This Week in Startups - Exploring AI's Pace of Evolution, AGI's Future, and Data Dominance with Adept CEO David Luan | E1855

Starting point is 00:00:00 How far away are we from AGI in your mind? I turn on my computer in the morning. I go to work, you know, whatever I happen to name my assistant Joe, and I say, Hey, Joe, what should I work on today? And Joe says, well, you know, looking at your email box, there's seven companies that have acute issues in your portfolio. And these three probably require a phone call. These four, you probably need some more information based on what I've learned.

Starting point is 00:00:23 So I'm going to send them requests for this information and then schedule them for tomorrow and Wednesday. Is that okay? And it just like kind of tells me what I'm doing for the next two days. When will that happen? I was going to ask you how you were going to define AGI. And if you define AGI by just what you said in that flow, I think that we'll be at a spot where you would be able to get that within the next one to two years.

Starting point is 00:00:44 But would you trust the recommendations? This week in startups is brought to you by Brave. If you're building AI and search-based applications, train your models with the Brave Search API. Get started for free at brave.com slash Jason. Arising Ventures is a holding company that acquires tech startups facing setbacks. Arising Ventures knows what founders care about because they aren't bankers, their tech founders themselves. Go to arisingventures.com slash twist today to learn more and connect with the team.

Starting point is 00:01:19 And LinkedIn jobs. A business is only as strong as its people and every hire matters. Post your first job for free at. at LinkedIn.com slash twist. Hey, everybody, welcome back to this week in startups. We're continuing our in-depth coverage of AI. It's moving out a crazy pace, and you're in for a treat today because we have David Luan, who is the CEO and co-founder of Adept AI.

Starting point is 00:01:43 Previously, David ran research and engineering at a little organization known as OpenAI back in 2017. He left there to go work on large models at Google, where he focused on Google Brain. And if you're wondering what Adept AI is, we're going to learn a little bit of about that today. But basically, they're building a machine learning model that can interact with everything on your computer. Welcome to the program, David. Thanks so much for having me. Okay. So you are not the highest profile person at Open AI, but you were a very key person. Maybe you could explain by background what you worked on at Open AI because it's pretty darn impressive.

Starting point is 00:02:19 Thanks. So my time at Open AI was really engaging and fun. I knew a bunch of the core researchers there from just the very tiny machine learning research community from back in the day. The analogy I always like to make about how ML used to work is, is imagine a world where the flat earthers toiled in obscurity for decades and then turned out to be right. That's basically the story of deep learning. And so that community of deep learners are actually pretty small. And I joined opening eye when it was about 35 people and ultimately grew it to about 135

Starting point is 00:02:49 before I left. And primarily folks in my org covered basic research. So things like QPT2 and Clip and Dali all the way to the supercomputers and some of the larger, some of the larger scale up efforts there as well. So when you were building those language models, maybe you could talk a little bit about what they were trained on. I know there were collections of data sources. Like there's the open crawl of the web. There are image libraries that were put together. How was that original dataset organized when you were in that like two and three phase?

Starting point is 00:03:23 So the thing about GPT2 that I think most people don't recognize as being two of its core contributions, the first one is actually not data set related, but just real quick. It's this idea that every single natural language understanding task could be reframed as simply writing more text. So historically, people were training these models for like sentiment analysis of tweets and all stuff. And you're training a model with the input is the tweet and the output as a score of as a positive or negative.

Starting point is 00:03:49 Then you just end up with this constellation of models that all do different things. But GPT2 said, we can just boil. was all down to one objective, which is like write more text and the next word is, is it a positive or negative tweet and you get it right. And the reason why that works is actually because of the data set. So historically, people training language models use things like common crawl, as you mentioned, which is like effectively, like it was initially made for making open search engines, right? It's just all the websites you can find on the internet. But most of them are trash. Like we looked at my colleague Alec Radford, who was the lead author and I like looked at all the data and

Starting point is 00:04:20 there would just be web pages and web pages of thousands of product. for Sony cameras and stuff like that. And the core insight that Alec had was that we actually live in a world where the open internet has given us these amazing tells as to whether or not an underlying webpages is smart or not. And that's called Reddit. So what he did was he scraped Reddit and found every single Reddit URL that linked to an outbound blog post or website or whatever that had more than three upvotes and said,

Starting point is 00:04:49 well, three humans said this thing was good, therefore it's probably pretty useful. So a human-powered search engine, exactly. Was the core here. You used humans directing to, or what you assume as humans, right? Yeah, it could be some spam or bots. But generally speaking, it sounds like a really good idea. And you use the three-vote-up mechanism to filter it even further. And so then you scrape those pages, you build the language model.

Starting point is 00:05:14 And that's what powered Gip-2, and that's why GPD-2, despite today looking tiny, was so smart for its time. Fascinating. And then what about Dolly and all the images? I know stable diffusion, there was built off of some collection of images that a lot of researchers use. So maybe you could speak to those collections of images and how that all were. So Dolly was interesting. On the dataset side, there was actually not as much, there was not as much sort of, this is not to discredit Dolly at all. Dolly is an amazing project, but the intelligence didn't come from the dataset side. It actually came from this fascinating thing. where this guy, Ditya Ramesh, owes an awesome researcher at OpenAI, he came up with this special trick that let you predict discrete codes

Starting point is 00:06:01 that correspond to images, even when theoretically, you should only be able to predict continuous ones. So it's like a very niche fact. Dali 1 architecture looks very different from the Dali 2 architecture, but I actually think that the Dali 1 architecture was particularly inspired in letting us do sophistication

Starting point is 00:06:17 at that level way back in 2019. And so where did the images come? from and how did that work? I'm curious. Yeah, just there was, I'm actually not sure how much I can talk about exactly where those images were sourced. Yeah, but there was, there was, we also licensed some data sets, which also made it easier. Ah, so I guess there, there is a little controversy there of like training of data sets. Just generally speaking, not talking about your time and opening eye and specific use cases, a lot of models that are being built on hugging face, a lot of open source

Starting point is 00:06:47 models, they just crawl the open web and they're trained on whatever it can find, yeah? A lot of the open source models are built that way. Yeah, it's like hackers trying their best to get their hands on whatever. And if they get it in their minds, totally fine, let's build the model. The challenge has become, once it becomes a corporate entity, like Open AI or the stable diffusion corporation, which is called stability. Stability, yeah. Then all of a sudden, the lawsuits come out and people are like, hey, you train my model with this. So maybe you could speak to what that means in terms of how this will all play out with regard to

Starting point is 00:07:21 such an interesting rabbit hole where that will all work out in terms of an advantage. So if you do this as part of OpenAI with Microsoft as a partner, those are two very big targets. A $90 billion company, a trillion plus valued company. It's a big target for lawsuits. And then stability, obviously, raise tons of money.

Starting point is 00:07:41 That makes it another big target. But open source, who are you going to target? A bunch of open source, you know, handles that, you know, may or may not have built this stuff. So is that going to give like this huge advantage to the open source community? I actually think a lot of the players in the open source community have been very buttoned up and forthright about how they're handling a lot of this stuff. So if you go see some of those models,

Starting point is 00:08:02 they're actually explicitly licensed under a Creative Commons non-commercial license. So there are ways in which people recognize that, you know, some of this data actually should not be used for commercial purposes. And the people who are then breaking the rules are then going downstream of these models that were actually in fact trained relatively responsibly. I think what's interesting right now is we're moving towards a phase where with the public internet closing down a little bit, right? Everyone building their own wild gardens, like platforms like Twitter, etc., making it harder

Starting point is 00:08:31 for models to be trained off of. Plus, this like increasing fragmentation in the internet ecosystem between the West and the Chinese internet ecosystem, I actually think that access to trainable clean data is going to be the number one problem, right? You make these models bigger. Every time you make them 2x larger, you need to scale the amount of training data by a similar multiplier. And people in the field are concerned will eventually run out of tokens or training data. That's why people are looking into can we learn from YouTube, right?

Starting point is 00:09:00 But ultimately, the models are one fascinating property of these models is that their maximum intelligence level, the way that LLMs are being trained today. And LLM's maximum intelligence level is like really, really, really roughly rule of thumb. the maximum intelligence level of the smartest training data in the corpus. If you want to get better at stuff, you need to be learning from smarter and smarter behaviors from humans. Got it. When you look out across the datasets that are out there, YouTube, pretty powerful, large

Starting point is 00:09:30 data set with their transcripts, and you have images involved in it. But as you're saying, who knows the Providence, copyright, all kinds of issues. Twitter filled with a lot of bots, also kind of staccato, but very much up to date. so that's pretty cool. Then you have Reddit, which has been baking for a long time. You got things like Cora, which has been baking for a long time, lots of experts on it, Stack Overflow. Where's the great wells, if we looked at these like oil repositories,

Starting point is 00:09:59 who's the Saudi Arabia, who's the UAE, who's the guitar, who's the Norway, you know, Texas of having Venezuela, Canadian salt flats. Walk us through like when you're, when researchers and, you know, people are building these things and building. models where they say, oh, this is the oil, this is the diamonds, this is the good stuff. To me, I think it all depends what you want to do, right? And what, you know, there are companies building fun chatbots, like the meta characters

Starting point is 00:10:25 thing, character.aI, all that stuff. You're going to want very different data than if you're doing what we're doing at Adept, which is how do we build like enterprise systems that help people be more productive at work, right? And so for us at Adap, the thing that we care about more than anything else is how How can we learn from the smartest knowledge workers in the world? And if you look at where the knowledge of the smartest knowledge workers in the world sits, it's actually never on the public internet.

Starting point is 00:10:54 And because of that, I think for Adept to be able to build things that let any end user teach adept a new skill at work in a very small amount of time, that's the kind of stuff that we really want to be learning from long term. Yeah, because what you do on your desktop or at your job is not published to the web. It might be in Slack. It might be in Microsoft Teams. It could be a notion. It could be in Koda.

Starting point is 00:11:14 It could be in a Google, whatever their Google repository. But you're never going to crawl that. Nobody would ever trust you if you crawled that. Yeah. Are you building the next great AI product? Well, if so, you know how expensive APIs can be for their model training data. Training AI is pricey. That's a fact.

Starting point is 00:11:31 We all know it. So you have to try the Brave Search API. Yes, I am talking about Brave, the privacy browser that I am obsessed with. Braves browser has 65 million users. Think about how much data that drives from. Brave Search, which is the only global scale independent search index outside of big tech. And that index is available to anyone with the Brave Search API. The Brave Search API can power your chatbots and train your models, inform answers to real-time queries, and it will serve images, web results,

Starting point is 00:12:00 and even rich text snippets. The Brave Search API features an easy, intuitive data structure, and its data is populated by real human interaction, not web crawlers, all for a fraction of the cost of the major players. free for up to 2,000 queries per month with paid plans for as little as a $3 CPM. That's cost per thousand. So if you're building a next gen AI app or chatbox, you got to try the Brave Search API. Get started today. Brave.com.com slash Jason. Oh, Jason's brave.

Starting point is 00:12:28 I like it. Brave.com slash Jason and get the browser while you add it. It is awesome. It's also got a VPN built in. That's pretty cool. Why don't you show me what you're building at Adept AI Labs? And thanks for the little diversion down history land there. So that's super fascinating.

Starting point is 00:12:42 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 flip through some demos of a release we actually recently did last week. The North Star for Dept from day one actually has been that in the long term, the thing that be the most valuable thing to build for work is an AI agent that does much more than reading and writing and drawing images, but can actually handle for you arbitrary work tasks

Starting point is 00:13:08 and workflows. 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 deeply understand not just the 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 to work on this training this model that could do anything a human can do on a computer. and we've been building effectively a product that enables knowledge workers to arbitrarily delegate tasks to the system. Here's a quick demo example.

Starting point is 00:13:50 In this case, let's say your response for paying invoices and you get your plumbing invoice. You fire up a debt that pulls up the invoice. It's all being done by the model right now. Pulls up the invoice. Reads the pixels in this PDF, realizes what it's about, store some interesting facts about this, and then pulls up QuickBooks and then correctly enters who is the payee, right? Savant plumbers, like, how did you pay? What category is this? Like, and it realizes the category was never written in the PDF, but it realizes a plumbing invoice, so it should go into

Starting point is 00:14:17 repairs and maintenance. And this task that you probably would have had to do like 10 or 20 times a day for your job, you show a depth how to do it once. And now every time you get a new email invoice, you just fire up a debt and it handles this task for you. It's all power. So where does it depth live? Is it in your system try there? Or how does it intercept this coming in by via email because you have this invoice come in via email. You got to get it paid. You're in purchasing or accounting. Boom. It needs to know. So was 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. And so I think the key of this is we're not forcing you to use this brand new system.

Starting point is 00:15:01 It's a helper for all the same workflows you are. It's a co-pilot. Yeah. Exactly. It's a co-pilot. Yeah. So it's sitting going to sit in the system tray, but right? right now it sits in your browser window or something. In your browser window, yep. 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.

Starting point is 00:15:22 And a lot of knowledge worker jobs is just doing that like ad nauseum, right? We had a, we're talking to some customers who 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 email from someone who's filing a claim. And Adept basically, once you show Adept how to do it once, automatically fills out all and populates all the forms involved there. You know, what's interesting 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

Starting point is 00:16:01 coming up is it's going to be about multimodality, which is the ability to, 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 series 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, if we were to explain them, these were like wizards, I think, in the early Windows days, you would create a, you would create either a wizard or a business process. And all 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.

Starting point is 00:16:43 And this is just taking that same concept. And instead of doing it with just brute force humans, offshoreed, millions of them working in lower income or lower cost of living locations, the person who's in the U.S. 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 these tedious tasks. But where we're really going with us and what I'm excited about is being able to do tedious

Starting point is 00:17:16 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 sees all the same stuff on screen. and what you do at your company and all that stuff. And then like 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,

Starting point is 00:17:44 arbitrary tedious things on your computer. How close are you to having this in market? Is this like in beta somewhere or people using it yet? Yeah. So, you know, what's been interesting this year is that, 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.

Starting point is 00:18:04 It's because most of these systems, the ones built on top of GPT are like 60% accurate. Like they work 60% of the time, one out of 10 times maybe it deletes half your records in Salesforce and you're like, I can't use this at work. And so with ADEPT, 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

Starting point is 00:18:30 once we have some custom fine-tuning data per use case. So we're excited to announce. There will be later an announcement about 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 that made them their stuff super accurate, and then just made it into a sandbox anyone can play with. And so this thing's called Adept Experiments. You can check it out at adept.a.i slash experiments.

Starting point is 00:18:58 And it's a super powerful first toy automation tool. The funny thing is actually this morning, someone sent me a post on Upwork, hiring people who were experts and using Adept experiments to automate workflows. So it's already getting some very cool attraction on the agent space now. So the idea is 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, we're actually very enterprise focused. So we're currently doing

Starting point is 00:19:34 larger engagements where we just come in and work with a company and figure out how the adept agents could just accelerate the knowledge work that's happening there. But this experiment's framework that we made 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. share them with people like ChetipT is doing? Soon, yeah. We'll soon be able to let you share. So far, that didn't quite make the MVP.

Starting point is 00:20:01 And so that becomes a business model, like an app store in your mind. So if I am 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? It's not the focus of our model. The focus of our model is these like making enterprises really successful. But what's really interesting about what you just said, though, is that.

Starting point is 00:20:23 is that like we hope that in an enterprise setting, you know, oftentimes like custom knowledge is locked in some people's brains, right? Yeah. There's one person, well, at our company, there's like maybe three people that know how to configure this particular infrastructural dashboard. And they could just teach a depth 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.

Starting point is 00:20:43 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 star. startups, that, you know, we're 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

Starting point is 00:21:02 and you're trying to get back on solid footing, we'll reach out to the team at a rising 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 zero to one million in A-R in just five months. What a save. In fact, two saves. Listen, arising 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 slash twist today to learn more and connect with the team. That's

Starting point is 00:21:46 arisingventures.com slash twist. Your shortlist of ideal people to use this in the first couple of years is who? Operations people, CEOs, sales teams, who are you targeting first? Because obviously there's many code co-pilots. You're not going to compete in that space. You're not going to beat, you know, GitHub or whatever. So who are you targeting? What jobs will become 30, 40, 50% faster? Yeah, we're really targeting right now operations. So like all those examples we showed earlier, processing invoices, dealing with tracking things, shuffling data from system to system B, customer onboarding, all of this, there are areas where you just got thousands of people who are spending their time,

Starting point is 00:22:31 like, instead of handling the higher level goals, just handling some of this low level flow. And I think, like, the reason why we really want to do this is because, I mean, like, so many of us spent half our waking hours at work, right? And if that time is, like, reinvested not in, not in more manual computer process stuff, but instead in like talking to customers or working on the next engineering project, I think it's a really big unlock. You can move up the stack. I've been dealing this with my investment team and I'm saying like, I wonder what low level things we do every day. We could eliminate or outsource. So, you know, really trying to figure out how do you automate it with AI? How do you delegate it? Offshore workers tend to be the work from home, remote workers in lower cost places. Or how do you deprecate it?

Starting point is 00:23:17 So I call this my ADD framework that I'm literally putting my company through. Automate, delegate, deprecate. Look at everything you do every day. And then if you do that, well, then you could call a founder on the phone and have a conversation with them, how things going. Or you could go meet the next founder that we might invest in. So there's really something here, I think, in terms of, you know, teams doing more with less or doing more important work. And that is super exciting. Absolutely. I think it's like, you know, there's definitely both of those, but sometimes we're hearing from customers, other value props I didn't even think about, right? Like in the customer onboarding case, like ADAPT can help cut down even time to revenue, right? Which is a metric I never thought that the stuff we were doing would actually impact. So time to revenue, hey, when we discover somebody could be a customer and when we then go do this, you know, close them as a customer. So explain to me when you pitched folks, you raised a ton of money, hundreds of millions of

Starting point is 00:24:16 as AI companies have been apt to do now. When you pitched investors, the argument I would have is, hey, there are verticalized solutions. So HubSpot, Salesforce, Slack, Notion, Superhuman, they're all looking at, hey, email's coming in and superhumans looking at, how do we respond to this and how do I draft your email? Outlook is doing that. Gmail's obviously doing that. So, and then if you're a HubSpot, you're going to be, I've seen Darmesh all day long, on Twitter talking about how he's automating HubSpot, it's going to go find you your next lead,

Starting point is 00:24:51 it's going to craft content for you. So how are you going to do a better job versus verticalized at-scale software companies? And why? I think this is the key question. So I think the most interesting pattern that we've learned from just observing lots and lots of people do work is that they use a million different software tools every day as part of their job. Right. The average knowledge worker uses something like 17 different software tools. And the most powerful and crucial workflows to those organizations are usually ones that span those different tools. And so one, workflows that span different tools, but two, custom workflows to that particular user, right? But even just taking Salesforce, for example, every company's Salesforce deployed looks actually can often look very different from each other.

Starting point is 00:25:35 And so there is no one size fit all thing you can type into a little text box. You'd be like, hey, I want you to go do this thing. Because even how you add a lead can be very different from company to a company. And so the power for ADEPT is recognizing that we should be focused on the highest value workflows and how an user could teach ADAPT a new workflow really quickly. And the second thing is that we should be focused on workflows that span many, many different software tools. So like other examples, like there's a, there's a case where someone wants to use ADAP to go do market research every day, right, like on the state of various different housing markets. So they're pulling up like Redfin and Zillow and stuff like that and just running queries and populate. them into a spreadsheet.

Starting point is 00:26:14 Like, whose job is it to make that happen in the vertical eye side? Is it Redfin's job? Not really. Is it Google Sheets job? Definitely not, right? Those are the types of things that we see lots of value in. Is that going to happen in my browser on my desktop? Or are you going to make me, you know, my researcher in the cloud and fire up a headless

Starting point is 00:26:33 browser and then just have that, have literally a virtual desktop on my desktop. And I watch my worker, my AI, you know, research. slave go through and do analysis 24 hours a day on, you know, properties on Zillow and Redfin and put it into documents for me. It could be either. We're doing the former right now, but we have had customers ask us, hey, like, why don't you just go spin this up in the background and we'll just monitor it? Yeah.

Starting point is 00:26:59 So both work. There's no, like the hard part about getting a debt to succeed is, is not in any of the scaffolding bits, but it's how the heck do you make reliable models like LLMs that read the screen, decide what to act on next and do that reliably. And so almost all of our challenge comes from that. How are you going to charge for this? Just be $100 a person per month, $1,000 a person per month. How are you thinking about it?

Starting point is 00:27:25 Yeah, we're actually seeing tremendous, like most, you know, when we started the company, I think it was a really valuable lesson for me. I always thought that I knew what this model was going to be and that we should just build for that model, which is choose a couple hundred bucks per seat and like upsell people and do enterprise capabilities down the line. What's actually happened is from the get-go, there has been so much enterprise demand, and they have know exactly what use case

Starting point is 00:27:53 they want to go deploy this thing in, and they are willing to sign up for relatively high ACV things off the bat. So we've said, let's just go do that. And down the line, when everything becomes more stable and mature, let's pull little chunks of that out that we can then monetize in a more proceed sort of setting. Got it. So you can go to a company and they really care about their accounting and purchase orders,

Starting point is 00:28:17 whatever. And you can say, hey, just give us $250,000 a year. We're going to eliminate 10 jobs or we're going to make everybody 10 times faster, whichever, however you'd like to look at it. I mean, it's all the same thing. You either, you don't have to hire any more people because even as you grow, people will just be 30% more effective a year. So what are the gains you're seeing per employee in the early tests?

Starting point is 00:28:39 What are your customers telling you they're seeing in terms of gains? because you're going to be able to charge more if you could make people more productive. So what are the gains like? Yeah, we're really focused on making people more productive and also in all sorts of like side objectives of making people more productive that we didn't expect, right? Like decreasing error rate or making it possible for more people at the company to go do a task, right? There's an interesting demo that we have where like Shopify is pretty easy to use.

Starting point is 00:29:07 But like as an admin, there are some things that like you might know how to do that like you want to make it really possible for anyone in your marketing organization to be able to tweak, for example, right, even though they're not typically the people who go, who go do that. And you can just teach a depth how to do it once, right? So it's like things like that. They're also really interesting side effects. So the time to revenue example that I mentioned earlier. That's the kind of stuff that were that we're really focused on. If you were to pick a number in the early tests of how, what percentage more efficient people were, would it be 10% a year, 30% in the early tests?

Starting point is 00:29:39 So I think it depends on the on the task. asking we'll probably be more, we're going to, well, we'll probably publish a case study at some point on this where we'll have more details, but like we're hearing things like a workflow that might take somewhere an hour and a half goes down to 30 minutes, for example. Got it. Okay. It's still very much something where we like, our philosophy for how to get the general intelligence involves agents, but it also involves a lot of human oversight. And, um, and so we've been the whole time, we only basically build human and a loop systems where you know what the models are doing. I think this is a fine way to look at it.

Starting point is 00:30:11 businesses are growing 30% year over year, 20% year over year, like Uber grows or Microsoft or whatever. They grow 10, 20, 30% year over year. And their teams now, this last year, and I think a lot of it had to do with AI and also people getting fit and maybe not hiring ahead. Their teams went down in size and the revenue still went up 30%. So I think what's going to happen is it's not job destruction or elimination. I mean, sure, some jobs will go away because of AI. that'll probably be a good thing because there'd be menial tasks that are arduous.

Starting point is 00:30:44 But you can have the same team size. And instead of having to add a thousand people every time you add, I don't know, $10 million in revenue or $100 million in revenue for some big group, you're going to be able to do, because the whole team can be 30% faster, 50% faster at these repetitive tasks, maybe you don't have to add anybody. So the company stays the same size, but yet can do more. Is that your thesis as well? I think giving, yeah, giving knowledge workers and companies,

Starting point is 00:31:10 lots of leverage through these systems is really the focus. All right. Congrats to the team at LinkedIn. They just completed their march to one billion users. So what does that mean for me and you? Well, we all know. Startup game is rough, now more than ever. And you need great team members to compete. Don't I know it. You need team members you can depend on. And there are so many great employees out there ready to interview for your job. And with a billion users, LinkedIn jobs has the best candidate pool out there. Hands down, bar none. And you can land. both active and passive job seekers. The active ones you know about, hey, they got laid off,

Starting point is 00:31:44 their startup shut down, they're actively looking for their next adventure. But what about those passive job seekers? The ones who, they like their job, maybe their boss is a jerk, maybe they're been there long enough and it's time to move on. Those are the passive job seekers. Some of those are the best in the world because they're highly sought after and they're not actively looking, but LinkedIn jobs will put your opportunity in front of those passive and active job seekers. So use LinkedIn jobs to find your next amazing hire. Go post an open roll on LinkedIn and you'll be 100% certain that you have access to the most qualified candidates available in the world. In fact, according to LinkedIn, 86% of small businesses get a qualified candidate within 24 hours. That's one day or

Starting point is 00:32:25 less. And guess what? First job listing, it's on your boy J-Cal. LinkedIn jobs helps you find the most qualified candidates you want to talk to and they do it faster. Post your job for free at LinkedIn.com slash twist, it's LinkedIn.com slash T-W-I-S-T to post your first job free. Terms and conditions do apply. What will this look like in five years if you're successful? So this is the part that I'm most excited about. I think as we were talking about earlier, the first part of what we're doing is enabling you to delegate things you don't want to do, right?

Starting point is 00:32:57 But I think where this really heads is as these agents become smarter and smarter and become better and better handling higher level things, right? Maybe right now it's like, hey, like, here the steps required to do this invoice processing thing. But maybe in like two or three years instead, it's like, I want to think about what I want to do in this part of the business. Let's figure it out and let's plan some scenarios together. Like, that is the interaction model that I think is going to be extremely powerful. And, you know, my personal background was I was always working on AGI. Right.

Starting point is 00:33:24 At Open AI, that was our North Star. At Google, when I was leading a large models effort there, we were really thinking about how do we scale up these underlying models and combine them with the other things we need to do to get smarter systems with adept. And we are building a super commercial company with a product that enterprises use. But the reason why we do this every day is because we actually think this particular path of building AI agents that can do smarter and smarter things for you at work that are interacting with and learning from the world's best knowledge workers and learning not just how to read and write, but the consequences of doing things that have a reward is actually

Starting point is 00:34:01 the critical path for getting to generally intelligent systems. in sort of the most, the most predictable way. And so what I expect to see from the adept product is that the abstraction level what you can ask you to do will continue to get higher every year. Absolutely fascinating. And how far away are we from AGI in your mind? I turn on my computer in the morning, I go to work, and, you know, whatever I happen to name my assistant Joe, and I say, hey, Joe, what should I work on today?

Starting point is 00:34:29 And Joe says, well, you know, looking at your email box, sounds like, you know, there's seven companies that have acute issues in your portfolio. And these three probably require a phone call. These four, you probably need some more information based on what I've learned. So I'm going to send them requests for this information and then schedule them for tomorrow and Wednesday. Is that okay? And it just like kind of tells me what I'm doing for the next two days.

Starting point is 00:34:55 When will that happen? You had to pick a date. So that particular flow you just said, I was going to ask you how you were going to define AGI. And if you define AGI by just what you said in that flow, I think that we'll be at a spot where you would be able to get that within the next one to two years. But would you trust the recommendations? I'm not sure we'll get to a point. Okay, let me say with the recommendations being as good as an MBA from a school and someone who gets paid 100 or 150K a year.

Starting point is 00:35:26 In other words, if, you know, to make this like a classic test, I wouldn't be able to tell the difference between its. requests and an MBA who was a chief of staff. So, you know, $150,000 a year chief of staff who crushes it would give, I wouldn't be able to tell. The AI from the chief of staff. I think with that particular flow you just said, I think definitely less than five years. And I think five years is conservative. Wow. So this idea of a chief of staff being able to watch an executive work and fill in all that connective tissue and advise them what you know where their attention needs to be just done completely by AI I think you're a couple a couple of suggestions of which maybe one out of three or two out of three hits I think that's yeah sub five years away yeah I mean the chief

Starting point is 00:36:16 of staff might give you five suggestions and you say like okay we're going to go with these three and the this one I would never do but thank you for suggestion here's a learning thing and this one you will consider it but let's put it on the not right now list yeah pretty amazing yeah I think the pace of progress, you know, I think right now, the field is still split between people that are like, wow, like, I see how the stuff is going to keep compounding. And then people who are like, well, just because the last three years has been crazy, it doesn't mean the next three years will be crazy. Capulability is slowing down.

Starting point is 00:36:46 Models aren't going to get too much smarter anytime soon. I think the first group is correct. I think we're going to still see tremendous progress over the next couple of years. Awesome. Well, this has been absolutely amazing. This was really fun. Thanks for inviting me on this. Like great questions.

Starting point is 00:37:03 And yeah. I mean, I've been talking to everybody and it's really interesting because I've gone down the agent rabbit hole a little bit and watching the desktop rabbit hole. And I think you're really on to something. And I just, I think it's going to just work. And then the question is, what can you charge for it? And, you know, which verticals can you actually carve a niche? I do like the answer to your question.

Starting point is 00:37:25 The answer to the question of siloed versus across your entire desktop. I think there will be excellence inside of superhuman or other apps. That'll just be amazing. You open up your notion. Here's what you missed. Yep. But then there's going to be a moment where it's like, oh, here's what you missed in notion.

Starting point is 00:37:40 Here's what you missed in Slack. Here's what you missed in Salesforce. I took those three, here's what you missed moments. Yep. And pulled them together. And I pulled them together. Here's what you missed in totality. Exactly.

Starting point is 00:37:52 They won't be substitutes for each other. They will actually both coexist very happily. 100%. I think the key with the agents thing is just. getting them to actually be reliable. And I think that's the key advantage that we're really trying to run at is you've got to control the whole model stack to do that. All right, everybody.

Starting point is 00:38:11 We'll see you next time on the swing and service. Bye-bye.

This Week in Startups - Exploring AI's Pace of Evolution, AGI's Future, and Data Dominance with Adept CEO David Luan | E1855

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.