Introduction

Hey, everybody. I'm Ben, founding platform engineer from Athena Intelligence. Excuse the presentation. I'm a bit better of an engineer than I am at making PowerPoints, and I'm not even that good at that.

So anyway, Athena Intelligence, what are we? And I guess the purpose of this talk, which is practical AI and how we're solving for it.

If you need to get in touch with me, you can send me an email. I'll try to respond.

What is Athena Intelligence?

So Athena is the 24-7 enterprise AI analyst. So Athena, the tagline, it automates the time-consuming tasks so that teams can focus on the strategic work. What does that actually mean?

Overview of Athena's Capabilities

So Athena is an AI analyst which sits on top of Olympus. which is essentially a full-scale data, unstructured structured data platform, which basically brings all of the traditional tools of something like the Microsoft Office Suite into a single platform and they're all AI native.

So I think this is a really good visualization. Web content, documents, structured data, email, Slack, GitHub, your cloud integrations, they all come together and they're able to be interacted with through these core UIs.

So we have chat, that's just like ChatGPT. Reports, very similar to something like a Word doc. Sheets, that's your classic spreadsheet. Drive is sort of like where you get to see everything and ask questions about unstructured data.

You have notebooks, which is essentially just Jupyter notebooks. Query Editor, that's your SQL. Browse, we actually can drop the agent in and have it control the different, basically like a web browser. And then finally Spaces, which is sort of like an infinite scroll thing, similar to like Miro.

AI-Native Collaborative Platform

The cool part about this is we built them all from the ground up, so they're all fully AI native. They can all talk to each other. And they're all collaborative. So in the same way where you might see somebody sharing the same Google Doc as you and you can see them moving around on the page, our platform powers all that. And I think sort of the coolest part is you can also have the agent in there working with you.

The Vision Behind Athena Intelligence

So the first question you'll probably say is, well, I've never heard of Athenian Intelligence. So you guys are probably a pretty small startup. How are you possibly building out all of these UIs in a first class way? And the answer is, because we have to.

We're making a concentrated bet that the person who wins the AI usefulness race is the person who has the most information. And so we have to bring it all into one platform. And once you do, there's really very few problems that the AI isn't able to like be unleashed on and solve as long as you like set it up with the right things that it needs to succeed.

User-Friendly AI Systems

So the first thing that I kind of want to talk about and sort of the reason why you're seeing kind of a lot of familiar faces in terms of the UIs that we offer is because for AI to be useful, we need to bring people systems that they understand. So you don't want to learn a net new interface to learn this new useful tool. You want to keep doing the thing you're already doing, but ideally faster and have AI not have to. I guess the reason AI is in the room with you is because it's making your life easier, not harder.

So you have to go and learn a whole new interface to interact with it or learn a whole new toolkit. it's really pushing out the ability for it to be useful much further than what we're interested in solving.

So we think AI is able to be useful today. It should drop right into how you're currently doing your work.

AI Integration in Traditional Workflows

If you spend most of your day in spreadsheets and you want to be powered by AI, you shouldn't have to relearn how to program or relearn how to put everything into the ChatGPT chat window in order for AI to be useful for you. It should just know how to use spreadsheets.

And so I have some examples here. This is a screenshot from our Drive view. That's the side of our application. You can see all these things live in the platform. And we bring them right to you.

You can interact with all of your traditional data types that you'd be able to, but inside Athena. People want to understand that people want the AI to be right there with them. But I think more importantly, they need to know what the AI is actually doing.

Building Trust with Transparent AI

So I think a pretty classic example that we tell customers a lot is even if you hired the most, graduated from Berkeley, math PhD in two years, he's coming in to renovate your data science team. and you give him all your Slack messages, you give him all access to all of your data stores, and he goes off for a week and he comes back and he says, I solved it. The problem to your business is it's 42. You would say, how did you get 42? What is 42 mean? You did a lot of work, but I don't see any of it and I don't trust you. So you got to think about AI in the exact same way.

The way to shift the Overton window on AI being like taken seriously at the cutting edge in defense contractors, in real investment banking, private equity, these industries where they're dealing with very important things, the AI needs to be able to prove to you that it got the right answer and it'll show you where.

Example of AI Transparency in Action

So this is another screenshot from our application. This is just a classic example of we have an environment put together with a bunch of F1 PowerPoints and information, because we're all big fans of F1. And you can see this is basically what our chat view might look like. It asks you a question. Hey, when was Ferrari founded? And it's going to literally say, here, this is the page. You click on that. It takes you right to this page. It says, look, this is where I found the info. same sort of primitive exists all throughout the platform. So anytime the AI is going to give you an answer, it's going to literally give you a link and say, you don't believe me here, click this and you can prove it to yourself.

We've seen this be absolutely transformative in terms of like how clients are able to get into it. They start using the platform, and as soon as something gets cited for them, they say, oh my god, OK, so we really can trust this thing. We really can give this thing all the power.

Is that the next slide? Yes, this is the next slide. So I've gone through all this. I think like the most important thing to get you and feel free to disagree with me on this.

The Future of AI in Business

Like I really do think this is probably the most controversial thing that I'm going to say up here. But it's kind of core to how we're thinking about the problem and how we're building our product is like the arms race for LLMs is like completely over.

I would say since probably for Turbo, we've been able to solve basically the absolute Pareto efficient problem set in terms of what needs to be useful for businesses. And what I mean by that is if you actually give an LLM the right context, not just like, hey, here's the code I'm working on. Here's the code I'm looking at. Fix this bug. But like... prompt it in a way that you would prompt an analyst coming onto your team to try to jump in and solve the problem.

The overwhelming majority of the time, the LLM is able to do the thing. If it can't do the thing, you probably broke off a little bit more than it can chew in terms of a specific forward pass of the model. And you need to essentially refactor how you're thinking about the process that's making the LLM useful.

LLMs and Business Utility

And so it's the engineering and in the product layer that really needs to catch up for these things to sort of come into their true form factor and completely transform how we're thinking about how we're doing business. So a great example of this is like when we invented the steam engine, there were probably a lot of people that were like, this is never going to work. It's not powerful enough. We need to make it bigger. We need to make it faster.

And there was one guy that was like, let's put it on the back of a boat. And that is what unlocked the, you know, trillions of dollars and the industrial revolution and everything that's happened since the steam engine. It wasn't the, you know, spending time in a research lab trying to make the steam engine better. It was putting it in the place where it's actually going to be useful. And suddenly, you know, the world has changed forever.

I think we're at a very similar impasse in terms of how LLMs are being thought about in the public discourse and in terms of how they're ultimately going to be unleashed on society and business. The product layer needs to get to a point where we're using AI and we don't even really realize it. So we try to solve for this in every single product that we build, and it really shows through in our customers.

We have customers that are doing a workflow elsewhere, and when they move over to Athena, they don't need to learn new buttons. They don't need to learn new ways to think about a problem.

If they have a big spreadsheet where it takes them forever to go and look up on the internet these specific things about these specific vendors when they release new information or something like that, it would be pretty untenable for them to incorporate Athena, buy a platform, and get right in if that was a huge Overton window shift in how they were thinking about using product. And so that's what we're really trying to solve for.

Programming AI Agents

And then that has the flip side of agents, which is how do you think about programming an agent? I think a lot of times when we think about programming agents, we think about it in the context of building this impressive sort of state machine, which is able to adjust for different parameters. And you have to solve the whole thing.

That is very useful and very necessary. But a lot of times, the thing that's going on in the node of the state machine is a total oversimplification of how a human would actually solve the problem.

So I think this is the part where the industry is like, maybe I don't want to say behind, but it's the biggest domino to fall, will be we need to start giving the AI the right process to solve the problem, which is hard. It requires careful introspection.

When you solve a difficult problem, you don't often stop and think, OK, What did I actually do? What it was the framework, the problem solving toolkit that I took, which humans basically learn over the course of going to higher education. You get harder math problems in your homework. You have to learn more diligent problem solving skills to be able to figure those out.

That is the thing that we need to be encoding in software and giving LLMs the ability to use. Because LLMs are effectively tiny little reasoning machines, but when they don't know what to reason about, I don't think humans really understand how many embedded frameworks we have for problem solving in our minds that you don't even think about it because it's sort of subconscious at this point. But LLMs don't have that. But they're really good at following directions.

And so if you can give them the right problem set, I personally, since I've started working on LLMs, have been absolutely astounded at when you set up the problem in the right way. is so incredibly capable. I really think GPT-5, Cloud 4, all those things are going to be great and those are going to be huge unlocks. But I think it's a misnomer to say that you shouldn't work on building intelligent product layers now because the next model is just going to eat all of that.

That may be true in some cases, but I also think if you concentrate on the fact that The processes that you're building, when you give that to a smarter model, will just be able to do it faster and make marginally better decisions. But it would still need the core framework to solve the right problem.

Importance of Context in LLM Performance

And so I guess to round all of this out, LLMs perform very impressively when you give them the right information. The problem is getting the right information.

And so all of these problems are, and I really, I don't enjoy the term like retrieval augmented generation, but like that is every single problem that an LLM has to solve is a rag problem. It's a rag problem in how you get the right information to the LLM.

I think a great example is tool calling. A lot of people don't understand how tool calling works under the hood. If you actually go into the Langchain agent and look at how it does tool calling, it basically just spits a giant window of context of, hey, this is the tool you have. This is the way you can use it, right?

And then when it has to return JSON, you're basically just explaining to it the idea of JSON. Hey, this is the format I want to put it out in. And if it doesn't work it, or if it doesn't work the first time, like the solution that is sort of leading industry right now is, hey, try again. Right.

And when you give it that useful framework and you put it on the software guardrails, I mean, tool calling is like the biggest unlock in LLM productivity that's probably ever happened. And that is literally just a rag problem under the hood. It's just sort of obfuscated behind these layers of how would you say it is sort of like this complexity is buried in some of the software packages that we typically use to interact with LLMs. But it's really the same problem at the end of the day.

Live Demonstration of Athena

So I'm going to give a quick little demo of Athena. And hopefully you guys will think this is cool.

So this is our drive view. And essentially what we can do here, oh, it looks like I already have some documents on my workbench.

So what I'll do is I'll just create a new spreadsheet. And hey, Athena, could you generate a template from the documents on my workbench for me? I've got to give it some more time. Athena, can you generate a template from the doc into my workbench for me?

Cool. So essentially what just happened there, Athena read all of these PowerPoints that I had in my workbench and said, all right, I'm going to put together basically a template for you to analyze the documents.

So now I can go over here and I can say, Athena, summarize this. And just click on that asset. And basically this just kicked off an LLM call.

So I guess like what this, I want to call out here is this brings back to my earlier point or like to use this, like this is literally basically just Excel dropped into Athena and it's powered by AI. So the AI understands it. It can put things in the right spot. When you want to call a function, you don't have to relearn some new syntax for calling functions or a new programming language.

I could give more on the demo. We'll wait for this to load. We'll see if my phone hotspot can figure that out. But yeah, that's my talk.

Conclusion

Thank you guys for listening and happy to answer any questions.