— Simon Asnake, The Data Guy

[00:00:02]So, my guest today is Simon Asnake.

[00:00:05]A very talented data scientist, but also data engineer, AI engineer, and data analyst just a lot of titles. Glad to have you.

[00:00:15]>> Glad to have me this week, man. Thank you for the time.

[00:00:19]>> Yeah, so I think to start off to most people, and by most people it includes me, the these titles blur together.

[00:00:26]Like, what is the difference between a data scientist and data engineer, for example? Is it just like a branding thing, or is there like a real distinction between between these titles?

[00:00:37]>> Okay, I think that that's actually valid question, because most of the people who are not in the world of data, that's like are they like some type of buzzwords, which do like the same job? Like, most people try to ask that question, so I think to clarify that, we would have to go to the most fundamental things. So, to make it extremely simple, data science is a field like natural science and social science.

[00:01:05]Like, you know, like that they're like different type of categories into it, but it's just a general field.

[00:01:09]>> It's just like computer science.

[00:01:12]>> Yes. So, you have natural science, social science, and as we know it, we have also data science. Then, in that field, we have certain type of roles created according to different type of tasks.

[00:01:26]So, those tasks sometimes intertwine to each other in different type of percentages.

[00:01:33]So, let me give you like a road map to them. A data analysis is a person who actually have the finalized data, the finalized product, and according to a streamlined data, they make an analytics and at the same time analysis.

[00:01:50]So, analysis is based on the previous data. They actually try to find the patterns and recognize what's going on.

[00:01:58]Based on that, they do the predictions, which is called analytics. Uh so, that's data analysis. Now, when you go to data science, uh it goes much deeper into handling the data. Uh data analysis don't do much of a data cleaning or, you know, like they already have like a silver the data in the silver platter.

[00:02:17]They just work on that one. So, uh but for data scientists, that's not the case. They actually have to go much deeper into it and handle the data. So, specifically speaking, the use cases in here are different. First one is uh not just prediction models, but categorization models, uh classification models, and these are just like text models. We have like vision, audio, image, and different type of models that work with them as well. So, data scientists work on those ones. So, when you go to the data engineers, it's a different game. Uh they're like the backbone of the data analysis and the data scientists because they're the ones who gather the data from different places. They organize it, they orchestrate it. They create different type of pipelines for the data analysis and the data science team to get the data.

[00:03:10]So, the These are like the unhidden people, which are which play the most critical roles at this point of the day specifically because data is becoming much more of a currency at this point.

[00:03:22]And right now, how you treat your data is actually how you treat your company.

[00:03:28]So, that's why we have like different type of hosts. I mean, you'll hear the those buzzwords like um forward deployed engineer or something like that.

[00:03:37]>> Yeah. Yeah, new roles are coming out every day. Yeah.

[00:03:40]>> So, these roles just add like a very small tiny thing into this fundamental foundations, and based on that, they just like try to create a fancy word out of it. Basically, a deployed uh forward engineer is what I would like to call it, just an agent orchestrator. You just babysit some agents. That's what you do, actually.

[00:04:01]So, I mean, you have also AI engineering. AI engineering is a concept born out of a vast majority of machine learning engineering. And at the same time, the we have theoretical analysis out of it and the applied. So, when you combine the machine learning, the theoretical, and the combined one, it becomes like AI engineering. So, you have a finished product of an AI product, which is like the LLM or the rack or different type of models that you're actually plugging in. And based on that, you actually provide certain type of values based on that.

[00:04:37]So, AI engineers, mostly they don't work on foundation models. They're not researchers. They're not the ones who create something from scratch. They used like semi-finished or finished products. And based on that, they actually continue it to the main ones.

[00:04:52]So, basically, that's those words.

[00:04:56]>> The Thank you for the very clinical response, but I'm just going to make the question a little bit harder on you.

[00:05:02]Do you think one can maybe be a good data scientist without without also being a good analyst or without being a good data engineer? Because of how how how self-sufficient does one person need to be? Like, can they really segregate these roles like separately? I imagine there's like so much overlap that you really just have to get good at like a lot of these roles in combination to get good at one.

[00:05:26]>> Yep. That's actually critical question because based on different type of experiences that I have seen in the field and most people's experience, they usually have the same conclusion.

[00:05:39]Not always, but usually, most of the time, they would actually have to tell you that since it's like a stack of skills that you already have.

[00:05:49]Then, based on the stacks So, one of the stack is analyst.

[00:05:54]The other one is domain knowledge. So, domain knowledge is let's say if you're working in fintech, the primary idea of how that region works on. And you like you know like you have to understand the core principles of that thing to actually work on it based on the data that you have. So, domain knowledge, then you have the foundational models.

[00:06:13]I've recent AI base type of like data cleaning, data preparation, and everything. So, as I've told you, it's more of like intertwined, but uh, what makes the role is the focus of what you're working. So, let's say according to the stack, if you're working on how to prepare the data, you become like the data engineer.

[00:06:36]And uh, if you work on the foundational models, the AI, uh, how you actually classify those data, and according to the domain knowledge of mix those AI models, it becomes a data science. Then, uh, if you have the analytical mindset of predicting something out of something like recognizing patterns, uh, finding the hidden things, or finding a different type of anomalies from a chaos and try to make sense out of it. That's basically what analysts do. So, it is much more of a combined suite of those. So, most of the people that I actually have heard will actually tell you that uh, you need a certain type of sufficiency in all of these stacks, and you just focus on one of them.

[00:07:24]But, it does not mean you don't know the other part. You actually have to know some part of it because you're still going to work be working on it. So, let's say as a data engineer, um, working on a senior data engineer at most companies, and uh, the way I prepare the data for the data analysis and the data scientist, I actually have to have a certain type of idea on what they want to actually prepare it that pipeline for them. So, which means if I'm creating some type of solution for the end product, which is the data science or data analysis system, I actually have to understand them.

[00:07:56]Uh which is why actually data engineers are um the most hardest and the most sought-after skill because data engineers are the ones that actually have to understand the skills of the data analyst and the data scientist. But that doesn't actually work for the data analysts because data analysts, they don't have to prepare the data.

[00:08:15]Or they didn't have to work on different type of foundational models, you know?

[00:08:19]So, it goes from data analyst, data scientist to data engineering. So, data engineering is like the one inclusive of all skills and the data analysis is not like a subset of it. You don't have to know a full pipeline understanding of like a small foundational thing will just be enough.

[00:08:37]But if you're working as a data engineer, that would be like a very concrete understanding of everything because since you're working for them, you actually have to understand what they actually know.

[00:08:48]So, primarily, that's why uh there's an interchangeable suite between them.

[00:08:54]>> I think that's pretty informative because I always used to think about it as um um each field has its own like um its own thing that it contains and some of the skills overlap. But what you're telling me is it's a stack and you know, data analyst is at the bottom and then a data science encompasses everything inside data analysis and encompasses more and it just keeps like climbing up from there, which is not the impression I had before that. So, thanks for that. And the second question I have for you is something I think you're very uniquely qualified to answer for two reasons.

[00:09:30]>> Okay.

[00:09:30]>> One is you've worked, like you said, across like various engineering domains.

[00:09:34]And two is you also keep up with AI quite a bit.

[00:09:38]And the question is um I imagine AI has taken a lot of the grunt work out of your day-to-day.

[00:09:44]And stuff like data cleaning, boiler plate um like even data formatting, you no longer have to do quite a quite as much as you did.

[00:09:53]And you know, what does what still eats up your time? What does the day in the life of data scientist look like in 2026?

[00:10:03]>> Um that's actually a pretty valid one.

[00:10:06]So, recently because of the recent AI changes, much of the um main data tasks that we usually do or take time on uh are partly automated if not fully or in some section of it.

[00:10:22]Mm.

[00:10:23]But the the main task that we used to do earlier and what we still actually do is finding the problem and converting that into a data-driven decision model so that we can actually understand the business problem at first. And by understanding the business problem, we actually try to find the patterns. And we actually try to understand uh how we can use the data to actually solve that business problem. So, like having the domain knowledge and correlating that to the skills that we already have and find find a business solution. So, that's what like what's the main uh problem that we're actually doing. So, uh the speed of the solution we're providing have like significantly increased. So, which means earlier before AI like let's say if we have the data before the skill and uh we have the business problem in different domain. Uh it would like take minimum of 6 months to prepare the data, analyze the data, create a warehouse, create the pipelines, and like have that data on silver platter to actually understand the business solution and provide something else.

[00:11:40]So, that have been quite drastically decreased specifically on uh like preparing the data for different type of business solutions. But, the I mean, the still the main problem that we're still facing is understanding what the client wants and converting that into something that's executable in data is uh still something else that we actually try to like find a solution for.

[00:12:09]Uh recently, I actually saw a post from the CEO of GitHub. I think they created something called GitHub Spark. We actually which actually do the whole pipeline for CICD uh starting from uh setting up the pipeline, doing the continuous integration, dealing most of the data intensive workers. Having all of the models and agents do the task and everything. And uh most of the CEOs under it, they actually commented and say, uh now since we have the foundational models working on the data, I think it's next to translate what the client wants into what actually what the data actually convert into. And that was actually a very funny tweet because we still couldn't do that. There's no AI model that can understand what the client wants.

[00:12:54]>> Yeah.

[00:12:54]>> That's like the biggest problem.

[00:12:57]>> I I I agree with you because I think most of what takes up the time is like managing the context of what the client wants, like organizing it in a way, and also you need to have like good judgment because there are certain decisions that need to be made. And one of the things I wanted to ask you and it's also one thing I'm I'm kind of embarrassed about is what is taking up a lot of my time these days is, you know, you prompt an AI and LLM, you maybe use cloud code, and then you tell it to make a change, and it takes like roughly 3 or 4 minutes to get back. Sure, you could open a new tab, and um you know, you could have a new prompt to work on a different task, but it is quite annoying how long these LLMs, well, it's not too long to finish a task in front of me, but still, like, I am taking a lot of time just staring at uh the screen waiting for an output. And if I'm I'm not sure if like progress is being made in making these LLMs faster for like uh how fast they respond. If we could cut that down to 3 seconds, I imagine I'd be like five times more productive. I it What do you think? Do you think um progress is being made here? Is this still like the same bottleneck? Because I think the human-computer interaction is like a huge bottleneck for me, and uh what what what's your take on this?

[00:14:17]>> That's a That's actually a valid question. Uh let me start by uh I think explaining a of recent research. Um so, scientists actually found a hard bottleneck on the large language models, which is the basic LLM, uh hitting the current computes power. So, if there's like Moore's law on basically the storage of what's going on, and the computes actually going hurt. So, uh the main problem that they actually facing right now is these compute power that they actually have is hitting a a certain limits on producing the large language models results, on predicting the highest accurate things, like, basically making it uh reason more, uh deep think, at the same time, uh try not to fall off in different type of system prompts, or uh fall under certain type of hallucinations. So, when they actually try to do that, uh what happens is there's actually a certain pipeline.

[00:15:21]On the top, you feed the data as much as possible in a certain distributed way, but what you actually get is becoming the same.

[00:15:29]So, the the more it reasons, the more it actually takes time. So, that's actually the current limitations that they are trying to change, which is changing it from what they call a GPU to an NPU.

[00:15:43]So, a GPU have like >> processing units?

[00:15:47]>> Exactly. So, they're changing things on the base level so that they actually hit a certain type of limit from the hardware basis.

[00:15:56]We can't do any type of software changes to make it faster.

[00:15:59]Unless you actually decrease the ability or the reasoning models of the LLMs, but recently they are doing much more research on NPU, so neural processing.

[00:16:11]Neural processing have a certain type of speed because the way it works is designed on how the human mind works.

[00:16:19]So, the human mind thinks in like seconds or microseconds. They like generating different type of things. So, it's a very complicated and at the same time time-taking process actually generate a text from a different elements into one like highly scored accurate word, which is basically what LLMs do, guess the next word. So, they're actually changing it from the neural processing units basis, but until then because there might be some time to actually find that gap or finish the gap, what's actually being happening is the way we interact with the LLMs have actually changed much a much a bit because an old one we used to interact with the LLM as if we're like prompting it very hard to get something out, but that have actually changed to give the LLM a certain type of goal, a methodology, and unexpected outcome, and let the AI prompt itself.

[00:17:22]So, different type of paradigm. Yes. So, the way we interact with one LLM is the same as doing it in parallel. So, it's the same as a parallel compute going on in different type of models. Uh so, you're not actually communicating with one model. You're actually communicating with like swarms of agents. It may be like 10, thousands of agents.

[00:17:45]So, uh the way we interact have actually changed, which actually drastically uh changes the outcome, not the time. So, like you have said earlier, uh we just collect take 3 to 4 minutes to do one task.

[00:17:58]The same 3 to 4 minutes actually stays.

[00:18:01]It doesn't decrease, but the outcome actually changes now, because since >> do more. Yeah.

[00:18:06]>> Yes. You get to do more. So, that's how uh current model has actually been running.

[00:18:12]I mean, we're seeing different things every week. I mean, I remember the time where Opus 4.5 was released. Not even under a month, then you have 4 Opus 4.8, which is like extremely crazy, right?

[00:18:25]>> New model releases are like coming up like they're fast, and I think um that's just from like one company, Anthropic. Uh imagine like like Gemini's doing a lot of great things, and um Google's doing Gemini's and DeepMind is doing a lot of great things along with um obviously OpenAI, and uh yeah, a lot of new developments, and I'm kind of satisfied knowing um at least those 4 minutes I'll get a lot more out of those 4 minutes that I'm get that I'm getting now, even if I do have to wait. That is that's not such a bad trade-off.

[00:18:57]>> Yes, exactly.

[00:18:59]>> It's not terrible.

[00:19:00]And I think moving forward from that, uh what are you doing lately? Any interesting projects you've taken up like in the background?

[00:19:09]>> Okay, so that's actually quite interesting. So, recently I have been more into uh agentic engineering. So, it's a different type of paradigm on uh having the most accurate results from agents and how we can actually accomplish most of our tasks is uh partially automate if not fully automate something on the way I could we actually wanted it. So, um the areas that I've been working on is uh internet research or deep research. So, I've been doing a certain type of uh like analysis going through different type of papers on how to actually perform or how to actually create the things that uh Anthropic and their deep research type of thing or even OpenAI uh or even Gemini. They have like their own deep research uh analysis going on. So, I've been doing a much deeper analysis on how to actually customize that into a different way that we can actually get the best results. So, one of the gaps that I've actually found is that most of the this deep research are still dependent on doing an internet SEO, which is like have a certain type of gaps because let's say if I want to have the best of something or the best of X, what they will actually give me is the most marketed one, not the most accurate one.

[00:20:40]So, that's actually creating a certain type of bias on the data analysis wallet. So, handling the markets or handling the data needs a certain type of analysis by itself. So, what I'm actually trying to do is having different type of AI agents search and analyze a source and validating it if it's market only or actually have a proven method of science like you'll actually have to do a hypothesis thesis expert that and actually fight the that thesis trying to not prove if it's wrong and actually have the right data.

[00:21:18]I've been doing a deep analysis on that one.

[00:21:21]Uh so I have like luckily I found the research paper on on topic and DeepMind have like their own deep research models on how to actually perform deep research. So all of the best things from this type of deep research method or methodology that they actually have I tried to combine it into one deep research and have like my own certain type of layers to actually validate that. So I created something called ultra deep. I know the name isn't still candid.

[00:21:53]>> Corny name but it's not too bad. It's not too bad.

[00:21:56]>> Better than the GPT series.

[00:21:59]>> Better than the namesake of it but >> Yeah. So the the reason I called it ultra deep is the layer of depths that we already have is not enough that we need ultra. So the name ultra have been used for marketing terms as I as you know it but it goes against that because the reason I call it ultra is so the models that actually create have something called a blue team and a red team.

[00:22:24]And a certain type of AI agents that actually on the go they score the resources.

[00:22:31]So what they do is you they can query not one but hundreds of website in a second. So what they do is they extract those data in a certain type of extremely simple JSON format so that they can actually validate it in the most easiest way. So let's say if they have a certain type of content in different type of paragraphs they can actually validate which paragraph is not marketed and which paragraph is actually talking about the data. So they actually can segment that and they can actually attack it in a knowledge-based type of graph. So, easily they can actually figure out and uh nitpick those data.

[00:23:08]So, let's say we have 10 websites and in those 10 websites, we actually have um let's say five of them are valid, five of them are not. So, what they do is they can actually nitpick those five ones in like a second. That's uh one of the things that I've actually achieved.

[00:23:25]I have used certain type of open-source models and fine-tuned them to the level that I want to actually perform that.

[00:23:31]The next one, we have something called a blue team. A blue team is something like the enforcer of the main goal. And we have like a red team. A red team doesn't have any context on what you do or anything about you. They actually don't care about you at all. And what they do is they will grill you as much as possible to make sure that you don't have any type of gaps on what you're actually doing. So, let's say one of the things I found on that research is that uh most of the AI models, if they fixate on one perspective, they don't actually change that perspective or see it in different type of That's why >> issue was memory features, yeah. Go on.

[00:24:09]>> True. So, that's why uh a certain type of over-engineering for a very simple because if you see something in the most different type of perspectives, uh you can actually find the most shortest path to something. Uh if you remember that we actually used to solve this in computer interactions. If you remember like there is a long way past the solution and the shortest path as well. If you remember?

[00:24:33]So, >> Yeah.

[00:24:34]>> seeing the shortest path will actually require you to see something in different perspective.

[00:24:39]So, that's what I actually try to do with the red team. They actually grill that thing from a different perspective.

[00:24:45]I mean, you would actually say like I want to learn this and they would actually say why.

[00:24:49]And you'll actually have to >> I've never had to explain my reasons to an LLM, but >> Exactly.

[00:24:57]>> Yeah.

[00:24:59]They're literally created to grill out more data from you and as much as possible drop any of those unwanted things to actually give you the best results. So, we have like a very classified red team. Uh those were created by Deep Seek, if you remember the Chinese model. They have like their own research engine. So, uh that's one of them. The other was by Anthropic as well. And there is this uh research paper that I found from Harvard. Uh they created something called a Storm AI, which is like AI-based research. They have like one of the most amazing research models in there.

[00:25:37]Uh so, I I actually combined all of that into that red team and it will actually give you a finalized outcome.

[00:25:43]So, I've tried to, you know, uh create certain type of custom solutions for my needs because I couldn't find them elsewhere. So, that's what you actually do is create them. So, that's what I've been doing.

[00:25:55]>> You're solving a very real problem because I know people who have same as there was SEO and now there's AEO. Like um there's people who explicitly optimize for like ranking for like the AI systems and they can only be exploited because uh the you know, the systems that fetch the information are not quite as good.

[00:26:16]And um if a solution exists that actually takes out the marketing fluff and just presents whatever you need exactly as you need it, I think it's a very real like it's it's a like I think it it would move the world forward in making sure like you take recommendations from the AI whenever you purchase something or you want to do something.

[00:26:43]And very, very useful.

[00:26:45]So, I've kind of yeah.

[00:26:47]>> Another I'm sorry, man. So, I've crea- I've created another one, which is called Market Scout.

[00:26:53]So, recently I've been trying to buy a 5G router and I would say like what's the best 5G router and you would find like the most promoted one, which is not like the best reviewed one. Like when you actually check the review for a market, it's not that deep. So, I created, again using the red team analysis, I just segmented that one and created it for the market. So, what I actually did is let's say if if you want to buy something, they will actually go on and the spec from different reasons and from different websites as well. So, let's say you're trying to buy a Mac or a laptop or a mobile for different sites and actually try to find you the best deals as well. So, one of the best tools that I'm actually using is archive.org. If you actually know that one, it will archive one website history throughout its entirety. So, what it does is goes into that website, check the listing of that twice and actually throughout time analyze that and give you the best deal if you're actually getting the most lowest one or the like the most expensive one. It will give you when to buy something, where to buy something. It will actually evaluate So, I I have found that it might be biased, but I have found that most of the reviews at Reddit are actually true.

[00:28:13]They're not usually targeted to certain type of user bases. So, It's actually another another aspect as well. So, that's one source. It's not the only one, but that's one source that actually do that.

[00:28:23]>> I think it's a good source at least for like the next year or so, but a lot of people are are targeting Reddit where there's like services being offered where there's like PR companies that have Reddit accounts that make posts, make comments and um Yeah, for now though it's really good and just like aggregating many sources and seeing how to process that data is probably like the best way to move forward on this. Which brings us to our next point, like looking into 2026, 2027.

[00:28:53]Uh which data AI projects do you think will pay off and which will kind of be which will kind of not? Which will be wasted effort? What's your take on this?

[00:29:04]>> Uh I think right now the markets or economy is leaning towards creating the most simplest solutions rather than creating the most complicated and the most sought-after ones. So, let's say if you have like 17 layers of a solution and a SaaS that that you're trying to provide all-in-one, that won't be marketed as much as possible because recently what we're actually seeing is you'll find one tool that it does it to the best level that they're actually selling too much in that niche. So, I'm actually seeing the most niched products are becoming more sellable at this point.

[00:29:53]And the most all-in-ones are not actually getting most of the the like the market fit for those are not getting much of a parent this way.

[00:30:03]So, it it trying to fix a solution or trying to find a very very good solution for a very very specific problem, that's actually going to sell a lot in the future.

[00:30:17]Um but trying to go all over it and providing like an all-in-one type of thing. I mean, it's still biased. If you're trying to provide for enterprise solutions, they still want that all-in-one type of solution.

[00:30:30]But, you know, uh how you actually scheme that and provide that in a different type of way actually matters.

[00:30:37]So, I believe uh what I'm actually doing right now is targeting one specific problem at a time. So, as I've told you, I've tried to attack the ultra deep one and the market scouts.

[00:30:49]I have actually created something called LLM Wiki specified for memory, agentic memory. And let's say if I I'm actually creating a certain type of marketplace plugins so that they can actually use these tools and marketplaces. They will be released in a very few days.

[00:31:07]Uh so >> Mhm. Congrats.

[00:31:09]>> Yeah, so creating a certain type of stack, so I call it the agentic stack.

[00:31:14]So, one agent can actually if let's say they want to do a deep research, they actually come to that ultra deep. But that ultra deep was not created in a way that it's an all-in-one, but we created in a way that we have actually targeted the most specific problem and actually solves that specific case. So, compiling that into all-in-one would be great. But trying to create an all-in-one from the start would be bad, so.

[00:31:40]Um this is what I'm actually seeing right now, but since this market is changing each and every day, I mean who knows what happened next.

[00:31:48]>> I I I kind of agree with you, but especially like if you're like let's say I have a small team or even if you're like a solo developer, it it makes like total sense to go after like very niche problems that like only a limited number of people are facing. And then you can charge like a very high price. Like it doesn't make sense to go for like those $20 task products anymore. However, on the other hand, like if you look at like the very big enterprise software companies like uh like Google, but like more like the CRM companies for example, like HubSpot and GoHighLevel. Like they've chosen the opposite stance where they keep expanding their product. And like you said, like enterprise clients prefer the all-in-one solution. So, I think for them maybe it makes sense. So, for different market segments, I imagine there's different ways to go go it. But like for the small guys, the small fries like us it makes more sense to kind of niche down anyway. So um it's yet to be seen and possible to predict like what's going to pop off in the future. So um I I guess we wait and see and make our bets and I I I'm I'm glad to see that you're making bets which is nice.

[00:32:56]>> That's how we do it. Yeah.

[00:32:58]>> And then I'm going to finally circle back to the first question I came up like what's the difference between like a data scientist and everything and how does one become like a good like what can one be like a good data scientist or like a good engineer without like a lot of the overlaps.

[00:33:14]Long story short, do you have any advice for anybody who wants to be useful with data? What do you recommend they focus on now?

[00:33:22]>> Um that's actually a good question. So data is becoming more and more valid recently because of the boom of AI.

[00:33:29]And that's this area have have actually become like quite highly paid and highly productive type of skills uh because of the huge demand on what they actually want right now. So what I would actually say is covering the most foundational and fundamental things is actually still valid even in the age of AI. I know that most people don't want to learn Python or SQL because the AI model I'll still say that um covering the most fundamental things is still the most valid thing right now.

[00:34:07]Because if your knowledge base is not bigger than the AI you're actually irrelevant at this point.

[00:34:17]Mhm. Do you get what I'm saying? So let's say if the AI model >> more. Yeah.

[00:34:22]>> So if the AI model knows more Python and SQL uh and let's say if it can actually do something faster The only thing that you actually know is how to prompt it, not how to actually create it. So, the main reason why we're actually saying that is, as you know, AI makes mistakes all the time. And for like far future, we're still actually expecting those AI to keep on making mistakes, even though that actually will be narrowed down into certain type of ways to capture those in the future.

[00:34:56]But, uh trying to understand that basic thing will actually give you much more leverage than the AI, so that you can actually be still valid on how to be solving something.

[00:35:10]And creating something to the core.

[00:35:12]So, having the most foundational models, and I'm sorry, foundational uh skills and everything.

[00:35:18]>> Yes. Will actually help you on growing much more faster.

[00:35:23]So, let's say if you want to solve something, since you already know the foundations, you're you can actually create something from the core.

[00:35:30]Which can actually on the most low level and fix something on the basic level, and you will have the most uh amazing thing created because you know how to do it. But, if you let Satya Nadella recently the uh CEO of Google actually recently said that don't outsource your thinking, outsource your time.

[00:35:54]So, your thinking means that what you know shouldn't be all gone or all given to the AI.

[00:36:01]Because what you know is what will keep you knowing in the future. So, let's say if you don't have the foundational uh skills, you can't actually keep growing on that.

[00:36:11]If you have already missed the foundations, how can you actually grow, you know? You cannot advance, you cannot grow or anything.

[00:36:18]Uh I mean, the AI will keep outgrowing you at some point that you become irrelevant. So, that's actually what's happening in the the uh market industry because most people don't actually know I'm sorry to say this, but they just don't know anything. They just they they just let the AI think everything, and that's it. I mean, when the AI think over you, you just become just a something that that's not >> you're Yeah, I think you're correct in so far as like if you pretty much outsource all of your thinking, all of your like skills, like it may not be necessary that you know the syntax to Python. Like >> Yeah.

[00:36:57]>> I've never seen an LLM get that wrong.

[00:36:59]But if you ever want to like move on to like higher levels of abstraction, like the the actual thinking you need to do, you first need to clear the lower levels. So like it is the pathway to like, um, move forward and learn and grow and expand. And just because like an LLM could kind of let you, um, skip some steps, you would kind of be hurt by it in the long run as you can't, um, >> Exactly.

[00:37:25]>> advance further once you reach the edge of capability the edge of capability for the LLM. And I think that's like, uh, very good advice.

[00:37:33]>> Yes, sir.

[00:37:33]>> Yeah.

[00:37:34]That's that's correct. And, um, I think that covers everything we have for today, Simon. You're You've been an amazing guest, very informative. And, uh, you've given me very You've given me stuff to think about, and I imagine a lot of people will will benefit from hearing what you have to say as well.

[00:37:53]>> For sure. Thanks for having me. Thank you for your >> And