Grow Your Business Exponentially with LLMs

Leveraging foundational Large Language Models to automate operations, elevate user experiences, and scale revenue.

Speakers: Yusuf Hilmi

Content author

22 Nov 2023

Video Transcript

⏱ 19 min read• 3,349 words

Read the full speech or switch languages

Hey everyone, let me jump right in. I wanted a simple, clean presentation. Let's talk through the headlines as much as possible. How do we actually grow the business? How do we speed it up? That's what I want to cover. So what do we do? At SearchBase, we help high-growth B2B SaaS startups build their own AI copilots. What we mean by AI copilot is basically a copilot with a natural language interface. A user shows up. They say what they want to do. They get the output they wanted. Shopify shipped this inside their product as Sidekick. We were already doing this inside our own product. And we realized a lot of startups and companies without their own AI team would want this. They already do.

So we started working on it fast. It brings a few things to the table. The first and biggest one kicks in the moment it's integrated. Once a user says what they want inside the product. We don't have to stare at events in analytics dashboards to figure out user motivation. Just reading what they asked for is enough. Our product managers, founders, and especially the exec team can decide where the company goes. On top of that, the biggest thing we give these users is less friction. You can solve every problem in the user experience at some point.

If you can take a single input from the user and run the action they wanted. So where does this translate? Where does it actually get used? Especially in BI tools and analytics tools. Most B2B SaaS products have a reporting feature. That reporting feature usually eats 15-20 minutes. Sometimes a full hour. And if there's a bug in there, it can burn hours. Our fix is simple. We take what the user actually wants. Then we hand all of it to AI'a on the back end. It figures out what the output should look like. We integrate that into the product. And the report just happens, right inside the product.

That was a quick intro on what we do. Now let's talk about the real thing. How do you grow your own business faster? And how do you scale using LLM'leri kullanarak? Here's what I always recommend. I hammered on this with Yiğit too. I don't treat LLM'leri like the old AI that never worked. I treat them like an actual person. You have to look at them as an intelligent being. Once you see them as intelligent, the second point kicks in. Say I hire someone new today. Maybe we'll work together on a marketing function. I have to pay them. And usually I'm paying by the hour.

But with LLM'lerle, I can run about 100 interns at once. And compute is so cheap. The amount we pay LLM'lere right now is tiny. Especially for the intelligence and reasoning capability we get back. We use all of this inside a company. We're a small shop, by the way. Two people. But for a year straight, no breaks. We performed at the level of a 20-30 person company. And there's one reason for that. From day one, we kept going LLM'leri kullanarak. Let's walk through these core business functions. I'll talk about where we actually focus. And how we use it. I'll start with sales and marketing. Most of the crowd here works on that side anyway. I'm terrible at that side, honestly. Genuinely bad.

We're trying to get better. And we are getting better. But me being bad didn't mean we'd be bad. Using these LLMs, first and foremost, I got one real thing out of it: knowledge. And along with the knowledge, it handed me the path forward. Most of the strategy, really. You just need to reinforce it. While doing that, my first tip is this. See the LLM as a person of average intelligence. Then go talk to actual experts. If you've found one or two people who are great in their field. Like Yiğit, send them your questions. Then you can focus on optimizing the LLM.

I have a few practical tips here. First one: personalized outreach. The importance of ChatGPT was really this. An interface that passes the Turing test, a model, went public. And it became the fastest-growing product in the world to date. So you can already do this in your own sales and marketing. There's one thing I love about it. Normally you'd have a writer produce a piece of content. Then you'd want certain variations. You send revisions back. And those revisions can bounce around for two or three weeks.

But with an LLM, right now, going through 100 variations myself takes maybe 20-30 minutes. Or less. And I can hand those 100 variations back to the LLM. Ask which one should go to which audience. Then target accordingly. This slide keeps jumping ahead, sorry about that. And the third point is this: on translation, we've used APIs so far. They helped in some spots. In others we failed hard. At my previous company, translation was our biggest problem. To solve it today. Almost all you have to do is make a call to OpenAI's API. Or use other APIs. Or run translations through your own model. One you build and deploy yourself, if you have an AI team.

And it doesn't stop there. Let me add a great tip I picked up from Yiğit. While you translate, think about who the audience is. Fold in that country's social and cultural context. That makes the translations much stronger. So it's not just translating the content once. You have to adapt that content to that audience and that country. Product is exactly what we focus on. I talked about friction inside the product. I talked about problems inside the product. And we talked about a natural language interface. A user shows up and says "I want to do this." And it just happens.

One of the first places this really pays off is onboarding. Yes, we're a SaaS company. But this applies to every company that isn't SaaS too. During onboarding, users have specific problems and specific questions. Right now, most companies handle it like this. If it's an important customer. I hand them off to one of my account managers or someone on support. Then I spend about 30 minutes, an hour, sometimes two. Sometimes it takes several separate meetings. Very few companies are working on optimizing this today. There are plenty in San Francisco, especially across the US. In Turkey, sadly, I haven't even seen the beginning of it. Everyone keeps working with the teams they already have. And a huge cost line builds up right there.

Cost-wise, you're constantly hiring more people. I met a startup in Turkey hiring 5 to 10 account managers a month. The only reason they do it is simple. There's no system that explains the product. And the actions inside the product aren't obvious. Neither is the value the user gets from it. So those meetings have to keep happening. There's a way we solve this. We provide a natural language interface. But at the core, we talked about an intelligent entity. Picture an AI that knows the product and knows the company. It has seen and solved the company's problems before. It has access to every other problem doc the company has. Just asking that AI solves most users' problems.

It works in Turkish too, not just English. I built the deck in English without really thinking about it. But if you're using OpenAI'ın GPT-3.5, 4 or Anthropic's cloud models. These models handle multilingual really well. There are even models you can fine-tune yourself. If you have an AI team, you can go deeper with those. Catch me later and I'll walk you through the details. Second point: conversion rates. In SaaS, that conversion rate is one of the biggest things for us. How many users show up, and how many do we turn into paying customers? We talked about time to value earlier. You have to minimize that time to value. It has to get so short that the user hits value the moment they open the product.

Say I pull a user off my website into my product page. If I make them wait 3 or 5 minutes, I'm in trouble. Or if they land in my UI and spend 10 minutes figuring out what to do. Odds are very high I lose that user. I lose them 80 or 90 percent of the time. So you have to run this tightly coupled with sales and marketing. A user should already have some context about the product before they arrive. And someone needs to be there to help them continuously.

You could hire a team of 300-500 people while growing fast. Or you actually use AI'ı. You cut that cost line by 1000x. Maybe 10.000x. And you solve it with one well-built tool. Closing that knowledge gap is one of the biggest wins. Anyone working in UX knows this cold. Inside a product, there's a gap. What the user knows, what I assume they know, what they need to know. Close that gap and the user sticks with my product. That user comes to my platform. They take the actions I want them to take. Those meetings I mentioned? They happen to close that knowledge gap. Content gets produced. Tons of it gets shared. We keep trying to land in front of the user across different platforms.

But once that user arrives, almost half of them. Per our current surveys, up to 50%. Either reach out to support or ask a question. Or 80-90% of them search on other platforms. They go to Google. They go to Perplexity. They hit other channels to solve their problem with the product. Operations. At my previous company I worked on the ops team. Working in ops is genuinely tough. Because you have to keep tabs on everything. I came in as an engineer. And as an engineer I picked up design, marketing, product API'ları, the backend, the frontend. We had no product manager either. So I was taking on a lot there too.

When I did this, I didn't have ChatGPT. I remember my pace back then. Handling all those functions took me a real week or two. Work that filled two weeks, I can now compress into about a day. Especially after GPT-4. Yiğit will cover those tools in a minute. But for your own ops team, here's the approach I'd take. Say I have a function in my business. Say I want to improve it. Don't ask how I use AI'ı. Ask how I'd solve this right now. What people do I need? What resources do I need? Then ask the real question. How do I get AI'a to make the calls that person would make? Those are the decisions a professional makes. Ask that, then chase it.

Because once you chase it, that smart entity can actually make those calls. Give it the right inputs. Make it use the right information. It really does decide correctly. I wanted my deck as simple as possible. But there's a paper. I'll get it shared with you somehow. Here's what that paper covers. They swap a company's CEO out for GPT-4. They describe the function to GPT-4. Then GPT-4 runs the team's decisions and management for a week. Nobody on the team is told. They tune the writing style. After that, they see almost no difference in performance.

They actually see a certain improvement. Because response time gets much faster. Normally a CEO takes an hour or two to reply. Here everyone gets that response in 30-40 seconds. A minute. Two minutes. So operations really means sharing information. When we first set out as a search base. We were focused on the knowledge-sharing problem inside teams. Shortening that knowledge sharing, speeding it up, is genuinely possible right now. And you don't just speed it up. You can accelerate it exponentially. Like I said, we're talking about a system that communicates with people. Right now you have to empower people. As these tools get used, we'll start seeing people who are 100x, 1000x more productive. There was an essay called "Where are all the hyper productive people" by an investor on the A16Z team.

I'd definitely recommend reading it. That article covers a lot of this. How these tools affect productivity, and where that productivity can go, is critical. We always talk about top performers. They stand out a lot on ops teams especially. People who perform at the top of a team show themselves clearly. We used to call them 10x or 100x developers. But now it's reached a point where this can go to 1000x, even 10,000x. And the only thing making that possible is access to intelligence.

Intelligence being accessible at a much cheaper price. Along with that, yeah, you can use AI'ı like an intern. Hand it every task you don't enjoy. That's what I do. Sometimes you have to step in and check. But I'm an engineer, after all. I love product. But honestly, I'd really love to focus on code. There's a maker schedule and a manager schedule. My calendar has filled up with meetings lately. Especially after we started growing. But to actually build a product, you need 4-5 hour blocks. You need that kind of focus. Creating that block is nearly impossible with a packed meeting schedule.

So I use AI'ı like an intern as much as I can. Where I owe a response, I try to go async instead of taking a meeting. Same thing inside the team, when I'm leading the team. And customer support. I've said it many times. How much cost customer support can create isn't obvious early on. You figure you know it, you can throw money at it. But once growth picks up, that cost line starts climbing fast. And it doesn't scale the real problem.

You can't manage a 1000-person support team well. Not even if you're a huge enterprise. It breaks down every time. User experience gets worse. Satisfaction drops hard. The way to fix that is to go human-assisted. I'm not saying replace people entirely. Here's what really matters: it'll be hyper productive. Give these tools to the people on your team. Using them, they'll perform 100 times better than they do today. Making those optimizations inside the company is the real point. And one of the critical things is feeding it the right data. Making sure it learns the right way. Right now, to run an LLM'i, you open ChatGPT or any other tool. You have to give it a prompt. You have to explain what you want it to do.

At some point you have to treat them like they have Alzheimer's. Nothing you didn't give them exists in their memory. This is one of Yiğit's examples. I love it. You really have to talk to them like you're talking to a person with Alzheimer's. Because it forgets everything you didn't give it. And it won't just forget. It'll start making mistakes. You'll see the problem we call hallucination. It starts inventing facts that don't exist. To prevent that, it has to consume that information very clearly.

Even if you have privacy concerns early on, you can run experiments through APIs. Later you can point your team toward building and training your own models. But it takes serious know-how. In Turkey, especially in this sector right now, that hasn't formed yet. In fact, in Turkey right now, outside the CEO. Few companies have anyone pushing AI'ı forward. There's something else missing too. Culture. On the AI side, San Francisco runs a hackathon every weekend. Every weekend engineers, product people, and designers show up. And they build something new.

And every week a new technology shows up that grabs our attention. We don't have that culture yet. We made a small start last weekend. I'm really curious where it goes. Still, like I said, the key thing is solving the human-assisted side. For operations and every scenario where you'll use AI. There are things AI can't do yet. If it could do everything, we'd probably be out of a job. So we have to position our own team right where it falls short. Position that team correctly and your productivity and output go way up. And like I said, you have to treat it as an intelligent entity. I'll hammer on this over and over. Because Sam Altman had a phrase for it.

He said intelligence on demand for OpenAI'ın API and GPT-4. Intelligence on demand. An intelligence that serves me whenever I want. An intelligent entity. Calling it an intelligent entity may sound high level. But right now, when I hire someone new, a junior. Onboarding them takes me a month. Explaining how my company works, all of it. With a fast learner it still takes a week. The questions they'll ask during onboarding and the steps we go through eat up my time. By the way. Take the juniors I'd hire onto my current team. I power them up with AI. I build my own AI knowledge base properly. I prep my prompts. I set up every tool my team will use. Then I can onboard that junior fast.

Today I can set this up instantly inside my team. And as we go, I can automate it. We already talked about teaching AI'ı. When you train a junior. Most of what you'd teach AI'a is data you're already producing. If you approach that process treating the LLM'i as an intelligent entity. You can teach it like a person. Just understand how it works first. Then teach it the same way. We talked about the memory problem. You solve memory with a different model type called embeddings. You store them in vector databases. That gives you exactly what you need.

It really lets you solve that intelligence's memory problem. Very simply, too. Right now it's a bit technical. We're trying to make that space easier, and other companies are too. This will reach the end user. And there are specific problems on the way there. Every company working in AI will hit optimization problems. Inference problems. Cost problems. Because getting GPUs is already hard. We talked about compute. OpenAI'ın has it. Here's why I say use the API: we can't find GPUs right now. Finding a GPU and budgeting for it is really hard. It's expensive, too. You've probably seen it. Companies in San Francisco raising a billion dollars spend almost 80-90% of it building their own compute clusters. Their own supercomputers.

So integrate AI'ı into your product, your team. Your company with small tests from day one. Over time, all the data you'll need builds up. That know-how builds up too. One problem I keep seeing with other companies I talk to: they want to jump in and do everything at once. Not possible. I got this far in AI and LLM'leri by starting with AI in 2018. Then working 24/7 for the past year. That's how I picked up the know-how and the insights. But there are plenty of companies working alongside me in this space. Plenty of startups. We all started from zero.

Honestly, when I first started, I wasn't even solid on semantic search. It was a space I'd never touched before. But now we've written an open source library for it. That library is used in over 2,000 projects. And it's growing fast. These things take time. Same goes for a company. Let's frame a team, an organization, this way. We're all organic computers. An organization, a company, is a computer too. One that moves as a collective organism. Let's treat it as an intelligent entity.

That organization has an intelligence. It takes actions. It makes decisions. Optimizing that takes time. My advice: start banking that time right now. Start using it today. Thanks so much for listening. I left my email. Write to me anytime with any question. Thanks.

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

Content author

Zeo's Guest

author

Contact us

How we handle your data

Open as a full page