⚡️ AI Talks by Zeo (Episode 4)

Comprehensive panel discussion on the trajectory of generative search, autonomous agents, and strategic forecasting.

Speakers: Yiğit Konur, Ozan Ketenci

01 Oct 2024

Video Transcript

⏱ 36 min read• 6,419 words

Read the full speech or switch languages

— We’re back together. It’s been over a week. Let’s check our initial rankings.

— No way. I really thought we wouldn't pull it off. I’m genuinely happy.

— You didn't trust me, did you? Definitely—

— No, I don't even trust myself with this stuff.

— You don't trust yourself either, right?

— I always flake out on things like this.

— Yeah, totally. I love the conversational vibe. No cameras, no pressure.

— Honestly, it might be the only style we can actually pull off.

— Exactly. It’s the current trend online anyway. I can’t stand overly formal content. Once we start talking, it just flows. We don’t even script these. We keep the prep minimal. We want to keep that raw, spontaneous energy for you guys. Yeah. We did chat a bit beforehand, but not much. We planned to cover some cool news and updates. I’m pulling from my own bookmarks today. It’s a mix of memes, motivation, competitors, YouTube, and startup stuff. I’ve got a few things saved for our "read this, use that" segments. I want to start there. I’ve been digging into some interesting tools lately. I’d love to share what’s been catching my eye.

At the top of my list is a deep research AI agent for Thinkbuddy. I’ve been checking out existing open-source solutions for it. There are a few different ways to go about it. OpenAI drops a feature. Someone immediately open sources an implementation. It’s like everyone is hungry for ideas. I started testing this out. My first impressions are pretty solid. Hugging Face has a similar product running in parallel. I’m trying to build an AI agent structure with these. This week—

— What exactly are you using right now?

— I always start testing research models with Flash. We also use LightLLM, as you know. It’s worth mentioning. That tool is a total lifesaver.

— It’s amazing. Open Router, LightLLM.

— Exactly. The best part about LightLLM is that it’s open source. You can host it on any server. You just set up a YAML or JSON file. You plug in your model credentials. It automatically converts all models to OpenAI format. You don't have to deal with AI Studio's weird formats. You can skip Google's proprietary SDK. This means you can use everything just like OpenAI. You don't need their specific code system. You just swap the model name. It offers incredible flexibility. The guys behind Open Deep Research likely use OpenAI API. That's the standard. This setup lets you run different Gemini models within the same system.

That’s a huge plus. Speaking of Gemini, are the new models on your radar? Any thoughts on Gemini this week? How was your first experience? I know you've been testing different LLM options lately. What’s your current go-to? What are you using?

— Honestly, I’m still a ChatGPT guy for daily tasks. I’m sticking with that. The 3 Mini High is actually really good. I don't really need the One Pro for my workflow. For quick tasks, I still use GPT-4o. I still use ChatGPT and Claude in the browser. I haven't fully adjusted to Gemini yet. But for API work, I’m leaning toward Claude and Gemini. Everyone is saying Gemini killed RAG this week. That’s because of the two million in-context token limit. There’s so much talk about it. I tried a few PDFs. I didn't get the results I wanted. I can't get a decent long summary. Not in Gemini Advanced. That really annoyed me about Gemini.

— Yeah, people are so quick to kill things off.

— Look, I wanted to mention this today. It’s like a tabloid press now.

— Right? Exactly, exactly.

— OpenAI pulled this, DeepSeek crushed that, OpenAI’s model beat DeepSeek. It’s constant. Someone dies, someone rises. We’re burying one, birthing another. We’re killing people off. Twitter has turned into total paparazzi.

— Seriously. The AI scene is just everyone chasing clout.

— Honestly, Twitter isn't what it used to be. It’s lost its spark.

— I’m actually more optimistic. Especially about the feed post-Elon Musk.

— If you like that content, sure. But it’s frustrating now. Even top creators are forced to write that way. The platform offers zero flexibility. Let’s not even get into the politics. I was pro-Elon at first. But look at it now. It’s fine for following AI, but everything else is pretty bad.

— Most social platforms are terrible for small creators. TikTok is the only exception. They use their recommendation engines and AI much more effectively.

— Speaking of TikTok, I have to jump in. We’re expecting a massive move from them. I was discussing this with friends. They’re building something that could end the influencer era. They keep pumping out video generation models. OmniHuman dropped two weeks ago. Now there's this new thing called Goku. None of them are open source, by the way. They just drop papers. The tech in Goku is insane. It’s heading toward a world where you can build your own AI agent influencer. We’re moving to a space where you don't need real human influencers anymore. We say TikTok gets what it deserves. But honestly, this could be a massive blow to them.

— Looking at local creators on TikTok, many won't like AI. Especially those making fun, short-form content. I love those creators doing funny street interviews with local accents. They’ll be fine. They’re safe for now. But let’s be real. I don’t think AI will kill influencer culture entirely. Sure, two LLMs could replicate our conversation right now using current speech tech. But there’s a difference between two humans talking and two AIs. People will still prefer real human creators for a long time. I saw a post about this on X recently.

People were making huge claims. Like Instagram dying in three years or AI making reality indistinguishable. Look, I don't care about distinguishing reality. I’m just here for Ozan. If Ozan sends a photo, it’s Ozan. Instagram should just end there. Twitter is getting weird. Let’s get back on track. We tend to drift. I saw something about LangChain using LLM for brainstorming. They prepped some templates with Streamlit. I want to discuss a different approach. Usually, prompt engineering means role-playing. People always say, "You are an editor, do this." That "act like" prompting is everywhere.

This episode offers a fresh perspective. It reminded me of something else. I want to show you that too. When you ask a question, you can use the Six Thinking Hats method. It’s a way to tackle problems. You look at data, risks, benefits, and creativity. It’s an old framework. They built a project called "Brainstormers" using this or the 5W1H method. It’s not a huge project, but I liked the flow. You should check it out. I’m trying to get it running now. One thing really caught my eye here. It’s called Anthous.

I discussed this with a friend before. It gives people frameworks to systematize their thinking. It tells you how to approach a problem. Or you could use the Eisenhower Matrix. It gives you frameworks for how to prioritize things. These problem-solving approaches are perfect for prompting. You can experiment with this. Take the Six Thinking Hats method. You create six experts to analyze one topic. Each expert has a totally different perspective. This helps you push the boundaries of prompting. You can upload these documents to an AI. Just tell it to create a system prompt that forces the AI to think like the document.

Exactly. I can generate a custom system prompt right here. How do you rate my Gemini fanboy level? I’m just testing the models. It’s pretty good. We took all of this. Use it to build a system prompt. If you have a complex task, create a custom tool. Let’s call it Six Thinking Hats. Ask it how to solve world hunger and water access by 2030. I told it to use those six approaches. It’s a solid decision-making framework. You’re basically forcing the AI to view the problem from multiple perspectives.

Look, it gave us different approaches. DeepSeek dropping Chinese characters always cracks me up. It’s so weird. It usually happens during the thinking process. But we see it in normal output too. It’s like a subliminal 25th-frame effect. It just injects Chinese characters. That can really change the output. That was a key point about prompting. Ozan, have you had any surprising experiences with prompting lately?

— Actually, remember when we started testing prompts at Latitude? Testing is the most efficient way to go. Hardly anyone does it right now. Very few systems actually use it. We don't use it in our daily workflows. But there’s always a better prompt out there. We can't check infinite samples. Prompts and system prompts are getting so long. Evaluating 10 or 15 examples is exhausting. Use tools like Latitude or Prompt Wizard. I have those bookmarked too. They help with iterative improvements. You can feed them frameworks like system thinking or problem-solving. That helps the LLM refine the prompt itself.

The main takeaway is to stay open to optimization. Always keep gathering feedback.

— Yeah. We started doing this way too late. We really regret not doing it sooner. It’s a huge hassle without the right product. That’s why I want to talk about Latitude today. Great setup. I’ll talk about how I use Latitude in my own workflow. I want to show you something here. Let’s call this syntax Prompt LLM. It’s a specific format. You define the model at the top. Then you write a system prompt. You can add examples too. Like user and assistant turns. This is the prompt for my voice dictation service.

It takes my Whisper transcriptions and cleans them up. It makes them sound like me. If I sent a raw message, people would ask if an AI wrote it. I don't usually write like that. I use spaces and lowercase letters. The prompt fixes that. I also use examples to teach it. Like when I say "Cloudy," it knows to output "Claude." Or "Light LLM" for "LiteLLM." Do you still use few-shot prompting, Ozan? You work with a lot of examples, right? Unless you’re getting JSON output, I’m a big fan of examples. I’m using over 30 shots here. It’s been super helpful.

I basically handle all my daily conversations through this now. It’s been a huge help. I use Latitude because it gives me an API. You just send your text over. It takes that message and drops it right into your prompt flow. It finds your variable automatically. This makes editing everything super easy. Let me show you a chunk of my earlier talk. Look, when I typed "Cloud," it knew to change it to Claude. It actually understood the context. Pretty impressive. It cleaned it up perfectly. The best part is keeping things organized and testable. If I paste some rough text and hit run prompt, can it actually fix it?

It didn't fix it. That's a writing error, not a Whisper glitch. It also runs an automatic evaluation for every task. That’s a great feature. Was the AI output actually good? It scores the results. Every request and response gets logged and evaluated. It says. "It kept Yiğit’s informal tone and style." It gives it a 5 out of 5. So what does this give us? We can constantly tweak our prompts. When we release a new version, we track the score changes. We can see how our median time shifts with each update.

You can analyze all of this. It goes beyond basic cost analysis. I’ve really been hyping up Latitude. I don’t even know them personally. There are other great features too. You could also use Langtail. It’s a similar vibe. We used to love that one the most. Honestly, they’ve fallen behind a bit. They’re still shipping updates though. They keep adding and removing things. It’s a bit hit or miss. Yiğit, you have experience there too. Exactly. Choosing a competitor over a tool you have a testimonial for shows real integrity. That’s ethical AI, man. Give credit where it’s due. I roasted them in two minutes. If they’d just linked my testimonial, maybe I’d be nicer.

They didn't treat us like a big enterprise, so we’re calling them out. Their playground is still top-tier. Honestly, it’s the best one out there. It’s the best playground for prompt engineering. Did we jinx it? Maybe. It feels like they broke something. They changed it up. Is that right? Does it still work? They tweaked it, but it shouldn't be like this. Maybe they changed it back. This might just be the public view. Once you log in, I think you’ll still like it.

Yeah, these are the links I have. Ozan, maybe send me a few. I'll shoot you something. You love text-to-speech stuff. I'm sending it via private chat. It's not a library, but a company called Zifra AI. I hadn't heard of them. They announced something called Zon of. You use 30 seconds of text-to-speech. Then you talk a bit. It starts voice cloning immediately. They say it's fast. I checked their site. It works with very low latency. How's the liveliness? We need to check that. It's nice. They added a Trump voice to the demo. It looks good. The argument that we can clone voices in 10 seconds is easily proven now.

I mean, it's not exactly a good thing. Cloning my voice from a 10-second clip is impressive. It's a wild field. We still don't have an open source equivalent to 11 Labs quality for Turkish. I really wish we did. The real-time performance is great. Emotion control is solid. Audio quality is good. Fine-tuning the speaking rate is nice. We should test it. Sometimes they work well in Turkish by pure luck. I'm thinking about projects we could build. There are tons of ideas. It just takes time. I don't know if anyone in Turkey is doing this yet. I haven't heard of any AI customer service or call center agents being used.

Honestly, I haven't heard of it either. I don't follow the Turkish scene that closely. It's not available in Turkish anyway. I thought maybe some generative models work language-agnostically. OpenAI TTS doesn't support Azerbaijani. But it can actually output it. I wondered if I should try it. It’ll probably sound bad. Still, that’s cool. Maybe we can use it in other products. 11 Labs is really expensive. We need more cost-effective solutions on the market. Exactly. I was going to mention something else about text-to-speech. Let me try to remember.

I want to talk about 11 Labs Reader. It’s a new app. It came out about a month and a half ago. It’s on iPhone. Not sure about Android yet. 11 Labs has the best AI generated voices in the world. I’ve had books I wanted to read for years. But actually reading them feels like a chore. So I take the books and customize the use cases to my own life. Then I turn them into my own audiobooks. It might sound a bit weird to some of you. But it makes the books listenable. It’s tough these days. I mean, who actually sits down and reads a technical book cover to cover? I’m talking non-fiction, not novels.

Honestly, I’m impressed. I take text generated by AI and send it to my phone. I listen to it while I work. When I’m coding in Cursor or Windsurf, I zone out. I just hit enter and keep an eye on the screen. Having a podcast in the background is a game changer. You should definitely try this. It’s a great feature. You drop in text and it turns it into a podcast. It’s just like NotebookLM. You can listen to the audio directly from there. It’s a really cool feature.

I wanted to mention that. Oh, did we see this? Let me send you a few things for reference. Karpathy dropped a new video. It’s a must-watch. I’ve only seen an hour so far, but I’m already blown away. Seriously, Karpathy, you could just translate this or have it summarized. Have it read to you, man. It’s perfect for your project. Let’s take this chance to plug our AI tool. I’ve put a lot of work into this. I don't know if I tried it on a three-hour video yet.

If it fails on the live stream, so be it. Karpathy’s content is top-tier, but it’s heavy. It’s super technical. It’s for people who really want to know the internals of an LLM and how they work. By the way, the Hyperbolic guys are trending hard after that screen share. This guy got 600k views in six days. The audience is A-plus talent. It’s a massive community. I bet they’ve changed a lot of lives. It’s talking about LLMs at a higher level, right? This is so good. It’s rethinking how we view LLMs. It covers the GPT-2 training process too.

It’s important for understanding the basics. Honestly, nobody watching this has a math degree. Nobody is sitting down to read an LLM paper. I doubt anyone is digging into the Deepseek R1 or R2 papers. But it offers great practical insights. Good call, Ozan. I need to check this out myself. It’s not really something you can just translate to Turkish and read. You have to sit down and dig into it. Watching it is better. Since we’re on the topic of papers, I’ll drop one or two. One is more important. It leans into a philosophical discussion.

Awesome. While you drop that, I’ll show you what I’ve been trying to remember. We talked about text-to-speech models earlier. Since quality is subjective, I wondered how they benchmark these things. They usually analyze quality versus price. The top right is for the most expensive, high-quality models. The top left is the price-performance segment. OpenAI is totally dominating that space. For Turkish, there are almost no quality alternatives besides 11 Labs and OpenAI. I wanted to show you how they score this data. It really surprised me. Artificial Analysis has this thing called Speech Arena.

They give you a text. They ask which one sounds better or more human. You pick one. You have to listen to the whole thing to choose. It tells you, "Hey, you liked the Google Cloud version." "You didn't like the Azure one." I think that's a great feature. It helps you build familiarity. They did similar things for video arenas too. It helps you gain awareness of these models. If you want to spot AI generated video models at a glance. Spend 2-3 hours here. You'll develop a real eye for it. Just a quick note. Oh, Ozan sent a paper. This is a good one.

I haven't read it yet, but I used Gemini to summarize it. It discusses an evolutionary approach to AI development. It’s a pretty deep topic. Did that actually work? I didn't send yours, I sent the Karpathy video. That thing is a beast. The model is excellent. I worked so hard on this. It’s a 3-hour video, 37,000 tokens. What are the main points Karpathy covers here? Let's see what pops up. Maybe some YouTube clips. Nice. Let's pick a few models. Perfect for someone who doesn't know how LLM works.

I don't think we need extra input. Yeah. By the way, we added something new. You can see the reasoning in our system now. It works with reasoning models. It's actually pretty cool. I read about this before. People with ADHD often scan the index first. They read the summary before the book. You can do the same with educational videos. Drop a YouTube link here. Get the text. Get the gist. Then ask questions while watching. Like, what does he mean by that? We fed the whole context to the AI already. It’s a super useful way to learn. Or if your friends like Ozan send you papers. You don't want to read them. It’s too much work. Honestly, we don't read them either.

You’re being honest about not reading them. How could I read all those papers? You’re right. Some papers are incredibly tough. There aren't even many math tricks involved. It’s more about the philosophy. It’s about how AI systems should take inspiration from nature. It highlights known unknowns and unknown unknowns. I told it to explain like I'm in high school. Don't be so technical. It used a test analogy. You know a tough teacher will give a hard test. That's a known unknown. But you don't know if you'll meet an alien on the way home.

It called that an unknown unknown. AI and machine learning handle known unknowns well. But they struggle with unknown unknowns. This is a huge issue for humanoid AI. Especially when we interact with robots. It looks at solutions and nature-inspired methods. Philosophy fans will love this part. It’s great. I love how I can question you while you explain. It’s a fantastic way to learn. But it makes people lazy. It kills the drive to actually consume information. That’s not great. You’ve probably seen that study from a while back.

They checked Google Scholar for ChatGPT patterns. They found 500 or 600 papers with direct copy-pastes. Academia is just using AI to summarize AI text now. We’re just chasing automated outputs. Stanford released a research module before deep research even launched. Remember when we talked about that? Universities need to pivot. Don't ban AI. Integrate it as a controlled resource. Track how students use it to prepare. It’s like that. Stanford has a tool for specific research reports. OpenAI has an education arm. Their enterprise side is totally separate. OpenAI Enterprise is one thing. Education is a different category.

They signed a massive deal with California State University. It’s their biggest one yet. Do you know what it is? How does the product differ? It’s not just standard ChatGPT. It’s different. Exactly. It seems like they’re providing fine-tuned models. It’s built for academic use, right? They aren't just giving you the off-the-shelf version. They’re building on top of it. Nice. Someone on YouTube is definitely using this. I’ll stalk them and pull some insights. I haven't opened your second link yet. Should we take a few quick questions? We have some great comments. Sure. Serkan, sorry we missed your additions earlier.

My bad. Let’s start from the top. Thanks to everyone watching. Thanks for the well wishes. Serkan roasted ByteDance. He’s right. It’s just a paper for now. Speaking of ByteDance, they built a competitor to Cursor. Did you see it? It’s called Trae. I haven't seen it. No. Everyone is jumping into this market. Seriously. Who was the biggest competitor to Cursor again? Windsurf. Right. Windsurf. I think there was another one too. Maybe something like Jet? Codium. Yeah, Codium. Okay. It felt faster and more efficient. Which one is the best? It’s subjective. Windsurf is making a strong move lately.

I can answer this well. I open both when starting a project. I run the same prompts in Windsurf and Cursor. I’ve tested this on 23 projects. Onur Özcan loves Windsurf. He’s been posting about it. You might have seen it on Twitter. He built an entire iOS app with it. He’s basically their local rep now. He’s really grown their presence in the Middle East. They had Windsurf ads everywhere in San Francisco. They must have raised a lot of money. They plastered the subways. I use Cursor more because my Cursor rules are dialed in. Honestly, Cursor is still king. People who don't use the AI agent mode often prefer Windsurf.

Many people don't even notice the difference. I have to point that out. The Cursor team has a weakness there. Even Danny Postma just discovered the AI agent mode recently. He said, "Wait, there's an AI agent mode?" He said he was just using the standard version. They replied, "Welcome to the internet, brother." It’s a game changer. If you aren't using it, prepare yourself. It’s like using a completely different product. There’s a Yolo mode. You can let it run commands without asking. I just set the prompt to execute whatever you want. I know it won't break anything.

I just kick off the task. I tell it to deploy to Google Cloud. It handles Cloud Run and builds the Docker files. It handles the entire deployment. It’s reached an insane level. Even with basic software 101 skills, you can build so much. Sure, complex data pipelines are still a pain. But most people aren't using these tools enough. Cursor is still my favorite. You should try Windsurf too. Both are incredibly cheap. They’re basically free for the value they provide. They’re definitely operating at a loss. As someone who uses a ton of LLM credits. I can tell you they’re losing money at 15 bucks.

Take advantage of this competition. Try both, but Cursor is still the king. By the way, there’s a slight lag between your questions and the stream. We see your questions a bit late. So our answers might be delayed. Someone asked how to get started with AI agents. Honestly, for any part of AI, I always recommend starting with NAT. You already started with NAT. Maybe we should get NAT tattoos? That could actually work. It’s heading that way. You can’t learn this just by reading.

That’s not how my brain works. I have to build it. I see it in NAT and just apply it. It’s super easy to learn. Jump into the no-code world. Build a small use case. Before you know it, you’ve learned it. It’s really not hard at all. It has its own interface. You can connect any model you want right there. Pick your models from the list. Drag and drop them. Add some classic memory. Or connect your own tool. That’s one of its best features.

You can do so much in NAT easily. Like, if this happens, send it to Slack. Or grab this from Slack. Just plug in your prompts and run them. It’s a huge advantage. It shows you that an AI agent isn't some massive mystery. You’ll see how easy it is to build once you tinker with it. I highly recommend everyone play around with NAT. NAT is something everyone needs to learn. There’s basically no Turkish content on it. That’s pretty sad. Honestly, you don’t even need Turkish content. If your English isn't great, it’s still fine.

It’s super visual. You just connect this to that. It’s that simple. Follow along with a couple of videos. Once you get the hang of it, it flows. People really dragged their feet moving from MACOM to NAT. We actually get annoyed now. Why aren't you using it? The best part is NAT is free. It’s open source. You can self-host it. There are almost no limits. Even the paid version has very few restrictions. We highly recommend building here. If we create content, we’ll probably use NAT. Maybe in the coming weeks, we can do a one-hour project build session.

Like and subscribe—just kidding, we don't care about that. There’s still debate around the AI agent concept. What actually defines an agent? If it doesn't do any decision-making, is it really an agent? People call simple chatbots agents sometimes. I don't think that counts. If there’s no decision-making involved, it’s not an agent. If it’s just answering questions, it’s not an agent to me. It’s closer to the general definition. It doesn't quite know what we want yet. Mehmet, maybe you meant a basic chatbot when you said AI agent. In that case, these tools aren't really necessary.

It’s a totally different ballgame. But there are great tutorials on YouTube, like the ones EAT showed. They’ve built massive agentic workflows. They plan calendars, send emails, or manage your day via Telegram. Start with a project like that. Everything else will follow quickly. Well said. Everyone can find something relevant to their interests. You can easily find a NAT tutorial. These have really taken off lately. It’s incredible right now. It wasn't like this before. Even six or seven months ago, it wasn't this advanced. We’re ahead of the curve as always. It hasn't helped us much yet, but these are prime companies to invest in.

It has a huge future. It simplifies software development from start to finish. It’s not just AI agents doing this either. Tools like Cursor make it so easy. If software creation becomes this simple, it’s a huge threat to things like NATN. It’s a different paradigm, but the goal is the same. It’s about simplifying or removing the friction. I’m going to try something new. I have some advanced NATN workflows, but I want to move to hardcoding. I plan to migrate the platform elsewhere entirely.

Ever tried taking NATN JSONs into Cursor? Have you tested that? I have. Definitely. How is it? Does it output anything good? It’s actually not bad at all. Take our prompt enhance in King Pad. Or wait, let’s use the waitlist check. It’s a simple script I wrote. It hits Redis to check some data. Let’s run a quick test. Let’s drop it into Cursor. Think of it like an AI agent. One of the best parts of NAT is how it gives you JSON directly.

You can select everything on the canvas and copy it as a JSON output. I usually start by saying. "Explain the flow of the script shown below." I’ve used that a lot. Then I’ll say, "Write the equivalent using Python and FastAPI." Thanks to Gemini. I don't even wait anymore. I just hit stop and move on. Look, it handled my email validation and waitlist API check. It got all the arguments and everything right. I usually run this code through Remix too. One catches the use cases the other misses. It’s a great workflow. You can prototype a flow in NAT and have an LLM build it.

You don't even need a deep dive setup. You can use it just as an illustration tool. It’s really powerful. Serkan asked a question about AI agent frameworks. I was thinking CRI, but I read a post about it recently. I can't find a good comparison of these agentic frameworks. If you’ve tried any and liked them, let me know. I’d love to check them out. Since I’m mostly exposed to CRI, I’ll start testing with that. I looked into OpenAI’s Swarm. I dug into the code with an AI. That’s the one I couldn't remember. It came out recently. Well, four months ago.

Time flies. There are some great examples there. Like a design handoff for booking flights. It supports what Ozan said earlier. It uses multiple specialized agents. One for support, one for refunds, and one for reservations. These specialized agents call each other. It’s like a tool call. A user connects and speaks. The system routes them to the right agent. It’s a solid demo. Agents are best for these complex, autonomous models. I checked out CrewAI, but haven't used it much. The founder is always hyping it on Twitter. They claim Apple uses them. They even list CRI in job postings. That’s a huge push for them.

Serkan and I looked at AutoGen, but didn't build anything big. I mostly looked at the early stuff. Remember those first experiments like BabyAGI? A Y Combinator guy started that whole trend on Twitter. We tried those things back in the day. You need a solid use case. I haven't done that in a while. My current work at Agenting Body is a web search AI agent. It’s a true agent. It has so many use cases. It breaks down user queries. It runs new searches, gathers data, and synthesizes it. It delivers one final answer. It has to work across all models.

These scenarios can be pretty challenging for AI agents. Mehmet, thanks for joining. Keep building cool stuff at NAT. Let us know how it goes in future streams. We’d love to hear about it. As NAT Turkey reps, we launched NATN. We love seeing the community grow. It’s a perfect product. Ozan, did we miss any questions? Yes, we missed one from Ilyas. He asked you directly. Did you see it? Yeah, I saw it. Hey Yigit. I had major biases against AI. I thought it was just machine learning. My mind changed over the last six months. Yeah, I see the rest of the question. It’s definitely way beyond machine learning.

It’s not just classic predictive modeling. As an AI maximalist, I would’ve convinced you earlier. But honestly, we’re still in the very early stages. For the market, CRI is perfect for those four experts. Or use NATN. You can build your own workflows there. Connect every tool to an AI agent. You can even trigger another NATN workflow right there. You can link them as tools. Maybe one agent handles refunds. Another agent handles reservations. Build them separately with their own tools. Position them as sub-agents under a main agent. You don't need to hardcode the logic.

Use an orchestrator. Don't think you need one agent for everything. If there are many decisions, don't limit yourself to one workflow. Use an orchestrator. It decides which sub-agent to call first. Each sub-agent has a role. For SEO, you might have a content agent. Or a competitor research agent. But you need a lead agent to pick the right one. The goal is simple. We used to pull data manually. Now we use natural language. The orchestrator must understand the intent. It routes the workflow to the right place.

Each decision-maker fits the NATN structure perfectly. It’s super easy. Connecting different agents is a breeze. You might struggle at first. But you'll get the hang of it in a day. It’s a very rewarding product. You start using it and think, "Wow. This is great." You realize what you can build. You keep stacking wins. It’s like a game. It’s fun. If you stick with it for a month, you’ll be twice as good. To really leverage AI, you have to use the API. I don't know if that's a bold claim, Ozan. Actually, no, it’s not. Web interfaces are fine for daily tasks.

They work well for most things. But for complex, chained processes, they fall short. Canvas or projects only go so far. You can't process massive data in a custom GPT. You need to build your own system. Exactly. Or scrape hundreds of pages to extract data. It boosts your creativity. You can build things without depending on anyone else. The biggest advantage of NAT over tools like Cursor is the visibility. Where did the code fail? I ran this 100 times and got 3 errors. It lets you debug those issues easily.

It’s a great way to grasp the core logic of programming. Let’s check the final questions. We’ve hit the hour mark. Let’s wrap it up. It’s 10 PM in Turkey anyway. Man, when is this 3-hour gap becoming 2? 3 hours is way too much. After living through that 11-hour gap in San Francisco, Lisbon feels like home. Yeah. So, Ilyas, you connect them like this. Not just tasks, but decision-makers. They handle their own jobs solo. Then they merge all the data. The final output could even be a separate AI.

Not an AI agent, but a final controller. Or you design a multi-stage system. You never use a single stage with APIs. Think about chatting. You write, then you write again, and again. We never solve things in one API query. It shouldn't be done that way. If you're doing something advanced, you check the web interface yourself. You verify if it's right or wrong. To see this in an API, you need a second AI node. You need to check if it's correct. Build structures like self-consistency. Break down your manual tasks into tiny pieces. Ask yourself: where can I use a rule-based system?

That part is solved with standard AI. Where is it not rule-based? Where are there too many variables to calculate? That’s where you dedicate an AI agent. That’s how you design the system. Exactly. Ilyas's point sparked an idea. Picture a main planner AI agent. It uses a tool to decide on SERP keywords. It then creates content briefs for those keywords. It has a tool for the SERP API. It pulls specific results. It has another tool to convert URLs to markdown. We could even assign that to a dedicated agent.

That agent visits the URL. It summarizes the content based on the query. It’s much more efficient. It extracts the core essence needed for the brief. You could also connect an API like Ahrefs or Semrush. It would pull search volume and scan related keywords. You can test this directly in NAT. You’ll get great results. Don't mind the clunky UI. Use a webhook trigger to run it like an API. Build something cool on top of it. Also, Upstash recently released something. Upstash is a fantastic company. I’m a huge fan. It’s great that the founders are Turkish. They launched AI chat components for Next JS.

My screen isn't shared, but check it out. You can connect this logic to build your own site assistant. Don't be biased against these tools. If you say you can't do it, you definitely won't. It pays to tinker. I push my friends to do it. Only one in a hundred actually tries. Most are scared of the jargon. These technical terms aren't that hard. We aren't building LLM models like Andrej Karpathy. We use existing tools to build smarter solutions. There’s a huge talent gap here. It’s even worse in Turkey. Get your hands dirty. That’s my advice. By the way, DataStax is a great product.

Check out Langflow. Is it called Astra for DataStax? Yeah. Look into that too. That’s it for today. We’ll be back next week with fresh news. We didn't even cover the gossip today. Elon Musk’s acquisition drama and all that. We’ll save it for next week. Scams and everything else. It’ll be a good talk. Let’s promise to cover it then. You’re welcome. Thanks for joining us. Take care. See you next time. Bye.

This transcript was generated using AI speech recognition and verified against subtitle style rules.

Yiğit Konur

Content author

Yiğit Konur

both

Contact us

How we handle your data

Open as a full page