⚡️ AI Talks by Zeo (Episode 3)

Building custom no-code/low-code AI agents to automate daily campaign optimization and reporting workflows.

Speakers: Yiğit Konur, Ozan Ketenci

17 Sept 2024

Video Transcript

⏱ 21 min read• 3,719 words

Read the full speech or switch languages

— We made it to another week, right Ozan?

— Yeah. You wouldn't believe it.

— Honestly, I'm shocked too. I didn't expect us to be this consistent. It’s great. This is my first long-term streaming project. I really hope we keep this momentum going.

— Yeah, let's see how long it lasts.

— We'll see. We'll find out together. The agenda is packed this week. It’s busy as always. But there are one or two big stories from last week worth covering. First up: Elon Musk’s xAI and their new model, Grok. If anyone wants to try it, we can test it together. Also, Grok hit the top of the latest benchmarks fast. That’s a huge deal. Everyone expected a big move from Elon Musk. He was actually lagging behind.

— I didn't see O1 Pro in the benchmarks.

— O1 Pro isn't in any benchmarks. I couldn't find it either.

— Okay, fair enough.

— They just don't include it. It's weird. Usually it's O3 or O1.

— Okay.

— I tried solving a few problems with Grok 3 this morning. Nothing major. But here is what I really liked. I asked a few questions yesterday. Let me check those. Hmm, no, not this one. I saw it somewhere else. Let’s check my history. Exactly. I was asking about this last night. Late last night, it hit me. There is a group that thinks AI will slash global costs. They predict a scenario of massive global wealth. The crowd saying everything will get way cheaper.

— The basic income folks?

— Exactly. Big names like Marc Andreessen are on board. He founded Netscape. They call themselves "techno-optimists." It is clearly the biggest shift of our time. They believe it will have a huge positive impact. Then there are the pessimists. They worry the economy will shrink. I threw that out there. I asked the question. Here is my question. Automation lowers production costs. We need fewer humans. Universal Basic Income might kick in. Will our living standards actually rise despite the income loss? We’ll be losing our income streams, right? I translated this into Turkish. It turned into some fun content. I really liked the result. Grok 3 felt a bit more advanced. The translation was spot on.

The Turkish translation felt super natural. It’s great for translation tasks. What caught my eye? They prioritized math and computer science in Grok’s training data. It performed better in other areas without specific training. They basically said, "We taught it math and CS. And that was enough." It was a pretty bold claim yesterday. I saw that.

— But it seemed to lag in coding. That’s what I saw in the examples. What do you think? Have you checked it out yet?

— I checked the benchmarks Westrod posted yesterday. I watched his video. It ranked first in coding benchmarks. I haven't tried coding with it myself yet.

— Me neither. I haven't tested it yet. You know those Twitter examples? Like the ones that pop up when you search "Grok 3." The ball-in-hexagon tests. It had some fails there.

— Exactly. You're right. But when I looked at it... I think there were other things here too.

— It hit the number one spot here too.

— Totally.

— Was it a shortcut? Something like that.

— It hit the market fast. That really caught me off guard. I usually track this stuff on Hacker News. I highly recommend it to everyone. Let's check Hacker News for the buzz on Grok and Grok 3. Is Elon Musk blocking us? Why can't we see Grok 3? It hasn't actually dropped yet.

— Look, Sarp Atil has a long tweet. They probably linked it.

— Sarp Atil's tweet was spot on. Exactly.

— He did some serious testing.

— Exactly. I read it yesterday. Let's touch on it quickly. Sarp Atil is a former OpenAI employee. He's a big deal in this space. He spent years at Tesla and in AI. He really knows his stuff. He has some fixed prompts. They designed a board game. They got the result they wanted. The only other match was O1 Pro. It outperformed classic O1 and R1. It matched O1 Pro’s reasoning capabilities. They tested an emoji-based storytelling task. It involved hidden Unicode characters. It failed that specific task. Then they uploaded the GPT-2 paper. They asked a really challenging question. It performed on par with O1 Pro again. It was a classic "vibe check."

It’s all about how the model feels to use. Sarp Atil didn't have any negative feedback. He seems to think it’s quite successful. It’s going to be an interesting time. I’m excited. I hope they release an API soon so we can tinker.

— Something new will drop right after. It’s like the DeepSeek situation. Grok 3 is out now. OpenAI’s plans are still unclear. They’re talking about a unified model.

— Yeah.

— Anthropic will likely announce Claude 4. The current "best" model could drop to third place. It’s moving so fast. Everyone is pushing each other. It’s a fantastic market right now.

— Totally. Let’s list the pros and cons as bullet points. Hacker News has a sharp crowd. I think we can get some real insights there. Time for a quick Gemini plug. Gemini 2 Thinking is my go-to model right now. It’s seriously impressive. That million-token context window is a total game changer. It makes everything so much easier. Especially for deep analysis. We’re about to feed a 93k token discussion into Gemini Flash. Let’s see how it handles it. We’ll catch up on the latest news while we’re at it. Let’s make it a "last of this" summary.

93k tokens is the limit for most models. That’s basically a 70,000-word document. We just dumped the entire Hacker News thread in there. This is how I consume content now. It’s great, right? No need to read every single line.

— Exactly. You’re actually consuming the content. That’s how you do it right.

— It’s perfectly bite-sized info. Thinking and Flash both performed great. Let’s not overcomplicate things. Let’s stick with Flash. It preferred writing in Markdown for some reason. Hacker News had a hot take. They claimed it outperformed Claude on coding tasks. Let’s see how xAI competes with tools like Cursor. I hope it doesn't just stay a Twitter-centric LLM. There are a few benchmarks out there. Specifically LM Arena and LMS. People are criticizing them for potential data manipulation. There’s also a lot of heat on benchmarks lately. The test data is public knowledge. You’ve probably read that model companies train on these exact sets.

— Yeah, exactly.

— It’s a real issue.

— I think OpenAI is developing a new benchmark.

— Yeah.

— They took 1,400 software jobs from Upwork.

— Praska?

— They tried solving them together. They published a solid benchmark. Claude...

— Yeah. What was that? I didn't expect that at all.

— I know. Claude is outperforming O1, if I'm not mistaken.

— Right. You mean the visual right there, right?

— Yeah, exactly. Spot on.

— Yeah. It’s really interesting. I didn't expect that much. Bold moves in a cutthroat market.

— Remember the benchmarks? They were over 60% on coding tasks. That’s how it looked. But here, with current tasks, it’s not even hitting 50%.

— Yeah. It’s fascinating. This transparency is interesting too. I looked at the benchmark. It has many different tasks. The scoring is dynamic based on market pricing. I solve a coding problem, and the score shifts based on its market value. Harder problems pay more, so they impact the benchmark more. It’s an interesting method. I wouldn't have pitched it, but it clearly works. They covered everything from bug fixes to major feature implementations. You can't just sit down and read this paper. I'd just drop it in AI Studio. Then I'd scan it. It's weird it's only on Twitter Premium. The EU access issue is basically fake.

Just log in with a US VPN once. You'll have access even from an EU IP. I found that while poking around. It's a really interesting point. Elon Musk criticized OpenAI for not being open source for so long.

— And then he dropped Grok 2 as a joke. He open-sourced Grok 2.

— Oh, really? I had no idea.

— Yeah. He said he'd open-source Grok 2. Maybe he did. But why not release the state-of-the-art model? If you're going open, just go open.

— Honestly, only the Chinese are doing this.

— By the way, Sam Altman picked O3 mini. Did you see that poll?

— I did. It’s wild. Sam Altman actually said it. It’s been ages since he mentioned open source. He’s asking if they should open source O3 mini. Should they open source an O3 mini level model? Or one that runs on GPUs or phones? Apple might get a solid AI without lifting a finger. After all their AI flops, that’s crazy.

— I just remembered this. Did you see that new lamp-like robot from Apple?

— Nope.

— They’re probably leaning into hardware. You know how it is. Like a smart lamp or a robot lamp. Apple even has a paper on it. Let me find it.

— Is it the Pixel Insight one?

— Exactly. Yeah, that’s it.

— Oh, I’ve seen this. It talks about how smooth the movements are.

— Yeah, it does everything. It can project your homework onto the screen. It can help you draw. You can even watch stuff on it.

— Is it a projector too?

— Everything. Not just a lamp. It sees, reads, and speaks. It can read your text and convert it. It can even query an LLM on the fly.

— That’s wild. I’m genuinely surprised. The AI world is usually just constant competition. We rarely see new hardware products. Human was one of the few unique ones.

— And that’s gone too.

— That was brutal. I even tweeted about it. It felt like they wrote that line just to roast them. They’re basically burying Human. It was supposed to be an AI pin for your chest. It projected stuff. It listened to the room. It was meant to trigger AI actions. Honestly, I wasn't a huge believer in it.

— But it was flashy. It looked cool. I actually had some hope for it. Maybe something good would come out of it.

— They gave up way too early. Take Rabbit, for instance. Was it Rabbit R1? That weird product we all mocked. I expected more of these to exit the market. They faced so many crises. People criticized it for just being an Android app. It was just running Android under the hood. I thought it would last longer. It’s a shame. HP bought them out. Now they’ll handle AI for printers and conference room gear. Honestly, it’s sad. I mean, they got their exit. I doubt they’re unhappy. It’s an interesting ending. Apple’s robot reminds me of this. If Apple commits to it, it could actually be a cool product.

But it doesn't have any grabbing functionality, right?

— Nope. These are classified as non-humanoid. They’re leaning into non-humanoid robots. The research points that way. They even published a deep research paper. If Apple really pushes into hardware, maybe... Didn't Mark Zuckerberg say they haven't invented anything since the iPhone?

— Yeah, exactly.

— It kind of built on that. Maybe that’s their goal. I don't know.

— Honestly, I’m ready for a new invention. I’d love to see one.

— Yeah, seriously. Everything is still just at our fingertips. AI doesn't feel tangible in our daily lives yet. When will that happen? I had high hopes for things like Humane. The demos looked great. But the final product was pretty bad.

— It’s not even grabbing attention, Ozan. Think about the Advanced Voice Mode. It made a huge splash. I’ll drop some links and explain basics as we go. Some people know, some don't. That Advanced Voice Mode is what OpenAI released a few months ago. It talks to you with a super natural voice. Have you seen any standout use cases for it? We keep trying to humanize AI and give it a voice. But even when it’s done, what’s the point? I tried using it as a psychologist. I even tried it as an English coach. But we just couldn't connect.

— It just doesn't work. You're stuck in a specific context. You're trapped in an interface. You lack a physical presence for real interaction. It’s just a coach or a psychologist on a screen. You end up doing things like vibe coding.

— By sharing your screen.

— If it were on your desk, or helping you cook. Or if it were a robot, that changes everything. None of what you mentioned actually translates to the physical world yet.

— I get your point. There are products like AI Studio'da or those AI friends. Honestly, beating the smartphone is a massive challenge for all of them. It feels like there's no real demand for another device. We don't want to carry more gear. The real shift will likely come with robots. I almost forgot to open this link. Let me pull that up now. I really like how Grok handles research. It cites academic papers across different fields very quickly. I hope it's not hallucinating. I really loved this specific term.

— Robot shepherds.

— Robot shepherds. When I code in Cursor, I feel exactly like a robot shepherd. They covered this well. After Universal Basic Income, we'll have global prosperity. Most people won't be working. I wonder about the workers. If people get paid well without working and live comfortably for free. What will motivate them to work? Where will that drive come from? They handled the worker motivation section well. I love how deeply Grok thinks about this. It’s a very abstract topic. It gets pretty philosophical. For instance, it said this.

If goods get cheaper, workers chase luxury perks. Status still motivates people. A plumber with AI-repaired homes might flaunt a handmade sports car. Flaunt it. I love that idea. It also mentions human competence. Working in a leisure world could be a badge of honor. These are deep, beautiful perspectives. I love discussing these sociological issues with AI itself. It gave me the feeling of debating a real, qualified person. I felt that strongly with Grok 3. Its Turkish translation was incredibly good. I started the discussion in English.

I switched to Turkish late at night. The output was high quality. I recommend it for these vibe checks. It’s very useful. Don't think you can't use Grok 3 in Turkey. Just get Premium and use a VPN once to access it. Any new agenda items, Ozan? We said this week wouldn't be too busy. If you have topics, let's dive in.

— I'm checking too. Same here. There was that Karpathy tweet. O3 mini going open source. The HP and Humane partnership. Grok. I'm looking for any cool new papers.

— While you look at that, let me remind everyone about AI Studio'da. Let's not just keep pushing ThinkPad. I've been doing some cool stuff in AI Studio'da lately. I'm building an agent for my own work. It writes project docs based on a massive code stack. I'm hitting the token limits right now. I feed the entire source code of a repo to AI'ın. It's our DevOps repo. Then I add my own custom thinking process in between. Let's get to the response section. I'm iterating on the thinking part a few times. I ask different models and merge their outputs. You can do some serious thinking few-shot here.

What does that mean? You have total control in AI Studio'da. You can edit system prompts, messages, and even AI'ın's responses. That's huge. We have it in ThinkPad too, but this UI feels extra smooth. You can edit a previous response as if AI'ın wrote it that way. Since AI'ın tends to mimic its own style, you can steer its thinking process. I use this to define a specific way of thinking first. Then I finish the thinking phase and move to documentation. I've been getting great results. I'm hitting 942,000 tokens. A single repo can easily take up 400,000 or 500,000 tokens.

I realized we can document our undocumented legacy projects this way. It was a cool use case. By the way, Gemini 2 Flash Thinking is totally free. I highly recommend using it more often. Just wanted to mention that.

— Can we do direct deployments in AI Studio like we do in LangTail or Latitude? Can we deploy the whole session?

— No, unfortunately not.

— So you have to build it from scratch if you want that.

— Exactly. You mean passing variables and sending data just like in Latitude, right? Like we discussed last week?

— Actually, no. I'll just leave the user at the very end.

— Sure, you can set that up however you want.

— I see. It looks like we can grab the whole session via GetCode.

— Exactly. That's what I do. GetCode pulls the entire session. Like this.

— I'm basically porting my few-shot over to production.

— For sure. But it's not like OpenAI. It’ll take a bit more work.

— Got it. I get it.

— It can be a bit annoying. It causes issues sometimes. Here is how I handle it. I usually save it like this. I go here. I add the source code. Run it and you get the output directly. It’s free with these two models. Gemini 2 Pro and Gemini 2 Flash Thinking. Both are absolute beasts. Just wanted to point that out. I checked one last thing. Let’s wrap up with this. We’re pretty much done. I sent the Reddit links to ThinkPad. People and Advanced Voice Mode. Look, they wrote this well. They’re finding it hard to find a reason to use Advanced Voice Mode. I had the same question. What are people doing with it? What can you do with Advanced Voice Mode?

Role-playing fans will have a blast. They enjoy a lot of things anyway. Interactive games are cool. That could be fun. Language learning and practice. We need to look at the use cases. I’m curious how they’re prompting it. Prompting by voice feels a bit weird. They mentioned recipes. Nice. I couldn't start the garbage disposal at my place in the US. I opened Advanced Voice Mode. I turned on the camera and showed it under the sink. That was a cool idea. Meditation is interesting. That could work. Sports and exercise are great too. You could chat while working out. Or turn the camera on for squats. It could give you feedback on your form.

— It definitely could. Exactly.

— Yeah, that’s an interesting scenario. Like you said, connecting Reddit or web data to LLMs is a great example. I wanted to show a demo of that. I couldn't find exactly what I wanted here. I didn't find a great output for language learning. It just said talk normally. Or scenario-based chats. Debates or negotiations. Pronunciation practice. That could be interesting. Yeah, it’s cool. There aren't that many unique use cases. Seriously.

— It hits a wall eventually.

— It’s great if you have kids. You can just let them talk to AI. Think about those kids who ask endless questions. "What's this? What's that?" You can just tell them to go talk to the AI. It’s a healthy way to handle it. I think that’s a wrap for this week, Ozan. Sorry for dominating the show.

— No problem at all. I’ll send you a few links. If anyone is diving into the operator side, please try these out. We’d love to hear about your experiences. I just sent you a link.

— Where did you send it?

— Right here in the stream. Exactly.

— Just send it over Discord.

— Sending it now.

— If it's not too much trouble.

— No problem at all.

— I'll send it in the background.

— Along with this parser. Sent it. OmniParser v2 plus GPT-4 has the highest score. It's leading in Computer Use Agent. We should start looking into these. Some tasks aren't automated because they just can't be. Most are web-based. You have to log in and do a ton of manual work. That's a problem. About 30-40% of marketing tasks are like this. You can't automate them with just LLMs. You definitely need a computer use agent. If anyone here is experimenting with those, please take a look. We’d love to hear your feedback.

— I'm curious about OmniParser. How does it actually work? It probably labels screenshots dynamically to find clickable elements. Exactly as I thought.

— Exactly, man. Yeah. You can think of it as a fine-tuned model. It's basically just another model.

— I think it's way better. It's more controllable. It's not a black box. You know how it works. You know why it did that.

— Plus, it's more stable. You'll always know exactly where it's going and what it's doing.

— Wow, that’s great, Ozan. I had no idea. I hadn't checked it out. I saw OmniParser but thought it was just OCR. Like those classic document parsers. It’s impressive. Identifying clickable elements on mobile is way easier now. Running mobile agents on a mobile OS makes total sense. If a use case pops up, we should test it. Maybe we can demo it in the coming weeks.

— Exactly. Let’s keep it in mind. Even for a small project. We’re currently working on internal automation. Can is focused on that too. The whole team is brainstorming what we can automate in our workflow. Maybe something will come out of that.

— My camera looks so dark. Sorry about that. I didn't notice until now. I checked for questions, but there aren't many. Great. Can we call it a wrap for this week?

— Let’s do it. See you all next week.

— See you next week. Take care. Bye for now.

— 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