, might be like, we need to release this in order to. I think it's interesting because we're all looking at Bing. But so Microsoft, my understanding is that in the last two weeks, 10 billion into open AI. And Satya, the CEO of Microsoft, he's made it very clear that AI in general now become part of their entire tools. So to me, Bing's like V1. And it's like V2 is like, how are they incorporating it into whatever, I don't even use Microsoft products anymore, but like Teams, Office. Yeah, exactly. That's where it starts to get super interesting. And you can see, and I do believe Google has some work going on there because you can see it in Gmail when it will complete your sentences for you. I mean, it's like very like whatever pedestrian, but it's like, this is getting very interesting. But to your statement about Google throwing a bunch of money into that, Anthropic is interesting. There's first to market and then there's first in market. I think open AI is first to market and like we're all talking about cheap, cheap, chappy tea. The next big thing in the valley will be who can actually capitalize this and ultimately make money off of it. I had a thought related to what you said, PJ, where you would ask people. you know, why didn't you use chat GPT for the base code of this? Does anyone or has anyone thought about working with their teams and doing like training sessions using these new AI tools? Cause I don't lead a team right now. It's something I'm looking to do, but I would definitely be doing that. That would be like a once a week, let's refresh on the state of the art to make all of our work more efficient. Can, can as part of that, can you go into a little bit more detail into like, what did you have to correct? And you know, what did you have to, how did you have to massage what chat GPT came back with? And do you think that it's like ready to be used? And importantly, did you get the job? You said it was part of an interview. Oh yeah. Yeah. So, so this was, this was for the dev advocacy work that I'm doing with, with the local base. So the take, the take home test for this was write a blog article, write a tutorial and create a video. Okay. So it wasn't due code, right? It wasn't code. Okay. So the blog article, I mainly, I took the blog article from chat GPT, I put it in a Grammarly. And I also put it into. Got rid of the watermark. And then there's another tool, Hemingway, I think is another, another editing tool. Right. So I ran it through bowl, corrected some of the things, made sure I was happy with some of the choices chat GPT had made. The main thing I do with chat GPT was I told it to, for the blog article, I said, massage the style of it, like make it cheekier. Because I saw somebody on TikToks, they added that into their prompt and then they liked it better. So I added it into my prompt and I did like it better. It was like a fun or a blog article. So the, the tutorial, it did very poorly. Yeah. It didn't. It did okay, but some of the steps weren't correct for how to set up liquid base and that's where it's actually kind of like code. It's liquid base configuration files and stuff like that. But it did, to your point, gave a really good outline for tutorial. And the steps you need to take and, you know, every iteration I did on it was like five or six steps. So I landed on cool. I'm going to do it in five or six steps and I'll just massage this a bit. And then for the video, I had it write the script for the video because I'd seen people on TikTok saying, I've got chat GPT. Sorry to be flipping everybody off. I've got chat GPT writing my, my my blog scripts now. So I had to write a script for me, which then I just kind of changed to be more my style. And then I didn't actually read that script when I made the video. I just, I had it in my mind and I, you know, was kind of more fluid about it. So that's how I used it. I did it enough, you know, it's a part time consulting role, but, uh, but I did end up getting the job and they were, they were impressed that I use chat GPT to create this and they were happy with my answers about what I changed and everything. But it probably made you a lot more efficient time wise. Yeah. And, you know, so why wouldn't they be happy about that? Right. Yeah. Right. You know, exactly. Yeah. Has anyone used it for code though? Yes. Yes. Oh, okay. Well, then why would you tell us? Well, uh, well, I was talking about in terms of the people management side in terms of helping your people learn how to use the tool effectively, because the same sort of skills that made me good at search in the mid 2000s make me good at interfacing with the AI now, but not everyone has those skills because people who grew up, even my sister who's here, who's 10 years younger than me, her search has always been good. She's always been able to just type in whatever and it's popped up. Whereas when I was starting to search things, you had to ask in a specific way. And that's the current state of AI. What I was able to do just like PJ, I've always been interested in funnily enough, I had the AI help me code an AI in neat. That's something that's been too technical for me. And there was the same sort of thing where I asked it and it said, here's the basic framework. It's missing a couple of things. And I said, okay, great, rewrite it with those things. And then it did that and it was like, you still need this. And so, yeah, four or five iterations. And it was able to spit out working code. And you pressed run and it ran. Yeah. So there is still the skill set that only works because I have, even though I haven't done a lot of development work, I've still done basic scripting and I have sort of the background that I'm able to read it and check on it. But that takes, that's a lot lower barrier to entry in terms of hiring than, you know, a full computer science degree or two to three years in QA testing or whatever it is. To your question earlier about like, as a people later, would you, would you, would you, would you, whatever make time for. to learn how to do this. So my default answer is no. He's turning your question around on you, man. Yeah. Because it's not like we make time today for like, hey, you know, David, why don't you spend some time learning Java? Like, it's assumed you can pick that skill up on your own. Right. I'm not confident that is the right answer going forward with these kind of mindshifts. But right now, I would not advocate, let's spend an hour a week learning about chat, GVT. I may provide some context to teams saying, hey, this thing's pretty big. You might want to look at it on the site, but like I wouldn't formally. Because there's a business we're trying to keep going. Sure. I would get somebody to do a bunch and learn on it. Yeah, something like that. Yeah. And not necessarily once a week, but you could do once a month and say, hey, these tools are immensely powerful. And they're going to keep getting more powerful. And for your own sake, as my employees helping keep the lights on, this will make you better at your job. So look into it. That's more powerful than saying, hey, one hour a week, I need you practicing on chat. Yeah, no, that's not what I meant or had in mind. I mean, you could. The best practices aren't known right now. They need to be discovered. Right. So I could see in a larger organization trying to say, we have competitive advantage if we discover the best practices. If we're sharing what we learn internally, we set up a community of practice or whatever you want to call it around how do we do this? And then we disseminate that information across the organization. And that's not a weekly training because there's no one to teach it. So you've got to open up some opportunity for people to explore it and share and build that. So you learn to write cheeky. And like, right now, TikTok is our best source. And there's numbers missing. Well, who said TikTok wasn't there to teach? You said there's no one to teach it. But what I would envision is actually a meeting very much like this, where once a month people say, OK, hey, I've tried to use this. I've used these tactics to massage the AI. And you can build best practices in your organization just from having people be aware of it. And like you said, picking up the skills over time. And I think that'd be an unbelievable advantage chain when you can figure that out. And I'm starting to use it in my own work. And so I work. Don't tell them. But I work like an hour a day because I can just use these tools to help perform. Have you found Millennial Nirvana then? I mean, I. You traveled the other seven hours. I took. Well, no, I don't travel. I can tell you, I have kids that are, I have twins that are 26, going 27. And I have a 35-year-old. And they're all kind of in computer businesses. And I'm hugely impressed how hard they work. And everybody says that they're lazy and such. But no, I'm just joking. This generation says that the future generation is. Yeah, no, I think they're actually working much harder. No, we're all in the same building downhill for the last 5,000 years. Yeah. So no, they're good people. Are you about to say we're at time? No, no, no. I want to throw one more. Final thought. OK, and then it's 11. So I would like maybe we'll continue this on the hallway. But I'd like to ask of you, who is interested in the idea of a sovereign digital twin, which is AI that will basically work on my behalf, right? Digital twin of me. Sounds like slavery to me, man. Robot rights. Hold on, Marsha. Let him finish. I'm curious where he's going with that. But one that I control, because like to Tim's point, right now I'm also very about talking to ChedGPT because it's building a model of me, right? I don't know what it's going to do with it. I would rather have this be a model that I control. Yeah, and obviously solar powered and the GPU chips. Right. So that's what I'm investigating right now. That's a great idea, by the way. It's a great idea. But then you would probably have to have a bigger brother to pay the taxes on yourself as well, because essentially you will not be able to sustain the amount of computing power that we have today in order to even fractionally represent you. And actually, I just saw the recent talk about talking about the human brain and some kind of a weird theory for which Steve Wojnar is great for. He's kind of off the wall. But he said on a neurological level, like the human body, human brain, we figured out all of the things where everything is. But he said, I could not find, nobody can really find the memory, right? And he said, when the kids are seven year old, they lose their teeth, and they lose the health of the childhood memories. And he said, maybe the teeth are the. All right, let's check out the board.