AI Talks by Zeo (Episode 1)

Deep dive into integrating generative AI workflows, agentic tooling, and automation within marketing agencies.

Speakers: Yiğit Konur, Ozan Ketenci

20 Aug 2024

Video Transcript

⏱ 26 min read• 4,508 words

Read the full speech or switch languages

— We’re back this week. It was a rough commute. Lisbon is rainy and chaotic today. Glad to be here with you. Welcome, Ozan.

— Thanks, man. How are you doing?

— I’m good. Doing great. It’s been a busy week. We’re halfway through already. I’m grinding through my to-do list. It’s been a wild week for AI. We’ve got plenty to talk about. How was your week?

— Man, my week was intense. I’ve been focused on Text to SQL. It’s a cool, interesting topic. I thought it would be simpler. But as the database gets complex, it gets messy fast. The agentic approach stopped working. It let me down on decision-making. So, I’m pivoting to rule-based testing. Response times are creeping up. Writing and running advanced SQL commands is tough. It runs the process and pulls data from the database. It builds tables too. It takes about 25 seconds. That’s not bad if a human wrote the code. But for a chat experience? It feels way too slow.

You hit enter, ask the question, and just wait. That’s the issue.

— User patience is dropping fast. As models get faster, expectations rise. They need instant feedback on screen. You’re spot on. Y Combinator even dropped a great video on the UX of AI recently. They cover this exact topic in depth.

— N8n has a built-in SQL agent. I tried it out. It ran like 32 iterations. It kept querying schemas and failing. It just kept looping. It spent two minutes grinding away with SQL. That is totally unnecessary.

— Isn't that N8n's biggest drawback? No streaming support is a huge issue. You build an AI integration and get hit hard there. By the way, this is a great video. Mirza Ming sent it to me. I really liked it. It opened my mind. We can even flag things here. They mention Polymet too. They want a polymeter here.

— Yeah, yeah, I saw it. I saw your post on LinkedIn.

— Exactly. I watched it. Who is first at Polymet? They do great work. It mentions that here too. UX of AI is a critical topic. It pushed you a bit, but if you get the output, it takes 25-30 seconds.

— It works. I ran a few tests. You know textosql.ai, right? It's very popular. One of the first to do this. It's widely used.

— This is Turkish, by the way.

— Is TextToSQL Turkish?

— Yeah, Mustafa.

— For real?

— Totally. They even came to our event.

— That’s awesome. The outputs are good. But it’s very specific to my database. I used it for testing. The infrastructure here is great. It pulls the entire database schema. Most tools don't pull that much detail. It grabs the whole schema. But it doesn't always recognize things correctly. There are errors in data types. Case sensitivity issues. The assumptions are usually right. But when it turns an assumption into a WHERE filter, it messes up the casing.

— You might have fried the AI's brain a bit.

— I actually did something about that. I tiered the approach. I shrunk the tables and their schemas. Essential schemas go in at the first level. If the data isn't there, it checks the larger schemas. Most queries only need a fraction of that. One table has 70 columns. That's a lot. But 10 of them cover 90% of queries. I trimmed it down. It became much easier to explain. It started working really well after that.

— That’s awesome. I focused on something else this week. I’ve been obsessed with that benchmark project I mentioned. As the number of models grows. It’s getting harder to choose the right one for each scenario. I’ve been working on that benchmark project since last week. I’m writing a report for Zeo. We want to build Turkey-specific, Turkish-specific, and industry-specific benchmarks. It’ll be great data for us. I used deep research quite a bit here. I had the chance to combine everything. I used deep research, OpenAI'ın, Grok, Perplexity, and Jina.

OpenAI is the clear winner here. No one can really argue with that. It might be surprising. Let’s dive into deep research. My output focuses on which LLM wins for each task type. We’re using this in ThinkPad'de for active routing. When to use Flash, how Haiku performs, and so on. I’ve gathered some context. I’m building an AI based on that. Once it’s finalized, you’ll ask ThinkPad'de a question. It will say, "Ozan. This is biology." It will suggest Gemini 2 Pro based on AIMA 2024 benchmarks. I want it to provide clear, data-backed reasons.

I’ve been working on that. I want to dive into deep research soon. Great news: Jina open-sourced their own deep research implementation this week. That’s a huge win. For those who don't know, Jina is an independent company. They build powerful APIs for various tasks. Their most famous one is the Reader API. You give it a URL, and it returns Markdown. They offer very generous rate limits, too. That’s how they made their name. Then they expanded into other areas. I’m a big fan. It streams the thought process for these broad questions. It’s a lot like Grok. I really liked the output.

I'm looking into how deep research works. That’s been my focus lately. Benchmarks and stuff. I had to make some business calls for the ThinkPad. I had it research competitor pricing one by one. I turned each one into a deep research doc. Then I combined them in Gemini. Specifically, Flash Thinking 2 is great for this. Say you have 10 docs at 10k tokens each. That’s 100k tokens total. Gemini 2 Thinking is the best spot right now. Its attention is solid. It circles back to every part of the text. It pulls out real insights. Plus, the output length is huge. It can handle 20k to 30k words easily.

Honestly, I don't like OpenAI'ın reasoning models at all. None of them. They’re pretty bad. Right now, Google...

— I wasn't a fan of their deep research either.

— I didn't like deep research. I actually thought deep research was good. I mean.

— So much stuff.

— It beats other sectors.

— But honestly, it feels like a lot of resources. Perplexity is way more concise. It’s not even close in length. But look at the sources it uses. Perplexity picks better sources every time. I’m asking the exact same prompt. I’m using the same prompt. I started testing Grok too. Grok is looking really, really good.

— Grok is solid. I used it for a task yesterday. I asked it to find input-output token prices for different LLM'lerin. The results were great. It nailed it on the first try. It went straight to the right spots. It checked each LLM provider's specific pages. I was asking a few things. I wanted it to fetch LLM'lerin input-output token prices so I wouldn't have to. I liked it. Especially the reference lists. The whole point of deep research is picking the right references.

— Exactly. It’s all about the references and sources. Once you pick and read, it’s all there. The key to deep research is finding the right sources.

— Totally. Grok is great at this too. But don't just read OpenAI's deep research. Take it and use another AI to build on it. Refine it or do something new. That’s my take.

— From what I’ve seen, O1 Pro writes the most detailed stuff.

— I’m not sure about that. I don't think you pick the model for deep research. It chooses its own model. You can even run deep research with GPT-4.5.

— I think you can pick it, though.

— I’m not positive. If anyone here knows, drop it in the chat. I’m pretty sure deep research is model independent.

— Could be. Maybe.

— We should check. I recall OpenAI published a paper on deep research. Exactly. We need to look into it. I think they're addressing it with an O3 mini fix. I've been torn between them. I've switched back and forth a few times. That was my latest take. Still, worth a look. Since we're talking model variety, do you use 4.5 on your Zeo account? Not on the ThinkPad. I check it out occasionally.

— Yeah, I tried it a bit.

— I liked it.

— I've been using it a lot lately. I'm happy with 3.7. It's fun to talk to 3.7. It can actually hold a conversation now. The responses feel much more human. For this database project. I used both O3 mini high and 3.7 to map out the schemas. Honestly, I still find Claude a bit ahead. It nails my requests on the first try. The coding side is also way better with 3.7. It writes nearly flawless code for advanced tasks now.

— I’ve used it too. I built a project in Cursor recently. Honestly, 3.7 didn't feel like a huge leap for my vibe coding workflow. But I haven't been coding much lately. Maybe that’s just me. Plus, Claude Code just dropped. Yeah, its agentic structure is getting a lot of hype right now.

— Exactly. People are already showing it off in YouTube videos with MCP and stuff.

— Totally. Someone even leaked the Claude Code prompt. I highly recommend everyone check these out. I always analyze the prompts written by LLM creators. I expect the best prompt engineering from AI researchers. They know the internal thinking of AI and their own models best. This was supposedly the Claude Code system prompt. Someone either extracted it or pulled it from the source code. Apparently, their source maps were exposed. It’s a CLI tool. Apparently, it has a source map. They open-sourced the code. That’s a nice touch. Definitely worth a look. Check out their prompts. I’m sure the engineers at Cursor and Windsurf put a lot of work into this.

I recommend this to anyone wanting to level up their prompting. I’ll drop the link in the chat. Ozan, you mentioned some new models released this week. Any favorites?

— There are two models. One is this Selen thing. I didn’t know about it. Not sure how to pronounce the company name. They have a small model for LLM'e zecacı. It’s not a massive frontier or SOTA model. They’re releasing it specifically for LLM'e zecacı. It’s even ahead of O3 mini. It’s currently beating all the top models for LLM'e zecacı. That’s the big Selen One model. The smaller one is for score-based tasks. The example is right below. Scroll up a bit. Just a little. See? It shows a simple medical system implementation. I want to use this. It’s great for self-consistency in advanced tasks.

Self-consistency is complex. No two outputs are the same. You need a solid LLM'e zecacı there. To decide which of these three is actually the best. I’m planning to use it for an SQL project. I'll give it a shot. Let's see the results. I'll test different models. I'll test different approaches too. I'll know which one is better. I've tested them myself. I want to check the results of this LLM'e zecacı.

— Great. These LLM'e zecacı tools are usually for benchmarks. Stop me if I get too technical. Or do the opposite. This LLM'e zecacı concept is for benchmarks. Or for your own evaluations. Say you wrote a prompt. You wrote three prompts for one task. Then you generated 100 data points. Now you test those prompts. You have three prompts for summarizing text without losing detail. You put an LLM'e zecacı in front of them. It does this. It checks the prompt outputs. Do they follow the system prompt instructions? Does it actually complete the tasks?

It scores them. Maybe you're evaluating all LLMs. Maybe you're evaluating your prompts. You put an LLM in charge. It says pass or fail. That's an LLM'e zecacı. They trained a model just for this. That's really interesting. I don't think there's a base model. It doesn't seem like a fine-tune either.

— That’s how I took it too. I couldn't find anything.

— Yeah, it’s not. Interesting. Honestly, that’s great news. We need more vertical models. Benchmarks just comparing foundational models aren't enough. I really want to see more vertical models. Like a dedicated medical model. Or a model for specific tasks. It’s great for our business too. As a company doing multi LLM, I’m looking forward to it. This is really cool. Let’s dive in and take a closer look.

— That’s why the DeepSeek talk was so important. If this is the path, then vertical is the way. Otherwise, how would you even train them? You’d need insane funding. You need to secure massive funding. Look at Harvey. The Harvey LLM. People in legal tech use it, but it’s not really open yet.

— Yeah, they’re the leaders in this, right? For vertical-specific work.

— You can't just pick this up and use it. Who knows how it works. But it's really taken off now. It’s a full-blown product. It’s been a while since I checked. They were still developing it back then.

— Still, I hate that "contact us" sales model.

— It must be super expensive, man.

— But the experience.

— It gives off that "don't talk to us" vibe immediately. Very bold. They raised a ton of funding. They’re bound to ship something solid.

— Yeah, they’ve been at it for a long time. They’ve been an OpenAI partner since GPT-3.5. Frontier models always launch with these companies first. They play a huge role in evaluation and labeling. It’s super successful. Exactly. They got startup funding. They’ve been at it since 2022. They built a solid product. I wish we had cool AI products like this here. It takes these companies forever to support Turkish. A friend of mine is working on something similar. I won't name-drop since it’s not public yet. I poked around it a bit. It was doing cool stuff with legal precedents.

— That’s a great mindset. That sector in Turkey is a total mess. It’s so tech-dependent and critical. Yet, it’s impossible to get real work done. You can’t handle legal tasks properly otherwise. There’s insane documentation. But it’s so disconnected from modern tech. The interfaces are awful. I talked to a lawyer friend about it. You wouldn't believe the systems they use. It’s painful. Your eyes would bleed.

— By the way, let me say something controversial to boost my engagement. With these models, practicing law is getting pretty easy for everyone. Anyone can act as their own solid lawyer now. If they’re a good AI user.

— It’s like that. Accounting became so easy with tools like Paraşüt. Writing petitions or legal notices wasn't that hard. Templates were already online. Now, we see it everywhere, even abroad. People are getting parking tickets canceled using ChatGPT. You see these stories all the time now.

— Yeah, I see these cases constantly. By the way, legal chatbots have really improved. This wasn't the product I mentioned. Someone else built it. It’s great to see, though. I read a lot of comments online. Text is most valuable in two fields right now. Law and coding. These are where LLM outputs provide the most value. It’s an interesting and promising space. Finally, you had a second model to mention. I haven't checked it out. Since it's Cohere's new AI, I assume it's open source.

— Exactly. It’s a vision language model.

— That’s awesome.

— It topped the LLM benchmarks. Especially the largest model. The smaller one performs great against much larger models. I’m doing a vision project with a friend soon. We’ll label a massive amount of data. We’re captioning for Flux training. We planned to use a Vision API. We’ll likely try this instead.

— That’s cool. It’s an interesting niche. The trend is consolidating power into a single model. Multimodal is the standard now. I thought specific vision models would be obsolete. GPT-4 vision preview merged everything into one. But Cohere excels at multilingual intelligence. That’s a huge deal for countries like ours. What can this model actually do? Give it a chart and ask who won. It extracts insights instantly. Show it Chinese text. It identifies it as product packaging. It’s great for specific use cases. Like identifying rare coins. Or diagnosing sick plants. It’s perfect for image-plus-text tasks. It works best for short, punchy interactions.

Thanks. I missed that. It might come in handy. I feel like inference companies will include this soon. Multi-model setups are trending right now. I want to talk about workflows. Everyone has their own unique style. I work closely with you, Serkan, and Can. You guys are actively chasing AI. Everyone has a unique workflow. The bad part is nobody shares it. There is just too much information out there. You do something in Cursor or N8n. Nobody talks about it. There are tons of videos abroad. I love catching these insights. It provides a great perspective.

Elvis says he prefers GPT-4.5 for brainstorming and planning. It has high parameters. It’s a model with trillion-level parameters. I don't think there is an official announcement yet. Let us know in the chat if you know more. Larger models understand text with more depth. Think of it as depth increasing with the billion count. That was a bit superficial. He uses 4.5 for brainstorming and planning. I used it on Zeo's pro account. It didn't hallucinate with large contexts. I thought it would get mad with 20k-30k tokens. I got great insights. I used Gemini to turn my deep research on ThinkBuddy competitors into nested lists.

I organized it into nested lists. I fed those lists to the LLM. It gave me pricing advice for my company. That’s the analysis I did for POE. I use nested lists a lot. It’s the cleanest way to feed data to an LLM. I handle messy data this way. I had a huge, messy deep research output on POE’s pricing. I start by taking that big output. Then I simplify it using GPT-4.5. I want to copy and steal its style. I take two of those and simplify them. Then I go to AI Studio or ThinkBuddy. I have Gemini 2 Pro do the work.

But I replace the Gemini 2 Pro responses with the 4.5 ones. LLMs work on next token prediction. So, the output mimics the examples. I use 4.5 to generate high-quality examples. You wouldn't want to see the original document. It was ugly and hard to read. If you can't understand a document, the LLM won't either. I used 4.5 to clean it up. I’m analyzing 35 other companies just like this. I want high-quality output when I combine them. If I feed 30 or 40 documents, the LLM might crash. Consistent structures are key.

It’s like paid subscription tiers for all competitors.

— I think that’s a big deal.

— Comparing them becomes much easier this way. Think of it like context optimization in prompt engineering. The better your input, the better the AI output. We’re jumping through hoops to get what we want from AI.

— Sometimes it bums me out, Yiğit. We get so structured. I feel like it kills all the creativity. It just turns into a parrot. Even with a top-tier model. I keep scrolling up to tweak my prompt. I end up forking the chat entirely. Eventually, you need a fresh approach. That’s why I hate long chats. I prefer starting a new chat once things get too cluttered.

— You’re totally right. I’ll add this. I did something fun and interesting this week. I want to show you guys. I’m moving to the States soon. To San Francisco. I need to grind my English. I have to get it to a pro level. My wife and I decided to learn by watching shows. I watch with English subtitles. It helps, but I want to catch the idioms. Ozan, you've been in London for four years. I bet you still don't know tons of idioms. It’s a tough topic. I use AI to build Anki cards for them. For example, look at "Dandy." I did this for a show I’m watching.

It’s a rare usage. Yiğit, it’s not the end of the world if you skip it. I generate these cards from the SRT. It pulls advanced English vocabulary for me. I don't know this one. I couldn't even remember it. "Dandy" is an archaic term for a flashy or vain man. AI didn't give the best translation. You can mark these as again, hard, good, or easy. Med students use these Anki cards all the time. Anki cards are standard for memorization. I practice with them constantly. I finish the flashcards before watching the show. Seeing the words in context makes them stick much better.

It’s fun. I sat down last night. I thought, how long would it take to build this? It turned into a product in 13-14 minutes. Should I tweet it? Nah, why bother. Maybe nobody would care.

— I think it's great. I'd use it too. Tweet it or push it to GitHub. I'm in.

— Using it doesn't matter, Ozan. Would you pay 5 bucks? That's the point.

— If it's 5 bucks, I'm in. I'll pay for two months.

— Let's give you 10 SRTs for 5 bucks. If you use them up, just top up. That could work.

— Once money gets involved in learning English, the math changes. You start comparing it to Open English or other services. You weigh your options. We can write great programs for ourselves right now. Even without using standard apps. It just takes a little effort.

— Yeah, it’s wild. AI is at a point where you can build your own tools. You can customize how you learn English. It’s a cool detail. Let’s move on. What else is on the agenda? I mentioned Cloudy Code is open source. It’s one of this week's interesting stories. I’m going to check the code. I’m curious how they built their agent. Maybe we can discuss it later. Checking if I have any other tweets. Sam Altman’s.

— Sam Altman had another wild tweet thread. Exactly. He’s talking about AI consciousness. He says consciousness is inevitable. Sam Altman loves saying creepy stuff like that.

— Yeah, it’s become a hobby for him. He just posts these random thoughts. He’s totally lost it. He’s trying to show off 4.5. Everyone is obsessed with the vibe right now.

— Wait, wouldn't 4 say any of this? I don't think it talks to us this way just because it's 4.5. That wasn't my impression.

— You start noticing that vibe at some point. It feels surprisingly human. I’m a certified GPT-4.5 hater. I hate how expensive this model is. But I have to say, I find Sam Altman a bit reckless here. He should be calming people down about AI. Instead, he’s making them more anxious. It’s a weird mindset. A very interesting take. I added it to my list to roast this week. Let me check my bookmarks for the final round. Maybe we can look at this next week, Ozan. Google has a new data science agent. I don't know if anyone here has used Julius AI.

— I tried using 4.5.

— Really?

— Let me jump in here. I gave a simple task to OpenAI, Julius, and Claude 3.7. It was an Excel file. About 8-9 columns and 50 rows. I asked for a report. Claude really crushed it again. It even built React components. It generated artifacts from the charts instantly. I have to praise Claude again. I sound like a total fanboy. But hey, what can you do?

— I’m the total opposite. It’s sad for the industry. Google entering the space and using Gemini freely is tough. They’re fine-tuning custom models. It’s going to make things much harder for everyone else. It’s frustrating for smaller companies. Another open source Perplexity killer just popped up. But they all just rely on a SERP API. Building Perplexity isn't that simple. I think Perplexity leans on that stuff too. We’ll find out eventually. Maybe our own indexes.

— I saw something else, too. It was more about visualization. Let me find it and send it over.

— Sounds great. Like you said, I’ll look into deep research implementation next week. Maybe I’ll build something with it. We said we’d talk about MCP, but let’s save that for next week. I've been deep into MCP clients and agents lately. Maybe we can pull something from that. We could talk about use cases. I think that covers the agenda for this week.

— Okay. Same here. Let's dive deeper into MCP next week like you said. It looks simple on the surface. It’s basically just tool calling. But what can it actually change? Why is everyone so hyped about it right now? We'll talk about that a bit.

— The industry goes through these phases.

— It has its own hype cycles, yeah.

— That happens too.

— You can't use it on OpenAI, for example. MCP isn't there yet. OpenAI doesn't support it.

— Yeah, it’s not really effective. Peter Levels had some interesting work this week. He did it with Claude.

— It’s one of those things that makes me question everything.

— It really is.

— I mean, network is everything. Sure, he had plenty of successful projects before building that network.

— Exactly. We respect Peter Levels, but that project was honestly meaningless.

— It’s not really about that. It’s about where things are heading. It’s about the community’s evolution. Attention is still everything. It’s about gathering people in one place. Someone recently asked ChatGPT if being smart and hardworking is enough. It was circulating everywhere. I even saw a post on Instagram. It was a screenshot of a ChatGPT response. It said no, that’s not enough. It claimed who you know matters more. It’s true in a way. We can see that here too.

— Prompt engineering is crucial. I was on a streamer's show yesterday. They took a ChatGPT hallucination as fact and totally spiraled. How you phrase your question really changes the answer. It dictates the quality of the output. People need to broaden their literacy here. It’s a critical topic. Everyone needs to master these tools. They’re becoming deeply integrated into our daily lives. There’s so much to build. MCP is a huge part of this. We’ll do our homework for next week. Maybe we’ll show some things from Claude Desktop. We’ll be using those features on our ThinkPad'de.

I need to put some work in there too.

— A good excuse for you.

— Exactly. Totally. Okay. Thanks to everyone for joining. Take care. See you later.

— See you. Bye 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