#19: The "Office Suite of AI"

Bindu, the founder of and CEO of Abacus, is one of my favorite people to follow on social media. She's always getting in there, punching up her her weight for open source AIs and talking about all sorts of topics from politics to AI. So it's a real pleasure to have you on today, Bindu.

Great to be on. Very nice to meet you again, Robert.

For people who don't, who are you? That's my favorite first question. Just to let people introduce themselves.

Yeah, my name is Bindu Reddy, CEO, co founder of Abacus AI. And if you guys haven't heard of Abacus AI, we think we're one of the first platforms in the world where AI, not humans, build applied AI systems. I think nowadays we're seeing everybody talk about how AI can build anything, become the new software engineer or our ambition is to get AI to build other applied AI systems, especially in the enterprise world.

Yeah, and you're building like a Microsoft Office of AIs. You're trying to do the whole suite. You're not making us go to, you know, LangChain or Pinecone or put part it all together and have an engineering team who figures it all out. You're trying to provide a suite of different kinds of functionality on one package, which makes you unique, I think.

Yeah, no, I think our whole point of view is that it needs to be end to end. You get your data and then you want to build your large language model app or your AI agent on top of it. Also ideally you want to build many of these. I feel like in the end an organization will have like what, a hundred of these LLM apps and AI agents. You don't want to keep like building everything in a very DIY fashion. You want to be able to do it automatically. As you can imagine with AI agents like Devin, actually AI is going to be building a lot of this. So what we do is we make it really simple for large companies, actually small companies too, to really build out an LLM app really quickly so you don't have to go and pick out 10 different open source packages and orchestrate them together. Instead you go to Abacus. You literally can use any UI if you want to build enterprise class vector stores, enterprise class orchestration. So it's a very, very powerful system which you can use very easily. That's really our value prop.

Yeah. If you're a big company, you probably have an AI team and have a bunch of computer scientists working for you that can figure this stuff out. But I'm talking to a lot of business people like Scott Jordan, who runs scottevest. And he has a whole bunch of data in his corporation, in his intranet, right. He might have a Slack channel, might have some corporate email that's going through. He might have a bunch of PDFs, he might have a bunch of a database here or there, customer service kind of thing. Right. And those people don't have computer science engineers working for them, but they sort of want to do something like a chatbot for their employees to make their business more efficient or talk to the customer and take care of some of the phone calls that might come in, or at least part of them. Where do you start with somebody like that rather than the company that has the resources to have a 10 person computer science team who's probably up to date on AI and is building their own stuff?

Yeah, I mean, that's like a really excellent example of the types of people who could use something like abacus really quickly and easily to build their chatbot. But even more than a chatbot, like people are now going past the chatbot, right? They really want like, even like larger companies, for example, they can use abacus to, for example, do things like talk to one system and another system and pull things together. Classic example. I want to run this report on Salesforce, get my report, draw a little graph and then post it on my Slack. You know, these are the kinds of things that you would employ data analysts for before. These are the things you would employ like, kind of like, you know, marketing analysts for all of that can be automated using AI agents and these things. You don't really need computer scientists to build these things. Right. And as you can imagine, you want, even if you had your 10 team of, team of 10 computer scientists, go build an open source LLM model because that's useful right now. But for things like this, like I see too many times like data scientists actually trying to still reinvent the wheel. So our whole point is don't reinvent the wheel as much as possible. Kind of use applied stuff. And yes, your friend can definitely use abacus. But our sweet spot really is with customers who have a lot of data, who may even have data science teams, but either who want to 10x the data science team or want them to go to other stuff so that they can like, we can start building this in a much more automated fashion. But generally speaking, I think when you look at the data science profession, right, I think everybody's afraid of losing their job or whatever. And I think the main thing here is not losing your job. It's how to 10x yourself. That's what I think is the key. In the next year or two, that's inevitable. You can't be the person saying, I don't want to use AI. That's not an option. You also, if you ask me, can't be the person who doesn't want to, like, you know, doesn't want to use tools like this. If you say, hey, I want to build everything from scratch, it's going to take you more time. There's no doubt about it, right? If you say, hey, I'm going to go to this place and that place and pull this together, there's no point. At some point, whatever you're learning becomes defunct. And a few months, whatever you do becomes defunct really quickly. So you want to go to production really quickly. You want to do a lot more in less time. You want to be that 10x. So our whole philosophy is try to be a 10x error by adopting all these tools really quickly because there's no choice. There's no choice where you say, I don't want to adopt it.

And now we find the difference between the San Francisco bubble. I mean, there's like six events in one evening here. All AI, you know, different kinds of things. Most people, when I talk to construction people, you know, people who run construction companies or other kinds of companies outside of San Francisco, and I start talking to them about AI, they're like, oh, I use AI and I start asking them questions, oh, you're just using Gemini as part of Gmail to answer an email or two. You're not really using all the AI that the San Francisco industry is developing. I mean, on X.com I have lists of more than 4,000 companies now that are building various different kinds of AIs for, you know, making videos, to making websites, right? Those people outside of San Francisco have no clue what's going on in this community. And I know you're watching it like a hawk because you're talking about all the different models, right? You're, you're embedded in, you live in San Francisco, so you're in the middle of the storm. Take the storm to other people to outside Silicon Valley. What do they need to know about these models? What do they need to know? Why is open source matter? Just cover some of the basics for them. Because I know there's a lot of teams outside the Valley that are, you know, starting to look at AI or starting to get educated about AI, but they, they're way behind a lot of the people who are your Neighbors.

I think the way to think about it first and foremost is I think the first realization is, look, AI is going to take your industry or even your company in some ways by storm sooner or later. And the sooner or later meaning in the next few months to a couple of years, not like 10 years later. So you have to start thinking about AI as soon as possible if you want it. The second thing is, hey, it's not just about you using Gemini or a chatbot or something. It's about if you're part of an organization, think about how you can develop an AI brain for your organization. That's how I kind of like simplify it and say, think about what you can do in terms of understanding the various different levers of your business, how AI can help you and so on. And one of the things you want to do is to go out and like, you know, you could talk to a friend or somebody who knows a little bit about AI, but you know, you want to have what, what most people are doing now. And most of these, I would say brick and mortar companies because those are the companies who are most of our customers. They do what is called a use case exploration, which is kind of like given large language models today. What are the different use cases that I can build out? And people are coming up with anywhere between 10 to 700 use cases, right, that are like large companies. So I think beginning a to think about, hey, how do I apply AI in my company, how do I create these use cases? And then going off and saying I wouldn't go off. And that's other big mistakes. Sometimes these kind of like I would say brick and mortar non AI companies will do is they'll say I'll hire one data scientist and somehow that person going to figure everything out. That's a much harder problem if you ask me than saying, hey, I'm going to partner with somebody and figure out how to deploy this in production. Because if you go hire this one data scientist, that person may or not be a good fit. It's probably not likely someone from Google because they're not going to come and join your company. So I do think that thinking about it in that angle is good. The problem with X is as you said, 4,000 startups, names of startups, even, you know, people that don't even know about. So it gets very kind of confusing and overwhelming. So I think it's better for you to look at it from your point of view like, this is my organization, I want to build my AI brain. What am I going to do to get there. What are my different LLM use cases? And then go from there to, okay, should I partner with someone? And again, a lot of times people are saying, oh, I'm going to go partner with Anderson Consulting or something like that, which costs a lot of money. Now if you're, you're okay. If you've got like billions of dollars and you want to do that, some of the.

I have a list of consultants and agencies that help business. Some of them are less money than that, but, you know, they still do cost something.

Yeah, exactly. So, yeah, my point is find someone you can partner with versus trying to hire someone at this point, especially if you're not a core AI company, understand your use cases. And most importantly, start thinking about it now if you aren't. Because I still see a lot of people actually outside the bubble, but we're kind of like, oh, we're still early in the AI journey, blah, blah, blah. And I'm like, but if you're going to be so early and the rest of the world is like zooming ahead, you're going to like really fall behind. So I would say that's really important. When it comes to open source, I don't know if an average company should care so much like open source or not or whatever. But like, if you forget about it, if you think about it more from a humanities perspective, let's put it that way, like, what's good for everyone is that we don't want power to be concentrated in a couple of different places. And the problem with being very secretive is this. Until 2021, even everybody was very open. Okay? Google, Facebook, even. OpenAI was sharing everything with each other because there was so much sharing and research was so open. We did what we did, which means we got chatgpt in many ways because of that open culture, because of open research. ChatGPT itself is based on Google's Transformer models, which they published out of their own, like, free will. And it was open. Right. Then what happened when people realized like, ChatGPT can be very lucrative? They started shutting things down. Right. And I, I actually think OpenAI did this first. They kind of shut it down. And because they shut it down, it's become very secretive. And so nobody actually knows what is going on in GPT4 or like in Germany or whatever. Right. And Ashley, so what's coming? Exactly. That's what I'm saying. So now we already don't know what's going on now what's coming next? We don't. But more importantly, we don't know how dangerous it could potentially get, even like this. Not enough eyes on this, if that makes sense. There's not an, you know, openness is transparency too, right? If you're afraid of AI, the best thing you can do is be, you know, is do it incrementally, which we have to do. There is no kind of like big jump into like some sentient thing, so. And do it openly and transparently so everybody knows what's going on. So that's why open source AI is really important. The other thing is you saw this with OpenAI too. They zoomed up from, you know, $1 billion, whatever their valuation was, to now, like almost $100 billion in record time in like 12 months. And the reason for that is because there's so much concentration of usage and power, which means that fundamentally they also can become this huge monopoly, which is really bad for everybody else. So we don't want that.

Yeah, there's a lot to cover. If I'm a business person and I'm trying to look at AI, I guess there's two points of view. One point of view is replace workers, right? I'm seeing this in the call center business, for instance. They're turning on AIs that are answering the telephones and laying off hundreds of people, if not thousands of people who used to do that. That's one point of view. The other point of view is sort of what you were sort of hinting at, which is you make your employees 10x ers, you make them superhuman, you make their jobs easier, you take away the shit job from them so that they can spend a lot more mental time on the really important thing that they're doing. I. My psychiatrist talks about this, right? She, she built an AI system that takes notes while she's meeting with patients. I'm patient number one. And it takes away the job from her so that she can pay attention to her patients and provide a much better level of care for her patients. And by the way, the AI is catching things that she misses right, over time because she's dealing with so many different patients. How can she keep track of what Robert Scoble is all about? Right? Tell me how your philosophy is when you sit down with a business person. What is the initial conversation? What are you trying to guide them to? You know, are you trying to guide them away from just, you know, let's look at killing jobs or tell me what your philosophy is there.

So I think honestly, that is, it's a combination. I mean, I think there's a lot of jobs which cannot be automated today. I mean, very frankly, for all the hype in the world, you cannot really automate a software engineer anytime soon. I feel like it's at least, I mean, on Twitter when I say it's three years away, people just freak out about that too, right? They're like, oh my God, it's going to happen in three years. So I think a, the first thing to consider is like, what is your ROI in terms of which LLM use cases or which AI agents use cases you're going to get the most benefit from from business standpoint. So if you are someone who employs like as an example, there's a very large customer for ours who employs like 3,000, like, you know, kind of like these consultants who basically do a lot of like, you know, everyday work in terms of pulling out reports. They have a ton of reports to pull out. And so nobody is saying, hey, let's get rid of all 3,000 consultants. All we're saying is, hey, we're going to automate those reports, we're going to pull those reports really quickly. So to your point with psychiatrist case, it's going to go much faster. Now if in that process they have other things to do, great. If they don't, and we don't need 3,000, you need 2,500, that's cool too. You know, you can go with that 2500 and then now you have that excess 500 for like a different job description or what have you. Right? So I do think there's a mix of that happening where you probably don't need as many folks at the same time. I do think there are very, very few professions where you can completely automate it. Even a writer, I mean, you can say, hey, GPT4 writes so well, but honestly, you do really need a human in the loop still now, I don't know when there will be some, you know, more automation, but I think it'll happen gradually. And mostly what people should be doing is upskilling themselves constantly adopted quicker, upskill faster, you know, find other like job description. I mean, increase your job description, all of that. So I would say nine cases out of 10, you're applying AI to be that superhuman and only like a very small percent. Maybe you're losing your job, but hopefully you find something you know, which is like adjacent where, you know, you're not easily replaced.

You know, I've been talking to people who work on factory lines and I ask them, how do you get AI onto the factory line? How do you Convince, you know, everyday factory workers to start using AI to do their job. He said you have to walk the factory line and build relationships and change them. Their education, you have to show them things. You have to do little tiny projects with them to get them on board and get them to see the power of this and the value. Do you agree with that? And are you doing all sorts of cultural change initiatives to get people, you know, when you get a customer that wants to do AI, you're going to have to get a lot of other people that maybe are resisting, right? Because they're worried about their job or they're just not, you know, they don't know about it, right. So they need to understand how to use it and how to incorporate it into their jobs and into their lives.

No, that's the single biggest issue, to be very honest. There's a lot of, like, there's two ways fear is getting into the market, right? There is a lot of fear in the market right now, right? Because one is like, people are like, I just don't want to adopt AI And I don't believe that they understand fully that that's not really a choice just to like, kind of like, you can't just close your eyes and just stay closed and hope will go away, right? And then some extreme of that are all these AI boomers, they actually, the real fear that they have is that their job, most of them, at least their job will go away, but they will say, oh, AI is going to kill us all or whatever. I mean, there is some very small fringe which actually believes this for a real thing, for real.

But most actually, I do, I believe that there's a possibility of an existential risk someday when these systems become really, really, really smart, right? That they could run. My show is called Unaligned because of the AI could go unaligned and do something anti human. Right?

Is that today? No. Right? Yeah. And if you were okay with open, transparent, incremental AI, you'd be able to see how it develops and make sure that it goes the right way. Right? It's only like if there's some secret crazy bunker somewhere, you know, nothing. Five years later, some crazy AI shows up and like, destroys us all. We have no clue, Right? That could happen. I'm not saying no, but this is why we're pushing for open source and everything else, right? And so, yeah, that could happen, maybe if it's all secretive and crazy and weird. But today we're nowhere near that. I mean, you can't say, hey, today, let's Pause AI because it's going to kill us all. Maybe sometime in the future.

No, it'll save a lot of lives first and a lot of dollars in your business.

Right, exactly. And so I think the first thing is like, hey, let's adopt it. Let's see how it'll actually make you more secure with your job than less secure. I mean, that's the first education. In fact, even this is way before LLM showed up. We've been in business for some time for three, four years. And the last few years, it's always been like, how do you make people adopt AI without making them feel insecure? And that's actually a very difficult process. Part of it is just getting them more convinced that AI is useful, because the first thing they will say is, no, it's not. And now with LLMs, it's very hard to say, no, it's not. I mean, because now it's almost delightful and magical. And then part of it is like, you know, basically, like, doing it small, show that it's useful in some small way. Do something quick, give them. Show them some roi, Then they start adopting it and then they start learning from each other. Right. And I think that's a necessity. And I think in terms of factory floor, I mean, there's a lot of talk about robotics, as you've obviously seen, it's still a little unclear, like, are we really kind of prime time yet? Because there isn't anything right now in the market, so we'll know more.

Maybe. Amazon has 700,000 robots, you know, rolling around its warehouses to move products around. But we're talking about humanoid robots where you can actually interact with the robot. High five. It handed a tool, or it can hand you a tool, something like that. Right. We're not there yet. Now, will we be in 18 months, maybe in some limited situation, 18 more months after that? I think we're going to start really seeing the stuff show up in our factories and our retail stores and our businesses.

Absolutely no. And I also think people also, one really small, I think, interesting point is people still don't know if AI is going to be like driverless cars. Meaning that we've been talking about driverless cars for 10 years.

Fifteen in my case.

Exactly. And so the question in everybody's mind is, hey, is this demo, where is this somewhat production, but not quite production? I mean, is this really going to be as useful as people are claiming? And that is still, to some extent, I think, to some extent, maybe the jury is out, but I'm much More bullish. I feel like this is not a 13 year thing, it's more like maximum, like an 18 month thing.

So yeah, a lot of changes hitting business people right now and that change is going to continue to increase. We're seeing all sorts of new capabilities. Let's break up some of the categories of things that business people are doing. I'm seeing a whole lot of companies that are going after marketing and writing, you know, helping people with emails. Right. Simple stuff where you help maybe somebody who's not very good with English, a construction worker or something like that, or write a report and get it to really a good quality. Dig into that. Are you seeing the same thing that a lot of companies are starting with a marketing effort with AI? Because it's a simpler thing to go at than trying to get a robot on a factory floor or something like that?

Oh yeah, yeah, yeah. I think the two big ones are, if you ask me, customer support and any kind of question answer thing. So customer support, tech support, that's obvious. That's already a lot of companies were there in the space even before LLM showed up. And then the second big one is a document generation, which is to your point of writing the writing the emails, the writing legal, you know, writing legal documents, writing like marketing reports, doing market research, that sort of thing. So these are the two big areas of focus. If you think about it from a business point of view, like the question answer and the document and generations of various different types of documents. I think what's going to be the next frontier is more like these tasks, what I'm going to call like these code execution tasks. Right. Build a dashboard, draw a graph, you know, pull this data from this place and do this other thing. So that's going to be the next frontier. So it's going to be incremental. It's going to be, you know, kind of like happening in like I would say before, I would have said through, through the next five years. Now I'm going to say through the next two years because it's going so quickly.

I'm starting to see hints that there's, I don't know that operating system is the right word, but it feels like an operating system where you talk to it and a whole bunch of things happen. Right. And we could break that up into AI agents, which I know you're, you're talking a lot about on X and other places and so is a lot of people in the San Francisco world. Maybe we should start with that. What is an AI agent and how would you use a set of AI agents in your business to get started doing some new kind of task.

Okay, so think of AI agents as an AI program which has a bunch of like AI, you know, AI calls, LLM calls like, you know, calls out to large language models to do a particular task. You know, that's the simplest way to think about it. Like, oh, my task is to pretend to be Robert Scoble in this zoom meeting. Okay, that could be an AI agent.

Or analyze the zoom meeting. Did Bindu say anything interesting or. Right. Or write a report about what we talked about, you know, and shrink it down so that somebody doesn't have to listen for an hour.

Right, Absolutely.

This is something, by the way, that I use on YouTube all the time. AIs can ingest a YouTube video, an hour long video in a few seconds now, and write you a report on it. It's really cool.

Summarizer. Just take that video summarizer. That's like an AI agent which does that task. Now depending on the complexity of your agent, that agent can be very narrow. It can just be a YouTube video summarizer. Or you could say, hey, I have a media agent which will take YouTube videos, which will take tweets, which will take XYZ. So there's various different levels of complexity. And what we think will happen is inside the company, people will want to build these custom AI agents based on that company's processes. So there's like a pharmaceutical say, and they're pulling together some sort of a clinical trials report. And they maybe do that every three days, let's just say. And now they're spending a lot of manpower doing that. If you can build an AI agent which will help you do that really quickly and really faster, that's an, that's an example. Another very common example is what we call an RFP response generator. A lot of Companies respond to RFPs. These are requests for proposals. A bunch of people have to respond to these requests for proposals. People will like bring those responses together. Now usually they'd hire like some very like, you know, competent person to do this. What you could do is build an AI agent which responds to your RFPs. So it has custom domain of your company. What your company's like. You know, pros are and cons are and has all the knowledge about your databases and everything. And it'll actually generate an RFP response for you from your custom knowledge. That's a classic example of an AI agent. Right.

And do you have to fine tune your agent to do that? How do you give your custom knowledge from your company, like the PDF documents or the emails that your company has sent around for years. How do you get that knowledge into an AI agent?

Yeah. So that's where the hard work is, Right. The problem is all of this data is somewhere. It's selling some data source or some PDF, et cetera. So you're building out the pieces. So you want to get those data source connectors. Maybe it's in SharePoint, maybe there's a bunch of Word documents somewhere. Maybe there's a bunch of other database tables that are sitting in a database. So the first. This is why you need a platform. I mean, it's not as simple as, like, Okay, I have GPT5. Can you create an agent for me to do this? It's just that GPT5 has no knowledge of your data, your custom stuff, nothing. So, A, you need all these data sources. B, you will have to orchestrate this AI agent, meaning get the data source, understand the data source, create a vector store behind the scenes if need be, and then have all these connections so that the agent can, when given this, when given an RFP document, let's say I uploaded RFP document and say, go create a response for this. It needs to know what to do, needs to read it, read your questions, know where the answers are, pick up the answers and create it. Right. That program, the thing which does that is the agent. That's what's written on the Abacus platform. It has various pieces to it. It has code execution, doc retriever, LLM calls. And LLM is an integral part of it, but it's not the only part. And the thing is, no amount of AI is going to overcome that hard problem of figuring all these things out together. Maybe an LLM, and then write that AI agent. AI writing AI agents is possible to do, but there's still this. Let's go and pipe all this. And that's where the platform shows up.

Last time we talked, we talked a lot about tabular data, which looks like a spreadsheet, Right. Something where there's a lot of rows and a lot of columns of data. LLMs are good at understanding the English in that because it's a large language model. But they might not understand how to do the math or the planning or the analysis of that. Right. You know, we were just digging into agents and how they really could help somebody's business, how to set them up, how to train them. Let's dig in a little bit deeper. Land the plane from 30,000ft to, you know, 15,000ft and dig in a little bit more. What are, what are groups of agents going to do for a business? And how do we get them all running and how do we manage them? How do we train them?

Yeah, okay, so like I mentioned, very simple way of thinking about agents is like think about agents as a, as a program which, which speaks to very different, different AI models to do a particular task. Task can be very simple. So you have a simple agent. Like, for example, summarize this video for me, is a simple task. Or it can be very arbitrarily complex. Like, you know, go and like, do the job of this particular person end to end. And this job involves like, you know, summarizing docs, writing. Docs writing, like a paralegal job. Say, like, you know, somebody who's like, you know, in the legal office. And that could be a very complex, very smart agent. Right. So there is, like flavors of these agents. Now, the very complex agent. I don't believe we really have that, you know, in production anywhere yet. The simple Asians, of course, are all over the place now, even for the simple agent. And let's take that simple agent example of say an RFP response creator, which is based on your data, your organization skill set, your organization's abilities. It needs to have the understanding of your data sources, of your knowledge of what the company does. And the fun part of it is some of that data is there and some SharePoint, some of it is there in some databases, some may not even be there. You know, like, you actually need people sometimes transcribing this stuff. And so even if you want to have a medium hard agent, we're talking about connecting to your systems, we're talking about understanding those systems, and we're talking about kind of like what the flow is going to be, right? I'm going to upload my rfp. It's going to go through each of the questions. It might even give me the answers. Then I might have to review it because you can't trust the AI to like, answer everything perfectly. I might edit it and then I might submit it. So that flow, that process has to be on some platform, on some system. And what we do, what ABACUS does right now is to be that AI agents platform. There's a lot of pieces to that puzzle for it to actually work out. So what you can do is build out that AI agent. And that takes time. And this kind of stuff inside enterprises is going to take quite some time. This is not going to be like, okay, tomorrow all of this gets automated because there's just some heavy lifting to do here, right? And then the third form of agent we went to a very simple agent to some medium hard agent. Now to come to like a hard agent, a heart agent usually needs a plan. Meaning Devin, for example, is what I'm going to call a hard agent, which is basically you say, hey, I want to create like a short novel that's, you know, that's a harder agent. But what do I need to put in that novel? Like maybe my prompt just says write a short novel which is a romantic comedy between these two people. Now I need to have a plan. So I need to have the AI first create a plan for the agent, then I need to exit, I need to review the plan and then execute it. Right. So there are various flavors of these agents. Companies are slowly beginning to adopt these different types of agents. But by the time you go to production, we're talking about talking to their internal systems, their security, all of that has to be overcome as well. So I do think it's going to be some level of transformation to get there. Having said that, there are some very cool agents, which companies are startups are kind of devoted to building the whole cool agent. Like the hey gen avatars, like hey, create an avatar of me. I mean what I just described to you so far have been like boring.

Translate to German or translate to Japanese or. Right. And then do lip syncing on your video. Right? That, that's a marketing task that a.

Lot of people complex. Exactly. Like you know something which is kind of like. And then I think the next iteration is going to be autonomous agents, agents which are going to listen in to like your meetings or listen into like you know, be a Slack, you know, be on Slack, listening to your whole channel, contributing, acting like almost a human. Right. Like when you look at remote work, what do people do? They're on Slack, they're on email, they're on Zoom. They check in code. Now imagine in the future there is a front end engineer agent. All you do is write to that front end engineer agent saying put in a new form on my website. The new form will collect this three pieces of information. It'll do some error checking and it'll put this in this database. The front end agent really listens to this on Slack, goes off, writes all the code for this, submits a code review and then you can read the code review and approve it and voila, in a couple of days that form is built, built out code reviewed and is on your website. Now that's your autonomous Agent, which I think is level four or level five of agent take abilities. And I don't, I mean, like I said, I don't know if we can do. We can do something simple like that if we really try hard, but we're not there there yet.

Yeah, the skill here. I hate using the word prompt engineer, but people who, for instance, asking to write a romantic novel usually doesn't lead to a good result. But if you ask it to split up the task into separate scenes, it can write a really nice scene for you, particularly if you give it a little detail. Oh, I want a scene in a coffee shop between these two people. You give a lot more context, a lot more detail, then it can write you a nice little scene. And if you do that over and over again, well, that's how you get your screenplay.

Exactly. Yeah. So you basically say, hey, write an outline assuming this ROM com, but some like, high level, like, you know, plot guidelines. And then it breaks it down and then you're like, okay, I don't like this guideline. Or I want you to like update this and that's your review, that's your plan review, and then you get to the next level. Yes.

Yeah. A lot of where business has done pretty well with AI is finding patterns. You know, oh, this store is doing real well because right next door is a business of this kind. And a computer can see that and tell you, you, oh, you should open up more stores next to laundromats because the laundromat increased your sales. Right? Something like that. I heard that from one of the AI companies. Tell me a little bit about how you might sit down with a business person, using agents to start seeing patterns in your business that are hard to see. You know, why is your sales up in one region? Why. Why is one person doing so well, Right. Looking for that kind of little pattern that might, you know, move the needle over time. If you start working on identifying these, you know, why is sales good or sales down and fixing the problem or using somebody found a way to sell something that you didn't know about, all of a sudden you identified it with your business, with your tools. Tell me a little bit about how to, how do you. How to find new patterns in business?

So most of the patterns or like, kind of understanding patterns in business usually has to do with predictive models, right? So LLMs are fun and all, but, you know, one of the things that are really powerful are these predictive models which operate on your data, like, which operates on your salesforce data which operates on your Usage data. So for example, a very typical one is, hey, can I look at all my usage data? And I have like millions of customers. Can you predict exactly which customers are going to churn out next month? And if I can predict that at like say an 80% accuracy, that's huge, right? Like I know exactly which 10,000 of my million customers might churn out next month and I will give them all a discount for the next three months to prevent them from churning out. And at least you'll save 5000 by doing that. Right? If it's especially next 3 months free, they might just stick around. So that's an excellent, kind of like understanding the pattern, making an assumption there and finding like, okay, how to like, you know, make that useful for your business. The other even more interesting one is, you know, very big sales teams, like people who have like 1500, 1600 salespeople, they get like so much incoming leads and marketing leads and sales leads. Tell me which percent of my sales leads are the most likely to convert into customers. Right. That's a classic. Another example. So what is it looking at? It's looking at types of customers you've converted before. It's looking at like who this person is, who's signing up, who's, what are, how are they Using my free platform tier. So it's looking at all these patterns and trying to find the answer. So I would say a lot of this pattern extraction type stuff comes with predictive modeling and forecasting. Now you can actually put an AI agent on top of that. So this is what we do a lot. You can take all your data, find your patterns, get your results, but now your AI agent can analyze those results. It can say, okay, I found these people are all going to churn out. But these are the ones who are most likely to churn out. This is what you, this is the decision you can take. You can say for these top 10%, do this and so on. So that's something which really, really moves the needle.

When I visit a new bank down in Brazil, they built their own customer relationship management system. So a phone, customer calls into the bank, they don't get a phone tree. They get routed to an employee that can personally help them and that increase their customer scores. Right. Customers are happier with that bank than other banks. In fact, when I visited, there was a six month waiting just to get a credit card from that bank. Now it's a public company. Warren Buffett's invested in all that. How can you use AI to do that to your business, to transform the Business to fundamentally change how you service customers, how you, you know, how you take care of problems, how you help people, how you, how your business is seen.

So the first and foremost thing I would say is data, right? I mean, a lot of times, a lot of companies don't record enough data because I can't help you if you're not recording data. Let's put it this way. Like, we don't know what, how your like, customer service agent, you know, treated some customer. Maybe they treated them very badly. We don't know that. And, or like, maybe, you know, somebody tried selling something to someone and that didn't work out. So we don't know that. So the first thing like customers ought to do and, you know, this was a problem two, three years ago. Now more and more people have become very savvy where everything gets recorded in some way in some database, in some place and whatnot. So the more you record, the better off you are. And honestly, over time, I would say, like, you know, the smarter your LLM will get. Meaning, as we get to GPT5, GPT6 and that sort of thing, what we're going to do is we're going to not just be able to kind of get these predictive models going, we're going to have LLMs kind of look at the data and suggest what to do next. Right. And just basically say, hey, we can build a predictive model from here. Like to extract the patterns. You might still need a predictive model, but the LLM can make smart, can be good at decision making. Right. What it can do is these LLMs actually tend to be pretty good reasoners. They're pretty good at reasoning, though. Like, some people will have a fight with me on X by saying, no, they aren't. But I do believe they do.

They can, like, and they're getting better over time, right?

Yeah.

Right. So even if it doesn't quite work today, you have to plan on a world where in a year it's going to work.

Exactly. But it will work. So the minute you start recording, so there are two pieces to this. The AI has to have access to this data. And that actually I think is a hard piece in some ways. Right. I mean, there are companies like Humane and Rewind, which are basically saying, we, like, record everything. I mean, that's a bit scary. A lot of people probably were not going to adopt that.

But you're being recorded by rewind right now, by the way. I use it on my Macintosh so it listens to my meetings now. It sounds freaky. Right. Because it's listening all the time. But it's only keeping all that on my local computer. It's not going to a cloud computer. It's not being centralized. Right. It's just being stored on my local computer and it's. For me, it's making me a superhuman.

Right, exactly. Yeah. So, I mean, maybe that's actually something which people have postulated. Right? I mean, will companies like, you know, record every keystroke of every employee? I mean, will people. Are people going to be okay with that? I mean, your case, fine. You mean you work for yourself, it's okay, Right. But, like, you know, let's say I work at Abacus still. Like, do I want everything I do recorded and put in a company server? And, and if we do that, of course the AI can get smarter now. It can listen in on what all you've done, whatever, you know, what's happened in this meeting or what is not, and then start suggesting next steps, like, okay, I saw this, or I have this data, let's go build this model and then start improving. Some version of this is probably going to happen. I mean, it sounds a little scary, but I think it's going to happen.

Rewind is real nice because it sort of says, here's what happened in the meeting. Like, you talked to Bindu about agents. And then it listens for, did you give me a task? Right. Did you assign me a thing? Did we talk about a project? Well, it separates that out. Yeah. Or if you said, oh, we need to do another interview next week, it starts suggesting putting. Putting that on my calendar or having an agent set up to talk to your agent about a calendar. Right?

Exactly. Right. Yeah. That's almost exactly what I think over time will happen as we record more of these things. Until then, it'll be like humans saying, okay, we need this or we need that, but over time we'll have AI doing that.

Yeah. What kind of question am I not asking? Because I'm not hardcore in the business trying to run a thousand salespeople or something like that. I'm not that person. And I also am not a real hardcore AI person. I don't build agents for every day as my job. I talk to people who build the agents and things like that. What am I not asking?

What are you not asking? I think the thing that maybe, you know, that people miss the most, I would say right now is a how. I mean, and you asked this question maybe how much should I spend, how much time of a company should I spend right now on AI? And how do I get started? Right. And I did address that a little bit by saying, hey, the sooner you can think about it, the better. And B, start small. That's always the case. Like, you know, go in and going in and saying, hey, I want to do this rewind for my whole company is a really bad idea. And people are not going to buy that. Right. But otherwise, you know, there are people, sadly, I would say 80% of the world still isn't taking this seriously and are freaking out about it in some way, shape or form. And then 20% of the world are trying to boil the ocean. Right. That's, I would say that's the only thing that maybe people should start thinking about this very, very quickly. The only other thing I will say which is maybe not too business related and that's actually like really, could I.

Could I wrap that up and say, look for a quick win. You know, look for something where you can spend a week building a little system for a little task. Look for a win, take it to your boss and go, hey, here's a new report I built with AI, Right. Look at how cool this is. Or show it to your coworkers.

Nothing happens in a week in these companies. So I would say look for a quick win.

A month or six weeks or whatever.

Take two months. Yes. Just look for a quick win. Do something. Look for a quick win. Get going. It's a very bad thing to be paralyzed. And I see a lot of analysis. Paralysis.

Yeah. So get started on something small. It's sort of like asking it to write a novel. It doesn't do a good job, but if you ask it to write a single scene, it does really well. Right.

And don't get stuck up with these hard tasks, which won't work. Exactly.

Yeah. They'll get better over time too. And you know, the tools are going to get better, the companies are going to get better. You're going to get better at figuring it out. Right. And, and you're going to figure out what you need to get to the novel. Right. What kind of data you need to find in the corporation to get to the big project. But you got to start small somewhere. Oh yeah. What else do you want to talk about? Because you're always like in there talking about all sorts of different things from the state of the art. Well, let's talk about state of the art stuff. You know, let's, let's go a little nerdier vision models, multimodal models. What are they? Because I've seen 30 different multimodal models. And we're starting to see robots like the, the Brad Adcox figure robot that actually could see everything on the table in front of it and then tell you and manipulate it. Right? Oh there's, there's a hint. Water bottle. Right. Tell us about multimodal and what that might mean for a business.

Yeah, I mean I would say like right now people are like still kind of like most, a lot of it is mostly language, right? Like that's text and documents and stuff. But the next step, obviously the obvious, very simple, easy next step is diagrams in these, you know, in these, in these various different text and documents. And then you're also seeing on one side like audio. I mean you just saw that open OpenAI voice engine which basically talks about like how do I mimic somebody's voice? And then there's obviously video which is like sora. So if you think about it right now these are three separate models. There's the SORA model which is doing video generation. There's the audio model which is doing audio generation. That is the text and image like understanding model. Right.

But I think there's a 3D scene model too. I mean there's.

Exactly. And then there's a robotics model where you're going to be like, hey, I'm going to actually like perceive things and like see stuff and then understand and do things. I also do think though as over time you're going to see a little bit of like these models coming together. Right. I mean like audio and video first of all will really collapse in that, you know, you will see SORA having video very quickly and like people talking. Right. You in initial SORA videos we just had like two people walking or whatever. Now we would have like a scene, like a Hollywood scene. And then you know, at the end of the day the holy grail is going to be like, let me perceive my environment. And then I don't care if it's like you know, a computer that I'm reading like out of or I don't care if I can see that hint bottle, understand that environment and then process it. Right. And that goes back again to that holy grail of having that humanoid which can perceive and understand and react to the environment. Because none of that happens with an LLM. The LLM cannot perceive the environment. You have to feed it. You have to, you know, when you feed it that information, maybe it can understand it, but it doesn't actually do anything to react. So I do think like in some ways before we say we have AGI, you're going to end up having these kinds of either really good reasoning robots which are like software engineers which can feed things from the system, or humanoids which will actually look at the physical environment and get going. Honestly, the real thing is there's a lot of hype today on robotics. It's still a few years away. The reason being that when you look at state of the art, you have to get data. One of the reasons language goes really fast or went really fast is we got lots of data. When it comes to video, we had YouTube. OpenAI says it didn't, but I'm pretty sure it did use these YouTube videos to train Sora and then audio also. We have some data. Right. The problem with robotics is we don't have data. Like we don't have enough data for.

Machine to run fingers. Right. We don't know how to do that yet. We know how to make the motors but not how to run it. Right.

So there's still a lot to be done. When it comes to data collection for robots and things like that, I would say that's where the future is going to be. Sooner or later, of course you'll get GPT5, but I think it's sooner. I, I would say by the end of this year even I feel like LLMs are going to hit the wall. Meaning there's not that much more to do. It already is as good as an expert human when it comes to solving problems. You can become a super expert human, but that's it, right? What else can it do? I, I mean, the problem is solved. There's nothing else do to. To do?

Yeah. When I talked to like Manjal Sh Shah who runs Hippocratic AI, he gave me a demo of his system. And his system is a voice system for nurses to run their nurse station. And I talked to his system for half an hour about a potential colonoscopy appointment I was going to have. And it, it answered the phone and I talked to it. It sounded like a human being. It was a real interesting business service. Are you helping build those kinds of services with abacus or are you going to let that part of the AI world be done by another company and then you're going to integrate it? Tell me about how you're going to work in. You know, because call centers are getting this kind of technology. Customer service representatives are getting this kind of technology because, oh, I, you know, I was just at a tire shop this morning and the guy had to answer four calls like boom, boom, boom, boom. Right. Well, he could have had An AI or that store could have had an AI answering those calls and taking care of customers and maybe answering most people's questions and scheduling things or whatever. Without a human info, how you seeing that world break down? Because I know Hippocratic's doing it for a medical system, but.

Exactly. So I feel like there's going to be a lot of Hippocratic AIs, right? And they're good. And they're good because in the sense that I call them vertical applications, they have a specific use case. They do that. Use case. Well, you buy that packaged solution from Hippocratic AI, it's done. That's great. You're going to see a lot of those companies and I think, I believe a lot of those companies will be very successful as well. And even existing companies who right now do X, O, Y or Z will have AI infused in there, then that's fine. The way I think, like Abacus will play is I don't think we're going to play too much in that vertical AI space. We're going to play more in kind of the enterprise AI where people are building a kind of like somewhat custom agents. Right? I mean, yes. Are you going to buy. Is there a vertical AI company which does RFP response generation? Sure there is, but you're not going to buy like 10,000 small vertical AI agents. The Hippocratic AI is doing something cool. Maybe that's useful, you'll buy that. But for things like RFP generation, or for things like, hey, I have my internal employees who need a chatbot, or for things like I need, you know, something which is custom to my company, all of that people are going to build on top of a platform. They're going to say, hey, I want these five, like really basic things, but this ten medium hard things. So I think this, like, you know, if you look up, look at the stack, there is that vertical package solution AI stack, which Hippocratic AI is in. Then there's this middle stack which I think we play in, which is kind of like, okay, how do I build LLM applications on top of all these LLMs? And we want to be LLM agnostic. We don't want to be like married to OpenAI or anthropic or even our own LLMs. And then there's the lower stack, which is the foundation models themselves, like the ones which OpenAI are building. I mean, we play somewhat in that, but we don't train models from scratch. But I see that kind of evolving because this whole space is huge. I mean, God knows How many solution companies are there? Like Rewind is one, Hippocratic is one, Humane is another. I mean, there's so many, right?

I got 4,000 split up into lists on x.com and by the way, I. I only know about 4,000. I can only find 4,000. One of my sponsors is AI Top Tools. They have 10,000 in their data set, right? So there's a lot of companies doing all sorts of weird little tiny things, you know, with AI, that some of.

Them will be mom and pop businesses, some of them will turn out to be these massive hits also, right? I mean, as you can see with any kind of like the Internet, when it showed up. I mean, how many companies do we have now? I feel like this is bigger than that in many cases.

Yeah, absolutely true. In 10 years, how many different AIs do you think an average business might have running? Because I could imagine that there would be a whole orchestra of different AIs doing different tasks and even checking each other's answers and. Or being built for different purposes. For instance, a smartphone can only run a small model, right? Well, I talk to like a Siri, it might have a small model that answers me real quick on my phone. But then if I start asking, hey, I need some code written. Oh, I can't. The small model is not good enough. It has to go to a big model for that, right? So there's a grouping of different sized models or different kinds of models all working together to make a business in the future. How do you see the orchestra starting to form right now? And then what does it mean in like five years for a business?

Yeah, I mean, my. Right now, believe it or not, we are talking to like dozens of AI models today. Like when we text, there's a model, to your point, the small model, there's a model on your iPhone which is correcting you. So the autocorrect is AI model, right? Then when you leave a voicemail, somebody is transcribing it. That's another AI model. When you kind of like go on Twitter, when you see the algorithm run, that's another AI model. So I think that every system, every software, every system is going to be AI infused. Everybody, like from Zoom to Slack to this recording software probably already will have some AI model in it. And I think it becomes kind of like software, right? You become kind of immune to it. You don't say, hey, this is an AI thing or anything. You just say, hey, this is a system. And systems will just be pervasive, way more pervasive than they are today. I mean, already it's quite. They're quite pervasive. I mean, you're like, you know, dealing with some screen or the other most. Most of your waking time. I think, you know, it becomes super pervasive. And I think the one thing that we are still missing is the listening in. I think part of that has been just the freakishness around it, like where the AI or the system is pushing things to you by listening in. That hasn't happened yet. I'm hopeful that that mode will also start happening where it's coming.

I'm starting to get little devices that listen in on you. You talked about the AI pins, and soon we're going to have lanyards and little devices hanging around just listening to you and interacting.

Your spectacles, obviously, will be very much AI. You know, you want your a. Your eyes are AI, your ears need to be AI. And then, like, you know, in some ways, your hands are AI too. Right? Because somebody doing your task for you, it gets crazy.

Well, thanks so much. You're helping a lot of businesses get into this new world, and I appreciate that a lot. So thank you. Where would we learn more about Abacus and you? Where do we follow you?

Follow me on Twitter. By now, I feel like a lot of people already do, so follow me on Twitter. Bindu Reddy, my full name. And then of course, Abacus AI, which is just simply Abacus AI is our website.