adept-ai

An AI platform that automates complex workflows across various software tools, enhancing productivity and efficiency.

Adept AI

Why don't you show me what you're building at Adept AI Labs? And thanks for the little diversion down, uh, history lane there. Super fascinating, yeah, for sure. 

So first, let me tell you a little bit more about what we're up to and why I'm really excited about it. And then we can quickly, uh, flip through some demos of a release we actually recently did last week. The Northstar for depth, uh, from day one actually has been that in the long term, the thing that'd be the most valuable thing to build for work is an AI agent that, uh, does much more than reading and writing and drawing images, but can actually handle for you arbitrary work tasks and workflows. 

Right, and those two things are very different. Right, like reading and writing is not the ability for you to say be able to delegate your entire, like, payments process to a neural network. In the latter case, what you really want is you want a system that knows how to use all the software you already have on your computer as if it were you. 

And in order to get there, you need to train these models that, uh, deeply understand not just the, uh, text but also the pixels on your screen and also what actions lead to what outcomes in the world. And so we've been hard at work on this, training this model that could do anything a human can do on a computer. And we've been, uh, building effectively a product that enables knowledge workers to arbitrarily share AI Agent Store with colleagues.

Delegate tasks to the system. Here's a quick demo example. Um, in this case, uh, let's say you're responsible for paying invoices and you get your plumbing invoice. You fire up a Dept that pulls up the invoice, all being done by the model right now. It pulls up the invoice, reads the pixels in this PDF, realizes what it's about, stores some interesting facts about this, then pulls up QuickBooks and then correctly enters who is the payee. Right, Savant plumbers? Like, how did you pay? Um, what category is this? Like, it realizes the category was never written in the PDF, but it realizes a plumbing invoice, so it should go into repairs and maintenance. 

This task that you probably would have had to do like 10 or 20 times a day for your job, you show Adept how to do it once and now every time you get a new email invoice, you just fire up a Dept and it handles this task for you. It's all so where does Adept live? Is it in your system tray there or how does it intercept this coming in via email? Because you have this invoice come in via email, you got to get it paid. Y you're in purchasing or accounting, boom, it needs to know. So how is it just sitting there in the background running? It lives as an overlay, as an extension right now, but we'll soon release a desktop overlay as well. 

I think the key of this is we're not forcing you to use this brand new system. It's a suggestion to share AI Agent Store with colleagues.

Helper for the same workflows you co-pilot. Yeah exactly it's a co-pilot. Yeah so it's sitting going to sit in the system tray but right now it sits in your browser window. Your browser window Y. And so you can delegate arbitrary tasks to it right now. 

I'll show you show you another example. It's actually a similar variant but like one of the most common things people do is people shuttle data back and forth between system A and system B. And a lot of a lot of knowledge worker jobs is just doing that like ad nauseum. Right um we had a we're talking to some customers who um their insurance agents have to go log on to five different software systems to be able to pull the requisite data to even get one quote done. 

And so this next example is one where you get an where you get an email from uh someone who's for filing a claim. And Adept basically uh once you show Adept how to do it once automatically fills out all and populates all of the forms involved there. You know what's interesting about about Adept is that it's actually been a really easy way for us to start working on this like what I think is going to be the next battlefield of AI. 

Right so far it's been about LLMs but I think what's what's coming up is it's going to be about multimodality which is the ability to understand images. And it's going to be about building AI agents because AI agents as defined as a model that could take a.

Serious of steps to achieve a goal is I think fairly clear to everybody in the field now. The thing we have to get right to get like tremendous value out of these underlying Smart Systems. Yeah, so agents, uh, if we were to explain them, these were like Wizards. I think in the early Windows days you would create a w, you take a, you would create either a wizard or a business process, and all of this is done offshore. 

Business process Outsourcing is a big part of this. Yeah, people send their accounting to India, they send their data entry to Manila, whatever it happens to be. And this is just taking that same concept and instead of doing it with just Brute Force humans, uh, offshore millions of them working, uh, in lower income or lower cost of living, uh, locations. The person who's in the US doing their desktop can just have it happen in seconds. Huh, it's a really easy way. 

Yeah, so our first step as a company is we've been working on some of these capabilities that let you as a worker every day just delegate um these tedious tasks. But where we're really going with this and what I'm and why uh what I'm excited about is being able to do tedious tasks is actually just a building block for what's even more valuable, which is effectively having an AI teammate that you can talk to at work, bounce ideas off of each other, get guidance. It has the same context as you do because it's uh, it's it it sees all the same.

Stuff on screen your and uh and your what you do at your company and all that stuff. And then like help helps you brainstorm and come up with the best ideas and maybe try some of them. And you're like well maybe this one's a little bit better. I think all of that lies on top of a foundation of being able to do arbitrary things arbitrary tedious things on your computer. 

Close are you to having how close are you to having this in market? Is this like in beta somewhere? Are people using it yet? Yeah, so um you know what's been interesting this year is that um is that the agent space has really suffered from reliability problems. If you go look at the space as a whole I think there was an information article about like the agent winter or something like that. It's because most of these systems, the ones built on top of gpte, are like 60% accurate. Like they work 60% of the time. One out of 10 times maybe deletes half your records in Salesforce and you're like I can't use this at work. 

And so uh with the depth we spent the whole summer unlike everybody else training our own foundation models in house that deeply understand the pixels on screen are tuned for generating actions. And that took us to a point where this fall we now have actually very reliable agent models. Once we have some custom fine-tuning data per use case, so um we're excited to announce there will be later uh an announcement.

About um one of our really large first deployments. But we also actually last week took everything we built for our Enterprise customers. Cut out the specific fine-tuning uh that made them their stuff super accurate. And then just made it into a Sandbox anyone can play with. 

And so this thing is called adep experiments. You can check it out at ad dep.

Aexperimental um funny thing is actually this morning uh someone sent me a post on Upwork hiring people who were experts in using Adept experiments to automate workflows. So it's it's already getting some very cool attraction on on the agent space now. 

So the idea is in you're going to build this platform but we as people operating businesses will make our own agents and for ourselves or we're going to make agents and have the ability to publish them. Yeah, so our main as a company um we're actually very Enterprise focused. 

So we're currently doing larger engagements where we just come in and work with our company and figure out how the Adept agents could just accelerate the knowledge work that's happening there. But um this experiments framework that we made is an is an easy way for you or me or any of our friends to pick up and just try what it might be like to go automate something. 

And then you can publish them and share them with people like uh CH GPT is doing H soon. Yeah, we'll soon be able to let you share. Uh so far that didn't quite make the MVP and so that becomes a business model like an app store in your mind. 

So if I am uh really good at accounting I can kind of make these tools build the classifier engine for where it should live or whatever and then publish it to the web and maybe share Revenue with you. Is that part of the model? Uh it's not the focus of our model the focus.

Of our model is these like making Enterprises really successful. But um what's really interesting about what you just said though is that like we hope that in an Enterprise setting you know um oftentimes like custom knowledge is locked in some people's brains. Right like there's there's one person well at our company there's like maybe three people that know how to config this particular infrastructural dashboard. They could just teach a Depp how to do that and just publish that workflow to everybody at the company. So whenever that needs to be reconfigured you just hit play on that and it does it for you. 

So I'm excited about sharing in that setting. All right you've heard me talk about ARising Ventures a bunch recently. They are a holding company. They acquire Tech startups that uh you know are facing some headwinds some setbacks and they give these startups a second chance at life which is awesome. So if you're going through some tough times right now and you're trying to get back on solid footing will reach out to the team at ARising Ventures. Could be just what your startup needs to get back on track. They've helped companies like UpCounsel. UpCounsel they took from burning a million a month and shrinking to profitable and growing. 

And Jive where they relaunched a shutdown company went from Z to 1 million in ARR in just 5 months. What a save in fact two saves. Listen Rising.

Ventures knows what Founders care about because they're not Bankers. They are Tech Founders themselves and they're here to help your startup get back on track. Learn how Arising Ventures can help give your company New Life by visiting arisingventures.com twist today to learn more and connect with the team that's arisingventures.com twist.

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

Adept AI Video