DataTopics: All Things Data, AI & Tech
Welcome to the cozy corner of the tech world where ones and zeros mingle with casual chit-chat. Datatopics is your go-to spot for relaxed discussions around tech, news, data, and society.
Dive into conversations that should flow as smoothly as your morning coffee (but don't), where industry insights meet laid-back banter. Whether you're a data aficionado or just someone curious about the digital age, pull up a chair, relax, and let's get into the heart of data, unplugged style!
DataTopics: All Things Data, AI & Tech
#36 Altman's Chips, Mojo Gets Fast and TUIs
Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.
Welcome to the cozy corner of the tech world where ones and zeros mingle with casual chit-chat. Datatopics Unplugged is your go-to spot for relaxed discussions around tech, news, data, and society.
Dive into conversations that should flow as smoothly as your morning coffee (but don't), where industry insights meet laid-back banter. Whether you're a data aficionado or just someone curious about the digital age, pull up a chair, relax, and let's dive into episode #36 titled "Altman's Chips, Mojo Gets Fast and TUIs", featuring Nemanja Radojkovic, an MLOps Lead and Educator, as our special guest.
In this episode, we explore a variety of cutting-edge topics:
- Text-based User Interfaces (TUIs) Rediscovered: Delving into the resurgence of TUIs with Ratatui and gping. Are we witnessing a TUI renaissance? Ratatui | gping
- The Surprising Length of ChatGPT's System Prompt: Unpacking the implications of a 1700 token system prompt. Is there more than meets the eye? Reddit discussion
- Mojo Outpaces Rust in DNA Sequence Parsing: A closer look at how Mojo outperforms Rust by 50% in benchmarks. Is Mojo the new king? Modular's blog post
- Sam Altman's Vision for the Future of Chips and AI: Examining Altman's ambitious plan to reshape the business of chips and AI with trillions of dollars. WSJ article | Reuters on Microsoft's AI chips
- The Real Challenge in Generating Code: Discussing the misconception that generating code is the hard part, with insights into the complexities of software engineering beyond code generation. Nick Scialli's blog
Follow Nemanja on LinkedIn and check out his courses.
Intro music courtesy of fesliyanstudios.com
Hello and welcome to the Topics Unplugged. After a few glitches here and there, thanks right. What is the Topics Unplugged? It's a casual, lighthearted, weekly short discussion, cozy corner of the web. From chips to mojo, anything goes. Today is the 9th of February of 2024. My name is Marilo. I'm your host of the day. I'm joined by always my the Robin to my Batman. I guess it depends. If you want to be Robin, bart Smith, hi, I'll let you think about that one. And today we have another guest. We have an Emania.
Speaker 2Hello, how are you doing, man?
Speaker 1Good, before we let the money introduce himself. Just letting you know that we're also on YouTube LinkedIn, twitch X, slash, twitter, so feel free to check us out on YouTube. If you have a question comment, feel free to also drop it in the. In the there We'll try to address it throughout the stream. So welcome and Emania. How are you doing, man? Good man, yeah, long time yeah long time.
Speaker 2The offices still smell the same.
Speaker 1Yes, yes, so you wanted to introduce yourself a bit for the people that don't know you yet.
Speaker 2What do you think they want to know? Everything, everything, but not too much. Should I say everything? Go out, maybe shortly a minute. My name is Nemanya. Nemanya Lekovic, born and raised in Serbia, came to Belgium around 10 years ago. I'm an electrical engineer by formal education and I switched to data science right about the time when I came to Belgium, and I'm now here in this wonderful, exciting, booming domain for the last 10 years. So first as a data scientist and lately as a MLOps engineer. Yeah, so that's, that's me.
Speaker 3Cool, and you are also an esteemed tutor at your host.
Speaker 2Host. No, not a host, your host, I guess. Lessons on data camp. Well, yeah, I had total two, in total two courses on data camp. One was archived, as they would say, it served its served its lifetime, and the other one is the MLOps deployment and life cycling which is, let's say, still live since, since last year, nice Data camp, both of them Data camp and also international conference speaker as well. Yeah, yeah, whatever, once in two, three years, let's say yeah.
Speaker 1And pen podcast famous as well.
Speaker 2No, Depends on how you define famous.
Speaker 1Among friends and family. Cool, cool, nice to have you here, and maybe you mentioned the booming field of AI in the past 10 years.
Speaker 3Yeah, I think 10 years, like Nemanja's data science OG.
Speaker 2Yeah, yeah, indeed, I think that was, it was still emerging. Then I think I remember seeing my friends finishing courses for like machine learning or like what is this machine learning thing? And then, yeah, then it became everywhere. But I actually realized that. We learned it at university. We learned all this theory but it was not called machine learning like in singles and pro single processing systems, modeling and so forth. But yeah, the data science title is still as undefined as it was 10 years ago and Python still is hard to package and ship.
Speaker 1I feel like machine learning is a sexier title. No, Machine learning engineer sounds.
Speaker 3Yeah, it's more sexy.
Speaker 1I think now we just talk about sexiness like AI engineer. Yeah, for the general public the busiest thing?
Speaker 2Yeah, indeed, but I think machine learning engineer is for me the most precise title. Then you really know what you're doing. You're doing machine learning.
Speaker 1What about a MLOps engineer? It's pretty pretty clear.
Speaker 2I would say still. Yeah, pretty clear. But it's it, that's still an emerging field. I would say I think it's still in kind of a crystallization phase. I would say still.
Speaker 1What about? It's not AI related, necessarily, but DevOps engineer. What do you think of that job description If you think MLOps engineers, pretty spot on DevOps engineer.
Speaker 2Yeah, I think so, I think so. I think it's very mature technology behind it, frameworks and everything. It's so mature that now we have difficulties changing DevOps to introduce MLOps.
Speaker 3True, but the new title. Yeah yeah, Think what you're a bit hinting towards MLO is that there is a discussion on whether or not DevOps is a specific role or that it should be like a way of working. Yeah, it's a bit what you're hinting towards, right? Yeah?
Speaker 1I think it's a bit. It's very opinionated Topic, I feel, because some people are like no DevOps engineer doesn't make any sense because DevOps there's no. Engineer is more of a mentality, philosophy. But then you can see the same thing about MLOps MLOps, but I do. If I today, if I hear someone is an MLOps engineer, it makes sense that like, okay, you're not necessarily building models, but you know what they are, but you're putting models in production, you're deploying models, right, mlops. If you say that it's the philosophy, methodology, mentality, whatever, okay, but I guess an MLOps engineer will be a MLOp employer engineer, I guess. Yeah, some of the place models right.
Speaker 2Yeah, I think it's often like that, and it's like that in our case also in my company currently about what we are working towards is to. I think that's that's the ideal way you should do. It is to be an in a neighbor, for the data scientists to have as much of self service as possible. That's that's how I see it, and I think it's the worst example is when the data scientist needs to depend on the MLOps or DevOps engineer for every, for every step on the way. That's, I think, the bad way to do it. When you were like have this complete segregation. I think that beats the whole idea of DevOps, which should be this famous continuum development to deployment.
Speaker 3I think it also is a bit of like a maturity curve linked with, like DevOps is a bit, or MLOps for that matter, and a more specific niche field is like, in my eyes, a bit of a set of best practices that you apply to continuous deployment of data products or machine learning models, and but if you, as a company or as a team, you don't yet have these best practices, like it takes a specific role to introduce this. Yeah, yeah, like it's maybe also as well.
Speaker 1Yeah, but I also think, like this MLOps, engineer kind of pattern, it is something that you kind of see and you read is like, okay, mlops, everyone should do it. The data scientist should be thinking of these things, but I feel like, in practice, the skills you need for deploying models, building systems, all these things is different from the ones like the mathematical thinking, the models, the leakage, all these things business.
Speaker 1So I think the skill sets are different. I also feel like the fields evolve somewhat fast, like the tooling and whatnot. So I think which one? The MLOps field, mlops field now more, I guess the data science one. Now it's a bit more stable, let's say Well.
Speaker 2I don't know how do you count chat, gpt and all these elements.
Speaker 3Yeah, exactly. Yeah, just a rapid API. Yeah, yeah, yeah.
Speaker 1But, then you have the LLM ops, like yeah, there's AI ops, right, there's everything else right PowerPoint ops yeah but that's the thing. Yeah, there's this. Oh yeah, yeah, indeed, right, but I do think it's like they diverge quite a bit. So I I feel like, in practice, you see, the successful teams, they kind of come up with this deployer enabler and the data scientists, and you know this and this, I think, in our team.
Speaker 2It works really well. So I think we still need a bit more. I think the ratio of data scientists to MLOps engineers is too high now. We still need to. We have like three to one now, three data scientists to one mellops.
Speaker 2Yeah yeah, yeah, but still there's a lot of data. Science works. It makes sense, but somewhere around that. For me it's also like three, three to one, two to one. I think that's a good ratio, but I think it's emerged as like a natural product of when you think about how many things you have to do to deploy, to build and deploy a machine learning system, you realize it's too much to ask from one person.
Speaker 1Yeah.
Speaker 2So either you have the same old school, I would say, way of doing it like of treating each project like a unique piece of software, and then somebody needs to build a model, somebody else needs to package it as a piece of software and deploy it, or I think what is the new common approach which you hear everywhere is like the platform development, where you should basically build something which is like a self service platform where the data scientists can make their models, push their models like click, release them as the API or something like that.
Speaker 2But I still did not see a case where you did not still need to do some programming on top of that.
Speaker 1You mean you as the MLOps engineer?
Speaker 2Yeah, yeah yeah, there's always some kind of extra logic that you really need to code, to code into there.
Speaker 1Yeah, but I think it's also an ideal right. Ideally, you want data scientists to be able to do everything and just click, and then if there's a bug, it's supported by them.
Speaker 2A lot, I don't think they necessarily have to do everything, but I think they should be left alone to think about data science. Yeah, because I remember when I was doing data science I was frustrated at one point that I had like 80 90% of work which was not the actual data science. So your start you make a random forest model. Oh, 80% accuracy. Okay, good, Good to go for. Now let's try to deploy it. And then you're working on like, oh, provisioning the infrastructure and making the pipelines and doing all of these things, and you're thinking, but should I have maybe made a model better?
Speaker 1Yeah, who's going to?
Speaker 2think about it, but I think when I guess what.
Speaker 1I'm thinking that data scientists should be able to do it all. I mean they should be able to push it to production, but not to build the components you know like if it's just like your package in a certain way, and then you press a button and it's there because I agree that data scientists should be able to make up the idea at least.
Speaker 2Yeah, that's the objective.
Speaker 1That's what we strive for. Yeah, yeah, yeah. And now there's the whole gen AI story, right Like the LM ops, like we mentioned which is a bit of a yeah, but so I coming but there's a bit of the coming yeah coming.
Speaker 1You have an alert like no, because there's the LM ops, I think it's, or the LMS, the gen AI, which I think it's an interesting one because it's definitely AI related, but it's not like a data scientist, traditional data scientist hardcore, that knows all the math behind these things he's going to have. He or she will have a hard time, probably right Like building a system, because in the end you kind of call an API and everything around that is more software engineering related. No.
Speaker 3That's the evolution that we're seeing right.
Speaker 1Yeah.
Speaker 3Like we see it very good LM's very much more large, generic models that you can even with zero shot to get a lot of performance for specific cases which are simply behind an API that you call indeed. But there is still a component to that, I think. Even if it is just an API, it's not a model you train like it's still, you need to do experiments with the prompt, you need to track a bit like what is the performance? Why do we think this is robust, like it's?
Speaker 2you need to define metrics Exactly. It's still you need to analyze where you have the error. Analysis like where is your model doing well, or the.
Speaker 1API, whatever which cases are, bad.
Speaker 3Yeah, I think you still have a lot. I think like take the black box philosophy model as a black box.
Speaker 1API is also black box, yeah, and then there's also the whole prompt engineering thing and the whole is that real?
Speaker 2But is prompt engineering real?
Speaker 3As a rule you mean prompt engineer. It's the sexiest job of 2024, based on prompt engineering says Forbes or somebody probably says something right.
Speaker 2I just said it. I just said it, I think, I think it's already a bit like this was 2023.
Prompt Engineering and ChatGPT Performance
Speaker 1Yeah, but actually I heard that it was a thing 2023, but it also died in 2023. Because I think also they started with the models that self engineer, the prompts, I mean, I think even the multimodal stuff, right, I think what we saw in the title prompt engineer pick popping up is with things like the earlier image generation models where you had to be very specific, like what are the?
Speaker 3if you add this word, if you add that style, if you add that number, then it does this to the image. Well, today we also see that as much more become natural language.
Speaker 1Yeah, and I think a big thing is like how descriptive you need to be, how descriptive can you be right, because there's all the context thing, right?
Speaker 2Yeah, I just don't understand it as a job. I don't understand that this requires a special position. You know, I see it as a skill, like like, like Googling, but you don't have a Google query engineer. You know like, oh, this guy just Googles. It's a soft engineer, bro, you understand what I mean. So, prompt engineer, if you would hire that person.
Speaker 3Yeah, it's a bit of a hyped up term. I can imagine that there are some jobs that do a lot of from them engineering for their day to day, like no.
Speaker 2But Copywriting, yeah, but even programming you can say you're always. If you're using co-pilot, you're doing prompts, but you're not. It's not your title.
Speaker 3No, no, but I think it was very hyped up. What we were saying is LLM experts, basically.
Speaker 2If well, what's that again? I mean LLM experts in using LLM, deploying LLM in fine tuning LLM. What's an LLM expert? I think what I heard is the AI engineer.
Speaker 1That's what I heard, I think from Versailles, even, or something. They were saying AI engineers are the people that build systems or applications using JANAI. So basically, just call an API to the prompt engineering. That's an AI engineer. But again, it's like it's how you call it, right.
Speaker 2Like I think, yeah, in the end you're going to do SQL.
Speaker 1In the end they're going to export it to Excel. Yeah, yeah. And about the prompts I've been trying to hint this a bit, but there was something like so. Actually last week I don't think we did it in the pod, we didn't talk about it in the pod, but we had a chat bar there was some theory, there was some conspiracy theories that they were training GPT-6, or 5. 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5., 5. 5.
Speaker 25., 5., 5., 5., 5., 5., 5., 5., 5., 5. 5.
Speaker 3So my experience and I think a lot of people experience that use chat GPT a lot is that chat GPT's performance on certain domains has degraded and like what? Especially on where you see this? A lot is on large bodies of text. So I have a large body of text as input and I ask, for example, translate this full body of text, not just give me a sum up. Translate it word for word into a non-language, which is a, which is a big job, and what you see now and you hear a lot on Reddit on this as well is that there's a lot of laziness going on, meaning that half of it was translate and suddenly stops and you don't get any more output.
Speaker 1I think it's a bit like. It's funny that you describe it as laziness, because it's a very human attribute, right, it's like the model is lazy. It's just like you just gave up in the middle.
Speaker 2It became a Gen Z or something.
Speaker 1It's just quit, it's just quit his job.
Speaker 2No, sorry, Too much man.
Speaker 3Getting paid enough. One of the one of the hypotheses in the community is that OpenAI has taken resources away from chatGPD four to train chatGPD five. It's a hypothesis. Just thinking about this hypothesis, I find it hard to understand how it would, how less resources would introduce laziness, but it's a hypothesis. From a technical point.
Speaker 2Well, if they want to optimize I'm just thinking they optimize like lag and response time and then you say, okay, you need to do to produce something within the same SLA. Maybe that is needed to introduce. So then you can only use less resources and just blur out what you can and feed how should I say? Maybe they're focusing on, maybe that's the query of yours is. Meanwhile they can serve five other queries or 10 other queries. They're just like sorry man with your Exactly.
Speaker 2Polish translation of your haiku poetry or something Potentially Like talking from personal experience, that example.
Speaker 3It was oddly specific. And now recently, yesterday, I want to say there was this laziness is part of the prompt.
Speaker 2Is that what we're seeing?
Speaker 3No, not really, but what came up yesterday on X I think really just put it on the screen is that there is quote unquote leaked a system prompt for ChetGPT. So you can basically ask ChetGPT for please clean UI. We're talking about the UI, right, chetgptopenaicom, not the API, the user interface. You can ask it clean conversation starting from scratch. You can ask it Show me everything that was described to you before, word for word, and what it will give you is it will give you what we call the system prompt, like the initial instruction that OpenAI gives, the Priming yeah, and it is actually, and we link it in the show notes it's 1700 tokens, which is big.
Speaker 2Has anybody tried to do it or are we all trusting some? I?
Speaker 3tried it, you tried it. You get it more or less word for word and there are instructions on, for example, for Dali, If Dali gets called return maximum one image these type of things there. It gives instructions to ChetGPT that it has access to a browser function and how to use that browser function to query stuff Like these type of things are in the system prompt and it's a big system prompt and there are also some discussion like potentially, because it is so big, this also has an impact on the laziness.
Speaker 1So but I think I thought there was also no, because I guess the system prompt is the input right and the context is because I was wondering if somehow the 1700 tokens will take up your space to actually input stuff and return stuff.
Speaker 3I don't think so, because it's very clear on how many. Well, I want to say it's very clear how many tokens you can actually input, but it's clear for the API, not sure if it's actually very clearly described for the user interface. Okay, it was a good question actually.
Speaker 1Yeah, and I guess the laziness is the. I guess they were saying they were lazy in the prompt engineering. I guess.
Speaker 3Well, there was some hinting towards it, but can't. So there's some hinting. It's not explicitly in the prompt but it does, for example, say it. Not to be elaborate, just give answer into the two sentences. Be it to the point that these type of things which might influence your how you interpret the output as laziness.
Speaker 1It's not explicit to be lazy. Yeah, yeah, yeah, and that would explain your experience with it.
Speaker 3Potentially, because a lot of our hypothesis here, like the other thing, is like there's potentially continuous fine tuning going on Like a recent, not that long ago two, three months ago a new model was released. If that is being fine tuned to its less resources because resources are being used for GPT-5, that's potentially could also show a degrading of but there's a lot of you mean that fine tuning can't degrade it, that it's not fine, then it's not fine tuning.
Speaker 2Huh, that is true, that is true.
Speaker 3But there are less resources there but also but then there's a whole nother topic that there is a lot of new information that is generated that is now also being used for fine tuning.
Speaker 2I think that's something when you think about, like 10 years in advance, what will be the output of these models, because you have all these garbage coming in.
Speaker 3It's cannibalizing it's like eating itself. It's a very generic language all over. Yeah, yeah, yeah.
Speaker 2Accelerating fascinating world. What do you think like when people become aware that, like, okay, if I write something original, chagipiti is gonna take it? Will then everything be paywalled? Will you say I will not give any like a token of my text without you paying for it?
Speaker 3What I think will happen is that if you're applied for a job 10 years from now and you make a typo and a motivation that people will say, yeah, this is a real person. Let's admit this one.
Speaker 1The more typos, the better, and then people are gonna be like prompting, add some typos. Yeah, yeah, yeah, yeah, yeah. Okay, that's a.
Speaker 2I am just thinking of like losing the skills as the society or whatever, as the industry.
Speaker 1You know, what is he?
Chat GPT Impact on Code Generation
Speaker 3He's programming skills in our industry, let's say, or thinking skills, if you just and people say, well, I'm not, okay, you still need to understand it, but if you look at the evolution that we've gone through it's such a short moment 10 years from now it's hard to imagine what it'll be. Yeah.
Speaker 2Yeah, my son did his. I caught him doing his homework with Chagy PT. Yes, I was like no, no, I didn't do it. I saw the history. No, I saw exhilarating there, yeah, yeah between my you know, between all my tech questions. I just saw something in Dutch and was like wait, I don't speak Dutch, yeah.
Speaker 1Oh dang.
Speaker 2Yeah, yeah, yeah, but you know he. Just he knows how to access it. He's 10 years old, he knows how to access it.
Speaker 1It was a long conversation.
Speaker 2But the funny thing was he was surprised. But dad, he does not know all these things. Like, for example, they had to do an assignment for something, for some kind of a song, and he asked what are the instruments in the song he's like? But got it all wrong. Like there is no. I heard the drums and it said there's no drums. Like why do you think it?
Speaker 3listens to it. How about it's funny? Yeah, yeah.
Speaker 2It's funny because you see the expectations.
Speaker 3Because there is natural language conversation, you assume that there is some intelligence.
Speaker 2Yeah, yeah, that it actually is, and this has this like false confidence. Yeah.
Speaker 1We're just saying yeah, yeah it has this it has that?
Speaker 2Does it ever say? Sorry, man, I have no idea, I didn't listen to this song.
Speaker 1Does it?
Speaker 2exist, because you know the examples with like bears in space. You know that. No, no, the way they asked, the guy asked it like how many bears did Russia send in space? And then he got answers like 60, you know, and there was never bears. There was like a dog, like a whatever, and they're like, okay, well, what were the names of the bears? And then there was like, yeah, ruskas, whatever, you know it made up like 60 names for the bears, which would sound like very Russian. You know, it made up the whole story. Yeah, so I think that needs to be really solved, huh.
Speaker 1Yeah, yeah, true, yeah. But you know, actually, like I had experience talking to a colleague that they were like I think I was talking about my dad how when he was writing his thesis, whatever, or he needed to know something, he literally went to the library. He asked for a book. He opened the book, go to the table of contents, go find information, read it, put it on the book. There's a reference. And I was like, yeah, man, like fuck Now, with like Google and stuff, you have like overflow information, you know Everything is there. I couldn't imagine writing my thesis without like having to go to the library for doing my research.
Speaker 2And then this kid was like Without control F, without control F, yeah, but then this kid he looked at me and was like oh yeah, I cannot even imagine writing my thesis without chat, GPT.
Speaker 1Yeah, I was like the hell.
Speaker 2Yeah, yeah, yeah, yeah yeah.
Speaker 1I was like you do-. It escalated quickly. Yeah, and it's like and actually we spoke in the past that chat GPT was released in November of 2023 as well.
Speaker 2Yeah, it's not that long yeah, but what I see it really is it really inflates the content without adding value in many cases. So people write yeah, like really long elaborate posts, and I'm like wait, I'm looking for two lines, you know, yeah, and the next person will-.
Speaker 3The next person will paste this and I'll say give me a bullet point summary.
Speaker 2Yeah, yeah, indeed, so you go from and then you need to forward it and you say and I need to be a little bit professional Like you- can paste it you can expand this to me, for me, and write a nice article about this.
Speaker 1Yeah, yeah, Maybe just a quick correction November of 2022. I think it's 2023. 2022.
Speaker 1Yeah, yeah, yeah, that's right, but interesting you mentioned, if you're losing the skills because I came across this article, the bar will put on the screen that generating code was never the hard part, so they do allude a bit to that. But basically, like TLDR, he's just saying that some of the things you also referenced, like chat GPT, will never ask a follow-up question, right, chat GPT will never say okay. For example, he has an example here like an adder function which basically adds two numbers and the prompt was write an adder function in TypeScript and you will do it. But the better answer would be how many numbers are you adding? What happens if you don't?
Speaker 2have a number. What's your budget?
Speaker 1Yeah, but like a lot of different things in terms of who pays, do you even?
Speaker 3need this function, but like, what are the specs? What is the feature that you're creating Exactly?
Speaker 1So he even mentions here that the value that he brings is not. The value that he brings is before the code is generated, right. So what are the things that are actually needed, like asking the questions why is this function even there in the right place? Do we need to put this function here? Is it part of there? What is a clean API? And he's saying that, yeah, chat GPT, like it generates code. But that was never why I was there. It was more to ask the questions what's the problem? What are the functions that need to be?
Speaker 3generated. I think this is a very timely discussion Like this is. Your response to this will be completely different today than 10 years from now.
Speaker 1Yeah, true.
Programming and AI's Impact on Future
Speaker 3And you will like, 10 years from now, like it's still valid that you need to have to be very specific on what is it that you actually want to build, but maybe you don't even need to look at the code anymore.
Speaker 2But sometimes the code is the problem really? You know what I mean. Let's be honest. Like my code no but it happens that I don't work very. I mean, I work daily. I have some bash scripts, but I'm far from a bash expert and I've always had to double check and think twice. And if their chat GPT was for me a savior.
Speaker 2Yeah, that's a really okay, just like a stupid thing how do I write an EFL statement and how do I fail it? The first failed command, which is like the biggest frustration with bash, it continues just plowing through the whole script. So I think, also learning new languages as a programmer, it will help you immensely. If you know one thing, okay, how do I do this then in Python or JavaScript or whatever? How do I make a structure for Android app and so that you get like all the boilerplate things?
Speaker 2Yeah exactly, which is probably where it should be really good, because those are the things which are like the patterns that we'll see many times. And if you say, okay, this is a good I don't know Python project structure with the PyProjectTomL configure to use rough and whatever, and then you just focus on the core logic that you want to, that you need to do, yeah, I think I also think that the systems will get better right and I think you will have a better understanding what your project is right, like get.
Speaker 1You have a lot of files and you had the files are gonna be related to each other and I think, understanding more the project and I think they are like source for it and for a hit of core pilot. This is already it's already the case.
Speaker 3It has more context than just what you're writing. It also has a lot of files and stuff that you do. Is this using GPT 3.5?
Speaker 2Isn't that the?
Speaker 3issue. Good question, I don't know to be honest.
Speaker 2Because I heard people saying that they get much better code from chat GPT 4 than from the co-pilot.
Speaker 3I think it depends really on how you're using it. So if you like boilerplate, I tend to use chat GPT 4. Say, write me some like starter for this or that project. While I'm working on the boilerplate, I tend to use co-pilot. You get a co-pilot, so really like smart auto completion. It's a bit my work flow.
Speaker 2Yeah, but that's also helpful. You know, like, when you know you had to change things five times, I remember I used the Tab 9, you know, yeah, yeah, yeah, that was also like immensely, immensely helpful. I was amazed also already then without chat, gpt, how clever it was. It's really predicting like five lines ahead what I want to do.
Speaker 1I also like to use GenAI. So there's a findcom Maybe I'll put it here on the screen as well that I quite like. So it's kind of like AI-assisted Googling, right. So how to do a I don't know Fibonacci in Rust, right? So basically what it does is that I think it does a Google search or something right, and then it fetches and it probably has, because this is for developers, right and then I think they put it in the context and they give something right. But sometimes when I describe my problem natural language, especially for things like when I'm trying to learn Rust that maybe it gives me a different direction that I didn't think of. Maybe I think this is the solution. The solution is like I don't know how to you cannot have subclasses in Rust, it's composition, but maybe it's like oh no, maybe you need traits, and I was like ah shit, true, I didn't really look into that right.
Speaker 1So I think sometimes there's these two kind of in a way it's a bit of extremes, because one is you know exactly what you want and you just need to get there. But I also think if you have a very like, you're a bit stuck, you know open playing field. I think it can give you a bit of creativity or like brainstorm, like oh no maybe you go in a different direction, right so?
Speaker 2I guess if I would have to write something with like a synchronous code in Python, I would use it a lot for starting because that's not something I do daily.
Speaker 1But the thing so and I guess the point of this article as well is how do you even know you need a synchronous code? Right, and that's what the author is arguing that that is where the software engineer comes in, Right? So he even mentioned in the bottom here of course, this is a bit unfair. Chatchapet is a generative AI model and not designed to be a software engineer. But that's the point, right, it's not going to be a software engineer.
Speaker 3But I think the question is a bit like what is the definition of a software engineer?
Speaker 2Ten, years from now.
Speaker 3right, yeah, exactly that is something that will change.
Speaker 2Yeah, yeah, it's a whole discussion now. It's like you see how many, how big a part of software engineering today is what I would call system integration. It's not actually coding, it's like linking, configuring this to talk to that, configuring this to talk to that. Like in the industry, you have like what they call OMS, like original equipment manufacturers, the hard industries, people that make like a motor or make a sensor, and then you have the system integrators.
Speaker 2They are the people that put the sensor in place, connected with wires, with a PLC, with something, and I think we also see this like separation of programmers into, I would say, system integrators and original software manufacturers. That's how to call it. Yeah Well, you need to know, we need to know both. Now I think it's not such a clear cut. Yeah.
Speaker 3But you still need a certain expertise. So that is a bit the definition of maybe change. Like today, if you make your own clothes and you want to sell them online, then you want to build an app for that. If you want to build a custom app, you're going to ask a software engineer, right, and 10 years from now you're still going to ask someone else, probably. And they're going to heavily leverage LMS, which is still need some expertise to build an online app, right.
Speaker 1Yeah, yeah, I mean, and also, how far are we going to go? Right, because I think this, like you said, it's a timely discussion.
Speaker 3But the thing is evolution is. I've never seen this acceleration in evolution.
Speaker 2That has been very like no one saw this coming yeah it's exactly like an impulse signal at what one DP backed style.
Speaker 1But I also heard, like another argument from this that the skills that we have for solving problems there's not software related necessarily. Right, like you have, I don't know. You can think even on an AI use case. Right, you're trying to predict churn and it's trying to. Okay, churn is this, but we need to frame it like this, and how are we going to frame it? Like this Is this acceptable? Is this actionable?
Speaker 1Like the skill of solving problems, it's something that if it's not going to be code, it's going to be something else. Right, Like I don't know, you can even think of programming languages. Right, how, before you had a lower level programming languages, you had to know about all these things and then all these things were abstracted and it's like okay, does that mean that all these skills are for nothing, that we're doomed Like? No, these skills are still useful, but you're going to be using a different context and even if chat GPT now can generate code and even like some sophisticated patterns and whatnot, the art of, like the skill of solving problems and knowing what you want to know and what's good and what's not good, it's also like that's not going to go anywhere. That's also what I another point of.
Speaker 2I agree. I think people will soon realize that maybe after a short honeymoon they will get into hard problems and they will see that maybe their chat GPT generated application maybe works for a week or two and then it starts crashing and then you need to debug it. Maybe chat GPT helps you once or twice but for many things can it, you know, and then try, you try refactoring. Can you refactor that? Do you have the tests? You know, and I think after a while we will settle in some kind of an optimal approach. I think it will be clear that it's not a magic wand.
Speaker 3See what the future brings. There's some discussion on this fast evolution, this tectonic shift, in the sense that there is a question whether or not the current architecture will remain scalable so far. Which are what architecture? The transform architecture.
Speaker 2Okay.
Speaker 3Where we've seen from beginning the major change with of improvement since the attention paper came out in 2008 has been throw more data and throw more resources. I think there is a. There is a. It's unknown whether or not this scaling will hold and to whether or not we have the resources for it, because yeah. Gpus, CPUs or.
Investing in NVIDIA and OpenAI
Speaker 2Yeah, then you come to the architecture of the computational architecture, which I think also something with people mentioning is how this is also legacy of previous decades, you know, and how this could also maybe be a big problem and also maybe we'll have to be reinvented in order to allow faster and more efficient computation. Indeed Meanwhile, invest in NVIDIA.
Speaker 1Or or open AI right.
Speaker 3Or open AI. We've seen a news article.
Speaker 2So if you lose your job, you get the money from the stocks and then you can leave from that yeah, you can leverage the money off of heteroscedarism.
Speaker 1It's called hedging it's hedging.
Speaker 2Yeah, so if it takes your job, yeah, but what is this?
Speaker 1You buy a beach house, huh, oh yeah, yeah, man. Just like or you just yeah, endless possibilities here, but what happened about open AI and chips and what not to part?
Speaker 3Well, we, there was an article. I'm looking at the screen now, this beautiful red screen, this beautiful. There was an article yesterday that some Altman tries to seek trillions of dollars to reshape a business, the business of chips and AI, which is, of course, big news. But there is not tons of details in the article. There is more or less. The objective is to greatly increase the global semiconductor manufacturing capacity. Where there are definitely a big, where there are a lot of shortages, they're looking for investors, the which the United Arab Emirates, and this is really to further let the landscape around AI grow, enable it to grow, to have the basic infrastructure and means in terms of hardware place.
Speaker 2Imagine the spot, imagine that anti-climax like we just run out of copper man. Yeah, yeah, after everything. Sorry, no more copper.
Speaker 3It's a scavenge, your old your old cards Neat indeed, and it's interesting to see some Alpen do this. And while I'm saying this, I'm wondering like is it actually under the umbrella of open air that they're doing the words? If it's some Alpen project, I don't know to be, honest. But it's something that we've seen from other big players as well Microsoft that launched it's, or announced it's, maya chip, nvidia competitor for for more server based for their cloud, basically services.
Speaker 2It's completely Wait. Is it competing with some Nvidia product or okay, which one? I would like to say 800. Not one of these.
Speaker 3No, the professional ones, the 800 and those ones I doubt that it's competing with the 800. So we see a lot of a lot of people doing this. I think the specificity or specificity of this industry as that there is basically one single player that does all the manufacturing. So you see the Taiwan semiconductor yeah, exactly TSMC, as like 90% of the industry yes, by TSMC Also wondering what's open, I will leverage. I think it's also very specific knowledge. I think it's hard to set up something yourself. I know in China there have been initiatives. So far they're not really competing with TSMC, but there's a lot of investment there and what you see is that a lot of the players today, like Microsoft themselves, I think as well AMD to some extent they all use TSMC. Yeah, yeah, yeah, yeah.
Speaker 2So it's also like the total meta, meta players. Yeah, yeah, you know. So, like NVIDIA, amd, they profit from AI and crypto and, like these guys, profit from all of them, you know. So, whatever happens, like ching, ching. That's crazy.
Speaker 3TSMC gets a buck. But so what they're the interest of open AI? What you're trying to do is interesting and I hope it also creates a bit more of a competitive landscape in the manufacturing domain. But let's see if nothing is an easy thing to solve.
Speaker 1Yeah, but there's actually quite a lot of. I mean we'll go ahead and see questions.
Speaker 2So 90% is. So there is 10% already. Ok, so somebody can manufacture chips already. It's not the only factor in the world. So why is it 90% there? What's the reason? Is it the price? Is it the precision? Is it like the technology. What's the reason why everybody's making it there.
Speaker 3So I'm absolutely not an expert in these things, that's fair. But how I understand that what is extremely hard to do is to go on to make the details of the chips or the very, very microscopic scale, a little logo, to put it.
Speaker 2Not a logo.
Speaker 3But to make this connection that really miniature scale, that requires a lot of R&D, a lot of IP as well, and TSMC is ahead of everybody else. And it's very hard for everybody, anyone else, to compete with it, and at the same time, like from the moment that they can go smaller and just put more on a single chip, they are again way ahead of everybody else.
Speaker 2So they remain a bit of a single player. They're just maintaining the gap and probably money is involved, because nobody's going to pay for somebody else For more money if it's the same quality or something, of course you're going to look for the cheapest one.
Speaker 3That's the thing, and I think it's also because it's extremely hard to reproduce what to do.
Speaker 2This is interesting, interesting.
Speaker 1Yeah, I see we're in the hands of Taiwan.
Speaker 2Yes, in that sense, I guess, unless the money.
Speaker 1you want to start something you seem very passionate.
Speaker 2Like how come no one is doing it? I cannot solder a resistor, I would burn myself. I have a soldering device, I can bring it. I try to solder the thing that detects studs in the wall. I have that. I bring the soldering you, but you bring the stud detector Halfway there, and then we play music or something we dance.
Speaker 1Yeah, I think a bit four-sealed segue here. This is the story about increasing speed and whatnot. It's like it's very relevant, not just in the AI space, but also Mojo. He's cheating, he's looking at my screen.
Speaker 2Yeah, so there is a. In my time we did not use chatGPT, we just looked over each other's shoulder and stole the homework.
Speaker 1Yeah, right, but yeah, in Python there was Python 3.11, which was the faster-see Python. There's the story. Then there is Mojo. But what is Mojo?
Speaker 3Maybe explain for people that.
Speaker 1Mojo is. I guess it's a programming language, a Python dialect, Python superset it's not yet a superset.
Speaker 1It aims to be a superset. Well, actually I'm saying this from months ago, right, but I doubt that they're already a superset of Python. But there are some things. For example, it's how do you say it? Like strongly typed, I guess, like you cannot, if you say x equals to 1 and x equals to 1.2, error. So, just by that very basic thing, like it's not every Python program, is going to run there Optionally strongly typed, or is it mandatory? Actually, I think what I mean is statically typed Like if x is a string x cannot be a floating point afterwards.
Speaker 1But so there are some differences, Like that's for sure. For example, we didn't even have a dictionary type. But the thing is that they were trying to overcome these challenges by making sure that you can have interoperability from module to Python so you can have, like a Python object that can be converted into a module object, which I also tried but didn't work out very well. So it's also early stages.
Speaker 3And one of their objectives is really performance right.
Speaker 1Yes, so that's basically like want to be a Python superset. That is way more performant.
Speaker 3And is it libraries?
Speaker 1Yeah, so they were working with that as well, Like how do you can import other things? The trickier part and that's not just exclusive to module is that Python. One of the reasons why it's popular is because you have stuff built in C++, C, Rust, right. So it's not pure Python code, Because if it was pure Python code, you could just kind of interpret it in module, right.
Speaker 3And is the premise then that, so for in Python, if we want to, if we need something performant, we need to build something that connects to Rust, something that connects to C, and is then the premise of module that you don't need to do this, you can build everything natively in module?
Speaker 1Yes, so maybe I'm just also disclaimer. I'm not an expert in module, right, I looked into it, I tried it out, I tried some personal examples as well. So not just going from the documentation to get a quote unquote, real experience. And yeah, that's the idea. So actually like module. So take another step back. It's a superset of Python because the Python syntax should all work and it should already be faster. But if you're really going for the super performant, you have another API. So, for example, you have def for defining functions, def and in module, if you want to have something lower level, then you can have fn to define a function, right, but then if you do fn, you are obliged quote unquote to use the quote unquote typing hints. So what would type in Python be typing hints? Module is built for actually. So module came out of the modules from modular.
Speaker 2Sorry, and you can combine it too. Can you have an untyped and a typed function in the same script, in the same program?
Speaker 1You could.
Speaker 2Ok, you could, I believe you could. But, in the end, is it compiled, it is compiled, it's compiled, it's compiled.
Speaker 1So even if you just have the Python syntax, it's compiled. So module is from a company called Modular. I have it also. Maybe you want to put on the screen.
Speaker 2Is it modular?
Speaker 1with a J.
Speaker 2Who no?
Speaker 3Oh, that would be a good one yeah, modular yeah.
Speaker 1But also the reason why it also made some noises is because the creator of Modular one of the creators is the guy that created Swift, and also Jeremy Howard. He's a big proponent of this as well. So Modular is a company that has other things as well, so let's check this out instead. Oh, max Platform, this is new stuff, but they have different products and the idea is that they had an engine that was for transformer models to run efficiently on hardware. So they had this, and to do this, they ended up creating Mojo. So Mojo wasn't the first goal, but it was something that they had to create. So even in one blog post, they said.
Speaker 2So that's the risk Whenever you do something you will eventually invent the language. But that's what they said.
Speaker 1They said Mojo, and I think it was actually smart. They said Mojo already has a success story. It's already a success because we use Mojo to create this platform. So you see some very specific AI stuff, like tensor types, built into Mojo. So, yeah, if you want the low level stuff, you can use the FN, you can use this, which actually kind of looks like rust syntax, to be honest, like you don't have. At least a while ago you didn't have the class keyword in Mojo, but you had the struct.
Speaker 2Right. So it's functional then when you go to the static mode, then it's a functional language, like a pure functional language.
Speaker 1Yeah, because you have this struct, you have the functions, but I think they wanted to add the class because they wanted to be a superset still Because class was not available. So they wanted to be true superset.
Speaker 3What I find interesting on the evolution it made a lot of noise in the community, even though I have to feel in the last months it was a bit more quiet, but it made a lot of noise in the community. But it's not open source.
Speaker 1No, so actually they announced that they open source the docs yeah.
Speaker 2So you can fix our typos. Yeah, yeah, yeah.
Speaker 1So you can ask them.
Speaker 2I'm honored. I'm honored.
Speaker 1And their reason was like we don't want to, they're shy. Yeah, I think they said they wanted to make speed. They wanted to increase the speed, crank it up, and their argument is it's easier to do it if you're in a smaller group, which, ok, I'm fine with it, but the promise is that it will be open source.
Speaker 2It will be open source soon. What do you mean? What will be open source? The whole language?
Speaker 1Yes, that's as I understand it. So why am I bringing this up? Because this, as I know, emanya, I know you also rust Inclusives.
Speaker 2Let's say you were rust curious. Are you a crustacean?
Speaker 1No, no, never. You said it so quick, yeah I was already talking about this.
Speaker 3That's what I was going to say no, no, no.
Speaker 1I know also someone that programs in Rust is a crustacean or a rustacean In Mojo. You know how the proposal was from the creator.
Speaker 2Please tell me yes, yes, I will not say.
Speaker 3I have an incorrect one, go ahead.
Speaker 2No, no, tell us.
Speaker 1Mogician. Come on, I got to give to them. I thought it was pretty clever. Also, maybe one other side note the FireMoji is actually a valid extension for the files, so if you do like mainfire, that's a valid.
Speaker 2That is too much. That's what we needed. That's what we needed. Man 2024, please.
Speaker 3FireMoji.
Speaker 2Yeah, yeah, this is too much marketing.
Speaker 3Just the tagline outperforming Rust. Yeah, yeah or if you go to the previous one where you were, this one the world's only extensible and unified AI inference platform.
Speaker 1They have such generic terms Like what does it even mean? They also had the guy from FastAI to come and do a pitch Like you know why you think this is the way to go.
Speaker 2So a lot of different ways it looks like some kind of a sound system. Yeah, it does right, it's like a 4D cinema, but you're a fan right Of Mojo yeah.
Speaker 1I thought I was a bit disappointed when I tried it out OK.
Speaker 3But just because it was early.
Speaker 1I mean, I think, a big part of it is.
Speaker 3It's a bit weird that they say that the use is internally to build this and when you try it out it still looks like someone that just started developing a language.
Speaker 1Yeah.
Speaker 2I think today everybody tries to sell it first and then build it.
Comparison of Mojo and Rust's Performance
Speaker 3That sounds like. You need to say that you outperform Rust. Basically, Tell us, Merida, what it outperforms Rust.
Speaker 1That's what it says here as well. But the thing. So maybe another thing, like because they equated a lot to Python but when I was looking at examples, when I was looking at this, like they have repos, right, but then you look at the code and it doesn't like it looks like a language that kind of looks like Python. Sometimes it looked like someone knew Python and they were inspired Poor brother.
Speaker 1Yeah, and then they're like they just came up with this language. Right, it doesn't look like Python code. But I also think it's because today the people that are interested in Mojo are the people that are more tech savvy, that they're interested in this low level stuff performance so they're not going to stay on the IT hipsters.
Speaker 1Yeah, but they're not going to stay on the Python API side, right, they're going to go in the FN, the struct, the registers, all these things, right. So then you look something's like, oh, this looks like a decorator. But then you see, register, passable, whatever. And he's like, yeah, but I have no idea, I've never seen that before.
Speaker 2Right.
Speaker 1So maybe it's also because, indeed, the people that are interested, the people that create, yeah, We'll see what you mean.
Speaker 3You use cases or Fair point.
Speaker 1Yeah, but anyways, this is something that caught my attention because Mojo is fast, that's the promise, but it's not as popular. And then someone decided to compare with Rust, which I think it's bold. I think Rust is a very big community and everyone is like, yeah, that's the selling point of Rust. Yeah. So here is like community spotlight, outperforming Rust DNA sequence, parsing benchmarks by 50% with Mojo.
Speaker 2Yeah, they didn't say who wrote it. If I wrote the Rust code, it would probably have performed by 1,000%.
Speaker 3Yeah, the thing with benchmarks.
Speaker 2Yeah, yeah, indeed.
Speaker 1Yeah, but like even here again, it's also another thing with benchmarks, right, I always have a bit of a healthy skepticism because it's always like, yeah, it's faster, but in one you're optimizing everything and the other one you just did the fastest thing you could, like just get out the door right. But here, apparently says Mohammed achieved 100 times benchmark improvements over Python and 50% improvements over the fastest implementation of Rust, which I also think is a big. How do you know it's the fastest?
Speaker 3You know, like maybe Well, the question is, of course, how much time enough for the deputant optimizing all these? Yeah?
Speaker 2But I think even Rust, as far as I could understand it from the little amount of time I played with it it actually I think it's not so popular because it's fast. I think speed is the benefit of it, but I think the type safety. I think that's what people say, Because usually you have people switching from C++ C where they have like hell of their life with all the whatever I think mainly memory safety, Memory safety leaks and whatever and so Rust sold it. So I think speed was not the reason why.
Speaker 3Well, if you compare Rust to, and how they tackle memory management with the borrowing systems, if you compare that to Python, which is garbage collection, garbage collection is not efficient. Ok, and that's why people typically say that Rust is way more performant than Python or something like Go.
Speaker 1Yeah, yes, especially with Go.
Speaker 2Yeah, but I think just the promise like this thing is faster. I don't think this is going to attract people just to you.
Speaker 3Yeah, I understand your point.
Speaker 2There needs to be a bigger problem. But also you mentioned like it's because you can paralyze for speed. You can do many things to solve. You can throw money Instead of having your. I'm always thinking about teams, not like individual people, but imagine like you do now need to introduce a new language in a big company. You know how much work is that. You know it's when you're talking about speed.
Speaker 1you're talking about computational speed as well, Right.
Speaker 2If you have.
Speaker 1IEO. If you have, like, a REST API and what takes a long time is the call, the latency, you can put Rust there, but it's not going to make a major impact. Yeah, if you have one VM that is overloaded with requests, that's like adding Rust will help. Yeah, but maybe better off, like scaling and stuff. But also like even the type safety, which I agree no-transcript, you can tell someone to use my pie in Python, right. You also had some type safety, but people don't usually like it because it's not a free lunch, right. It's not like you're just adding it and you get all the benefits and it doesn't slow you down. It will slow you down.
Speaker 2Yeah, but I see Python moving in the direction of changing many things which we thought is unchangeable. To talk about, like removing the gill, talking about also adding, like support for, like static typing, and this is all like somewhere in the discussion, in the pipeline.
Speaker 3So and the question? Is like how fast can Mojo go If tomorrow it becomes a superset of Python and it is by default X percent faster? Yeah becomes very easy to transition, like it needs to have a clear out of value.
Speaker 2Yeah, yeah. But I think in many cases, like if you have a bad programmer it doesn't matter.
Speaker 3You know, if you write crappy code, then no, no, that's true, but it's interesting to see like these new languages popping up. That is more or less a replacement of something that already exists.
Speaker 2Yeah, I think it's interesting for companies which are really big, which, like, if they save I don't know 10% of the electricity bill from their servers, like Facebook, like I think for them it sounds like also something coming from the Silicon Valley or something, but I think they need to know. Okay, does it make sense for me, for my company?
Speaker 1for my career, you know.
Speaker 2And they're gonna spend how much time learning this, are you really? I mean, if you're a data scientist, are you really gonna use?
Speaker 3maybe you will, but yeah indeed, and you see these parallels in, for example, in the more JavaScript-issue domain as well, and with Node, where you have now competitors in Buden and Dino, which is Never heard, which are more than just a language you could argue is also runtime. But it's also like something that replaces Nodejs. Yeah, and I see a bit of the same feeling here, like it's an attempt to replace something that exists and like you need to come with very good added value to your users to build a strong community around it.
Speaker 1Yeah, but I think Modra's case in particular is like I feel like they start with Python and they're like okay, but we need this. Okay, let's make that one exception, okay, but we need that. Okay, let's make this alternate syntax for this, okay, but we need this. And I feel like it's further removed. You would say it compiles, I mean yeah, like it compounds the difference, right. So I mean these are just a few examples, right? Like you have this thing that looks like a decorator, always inline, okay, and then you have FN, so you're not deaf, and then you need to specify types and type is D type, and then you have TensorFlow T and it's like, yeah, you kind of like you squint your eyes, okay, you think it's Python, but when you actually look at it, like you know like or the same thing here, like this is a SC and S-I-M-D vectorized right Single instruction, multiple or something.
Speaker 2Yeah, yeah.
Speaker 1And then you see more stuff. You know, like the TensorFlow T, and you see the alias and let and all these things and it's like the premise was you don't need to learn another language because this is a Python superset. But I feel like in practice you kind of do right, because there's no free lunch, like you cannot get to, to let is immutable, but like you need to know that, like you cannot make it just more efficient by making it more efficient, right, like it's. There is a trade off there, always right. So yeah, let's see.
Speaker 2Yeah, but we see other languages. For example, I don't know, is Julia still alive?
Speaker 1Ah, actually I saw a blog post, someone saying that Julia is faster than.
Speaker 2Yeah, but there was the promise of Julia, for example. It's faster and like convenient. Like that it should marry the best of Rust, r and Python right and I think the original.
Speaker 3That line was something the syntax of Python with the performance of C, something like that, or C++. Yeah, yeah indeed.
Speaker 2And so I never tried it. I believe that it's probably a good language.
Speaker 3I didn't expect it.
Speaker 3You see the importance of the ecosystem through it's, in the end, this inertia and the difficult thing is also like, where do you see a language really being adopted from the moment that companies start using it and as a company? Let's take Julia. You have some companies that are built on Julia, but if you need to start from scratch, you're gonna think, okay, is this language the right one for me? Does it have the right community? Can I even find people that know about this? Like it's much more than yeah, does this language really thing for you?
Speaker 2Yeah, but I think, for both small and big, I think the best, safest choice is to really select something which is yeah, I agree. Yeah, and.
Speaker 1Yeah, I agree. I think, talking about move and shift, I think we saw a couple of articles this week that we didn't cover, but like, looks like Microsoft is looking for investing in Rust for the interoperability with C++ and whatnot. Rust seems to be gaining a bit momentum as time progresses, but also the Rust is bigger. Yeah, I think.
Speaker 2Rust already has earned its place. I said I don't wanna use it.
Speaker 1I don't wanna use.
Speaker 2No, I mean I'm working machine learning and you cancel him, but you probably use it. You mean I'm using something built in.
Speaker 1Rust, that's a different thing.
Speaker 3I'm using.
Speaker 2Rust every day, but I don't want to code in Rust. I don't want to do it and I think it's a waste of data scientists' time and brain cells. But obviously it has proven its value and so many things have been rewritten in it. I obviously I would call it pure software. Engineers love it and so you cannot argue with that.
Speaker 1But the thing is a bit A new semi-something this morning. I wanna say that Rust is a bit advertised as general purpose language or something I can actually try to find that Rust is a. Everything is a general purpose language.
Speaker 2Yeah, but that's the thing, right, Bash is a I know people. Oh, you can do it in Bash. Everything can be done in assembly, assembly, is a general purpose language, then right, come on.
Speaker 1Yeah, let me see if. I can find it quickly Because indeed it was People, because, for example, you do see stuff with. The thing is.
Speaker 3What I sent you is something from a few years ago from Theo de Raat, the guy behind OpenBSD, and this question was actually on ZIG.
Speaker 2A statement on ZIG Sorry there are all the languages I hear about during the advent of code.
Speaker 3Yeah, no, it's fair.
Speaker 2ZIG, OpenCaml, OCaml, whatever.
Speaker 3All these things like who?
Speaker 2uses that Haskell.
Speaker 3Yeah, it's here on the screen. So ZIG is a general purpose programming language and a tool shape for maintaining robust, optimal and reusable software. So it's a big statement, right? And he criticizes a bit wow, that's a big statement, Like we're just proof of that. And he says one of the things that Rust has done is to teach everyone that it must lead with a lot of propaganda, which is yeah, yeah the module thing story.
Speaker 2Yeah, yeah, but I see really people say, yeah, you should use Rust for everything you should use Rust for everything, and then you feel like a bit stupid if you don't do it.
Speaker 3That's why I tried it, but for a couple of months, like basically there's a big community about it, like even like going as far as, like, if you're like being involved in open source, if you contribute to open source stuff, if you're not contributing to something Rust based, you're not cool, right, yeah, like I mean, you have these drivers as well, Like a beautiful driven driver.
Speaker 2Too old to fall for that man? No.
Speaker 1But, yeah, but there's a thing, right. Like I think Rust is kind of.
Speaker 3But you're the data science OG. You've seen everything come and go right.
Speaker 2Yeah, man indeed, it's maturity. I've been there, gandalf, yeah.
Speaker 1But indeed right, Like you kind of see Rust all over the place for like low level stuff or data frames for data lakes or whatever, and even to East.
Speaker 2Two is what to is so what is that? The travel company to?
Speaker 3This travel company is well, yeah, no, no. What does this time for Marilla to a?
Speaker 1terminal user interface. What is it so? Terminals like a, where you would put your CD, your tape, and, like you, open a terminal is an application in your Mac and then you can actually like the command line. I guess, like CLI is for the command line interface, a two is basically Interactive, I guess.
Speaker 2No, think that we're using to is huh.
Speaker 3Aren't we using to is all what kind of two is you use?
Speaker 2well, paint if you open like also. But if you open like the post-dress shell or something's not that to you also Terminal user interface, you open it and then you know when you work.
Speaker 1Ah, it's your sequel.
Speaker 3Don't think so I think it's like a ripple.
Speaker 2You think it's more when it imitates a rebel. Yeah, but it more or less imitate. Yes, repl Correct, hey something like a window or input boxes or the old like does it's a UI, but actual I was called the old, the dust file file, the commander thing about like North to command, north to command it or is that?
Speaker 3but if you know, no, no.
Speaker 2Yeah, it's a it was totally a total commander.
Speaker 1It's something like that you could copy left and right, yeah, that's a 2, yeah, no, but Vim is a 2.
Speaker 3I.
Speaker 1Yeah, there we go.
Speaker 3Yeah, that's maybe you don't know what does is anymore.
Speaker 1No, I just want to be included in the conversation.
Speaker 3He's now use the new of him. No, that's true. Yeah, definitely. You're dating new things the way I know.
Speaker 1But yeah, and we're talking about rust, how this?
Speaker 3so new of him or them is a, do we?
Speaker 1yes, and.
Speaker 3And it is a user interface in your terminal and you see a lot, a lot of hype on this. I think you have Framework libraries and all languages. I think textual for Python Ratatouille for rust actually looks visually very appealing.
Speaker 1Oh really.
Speaker 3You have a hot snake and go again. Bubble gun, bubble gun is a bubble gun.
Speaker 2Nah, uh, that's another thing.
Speaker 3Oh, it is something like that bubble tea, bubble tea, thank you. And it's really like you see a lot of the community as well. It's. But in the party community, every other day you see something in a new in textual that is being built and it looks very rate one and looks very giggy and I'm wondering, like, do we actually need all this? I think it's fun. Hmm, I'm wondering what the Usability of it before we get there.
Speaker 1Which do you? Because I looked at the. I'm looking at the Ratatouille. Actually, I have it on the screen. If you want to put it there. Yeah, which one do you think? Well, and I know there's a, as I understand, they're different, like some of them can access the graphics card, so some of them look nicer than the others, like the more pixelated, and one, not. Which one do you think is the most visually appealing? Because I think you mentioned at the two, but I thought that the, the bubble gum one is there.
Speaker 3I like bubble gum as well very much. I think the Company behind bubble tea is. I think it's charm charm bracelet Is a very creative company, and I think that's a visual, visually creative, like if you go to the website as well and then you see it as well in there too.
Speaker 1We Very colorful, very yeah very cute, very cute.
Speaker 2Yeah, are they Korean?
Speaker 1You know, bart, my money is gonna get us canceled, but look, this is really cool and there's actually like the terminal you can do quite a lot of stuff.
Speaker 3You can do lots, but like you can go as far to as to and relate more or less of who window manager, right?
Speaker 2Yeah, no, but where it makes sense for applications which you would run in the terminal, that's it. So it's like what do you need in the terminal? Why?
Speaker 1what do you need? The terminal. That cannot be in a web web app, or aside from like it's fun.
Speaker 2Yeah, yeah, aside from it's fun because, imagine, I'm not, I'm not like a systems administrator, but I imagine these guys like Doing security and whatever, sshing into different machines and go yeah they could, they could eat something like that, like. Yeah, like Network engineers and so forth, that they would need something portable, something which is everywhere and that they don't have to access From a web browser or something. So I would say, those dwellers, linux terminal dwellers, every day, all day.
Speaker 1Yeah, but I think he's like. I mean, Maybe there are use cases, but I do feel like the most of the development is because people think it's cool, because it's good, because paint paint in the term yeah.
The Challenges of Python Packaging
Speaker 3Yeah, actually, you see that part yeah and then maybe you can show it, Show what it's gonna show paint in the textual. It's textual base right.
Speaker 1Yeah, yeah. Lip gloss so someone really, really created a copy, a clone, of Microsoft paint in the terminal.
Speaker 2You might have the pointy click, adventures man.
Speaker 3Huh, so this is Microsoft paint Rebuild in the terminal. Yes, I think this is a very good example of someone that just thought can I do it? He just has a lot of time, like never thought like should we need this?
Speaker 2should I do it? That's like this is clearly like, not a hero we wanted, but a hero we need I mean.
Speaker 1I mean I also think it's fine, like it's not a problem, Like if you just want to build something because it's fun. Not everything needs to have theoretically sorry, sorry.
Speaker 2If a guy came For a job interview, yeah, I said what do you do? Show me a portfolio. It's like, yeah, I made this to be terminal for my miss paint. Hmm, I think it'd be cool.
Speaker 3I would let them do an exercise. I would say this is python base textual.
Speaker 1Can you?
Speaker 3now Wrap this in wasm and run the natively in the browser. Hmm, and then this circle is round.
Speaker 2You wrote the browser in your texture.
Speaker 1There's a textual textual web here.
Speaker 3Yeah, check it out.
Speaker 2Textual web. These brain will probably implode.
Speaker 1Yeah, look, check it out, See, you can actually have a good story bro, can you put it in a browser? But check it out, like there, people that did this, like yeah, I even seen people that took textual as a back end to build native applications like Mac.
Speaker 3I had some new one.
Speaker 1But then it's like and again it's like people Thought can we do it instead of should we? Because even Python I think even for CLI's and all these things is not of. I'm not sure if it's a very good language right In the sense that it's interpreted right. Sometimes I have issues, I don't know. I know the G cloud, tli, aws, cli, dnf all these CLI tools were actually built in Python and I even saw a lot of presentations that from Red Hat, for example, that the guy was saying that if someone went on red hat and the pip, pip upgrade, global pip or something, it will mess up everything.
Speaker 3But I would never build a CLI to them, byton. Oh, never be careful. But that is so hard to distribute. Yeah, that's, but that's the best. The point is, biden doesn't compile to a single binary.
Speaker 1Yeah, that's my point. Like I had, like I think I had like the Google Cloud CLI and then I was trying to upgrade. And every time I'll get like a weird prompt, and Then it's because, like I changed the global Python version that I was using to 312 and that wasn't supported by this version, because like there's some dependencies in the back, and then I had to upgrade the CLI to or downgrade.
Speaker 2You know we're back to Python distribute, distributing Python applications which, like the, might be always be.
Speaker 3This is something that, ten years from now, we will still discuss. Yes, yes, indeed indeed.
Speaker 1Yes, Actually that was another article like the state of packaging one year later. Right so, but we didn't cover. But indeed, indeed, indeed.
Speaker 2Yeah, and whenever I see something new and in Python, like a walrus operator or that's already old, but whatever kind of new proposal, what's called a pep for it, I think can you just solve the packaging first, man, can you just deal with that and then leave all the rest? You know, make, make a standard. You know it's like man, we cannot move virtual environments. You know it's still. I check every day. I'm like, am I, did I dream it Like can we really not move the folder of virtual environment? And that it still works? No, it cannot.
Speaker 2Yeah you had to do magic. You know I have like commands when I move the van. You know I need to do some said voodoo, yeah, to change all the paths. It's like it's, I feel, you know, dirty, not that it doesn't bring me over over that line.
Speaker 3Yeah, but I also the feeling like every other month we see a new Python package management solution we saw last week. It was a lot of.
Speaker 2I mean like like wheelbuilder or, for example, like a project. It's always like poetry, like Sometimes it's packaging and building a wheel.
Speaker 3That's related stuff. Sometimes everything included. This week we saw rye, a lot of the stuff around right, we saw the front, pixie, there's, there's so much stuff. I have the feeling that if you are, full time a Python developer like full time by developer. There comes a time in your life so more between 40, no 35, 45 that every Python develop. If you full time, you're gonna take this, tackle this project. You say gonna build a Python package manager.
Speaker 1Like that is the state of the language like the. You know like this is crazy.
Speaker 2We have so many standards, we should create one to all the standard and then like new states, like plus one.
Speaker 1I have 26 standards, yeah, yeah let's see if I can find it a little bit.
Speaker 2If Guido was the benevolent dictator, that that's. That's the one thing you should dictate, man. It's like how to get end to end because in some ways, after using rust and we can't coming back to to Python, I was actually I started appreciating more Python. I was like look how simple it is, look grids. It's really brilliant how the language was built. But, man, it misses that last mile. Yeah, come on, man, you know. But I also heard like people solve it. Huh.
Speaker 1The. I heard from Guido, I think, like you know interview that not. Not, we're not friends, but I know interview.
Speaker 2Because you don't respond to his calls.
Speaker 1He told me that the voicemail no, but so before Python they literally was zip files and send it to each other, because it was all Python, pure code. But the problem with Python now is that we have no, no, no.
Speaker 2You would put it on a desk, on a floppy disk or the big one you know wheel, because it will be in the shape of a wheel and you would push the wheel you know that's what it was, an egg before so yeah. So you came to your friend and you throw the.
Speaker 1Yeah, so it's like, but when it was just Python code, it was pure Python, it was just simple like yeah, there's even today, there's a zip import right If you have a zip a Python project.
Speaker 2You can open, as that's actually useful. That's actually useful. I mean, I think, it puts it together, you know yeah.
Speaker 1I use the one time, but I felt dirty after it.
Speaker 2Yeah.
Speaker 1But yeah, like I think the issue today is like now there's C stuff, now there's rust stuff and it's like now they're shipping pre-compiled binaries, right. So because sometimes you have, they keep you to pip install. Cryptography was one of the first ones to migrate to rust and people complaining because the binaries were shipped so they needed to now install rust to compile the thing, to install the package, right. And I think because Python tries to accommodate everyone, then it becomes hard wait.
Speaker 2The wheel should be a pre-compiled Package. I think with wheel you don't have to build anything, if you still have a setup by I think you still. It will still try to build it If you have like a tar gz. If that's what's on pipe I, if it's a tar gz or something with that pie, it's gonna, I think, locally.
Speaker 1Then you have the gcc issues and I think, I think the wheel is almost like a zip file, like if you change it, it is a zip file, but I think within it everything is already Built.
Speaker 2It could be yeah, there's no need for you to have any built tools if there is a wheel.
Speaker 1Yes, it should be like today.
Speaker 2Yes, yes but that's the whole revolution that wheel wheels brought yes to the world.
Speaker 1There's like I think I played a bit with the for trying with rust like there's the DLL or dy, live or something that the libraries and then like Python knows because uses the API, the application binary interface and whatnot, so it should be all there. But then it also means that if you're building a Package for Python that has rust in the back, you need to pre-compile to all the distributions to expect so in your side you pay the cost once before clothing. But yeah, and I put here on the screen the, the, the house standards proliferate, right, so there are 14 competing standards which we could easily replace this by Python packaging solutions, and then 14 ridiculous way to develop a universal standard that covers everyone use cases. Yeah, situation, they're 15 now Sorry and what does Mojo promise?
Speaker 2Hmm, because you see, this should. They should not say we are faster than rust. They should say Python, which you can package and ship. If they said that come on man, let's go for.
Speaker 1Mojo, I also feel like maybe there's there should be a project that just compiles Python so you can foresee allies. Will be you?
Speaker 2ease that a bit yeah, but how do they solve it? So you see, maybe the most interesting Woo-hoo bam need some help.
Speaker 3Yeah, I'm fine.
Speaker 2I'm fine, I'm together control because If they claim it's a compile language and you can type Just Python, yeah, at least some pieces of the Python syntax, then it means it can compile it.
Speaker 1So true, I agree, I mean, but I also think the other tools that are like code on or something that they may be you, but I think it's a. It's a topic that I have in my head in the back, yeah cool anything. Anything else we would like to?
Speaker 3Demania, I think else.
Speaker 1I'm done If you're a long time listener of the show. We have the code or not, but we're not gonna play this time, we're revamping. We have a 2.0 version of it.
Speaker 3It's gonna be back bigger and better bigger and better indeed.
Speaker 1So are we okay to call it, call it up on we. Are you want to press the button? I'm not gonna say that. Can I, can I, can I do? Yeah, do it with the question.
Speaker 2Is the top left yeah.
Speaker 3Thanks everybody for listening.
Speaker 1Thank you, thanks for the live stream as well. See you next time.
Speaker 3Demania, thanks for joining us.
Speaker 2Thanks guys. Yes, it was a pleasure.
