Name: Aomni
Uploaded: 2024-12-17T15:30:14.299Z
Duration: 2 min 33 s
Description: An AI-powered platform that accelerates account-based sales by generating personalized value propositions and sales materials tailored to each stakeholder's unique needs.

aomni

An AI-powered platform that accelerates account-based sales by generating personalized value propositions and sales materials tailored to each stakeholder's unique needs.

Aomni

Uh our first presenter is going to be David Zhang. Uh showing off his Aomni auto GPT research bot. David come on over. 

All right hello everyone. So I want to show you guys uh the product I just launched. It's called Aomni and uh it's essentially it's an information retrieval agent that's sort of meant to find and process data for you on the internet. So what does that actually mean? 

Uh so this is the agent. So it's sort of based on you know the um the auto GPT baby HR architecture that sort of recently came out. It's been super popular. I you know I started working on Aomni to solve a specific problem I have which is I want to you know have a navigator web and find stuff find information for me. 

So Auto GPC is a great architecture for this but as I was building it I essentially started rewriting it one you know one file at a time until pretty soon the entire code base is rewritten. So this is basically a custom other GPT implementation for information retrieval. 

So what why uh why information retrieval or why use other GPT for information retrieval? So uh this is sort of this is actually quite interesting. Um if you sort of compare the architecture of Auto GPT versus the architecture of a database you know so on the left I have my Aomni architecture on the right you have a database and you could you could kind of you know you could kind of look at.

The big high level boxes you know you have a user submit a curry. The curry goes to the parser. Parser goes the planner. Planner executor in Auto GPT this is sort of this recursive process because the data you're finding is unstructured. And after that you sort of uh you know the query outputs. This is actually the exact same sort of architecture that you will see in a normal database. I literally took this slide from I saw my mysql's website or something so it's like literally exactly the same. So that's why I think it's just such a great architecture for information retrieval. 

Uh so I launched it you know three weeks ago like like I kind of just tweeted it randomly as like a tech demo and it sort of just blew up. Um so so far you know we have 13k accounts. The craziest part is we actually have uh 2 000 people on Discord. Uh which is pretty nuts you know like nothing motivates you more than 2 000 people on Discord sending you feature requests every day. 

So okay so let's kind of diving the product a little bit. So I'll actually just show you guys uh the product. Um so here's what it looks like. So you know go to Aomni.com here's our homebo website. We could run a search live right now. Uh do anybody have any suggestions for what to find if not okay so now just I'll just do a search for let's say so you know so now this agent is kicked off it's.

Gonna you know make a plan and it's gonna start just navigating the web on your behalf and collect data and you know I'm Consolidated for you. It actually takes five to 15 minutes so we're not gonna sit through all that. Uh but I actually saw what would be cool is I could show you guys some of the uh results from uh kind of our actual users. 

Okay since those are actually pretty interesting so here's one use. I mean the really simple use cases you know you could use it to sort of generate podcast script for you right. So you could have it uh see all the latest posts on Hacker News and process it into podcast script. So it's like your sort of default Auto GPT demo right now. 

You can kind of go a little bit deeper here so let's let's look at this one. So this is actually a pretty very interesting one and this is sort of where Auto gbt and that sort of database architecture really sells. Because you could start doing these multi-step queries. 

So you know you could say find other news uh find auto post on Hacker News that's a GitHub repo and then go into each GitHub repo and then summarize repo and give me the URL. So you have this multi-step queries and start chaining it up almost like a database query. So this is kind of where you kind of start seeing sort of the power of using Auto GPT for information retrieval. 

Uh this is actually a very cool suggestion to share AI Agent Store with colleagues.

One so this is um just for another user uh so he wants to find um wind turbine for some random County in the UK. So you can kind of see how you know how Auto GPT actually um or how only actually figure this out. 

And it actually started by looking at the average energy consumption of that exact County in the UK and then figured out the wind conditions. And I use that to calculate what your energy consumption is and how much energy a wind turbine will actually produce and use that to filter out the right turbine. 

So there's actually a little bit of critical thinking here right it's able to sort of look up this one sand which is the all the turbines and work out this other information which is what's the power consumption is a households. And it's going to do like a database join to find you the right information. 

So this is actually where it starts getting really really cool. I'll do one last one uh so this one you know maybe not as cool but this is super useful which is we have a lot of international users on our product. 

And this is actually a very cool use case here which is that when an international user use the product uh it knows to it knows that they're asking a question that's only available on English web. So actually automatically translates that to English you know find all information on the English web and translate it back to.

Chinese in this case to present it back to the user. So this is actually you know super super cool use case. I think like probably simple technically but it's actually providing a lot of value to users. Uh of course you know you can't talk about Auto GPT without talking about uh all the uh interesting failure cases. 

Um what's cool about Army is actually um so here's like here's some um I think it's like one of those trick questions people post on Twitter. You know like how big a gorilla you need to feel to fleeing a car into space. And you know you look at chat GPT or bin or whatever they were just hallucinate some random answer. But for Aomni um so this is kind of like maybe it's a failure case but maybe it's a success. It actually just says I don't know and I could give it uh you know try it really hard and just I don't know. 

And you know and the last I'll I'll start I'll enter the last one which is at the uh it can read PDFs. So here's a here's a cool case where somebody asks you to read a PDF it can't read it. And then so it's actually this is actually just some random random info but I actually tried really hard. It actually tried to find the same information that would be in the PDF because I know it's titled the PDF. So it's actually trying really hard to find information. So this is also pretty cool because this agents are actually.

Quite persistent and they're you know and they use they have a lot of good reasoning skills. So yeah so that's the product um you know uh so far lots of feedback and uh it's a full company now. I'm hiring and we're also raising money. So if you guys could contribute to either please let me know huh thank you David. 

Um we're going to take audience questions think about time for one maybe two questions. Uh anyone have a question I can come to you uh awesome okay one second here uh raise your hand again some uh okay got you right up here. 

Uh tell us your name and your question. Hi David Bob Macklin great presentation thanks good luck thank you after a great to start. 

Uh question uh is regarding the time you would mentioned when you did that impromptu prompt it's going to be 5 to 18 minutes you know versus like a pre-trained model that gives you in seconds what's taking place in order to take that long at this particular phase in what would you need to do to speed it up so you can get let's say 30 second response. 

Yeah so I mean it takes so long because it is actually browsing the web on your behalf and uh what I realized is actually there is this huge market for sort of these asynchronous information products where people are actually willing to wait to get a um to get a more complete answer which is sort of opposite what you would see with.

Like child GPT or Bing or you know these other products. I'm actually quite surprised. So our users are actually asking us to make reports more detailed and they want to wait longer because they want us to sort of go all sorts of direction and make it as comprehensive as possible. 

So we're actually, if anything, we're sort of going the opposite direction where you know maybe it's a 30 minute report, right? Which would still actually be a lot faster than if you will hire a human to do this research for you. But you know, and a lot cheaper. But you know, maybe but now that's done by agent we can automate it and we can send it to every day. So that's actually that we actually kind of position yourself as the opposite to these existing products. 

Awesome, thank you. I got one more question. We're going to try and keep it a little faster here because we can move on, right? Thank you, David. Um, so one quick question is what are some of the interesting feature requests you're seeing and what do you think about focus on specific vertical versus just a broad tool that gives general answers? 

Interesting feature requests, let's see. Um, I mean we have, I mean I don't know if any one of them is particularly like surprising. There's a lot of data export requests. There's a lot of support for extra for new sources, PDFs, videos, whatever. Uh, I think there's a, there's just

A lot of demand for people you know that wants to you know be able to ingest a large amount of content as you know is true right whether it's on the Internet or it's on the internet. Um and that sort of leans into where this product is going which is this is a B2B SAS product. 

And uh and you know and what we're the market we're targeting is really an or the Persona we're targeting is really everyone who works in PD. Right because those are the people who does a lot of these sort of manual research on competitors and products and reviews you know whatever. 

So anyone that works in BD is ideal customer for this product. Awesome thank you David thank you.

Note, that there may be slight differences in text and video. Watch video for the original content.

Aomni Video