Building the Ultimate AI-Powered Development Environment

**Farhath Razzaque:** [00:00:00] Recently we saw o1 come out in December. There was so much fanfare about it was because it was scoring really high, or at least the pro mode with extra test time compute, so it was giving more time to answer questions, it was doing really well on RKGI's benchmark.

The creators of the benchmark came out and said no this wasn't a suitable test. It never was. Now we have to create something else. Yeah. So I feel like thinking a little bit further ahead and creating tests that are going to stand the test of time. 

**Simon Maple:** You're listening to the AI Native Dev brought to you by Tesla.

On today's episode of the AI Native Dev, we're going to have a very practical session where we're going to be talking with with Farhath here, who's going to be telling us about his perfect development environment, how he uses AI today in his flows some of the most amazing tools that [00:01:00] he's used and hasn't used because he hasn't had time or maybe they don't exist.

So we're going to have a deep dive into what a good AI powered developer environment looks like. I'm hoping there'll be a ton of practical tips. Farhath, first of all, tell us a little bit about yourself. 

**Farhath Razzaque:** So I'm a freelance software AI engineer so I'm working on multiple different products and I've been using AI extensively for the last year and a half, looking into different tools, testing them out, comparing them, playing around with different models, and yeah, just trying to it use it as much in my development workflow as possible.

**Simon Maple:** And every time, every time we chat, there's always some new tool that you mentioned on. I'm like, Oh yeah, now I have to add that to the list. You're one of the people who is always like on the latest cusp of the news and trying out new tools.

So I'm really excited to have this chat with you today. We're gonna say, let's kick off with the scenario, picture the scene you're about to set at your desk you're about to develop a new application. Let's talk about what your [00:02:00] environment looks like. First of all, your physical environment lights on, off, low when you code. 

**Farhath Razzaque:** During the day I'm getting enough sunlight. I'm working during the night I just have a small lamp light up, but Mostly dark. Yeah. 

**Simon Maple:** Proper hacker mode. Okay. Hoodie as well. What do you, what's your ideal coding drink? Water. Just water? No, no caffeine. 

**Farhath Razzaque:** No, I try to avoid it. I try and I don't want to become dependent.

I want to rely on staying healthy, staying fit and healthy. And I keep a couple of liters of water. Wow. I'd probably go through them probably up four times. Jeez. Wow. Wow. 

**Simon Maple:** Snacks. Why you code? I don't. No snacks. I don't like getting my keyboard messy. Pure. Okay, I like that. Oh, you've got a Keytron keyboard as well, right?

Is it a clicky one? Yeah, it is. I once I thought now I'm working at home. This is my previous role. Now I can get a clicky keyboard and not annoy anyone in the office. And now I realized how annoying it is for everyone on zoom as I'm tapping away. I've noticed that as well. Okay. [00:03:00] When you code.

What's your favorite distraction? Do you play music? Are you complete silence? Do you have a TV going in the background? 

**Farhath Razzaque:** So typically I'm in silence. I enjoy just being able to think and write down notes and that kind of stuff. But I do also still work when I'm easy if I'm on like a lunch break or something like that I'll have a what I've started doing recently is having the TV on.

Yeah, and I used to do research like having my references on one screen and my notes on another now I'm doing is with advancements in AI tooling, I've been using. In, for example, a Cursor ID running an agentic workflow on a easy ticket. So I'm splitting my tickets into more difficult ones that I'm definitely going to have to pay a lot of attention to, but smaller things that aren't, that I know the AI is capable of, I'll just give it to it while I'm having a snack, eating and come back to in a few minutes.

If it's great, it's done perfect. If not, then just run it again. So I can even be productive when I'm. 

**Simon Maple:** Yeah. Yeah. Awesome. So let's jump into that. And what is [00:04:00] your actual development environment look like? You mentioned Cursor there. Mac or PC? PC with WSL. And on top of that, then you mentioned Cursor.

Is that your favorite IDE slash AI dev assistant? 

**Farhath Razzaque:** So typically I'm using Cursor nowadays. I see myself coming back to it again and again because I just started with it and they do release features on a regular basis, but I've got a few things a few I'm looking at right now. Winsurf recently released their Wave 2, so they released a bunch of features together and they added a web search and they've already got agentic workflows where it can do a bunch of stuff for you, but now we can do a web search where you can ask it to implement a latest API.

Yeah, you can go into a web search. Find the top links, extract the information and then implement it for you based on the docs that it found and all that kind of stuff. So that is, could be a game changer. In Cursor I have to upload docs. It does have a web search, but I don't think it's as intuitive as Windsurf just made it last in the last week and the last couple of days.

So I definitely want to [00:05:00] check that out. There's also Cline, which is an open source VS code extension. And it very much does similar things, but they recently released the feature where you can now select, so you've already been able to select your LLMs, which I love having the choice of what large language models to use.

They've added something called architect mode and edit mode, which is stolen from another open source tool called Ada, but that's a CLI tool. But architect mode allows you to, so you can select a module to be the architect for this quite good options. It seems like people are gravitating towards OpenAI's o1 and recently DeepSeek's R1 because it's good at thinking and it can really create a good plan of what's great and for the edit mode people are still defaulting to Sonnet 3. 5 called Sonnet 3. 5. It just seems to just be sticking around even though they really, it was released quite a while ago. I'm really excited to see what they're going to release next. I'm waiting on Claude 4 for sure. 

**Simon Maple:** Yeah. Interesting. And you mentioned. So Cursor, because perhaps it was one of the [00:06:00] first ones that you tried and first ones you were you a VS code user before that?

Or you were so it's very familiar in terms of the IDE, the UI as well. And then you landed on Cursor. What do you feel would be the killer feature that would take you away from Cursor right now? What is your biggest need from the IDE right now that you don't have? 

**Farhath Razzaque:** I'm seeing a lot.

So the stuff I mentioned already are very attractive features. There was a fork of Cline recently. It was called Rue Cline. They just recently renamed it Rue code. They've added prompt enhancement which is pretty nice, which I think everyone should do. So you type in a prompt. click a button and it can improve it for you.

But they also added, which is even cooler, I think is like personalities. I don't know if that's exactly the name, but you can ask it to be a QA engineer and you can create a personality for that. You can ask it to be UI/UX designer that folks that understands accessibility really well. And it can create a personality for that.

So you can you can create personalities and save them that can behave in certain ways like certain roles within software engineering itself. Yeah, that's something really cool something i'd [00:07:00] really like to see that I don't I haven't seen anyone do yet. i've thought i've been thinking about this for quite a while now is style references so Midjourney, the image generator.

I used to play around with it quite a lot. Last year, a period I was doing every weekend, just messing around and stuff. They had this cool thing called style reference. So you can generate an image and then provide you another image as a style reference. So it's like format painting in your word processor.

You select some text with certain formatting, select another piece of text, apply that. I think it'd be really cool to have that for code. Yeah. So I could copy that. So you've got in Midjourney, for example, for the images that copied. This artistic style, the colors that kind of stuff for code. It could be the language how you structure your code, the level of commenting you do or some other kind of stuff.

**Simon Maple:** And even the architecture or the design, how you would go about designing it. So that components that you build as one and have a similar architecture, whether that's good or bad. 

**Farhath Razzaque:** Cause every artist has their own unique style. It's the same for. developers, right? We are [00:08:00] based on creativity to it.

There's science to it. There's some creativity to it. And I'd really like to be able to copy that. Also when I'm looking for tools and libraries, I always find there's so many options. And this is something that's going to help it. But what would be cool a tool that can analyse different libraries that do the same thing and take the best features from each and like that style reference I've said before.

**Simon Maple:** Yeah, 

**Farhath Razzaque:** do it in the certain language that you need it. 

**Simon Maple:** Yeah, or almost merge those to create something that you is ideal for what you want your needs. 

**Farhath Razzaque:** Yeah, so this is using the fastest language for it. This one's got this incredible feature. This one's integrations, all these other things.

All this stuff is out. There's so many great devs creating even I'm seeing repos that are huge, but not the best tools and smaller repos and even really niche things where a lone developer has created a really cool feature, but it's just not gaining traction. Yeah. Or if you could find all of this and then just have a tool that merges them all together to be the perfect 

**Simon Maple:** [00:09:00] Yeah tool, nice. Let's talk a bit deeper into your usage of Cursor. I guess do you use it for project creation from scratch? Do you use it more for just like slight changes to existing projects? What's your typical, or do you do it irrespective of what your goal is?

You just go ahead and use Cusor for everything.

**Farhath Razzaque:** So I don't actually start with Cursor. Usually I go to Perplexity and do some research on what stack libraries and tools would be useful for my project. I so this is why the tool that I mentioned earlier that I like to create would I would love to exist because I manually go through the different tools and libraries and look at them, analyze them, see if they're going to be good for what I'm trying to build.

After the stack, I'll probably then go into Cursor and build a Cursor rules file. So give it some context of my project and the rules I want it to follow. So how it should go about writing comments, how it should make edits. The style like comments that you mentioned. Exactly. Yeah. All of that and [00:10:00] then I'll probably start coding and then as I'm, I usually tend to start with the back end. So some people like to see what they're building first. I think that is nice. Yeah. But, I feel like the backend and the business logic is the meat of the program. And actually for the front end, I actually do that separately.

So what I typically do is go into Claude and start experimenting, generating UIs. And it has a nice preview in the artifacts screen. And then what I do is after I've got. a sitemap with a bunch of different screens generated. I asked it to give me this in text form. I go to0ro and ask it to generate me these pages with these elements using a certain UI library like shadcn.

And for example, is one that typically go back to it again. That way I get really nice, consistent pages that just have, because what AI tools don't have is taste. Yeah. Provide, providing them a tool. Providing my UI library like that. . I can ensure a nice quality UI. . I usually take the UI code of the CS generated the backend stuff and at some point I'll merge them [00:11:00] together, ask it to boot together.

It's quite a bit at that. 

**Simon Maple:** Yeah. Yeah. When do you think about the data model? Do you do that before the backend or you build the backend and the data model together at the same time? Data model first. Data model first. So you do data model, then start building the back end to address the data model and to connect with data model and the UI on top of that.

Exactly. Yeah. Amazing. And so everything you're doing in Claude is throwaway, it's prototype only and just trying to like, almost like pulling out a doc that, that kind of describes what you're trying to do. Is it a learning thing? 

**Farhath Razzaque:** Yeah, I would say it's pretty throwaway. I, it's more for experimentation.

Yeah. I think AI is really good for helping you ideate. Even if you have it all in a forward projects or I know ChatGPT has the same thing now as well. Just opening a new chat and asking it for, So it knows it based on the documents you provided about your project, it knows what you're trying to create.

It knows, maybe if you've added features that you want, it knows about those, but it can help you expand on those. What features do you think would be good for this these users, depending if it's a consumer product, business to business, Yeah. Or just a utility used by a [00:12:00] company internally.

For experimenting with ideas, it's really good. Yeah. And even so something that I've created is a sort of rubber ducky tool. So sometimes obviously everyone, as most of us knows, it's quite nice to be able to talk to someone or what developers do is talk to a little rubber duck to help them work through problems.

What I've done is use OpenAI's real time API to create a senior engineer rubber duck for me. So when I think something might be the right way to do it, but I know there's other options, I ask it. Okay. So what are the pros and cons of this this way of doing it? Do you have any other suggestions of how I could improve this?

Are telling me to act as a very high level engineer. Yeah. It gives me very good suggestions and I've found that I'm writing better code because of it. Yeah. And I think engineers of every level, even someone at a very high level, having the extra advice or that extra perspective is really useful, but especially for junior developers for learning when you don't know about standards and how you should be going about things or what technology you should be using.

Like [00:13:00] when you're that early on, I remember being there, you don't know about anything that exists. You're trying to write everything from scratch. Yeah. Yeah. And there are incredible libraries and industry standards that you should be using. Yeah. Convention. So asking it about best convention how to keep your code dry is really good.

**Simon Maple:** So let's summarize where we are right now, then you got you're about to start a new product or a new project. Sorry. So you dim the light slightly. Great. You got yourself the six gallons of water behind you. So you're drinking some water, no snacks. You're hungry at this point. You start off with Cursor, not cursor, sorry, Claude, a little bit of Perplexity, do a bit of prototyping.

You build that in, get a little rule set in Cursor, then you start building in Cursor. Data model first, which you should already know a little bit about from the, from your prototype. Then you start actually building code from there. In terms of then how you then build tests, how you build documentation, all those other kind of like parts around that, are you using Cursor for that?

Is there a second tool, third tool that you use for the other kind of like pieces there? 

**Farhath Razzaque:** So in [00:14:00] terms of documentation, currently I quite like, so Cursor is not just good at code, it's really good at text, right? So I'd go and start writing something in, maybe make, yeah, make a docs file, start writing whatever that specific file is about.

**Simon Maple:** What model are you using under the covers? Is it Claude again, or is it? Yeah, it is Claude. Okay, which tends, a lot of people tends to say that's that little bit better at things like documentation and the kind of like the written word as well. 

**Farhath Razzaque:** So sounding like a human. Yeah OpenAI's models, maybe not sometimes. Yeah. I maybe start writing the docs and then I can ask it. to write the rest based on my project, I can pass the context on my project and ask it to write the rest or if it's writing the, if I'm actually for the data models, actually for my databases, I'm typically starting off with a general idea of what I want, asking it to expand on these and give me suggestions and then I accept some of the suggestions, change some of the suggestions and yeah that's how I go about it with documentation.

For now, I know there are better tools for the job. . I haven't [00:15:00] started using them yet. But it's definitely on the list of something. Yeah. I want to start incorporating. 

**Simon Maple:** Yeah. We had the Swimm guys on the podcast earlier actually, which is one of which is a tool that I see growing popularity.

That could be one to an interesting one to look at. Then when you're thinking about other things so test, for example, do you use Claude? Do you use a cursor to also write tests and things like that? Or are you using any of the test AI test frameworks and tools. 

**Farhath Razzaque:** Typically right now I, 

**Simon Maple:** or do you 

**Farhath Razzaque:** test in production?

That's cool. That's a cool answer. I start off not writing tests. I find it easier to change the code a lot early on, like early on when I'm signing off for a project, I'm changing the code a lot. So I don't want to write tests that are breaking constantly. But afterwards I'm typically using.

Cursor right now, but again, I want to experiment with a lot of other tools. 

**Simon Maple:** Yeah, 

**Farhath Razzaque:** I know there are a lot of really good options for that. 

**Simon Maple:** Yeah. 

**Farhath Razzaque:** Yeah. 

**Simon Maple:** Okay, cool. Then you go to deploy any devop style deployment tools, AI tools you use. What's your typical path to production? 

**Farhath Razzaque:** In my deployment at the moment, I'm not using many [00:16:00] AI tools but I have seen some good tools, for example, eraser. io, which is quite good for diagramming. So if you are trying to build a whole cloud architecture, it can be quite good for visualizing what you're trying to build. 

**Simon Maple:** Yeah. 

Yeah. Interesting. Okay, cool. When we then think about I guess an interesting one, which is maybe a slight choice or maybe not a choice these days in terms of the language that you use to, for your application.

Do you care about the language as much these days? Would you dip into languages that you are not as familiar with, more likely to because of AI tools? 

**Farhath Razzaque:** Yeah, so I've always felt like I'm pretty language agnostic. I started mostly with Python. I did Java most at university and I do mostly JavaScript now, but along the years I've done probably another half dozen played around with a half dozen other languages as well.

So I've already been quite language agnostic, but I'm far more likely to dip my toes into a new language now because I know [00:17:00] there's a lot less friction in being able to learn it. Back in the day for example when I built a Swift app many years ago, this is when Swift just came out. There were no good tutorials on how to do anything.

And they were changing the syntax every week. Yeah. Code that I wrote one, one day would be broken the next. Yeah. And this is happening constantly. Yeah. Now with AI tools, they can analyze the new docs instantly, as soon as it's released, they can update the docs that they've got stored, that they've indexed, and they can help you, so you don't need to wait on other people, other developers to create this content for you, you can go straight to the docs, the source of truth, and use that to start climbing, so it's 

**Simon Maple:** just all around a lot easier.

Yeah, interesting. One last question before we jump into kind of more advice based in terms of the, in terms of the flow that you've got when you check code in PRs, those types of things, any advice you'd give in and around various PR AR tools? That, that kind of like maybe do code reviews or those types of thing or do you heavily rely on that senior developer duck as your code review?

**Farhath Razzaque:** I [00:18:00] think. Right now I rely on the advice I get from talking to AI different chatbots and stuff like that. Yeah. It's something that I, again, want to start implementing a code review in my CI CD. So it automatically views it, gives advice, and maybe even creates a. a commit with changes on a fork.

So yeah I've seen a bunch of these tools. Recently I heard about Sagittal AI's Neo at a event. So these are things I'm definitely looking into starting to implement now. Yeah. 

**Simon Maple:** Awesome. In terms of advice, like practical advice that you'd give for developers who are trying to build AI into their development process and workflow what's like the top tips you would give a developer who is trying to be more proficient in using AI today in their workflows.

**Farhath Razzaque:** So when it comes to building with AI, I think it's really easy to prototype. But quite hard to get into production. So for those just starting out, I would start playing around with [00:19:00] different models. Yeah. And cause each model feels like its own personality. Imagine them like people, they respond in different ways, how they talk the limits of what they're allowed to speak about like the guardrails are in place.

And compare the speed, the quality of the responses and accuracy cost as well is a big one. I'd say OpenRouter is quite a good thing to use. So it basically escalated APIs for all these different models. And it means that you can just use put some credit into your OpenRouter account and start using all these different models in your applications.

Yeah. So that's definitely something I'd recommend. Play around with both. Open source and proprietary models. So most people start with OpenAI. It's definitely, OpenAI's models, definitely a good starter. But there are, there's a whole world of open source models out there that are incredible. Yeah, DeepSeek for example, that was just released matched OpenAI's latest model o1 and performance, it seems from the benchmarks and a lot of testing that people have done. Yeah. So play around with open source [00:20:00] models. I feel like they're going to be very big this year. You can use, if you want to play around with them locally on your computer, you can use tools like old armor or LM studio.

So there are a bunch of libraries that help you build with AI. For example, Langchain is probably one that everyone's heard of. Yeah. Some people say it's a little bit bloated, so you can absolutely write these things from scratch as well. And because we're at quite an early stage, it's quite, it's a really useful way to learn about how all this stuff works with AI engineering.

Yeah. Learn how to prompt engineer. This is a really big one. You can get, you can massively increase the quality of what your responses you get and the products you're building by learning how to prompt engineer properly. Yeah. And then play around with, I don't know. It's like RAG, which helps you stop hallucinations.

Use your own data tools great tools and function calling, like giving access to the web search and other sorts of things like this. And then there are agent frameworks as well, where you can start playing around with to do more advance things. They're really good to start dipping your toes into agents and this opens up a whole world of possibilities. Yeah. And then also for [00:21:00] production, you really want to start monitoring you want monitoring observability into the inputs and outputs of your LLM LLM requests. You want to monitor what's working with your prompts and iterate on them. So improve them over time to get better responses. And also you want to run these like, thousands of times. And also you want to try and be, try playing around with adversarial prompts. Is your application breaking when someone's trying to break it?

Like you cannot rely on your users to always have the best of intentions with your application. Yeah. production. There's so much more you need to think about. 

**Simon Maple:** Yeah. 

**Farhath Razzaque:** Yeah, that's my advice. 

**Simon Maple:** Wow. So there's a whole bunch of like things that interesting. You start with models first, do you think individuals will have two different developers having a preferred different model?

Is that based on style? Is that based on accuracy? And people have different different expectations. Why do you think different developers will choose different models? 

**Farhath Razzaque:** So developers should try different models, and they'll have different preferences based on their use case. So you have smaller models, which are faster, [00:22:00] cheaper and maybe don't have, for example, Microsoft has their five model.

Yeah. Five, four was recently released. It's very good. In my use, even it's hallucinating. It told you that he doesn't know the answer. So these small models that typically don't have the world knowledge for a lot of applications, you don't need those. So it's better to save cost and and to get faster responses to go for smaller models.

Larger models, maybe you do need more creativity. And there are different different models tell stories in different ways, something more creative. 

**Simon Maple:** It's almost depending on, it's almost like the usage patterns of how you would use that model, what is most important in the UX of a developer using that model, which is one of the nice reasons we're seeing more and more tools almost offering which model do you want to use for this tool, which kind of makes it that much nicer 

**Farhath Razzaque:** on that point.

Yeah. That's another thing. I don't think we're going to have one model that's going to rule them all. Yeah, I think we're going to have different models for different use cases that specialize on different data sets. And that's really why you want to be experimenting with different models. Depending on your use case, you're going to be using wildly different models.[00:23:00] 

There's so many coming out now, play around and experiment. 

**Simon Maple:** Yeah amazing. So we talked about a ton of different tools. Let's talk about the tools that don't exist. This is the harder bit, right? If there is a tool or if there is a problem that you have today, where you think, you know what?

I'd really love for an AI tool to exist that will take this pain away. What would that be? What's the pain? What's the tool? 

**Farhath Razzaque:** I think I like to see better benchmarks. So oftentimes we're seeing models perform better than another model on a benchmark, and then that's not translating really well to the real world use cases.

Once developers get their hands on it and people start experimenting with the outputs and for their use cases. So recently we saw o1 come out in December in the 12 days of shitness. And part, part of the reason why there was so much fanfare about it was because it was scoring really high.

Yeah. Or at least the pro mode with extra test time compute. So it was giving more time to answer questions. It was doing really well on RKGI's [00:24:00] benchmark. And RKGI came and said So for the longest time, this has been seen as if once you start scoring really well on this, it really goes to AGI.

And The creator of the benchmark came out and said no this wasn't a suitable test. It never was. Now we have to create something else. Yeah. So I feel like thinking a little bit further ahead and creating tests that are going to stand the test of time. 

**Simon Maple:** And benchmarks are always hard, right?

Because every time someone puts a benchmark out, that benchmark's there to be played. And I think models are going to be no different in that respect. And it leads on from what you just said as well about. Developers choosing the best model for them if they have some level of realistic expectations to what they can expect based on their needs, that will give them a short list, at least of models that they can try.

Yeah, no, I think that's a nice a nice request, a request rather that that would benefit us as an industry for sure. What about tools that you. Existing tools then today that you really want to play with, but just haven't had time, what would be on your list of, if you had a week to play with the various tools, what would you pick?

**Farhath Razzaque:** [00:25:00] Something that I'd really like to play around with is Anthropic's MCP model code context protocol that they released end of last year. So it allows different data sources to be connected and it allows you to use your AI tools with them. So for example, you create these things called servers. It comes with a few, so you have clients Claude, the Claude desktop app is an example of a client.

It's not the only one you can actually create your own, but he has access to these tools. For example, a GitHub server, you can essentially chat with Claude in the Claude desktop app. I can, you can do stuff in your GitHub repositories and create them for you and modify them. But there are lots of examples of tools you can add to it.

So for my development workflow, it'd be quite useful to, for example, to connect it to Jira or my, my ticket monitoring platform. And I typically write what I need to do in lists. I could get it in one go forward to. [00:26:00] Make my list suitable for a whole, for my team to look at clean it up and then send it to Jira and it could have another tool that allows it to act on those tickets.

So really agentic behavior is something that you could enable. That's definitely something I want to play around with. Pair AI, it's another VS Code fork IDE. I think what they're doing a little bit different is they're not building the piece of themselves. I think they're trying to be the modular alternative to some of the other ones like Cursor.

They allow you to bring other tools. So for example, your code completion the tab thing everyone's been doing about your code completion, they allow you to bring that in something else for your chat. Something else for essentially every piece that where AI is included in the IDE, they allow you to use different services.

So you can use the best. Of each best in class software for each and maybe, build your own solution. So that's something that I've heard about, read about a bit, but definitely want to play around a bit more. You

**Simon Maple:** mentioned DeepSeek as well. Is that something you've had [00:27:00] much chance to play with just yet?

Or is that a a future as 

**Farhath Razzaque:** well? So 

I've played around with it in the chat interface on the website, not the API so much. Yeah. Deep Seek came out about a week ago at the time of recording. So I think the reason Deep Seek has gone so viral partly is because they're offering the API for free. So DeepSeek on the app store has hit number one on both the US and European app stores.

But it's a model trained for under 6 million. The DeepSeek's R1 model was trained for under 6 million. And it seems like from the benchmark tests and from some user testing that it is on par with OpenAI's o1 reasoning model, which is trained for people suggest people are thinking in the hundreds of millions.

So much cheaper it came out of China and I think it's I think everyone's really heard about this. I think part of it was because it caused quite a ruckus in the stock market. So NVIDIA went down 16%. Yeah. I actually saw Meta rise. I think partly because they are [00:28:00] also doing open source models with the Llama series.

And to be honest, I wasn't too surprised by this. So what we've been seeing over the last last few years is that open source has not been too far behind proprietary. So there was like that famous quote, we have no moat from a leaked Google memo about how the models themselves are not not enough to give you a competitive advantage, especially when big tech has been spent chucking so much money into this.

Yeah. Yeah, it's it's quite interesting. Yeah, but open source has been around. And we started off six months behind, so it goes to three months and then Open AI's o1 was released last month, R1 released this month, a month behind. We've been seeing the gap close over time. So I'm not too surprised.

Yeah. And again, with the cost we've seen models train, the first model to train at a certain level state of the art, costing tons of money. And then pretty soon after we're seeing lots of smaller AI research labs come out with models that are equivalent. [00:29:00] But trained at much smaller costs. This has been happening for quite a while already, and I think people are still drawing the wrong conclusions.

People are saying that, oh, look the American research labs have been spending so much money and they are completely screwed. Not obviously, maybe, but I think that's the wrong conclusion to take from this. So DeepSeek's research was open sourced. So they've written papers about it, how they went about creating it.

This knowledge is going to be taken by all the American. Research labs big tech, and they're going to recreate it improve their own models, make them a lot cheaper. And they also have a hardware advantage because of the embargoes of the latest chips not being allowed to be sold to China.

So there's still a massive hardware advantage. The US has and the big tech companies have, and what we're going to do is take this research, make their models better. And there's something called Jevons paradox. Have you heard of it? So it's when something that's more efficient, you'd expect us to [00:30:00] be using less of it.

Yeah. So in terms of compute, when it's getting cheaper in terms of compute, you'd expect us to use less. Because the price is coming down so much, it opens up so much more utility that usage goes up like crazy. Yeah. And now we have even a higher demand for compute. Yeah. So Nvidia stock price going down, I think that makes no sense.

Yeah. People saying that big tech is in a bad position, I would say no, because they're not going to stop here. So they if we were just going to stop at this level of models, maybe people would have have a case. But the whole point of these research labs is to be pushing the limits and building the best models that they can.

And for that, you require compute. Yeah. So I think yeah, DeepSeek is great. But, and it's amazing that this is happening all happening in January. It's pushing more competition between all the different labs and different companies creating these models. And I think we're just going to see, I think this is just great for us.

It means that we're getting better models. OpenAI made the decision to, the o3 [00:31:00] mini model is going to be released soon. And they said that's going to be free for everybody. And I believe that's in part due to the pressure from DeepSeek. Yeah. 

**Simon Maple:** I feel stressed if I work for a model company when you just constantly feeling like, Oh my gosh, it's constant jumping, who's jumping ahead at which time and you haven't always catch up or try and keep that lead, right?

This must be pretty stressful to keep pushing. Let's talk about one of the other things actually, just in terms of Hackathon projects because I know hackathons have been wildly popular again now with just seeing whether it's a internal company hackathon where people are just trying to work out how they can use AI to improve their processes or automations and things like that.

Any good hackathon style projects that you've seen recently? 

**Farhath Razzaque:** So I did see a project recently where these two guys, they built using Gemini 2. 0 Flash, which is a multimodal model. So it can access, it can view, understand images video and speech. So what they, as well as text, and give responses as well in audio or text form.

What they [00:32:00] did was they built an app you'd use to film your house, so it would analyze everything inside your house using another object detection, or I think that they use, and it makes a catalog of all your belongings, valuable possessions in your house. And the use case for this was actually insurance.

So if usually you'd have to estimate the value of your property and possessions, if you undervalue this, if something, God forbid, was to go wrong, then you're not going to get full compensation. If you overbid, then the premiums you're paying are too high. They built this in two days. 10 minutes, you walk around your house, it catalogs all your items, gives you a valuation.

And now this can be used by insurance companies obviously with more development, like I said, production takes a while to do, but it can be used by insurance companies to value your properties. 

**Simon Maple:** Yeah. Yeah. As much as that sounds useful. It also sounds like a security nightmare, but in a couple of days, it's amazing [00:33:00] what you can do with that kind of thing.

So that's brilliant. Yeah. Final question. You obviously read a ton watching YouTube, watching clips of the new developments of various AI tools and methodologies. Where do you go, though, to maybe pick two or three sites? What are the sources of news and good in depth content for you? 

**Farhath Razzaque:** I use a huge variety of sources.

Probably what I use the most is YouTube. Yep. So my few channels I'd recommend are for developers, IndieDevDan, and WhatAboutAI. Yeah. I feel like they really do their due diligence when it comes to using the latest tools, models, and having good insight. And there's also AI Explained, which I feel like he talks more about the broader implications of the developments that are happening.

So to stay more up to date with the news, I feel like that's another good channel. I also use Discord, so following different project Discords. There's usually a lot going on there. People are building all types of interesting projects. Twitter and Reddit are [00:34:00] really good for getting people's perspective.

Like the latest news, when someone hears about something they were posted about on Twitter when people have been experimenting about stuff they'll have discussions about it on Reddit. I feel like these are very good platforms to be using and a lot of the top researchers and companies are all on Twitter as well.

Yeah. Constantly posting about stuff. So it's really, and researchers, so it's really, it's that first source of information where it's first. Yeah. Where it hits. Yeah, it comes to the first. Product hunt is a really good resource as well. So these are projects that are people have really thought about.

So I'm not just demos. People are really thinking, how am I going to get this production? And tons of really innovative stuff there. And plus it ranks itself, right? So people vote and ranks, when you see the best of what's coming out. 

**Simon Maple:** Yeah. Yeah. Oh, that's cool. Lots of great tips there as to where people can go.

Amazing. Farhath, it's been an absolute pleasure chatting with you and and yeah, great to have the practical tips and the user's perspective as to what they're using. So thank you very much for sharing. And for those of you who are interested in our Discord and YouTube, of course, got the AI Native Dev [00:35:00] on YouTube as well as Discord, where we can talk about a ton of news tips, tricks and various new tools and things that people are trying AI ecosystem as well. Make sure you subscribe if you like the content, enjoying the content, and thanks very much for listening. Tune in to the next episode. Thank you. See ya.

Thanks for tuning in. Join us next time on the AI Native Dev, brought to you by Tessl.