.png)
Develop Yourself
To change careers and land your first job as a Software Engineer, you need more than just great software development skills - you need to develop yourself.
Welcome to the podcast that helps you develop your skills, your habits, your network and more, all in hopes of becoming a thriving Software Engineer.
Develop Yourself
#230 - AI Will Replace Most Coding: What Chief Scientist Laly Bar-Ilan Says Developers Should Learn Next
Send a text and I may answer it on next episode (I cannot reply from this service 😢)
Laly Bar-Ilan has been working in AI and natural language processing for over two decades—long before it was trendy. She’s now the Chief Scientist at Bit, where she’s helping build composable, AI-first developer tools that might just replace most of the coding we do today.
In this episode, we talk about what staying ahead actually looks like for developers in an AI-driven world. Laly shares what skills are becoming obsolete, what new ones are emerging (RAG, agents, prompt engineering), and why “vibe coding” alone won’t cut it.
If you’re wondering how to future-proof your dev career—or whether learning to code still matters—this conversation will give you both clarity and direction.
Connect with Laly here: Laly's Linkedin
Check out Bit to build composable software in the cloud here.
Shameless Plugs
🧠 (NEW) Parsity's The Inner Circle Program - a highly customized roadmap to take you from 0 to hired. For career changers who want to pivot into software.
💼 Zubin's LinkedIn (ex-lawyer, former Google, Brian-look-a-like)
👂🏻Easier Said Than Done Podcast
Already a developer? Check out 👉 Not Another Course
Serious about joining Parsity? Schedule a call with me ☎️
Welcome to the Develop Yourself podcast, where we teach you everything you need to land your first job as a software developer by learning to develop yourself, your skills, your network and more. I'm Brian, your host Today on the Develop Yourself podcast. I'm joined by Lali Bar-Ilan, who has 20 years experience in natural language processing, machine learning, data science. She's a big old deal. She's held leadership roles at different companies been VP, I think, cto, chief Scientist and you work at Bit, a company I'm familiar with, and you're building composable software, and you've done a lot of stuff like from academic research to cognitive science and AI stuff, which I'm really interested to talk to you about today. Welcome to the show.
Speaker 2:Thanks for having me on. It's a real pleasure and I'm really excited about that.
Speaker 1:You must be very popular nowadays. You were doing natural language processing in AI before. It was cool. Yeah, it appears.
Speaker 2:Exactly For like the past 20 years almost.
Speaker 1:And what do you? What do you do at Bit, like right now? Can you tell us a little bit about what you're doing at bit and what? Yeah?
Speaker 2:of course. So we're bitcloud, um, and we're a platform for composing software, uh, using ai, basically. So, okay, um, and it's for anyone. It used to be more for developers, but now we're really like a lot like lovable and bought, but we're doing it using composable architecture, which really changes everything.
Speaker 1:What I really want to talk to you today, though, about is how developers can stay ahead in an AI driven world. This is something that I speak to a lot of younger developers about. I think there's a lot of like fear around it, and people either wanting to put their head in the sand and like pretend it doesn't exist and like go really old school which I think is kind of weird, if I'm being honest or people that think that, like, I don't need to code at all anymore. I'm just going to have AI write everything for me and I'm going to be like a vibe coder as a profession, and my opinion, honestly, is unmade. At this point, I don't know. I don't even know what I don't know at this point.
Speaker 1:But, you do. You know a lot about this, so I'm really curious to hear your thoughts on this. So if we just start at like, what do you think staying ahead even means?
Speaker 2:in an.
Speaker 1:AI-driven world for software developers.
Speaker 2:Well, I've been thinking about that a lot, especially since I'm developing a code generation tool, or part of a team that develops a code generation tool. You know, like basically cutting off the branch in which we stood, or however you say that, however you say that. So I know the capabilities and I'm noticing how different tools are just developing and progressing by the minute. Right and the base models. You know the infrastructure. I mean, this is amazing progress. This is ludicrous speed. And look, I do think that AI is going to take over most of the tasks that we do today as software developers. Eventually, look, it learns how to do anything which is repetitive. So it's not just a function that it can generate. Eventually it's going. It's not eventually. Actually, what we're doing now is even on the architectural level. We're able to build whole architectures, whole platforms so that's pretty wild.
Speaker 2:Yeah, so I'm noticing this and I'm saying, okay, so this is already taking over so many of our tasks. So that's on the doom side of things. Yeah, yeah right, there is a bright side, though, like there are a few bright sides sure first of all, I think we're going to take on more roles that have to do with integrating ai into existing systems. I think we'll see more and more companies and we're already seeing that more and more companies um that that must have ai right now.
Speaker 2:You know everyone is, everybody's been told they have to do it Exactly, and all their competitors are doing it, so they have to integrate AI into their systems. So there's a lot of work to be done there, okay, so I would say first of all to junior developers first of all, you do have to learn how to code, because you still have to evaluate the code that's being generated, okay, yes, and you still have to know architectural principles in order to oversee, you know, whatever output is generated by AI. Okay, yeah.
Speaker 2:You have to evaluate it, so you do have to have this kind of developer thinking okay, but you do what I would do if I was starting off and you know what, not just starting off. Even with 20 years of experience, I still had to learn so many new things, like in the past two years. Yeah, I had to learn RAG, which didn't exist right. I had to learn how to fine-tune models. I had to learn how to do guardrails, to create guardrail. I had to learn how to deal with output which is not deterministic and to integrate it into my very deterministic system.
Speaker 1:Yeah.
Speaker 2:You know and how to evaluate these systems. There's so much to learn with regards to integrating AI into existing systems, so these are definitely things that I would go and learn.
Speaker 1:That's really interesting. I just left a company where I was introduced to RAG, retrieval, augmented Generation for people wondering, and it was like, oh, this is. It was so cool. I'd never done anything like that. I was using like vector databases and like using OpenAI as API and building what was, you know, a fairly a moderately complex rack, you know. But to me it was like really new stuff and I'm like this is so cool. But then I got into issues like well, how do you write tests for this kind of thing? Because the output is non-deterministic, right?
Speaker 2:So yeah how do I?
Speaker 1:write tests based on this, and there's some interesting people out there. I think Martin Fowler has an article about about how you write tests for things like this now and guardrails, like you said, or even fine tuning models. There's so much cool stuff too I think that we have traditionally not been exposed to Like. I've been writing code for about 10 years now half the time you have been but in that time this feels like the most fun time and also the most different time, because I write like 80 to 90% of my code with AI. I use Cursor and I'm writing most of my code with AI, but it's rarely right on the first try. With smaller things, it's usually right on the first try. It usually gives me really good answers. Larger things, more context it tends to kind of fall apart. So if I just took what it said out of the box, I probably wouldn't have a job.
Speaker 1:I probably would probably blown some stuff up pretty poorly, so that's cool that you're dealing with the same thing. That's.
Speaker 2:That's really cool to hear actually yeah, and agents, of course, which I forgot to mention agents you know, you have to learn that that like this is.
Speaker 1:This is how you build systems right now you know, and so you're kind of saying for for people, hey, yes, ai's gonna write a ton of code, but at the end of the day, it's code and you have to have somebody that understands that code to to work within it and the systems that it lives in as well. It's not like this. You know that the whole vibe coding thing, which I think is cool, to be honest. Are you familiar with the term vibe coding?
Speaker 1:yes, of course um okay, but I'm also familiar to be with uh, with its downsides, you know. Yes.
Speaker 2:What do you think about that? Like, what do you think about vibe coding?
Speaker 1:I think the original author the way he put it out there was smart and I agreed with it. I'm like, yeah, it's great to get started or even more than started on a project, and if you're doing something fun and you want to just go off the rails and build something really quickly, it's great Now. But this was coming from like a really experienced software engineer, like a genius, damn near the guy's, like he was working with AI forever. So I think people took this and said, oh yeah, I'm just going to run with this and I'm like going to either extreme to me is a little crazy. Like saying, now I don't need to know anything and this will vibe code everything for me and I don't need to know. I'm like this is dangerous because if you don't know, just basic security or exposing things like API keys in your software.
Speaker 2:Exactly like this guy did, which is kind of funny. Yeah, yeah, yeah, yeah.
Speaker 1:And then just get wrecked. I'm like, well, that's not good either, but I'm not of the team that like I don't use AI for code. I'm like I want to use AI for code. I don't see us going backwards. I see us only using our primary language, english, for coding and then being kind of like a little manager over my little AI tool is the way I kind of see it. That's my mental model at least. I agree, is that kind of yours as well?
Speaker 2:Yeah, I totally agree. If you think about it, it's just the next level of abstraction. I mean, we went from punch cards with zeros and ones really close to the hardware, and then we gradually shifted from there with assembly languages, then with object-oriented languages, then with scripting languages and now we have natural language. It's just the next level of moving away from the hardware, of telling computers what to do you know?
Speaker 2:But the thing is that the fact that it's natural language actually democratizes software development, or will democratize because of, you know, because bytec is still not like most vibe coding tools. They don't build, they don't kind of mock, I would say, maybe something likely better than locks, but I don't think that they're building maintainable digital assets you know that's a really good point.
Speaker 1:Yeah, like the stuff you, it's great. It's like a prototype, like an mvp potentially, yeah, but it's. It's hard to take that to a place where you should feel comfortable releasing that to the public. And it scares me a little bit when people say, oh, I'm deploying this, I'm getting people to put in their credit card information and I'm paying for a service I'm thinking, oh, I guess that not. I don't see that working out too well for you.
Speaker 2:Exactly, or the people putting in their credit cards, you know.
Speaker 1:Yeah, that's kind of yeah, Democratizing software across the board to people that don't know any coding is interesting to me. We'll see how long that trend continues. So you've been working in NLP machine learning. Like we said before, it was like a trendy thing really. I actually remember I wanted to switch careers at one point about seven years ago and I was like, oh, I want to learn machine learning, but I'm like the statistics and the linear algebra and calculus. I'm like this is a lot of math I need to learn. For people that are interested in this right now, because it's a ton of people want to switch into machine learning and AI, whatever that means. It's like a big, nebulous thing. I want to be an AI engineer. I'm like what does that mean?
Speaker 2:Yeah, exactly.
Speaker 1:I don't really know what that means.
Speaker 2:You know, I have a friend who teaches machine learning in college.
Speaker 1:Okay, cool.
Speaker 2:And she teaches like the fundamental. You know gradient descent and how transformers work. You know with the actual basic and I keep telling her listen, it's a good thing to know theoretically, but they're not going to use it. What you should be teaching them is how to work with AI, how to do prompt engineering properly, because there's a lot to know there. How to do everything we talked about, how to write agents, how to evaluate AI output and everything we said before. I think this is what you should teach.
Speaker 1:Okay, this is super interesting because I've, like I picked up a book on large language models and I've been like doing a lot of experiments with just like okay, like basically what you said, like building stuff, mostly honestly in the front end, like I'm I'm working, like you know, building agents that can talk to other agents, or kind of figure out like okay, how do these all work together? I'm trying to learn things like model context protocol. I did not expect you to say that. I thought you said, yeah, you should learn linear algebra and you know, like Bayesian naive, whatever the all the terms that I've since forgot about that. Are there some like principles, though, that you think that people should be aware of, or even math that people should know, or not? Is that for a full stack software engineer right now? Is it worth like learning the math behind machine learning in natural language, processing whatever, in order to be an effective quote unquote AI engineer?
Speaker 2:I think it's kind of similar to maybe learning assembly languages to understand what Python does underneath. You know, I'm sorry to say that I mean, I feel like maybe I don't know, but I'm I'm pragmatic and yeah, I'm thinking about the job market and its demands and there are going to be, I think, fewer and fewer people actually developing base models.
Speaker 1:That's what I would kind of assume. Right, it's expensive to develop these base models. It's not like cheap to train a whole model, is it?
Speaker 2:Yeah.
Speaker 1:I assume not.
Speaker 2:It's not cheap, it's usually done by the largest companies yeah it's, it's scraping, and yeah, and then it's not just the scraping, it's actually the compute that's going into training the model itself once you have the data, and so it's not just dead. But it gets even worse because I think ai will also replace a lot of data scientists doing that, because, at the end of the day, this is also repetitive. You know, tuning hyperparameters is something that AI can do.
Speaker 1:I didn't think about that.
Speaker 2:And I'm sorry to say, but I do think that both data scientists and developers will still have a lot to offer in the job market regardless. First of all, there's probably going to be new jobs created by ai, as every new technology creates new jobs, um, and there's going to be shit, you know. So there will be prompt engineers doing it full time and people who evaluate models and AI tools, and I don't know people doing everything that has to do with ethics and security aspect.
Speaker 1:Yeah.
Speaker 2:There's so much to do there.
Speaker 1:That's an interesting take and that's honestly where I'm kind of guessing. It's so hard to tell the future now because I feel like things are moving so fast.
Speaker 2:Yeah.
Speaker 1:It is really difficult to know, like, what am I going to be doing? Because, yeah, just this year alone, I'm like okay, my job as a software developer has dramatically changed in many ways To me for the better. I've been excited to use these tools, but every once in a while, I mean, I meet more than a few developers a non-insignificant amount that are freaked out and just want to like, avoid using AI, and they're even going so far as to like I'm not using AI text editors and I'm thinking why. And also I'm thinking where are people going to hide? Exactly? Because I think that's the other part. I see a lot of younger people especially. I've gone to some colleges and I've spoke to younger people and they're saying I don't want to get into software anymore. I'm in computer science, I want to get into what. I'm like, where are you going to don't? I don't know if you have an opinion on that. That's kind of off topic a bit, but like I don't know if you have kids I do, yeah, what would you tell them?
Speaker 2:I have two kids, and what do you tell? Actually, my son is a software developer as well.
Speaker 1:Oh, yeah, okay, cool he's 19.
Speaker 2:He's a software developer. He I taught him how to code when he was like 11 or 12, and so he's been doing that ever since.
Speaker 1:My daughter not so much, he's not into that yeah, I have three and none of them want to do it so far. My daughter's still really young, so maybe I'll, maybe I'll catch her. I'm hoping she can save, save me yeah maybe.
Speaker 1:I'm just curious, like what do you tell you? What do you tell your son like, would you still recommend that he even does it? Is he scared? Or what do you tell your son, would you still recommend that he even does it? Is he scared? Or what do you tell him Like, hey son, go be a bricklayer or go learn how to weld to save yourself from the AI.
Speaker 2:Yeah, we actually joke about it a lot, but what I? Do tell him is to be an entrepreneur, to use AI to his advantage to build product, and that's the thing. I think we have this solopreneur trend right that's been going on in the past.
Speaker 2:I don't know few years, and I think we talked about the democratization of software development. I do think that we still have a huge advantage as software developers dealing with AI and building digital products, and I don't know, I think we will see a lot of people opting for freelance work and for multiple income streams and for developing their own digital asset products, services that are based on AI.
Speaker 1:Yeah.
Speaker 2:And this is what I'm encouraging him to do.
Speaker 1:That's really smart. Actually, I think that's a really good take in general, like we are in a unique position. Also, we're at the beginning of like this big wave of technology where I feel like there's going to be a lot of winners that are early adopters and people that are jumping on this sooner than later. And, yeah, like entrepreneurship and learning how to market yourself, like that's a skill I've kind of learned over the years. I didn't start off being like an outspoken software developer. I was a shy person. I wanted to be in my little code hole and I thought I want to get into software so I don't have to deal with people. That was naive and, of course, as you know, we deal with people all day long, and now you're on a podcast talking to me. I mean, a lot of our jobs are changing, I think, in many ways. When did you know the term DevRel? And for people out there that don't know what DevRel is, can you explain? Yeah, yeah, first, let's start off with that. What is dev rel?
Speaker 2:so dev rel is developer relation. Um, I wasn't really familiar with it until I got, like years ago. Um, basically, if you're, if you have a product that's dedicated to developers, then you have people like me, uh, that you know, go on podcasts and do YouTube tutorials and do presentations and live demos in conferences to. You know, introduce this product to developers.
Speaker 1:And you have to be a developer to do that job now, at least now you do.
Speaker 2:I think you do.
Speaker 1:You kind of have to be a developer to do it. You can't really speak the language or like code live if you're not a developer. If you had a bug or something, something bad happens during a demo, it'd be really hard to fake that or just kind of talk around that People would easily see like this person doesn't know what they're doing.
Speaker 2:Yeah, and you have to answer questions. It wouldn't work probably.
Speaker 1:You got to answer questions and we talked about this before we got on camera. No-transcript is easier to create, it gets democratized, as you say, and we need more people to show either people that don't know as much how to code. Maybe we're teaching the vibe coders how to use the software that we are putting out too. I don't know. I could see all sorts of interesting things happening at this point.
Speaker 2:I really don't know at all. I agree, I agree and I I think you're right. I mean people who vibe code will eventually learn, will need to learn. You know at least the basic principles of code and how to get bug, how to vibe debug and how to vibe deploy. You know how how to vibe maintain.
Speaker 1:Yeah, we could add a whole other umbrella of terms under there. Actually, I like where that's going Back to the composable software aspect of what you do. Composable software is probably a new term for a lot of people that are listening to this. We went over a bit about it. How do you think composable software is going to change the way developers build applications? Do you see what Bit is doing as potentially the forefront of what developers could do, because it's currently a developer platform? Right, I mean, it's made for developers. I see there's like an SDKs and there's all the docs in GitHub and how to use it with React or Next. But, yeah, how should developers be using something like Bit or other composable software platforms to think about how they build applications?
Speaker 2:Right. So when you take most AI tools today that developers use, like Cursor or Copilot, they're integrated into your IDE and they give you code snippets and they help you generate lines of code. We don't work like that. First of all, it's a web-first application and we think of it top-down. It's a web-first application and we think of it top-down. We think of it from the architecture first not from the code snippet Most AI tools today, code generation tools.
Speaker 2:they get as context usually your local repo right, right, what you have on your local repo, okay, but um, that's not necessarily the right context for them. For example, if you want to create I don't know a header, uh, yeah and you have a header somewhere in the code base of the organization, just not on your local repo, then the AI will gladly generate a new header and you'll get code inflation Right.
Speaker 1:Yes.
Speaker 2:Moreover, let's say that you don't have any header in your code base, but you do have some components that you can use to create a header. Okay, in your code base, but you do have some components that you can use to create a header. Okay, so you have search box and you have logo and avatar and menu. Okay, so you have all of these. What Hope AI does, which is our AI, is take these existing components and use them to build the header that you asked for, so we optimize for reuse, which is something that, as far as I know, only we do.
Speaker 2:We use your existing code base.
Speaker 1:It's awesome.
Speaker 2:You know, we see the code base as like a live organism. Okay, yeah. Basically and we want to extend the functionality of the whole organism. We don't want to just generate lines of code, we want to extend the functionality Only if necessary, okay yeah, and not repeat the same functionality.
Speaker 1:So that is really smart yeah.
Speaker 2:So basically, what we have is a graph of component. Smart, yeah. So basically what we have is a graph of component. It's like a dependency graph that shows you which component depends on which okay, and then you have like a live map of all the business and product functionality in the organization. Okay, so you can always know what functionality you have and use it.
Speaker 2:Okay, so we also have implemented retrieval augmented generation yeah so, for example, if you ask our ai to build you a header, it will first go and see if there's already a header component and if there isn't, it will go and find header. And it will go and find our menu and search box and avatar and all of that and compose a header Not from scratch, not just write it from scratch.
Speaker 1:That is really smart. Yeah, that's a really cool use case Because, you're right, like in the code bases we work in, they're massive. It's like if you're new, or even if you're not so new, to a code base, you have like 1000 files or 10,000 files or whatever. It's really hard. Sometimes you'll realize you duplicated something without even knowing it half the time.
Speaker 1:Exactly oh this, this hook already existed or, oh yeah, this component. I could have just composed this from other components. What a smart idea. When we talk about rack, is that the? Is that probably the most popular use case you see for AI? That's been kind of my bet that it's going to be likely the most popular use case, and that's the thing that full stack developers likely want to kind of dip their toe into, to become more familiar with, like LLMs and AI and things like that.
Speaker 2:Oh yeah, for sure, I think RAG and agents. Usually people ask about RAG versus fine-tuning a model.
Speaker 2:And definitely the most popular use case is RAG, because, eventually, what you want is for the AI to get to know your knowledge base, to be able to talk to your database right, see what products you have or what uh, I don't know customers you have, right, um, and and, and. So this is rank, basically, fine-tuning a model is something completely different. It's expensive and basically it means that you, um, you make the model do slightly different things than what it used to do, so you actually change the weight of the model, but this is not called for in most cases. You just want the AI to speak to your database. Basically, yeah.
Speaker 2:Right.
Speaker 1:I'm trying to train one right now. I write a lot on LinkedIn and I'm trying to train one right now to I write a lot on LinkedIn and I'm getting kind of tired of always doing posts. I'm like if I trained, you know, made a database of all my posts, vectorized them, and then I can then ask questions to the LLM and it can then give me posts that are similar and then rewrite them using my tone and voice and things like that.
Speaker 2:Because I have like hundreds of posts. I can't dump a thousand posts into chat gpt yet. So this was a cool experiment for me to run.
Speaker 1:Yeah, to do so this is where I like this idea is actually a great use case.
Speaker 2:Oh okay, yeah. So first of all, if you don't need that live data, because our databases get updated all the time- oh, they update, okay, yeah if you don't need that live data, then and and you want the model to be able to do something that is specific, then you you can try fine-tuning it. Why not? It's something that you do every once in a while. You know it's not something you do every day okay cool.
Speaker 1:yeah, no thanks for for that um, for that suggestion, because, uh, yeah, I've been like trying to just explore like ways I can begin integrating more ai tools into my work and kind of learning still just a little bit under the hood. You know, I'm not going to become a machine learning engineer or or get into like data science at this point I'm like that ship has sailed. But I really appreciate you keeping it super real. First of all, about your thoughts on like AI and how developers can actually stay ahead and the things we should be learning that you think are important, and also just learning more about bit and composable software Before we leave. Is there any place people can find you online? You seem like you have a lot of cool stuff to talk about and say Thanks.
Speaker 2:Well, they can find me on me, specifically on LinkedIn.
Speaker 1:Yes.
Speaker 2:Lali Barilan on LinkedIn.
Speaker 1:Link to that in the show notes, by the way.
Speaker 2:That's great. Thanks, and, of course, Bit. You can find it bitcloud and check out Hope AI, which is really cool.
Speaker 1:What's that called Hope AI.
Speaker 2:Hope AI. Yes.
Speaker 1:Okay, cool, very cool.
Speaker 2:Check it out.
Speaker 1:Yeah, all right, awesome. Thank you so much for being on the show. I really had an excellent time speaking with you.
Speaker 2:That's so great to hear, and thanks so much for having me on. I had a really great time as well.
Speaker 1:That'll do it for today's episode of the Develop Yourself podcast. If you're serious about switching careers and becoming a software developer and building complex software and want to work directly with me and my team, go to parsityio, and if you want more information, feel free to schedule a chat by just clicking the link in the show notes. See you next week.
