Sina Yamani on Large Action Models and AI that executes online tasks

0:00I'm Ashton Addison from the Crypto Coin Show and today on blockchain interviews we have Cena Yammani, founder and CEO of Action Model. Here to discuss AI, large action models. Why it's important to have community-owned AI and where AI is headed if we don't have decentralized versions of this. Cena, thank you for taking the time.

Pleasure is all mine, Ashton. Thanks for

0:27For having us. Yeah, excited to dive into this because it's a super important topic. Not only does almost everybody I know use AI, they might be wiped out by it as these centralized giant companies start using it to take over people's jobs. And we're already seeing it happening a little bit, but I feel like the acceleration as we get into

0:49automation and actions that the AI can take, it's no longer ask a question, copy paste, I do it manually. It's like the AI does it all and I've been taken out of the equation. So, I know that you've been deep in this for many years and in crypto. I would love to start out with what you and your team have built at Action Model. What is the difference between LLM and large

1:14action models and why it's so important to have community- based AI?

Yeah, absolutely. So if we think about AI and where it's been over the last few years, we've had LLMs, large language models. These are essentially the chat models, right? And the way they work is you speak to it and it gives back an answer. So we've been in this advice era for the last few

1:39years, you know, you tell it, I want to do this, rewrite this for me, do that, and it will it will write that structure out for you and then you have to go and do the work, right? So that's kind of the chatbot style. We've obviously got images and videos now as well which are all you know frighteningly good. And the next stage is that the AI actually does things for you, right?

2:03It's not the fact that you can ask it to draft you a contract or an email. It will actually go and do that work too. You know, it will read through your emails, make decisions because it understands your business and it will respond. it will then follow up, do whatever work was required, draft the actual contracts, literally everything. [snorts] And so this next wave of

2:26technology is very disruptive because it's pushing past the I'm a cool tool for the person to use to speed up their work to does the employer actually need the person at all anymore, you know? Does the does the company need all of this staff when you have these incredible AI employees that can do everything? Yeah, I know it's it's getting crazy and we're starting to see hints of it in

2:55these large langu language models. I'm sure they're going to adapt with you know some automations I've seen inside of ChatgBT. I think Claude has things that they can now sort of start taking actions for you. We've seen the Claude bot come up recently. But maybe we can dive into, you know, why it's not such a good idea for Chad GBT's automation to be taking over and to look

3:23at the decentralized or community-owned perspective of this and what happens if the same sort of fang companies of AI just control the whole AI space. Well, I mean, it's very obvious why it's a bad bad thing that can happen. I don't need to really explain that. The situation is that there's about a billion people around the world that are employed to work behind a computer

3:49screen, some kind of white collar job. And these aren't necessarily very technical people. In fact, a lot of it is repetitive tasks, data entry that can easily be automated. And AI can do all of this stuff today, right? The next stage in the next few years is enabling the AI to be able to understand what is the business and their needs and then it

4:15will go and do all of that work. So you know if you're the employer, right, the majority of the world and employers, they all want the same thing. They're just trying to make more money and be more efficient. And when these tools pop up and the tool can do 95% of the work for 10% of the price, it's a no-brainer for these businesses. So unfortunately,

4:42unless the government acts right now and puts deep and strengthy regulations in place, then this is inevitable. It's it's going to happen. Mhm.

And as it stands, almost all of that value transfer is going to go to these big tech companies. We're talking about a handful of trillion dollar companies that will grab the value of nearly a billion people over the next 5 to 10

5:11years. You know, even this CEO of Anthropic said, I think it was a few weeks ago, he said that half of white collar jobs will be gone within five years. And I personally, I think that's an understatement. I think if he said what he really believes it would be a much more frightening statement.

so you know we know what happens when individuals we're talking about the

5:37wealthiest people in the world these billionaires all these trillion dollar corporations own and control so much of the world and this is you know this is like 20% of the planet we're talking about here. So it's clear that this needs to be either regulated in some way or it needs to be in part owned by the people.

Mhm.

And that is what the action models focus

6:02is around.

Yeah. No, I agree. And obviously being in blockchain for 10 plus years and yourself as well, that's sort of been the goal is move our finances owned by the people for the people for yourself. when we translate that to AI in what the team has built at action model. So from what I understand these LLMs are also implementing action as

6:28well. So what's the difference here in how action model can stop you from losing your job or you know stop this transition from from at least take a piece of the pie from these corporations laying off a billion people? It's a good question with a pretty technical answer. And essentially these LLMs, they output text. That's the main thing they can do. And

6:55and the way they work at the moment with regards to kind of connectivity with applications is all through APIs.

And fundamentally around 1% of the internet has APIs available. That means that 99% of the internet is only accessible to people who use the mouse and the keyboard. And so that's these LLMs. The next generation of AI actually uses the mouse and the keyboard. So it

7:24can use the computer just like a person does. And that means that it doesn't need to integrate with anything. doesn't need access because the platforms, applications, websites, they have no idea that this isn't a person because it acts and navigates through the internet just as a person does. So, so they're becoming just as capable as as as people

7:49are.

Yeah, that's a great point about the APIs. If the AI is in the operating system and it's moving the mouse, it doesn't need to work on APIs and that might have to shift the business model of some of these centralized AI platforms. So can you explain the community- based part of action model? Can I go and on action model and you know it's very similar in appearance

8:17to an LLM but I own the data. How does it exactly work? So the action model is a community initiative. It's basically around so this change is coming, this disruption is happening and individually there's nothing we can do about it but collectively in the millions we can definitely do something about it. We can fight back. We can resist. We can own a stake and the whole

8:45part of the community is that we own a piece of it together. So it's fractional ownership distributed across millions of people who have come together all facing the same direction against a common enemy I guess right and it's all about contribution to that ecosystem so there's different ways that you can contribute and the you know the easiest

9:07way is by downloading the browser extension you don't have to do anything in the background you essentially turn it on and you're helping to train the action model and That training improves the AI that knows how to use every single website. And you know when you have enough people using that, you end up learning how to use every single website.

9:30

Very interesting. And so the Chrome extension, you know, I see some of these LLMs, they've just come out with a browser. They now have desktop apps for claude further integrating it. Chrome extensions. Is that available right now or how where's the team at with the development of it?

Yes, we have all of that already. So if you know go to the

9:52documentation, you can see actionist and you can see the demos. We have got dozens of different use cases across marketing, sales, operations, HR. It literally covers any kind of business department. And we've got it in beta with a number of different test clients right now. So, it runs locally on your on your Mac or Windows or also in the cloud if you don't want to leave

10:20your computer running 24/7 because that's what these new era of AI employees will be doing. It's not I'll speak to it and it sends me a message and it's finished. It's these AI employees. They have their own calendar. They control their own computer

and they can work 24/7. So, you know, a single one of these can work probably at the same speed and

10:44velocity as five members of staff. That's what that's what makes this transition so inevitable. Because businesses will 100% adopt it when given the chance.

Yeah. Yeah. I know. It's it's moving quickly and you know what I'm guessing that the early adoption of this is going to be crypton natives or you know libertarians but more people that are in

11:12the blockchain industry and then you have to break out from there. Is that sort of the case right now?

No, I don't think so. I think that I think that with the action model in the community I think that 95% of our community will be web two you know normal people that might not have any involvement with crypto

you know in my last in my last

11:35company I ran a fintech in the UK and we were one of the first companies in the UK to bring QR code payments in. So we had a banking payments license and that essentially enabled us to connect to banks directly to move funds without Visa and Mastercard. So you know our whole narrative in that company was when you pay with Visa Mastercard they take up to

12:035% from the business but when you pay with us [snorts]

we don't they don't have that fee right

so through word of mouth that grew to millions of users within a year just because people wanted to support the resistance right people like to resist against something they like an upright ing David versus Goliath, right? And in that case, it was Visa and

12:29Mastercard. Even in a scenario when people didn't actually lose anything, it's the business in that scenario that lost the transaction fee, not the millions of users.

Mhm.

But here we have a billion people that are at risk of losing their jobs to AI. And so not only do we all have something at stake and something to lose here but the way we've structured the action

12:55model is you actually own a fractional share of it as well. There's an incentive to contribute and participate. So we actually launched two days ago. And we hit 60,000 users in the first 48 hours. the platform's invite only, so you can only get access if you've been invited by someone else. And really blew us away. Well, the team and the

13:23developers because we didn't [clears throat] really expect those numbers, but it really shows how how powerful that narrative can be. And we haven't done any real marketing in web two either. We haven't really pushed to anyone. It's mainly been a very small handful of early access communities that have have grown this so far. So I think that you know this

13:48will reach millions of users because our cause is so important. You know our ethos as a project and as an ecosystem is that everyone should be able to participate in AI and own a piece of it not just a handful of billionaires. Definitely. I agree. And that's some nice growth coming out of the gate with invite only. I would love to get an invite and maybe if we have some for the

14:15audience, we can try and [clears throat] get some links to that. And you know, it's a great example with the Visa, you know, yeah, it's huge to be saving 5%. That's most of the margin of most companies that have you know, thin margins. It's like it's the difference between not making money and having your business survive. But with the

14:37at least with the you know the chat GPT moment when they launched it made it so easy to just like type and get the answer that you know grandparents were able to use it and I think that was a huge part of the adoption and now as we move to autonomy and AI agents there's definitely more of a technical barrier right now even myself exploring through different platforms like the

15:04claudebot that came out for example you know you have to use lines of code you have to like know NodeJS or Python and even for tech you know these white collar workers as you mentioned a lot of them they're not very technical they're they're sort of just doing low-level tasks somebody has to do it they it's hard for them to set up an AI agent that

15:24requires code

so does action model solve that

absolutely you're completely right and there will always be a thriving open-source community and I mean look at the open source community for LLMs right now. We've got thousands of models.

firstly, none of them come close to the big boys,

but secondly, look at the size disparity

15:50in user audiences.

In terms of AI users, I would say 99% of them are paying for it via the big tech companies. And it's only the skilled technical people that are even bothering to test the open source models because inherently it's quite complicated. Mhm.

So we are solving that like we're making it extremely easy for businesses to use this and in part that's one of

16:19the functions of our development. But also the community as well. You know the biggest challenge that's going to come with this next stage is actually the onboarding. Mhm.

If you think about hiring an employee, you have to on board them for one to two months. Teach them everything about their job. You know, what platforms they have to use, how to use them, when to

16:45use them, policies, processes, procedures,

takes a lot of time and a lot of manpower to do this.

And if you want the same level of output from an AI employee, you need a similar level of investment, right? So, we've built a tool that essentially does this and it's very simple to use for businesses. It's called the onboarder

and you essentially turn it on and it's

17:11completely optional by the way. This is for what I think the majority of businesses will use because it's so easy.

And you turn it on and it starts to shadow you and it learns from what you're doing. And you can leave it on as long as you want, whether it's for five minutes to show the AI a very specific thing that you do and then you can schedule that manually in the calendar

17:33of the AI employee or you can leave it on for a month and it will learn everything that you did in that process. And by the end of that period, you'll have an AI employee that's built a brain of all the context, all the different variables, where all your files are, what platforms you use, when to use them, what kind of workflows, schedules, calendars, everything, right? And

17:58and so that's, you know, one of the tools and the community aspect of it is the workflow creation, right? So we have a marketplace within the ecosystem. It's one of the ways to earn more of a fractional stake and it's essentially the fact that you can create workflows and then use them yourself or you can publish them to the marketplace and when

18:20you publish them anyone can go and find that workflow and within one click they can attach it to their own AI employee and it onboards the workflow within a second. Right? So, it's super super simple and it makes it easier for businesses to be able to use this.

Wow, that's very cool, Cena. And the onboarder. Yeah, that's that's a huge time drag for most of these

18:44traditional companies, especially anybody who's an employer and they lose an employee, it's like, oh, we got to train someone else. That's going to be, you know, the documentation you have to go through to set that up and then they have to follow it all. It's very laborious. And but at that point if the AI has onboarded, you know, done all the documentation for onboarding, they

19:06could just do the job themselves at that point.

Exactly. [laughter] That's that's what's coming like and that is how easy it's supposed to be for businesses.

You know, businesses are not going to be able to manually configure these things. If you look at, you know, the rise of NA10 and make and Zapia and these tools, I mean, they're fantastic. They have

19:30their limitations, don't get me wrong.

You know, the biggest one being that they can only use APIs

and therefore 99% of the internet's functionality isn't available to them. But the bigger challenge is the fact that a they're it's RPA, so they're brittle workflows. And secondly, they have to be built almost entirely for a specific use case. You know, a business

19:55can't just one click install and get them up and running. They have to be configured one by one. And it either requires the business to spend a lot of money on someone that can do that

or they have to

learn with like a one month learning curve how to do it themselves.

Mhm. Yeah, I've been through that going through Zaper and the automations and

20:19there's a lot of limitations. It's great once you get it set up, but yeah, if the AI could just control the computer keyboard and mouse, then you would save all of that. And I think those automation companies are also realizing that and they're integrating further AI and they're they eventually they'll probably have to shift their entire business model.

20:38

I don't think they're going anywhere. You see this these these new automation AIs they can use the computer. It's not going to necessarily replace a Zapia or an N10.

It's going to know how to use them.

Right.

Yeah. So these these workflows through these workflow tools like Zapia, they're great for a specific use case that you might need to run a thousand times a

21:04second, right? And instead of you knowing how to build it, the AI can actually go and use Zapia to build it, test it, configure it, deploy it, and maintain it.

So that's the difference that's coming.

Definitely. So you talked about the onboarding, you talked about the workflows, which I wanted to dive into on the workflow because that sounds

21:30like I'm like I'm trying to understand how to own more of the community-owned AI if you're the one presenting sort of like a library of the code or like here's something that the AI can do to automate. Here's an onboarding process for my business that could be applied to other businesses. people can use your code per se and then you're getting a share of the revenue from that.

21:54

Yeah, that's exactly how it works, right? So, as a as an individual, you can create workflows like I said for yourself and if you want, there's a nice shiny button. It says submit to marketplace. Once it goes through review process, it gets published. We the action model team as well can also feature certain workflows

especially if it's a really good

22:18workflow in a particular domain. So the marketplace is split by platform. So you have workflows for Gmail, workflows for Figma, workflows for Salesforce, any platform that exists, you can create workflows for it. You don't need any integrations, you don't need APIs. This is the other thing that our engine is really really good at. Zapia and NA10 as

22:43you know they are still technical. You need to learn with you know a significant learning curve of at least a month if you want to build anything you know reasonably complex. They're technical and ours you just need to be able to speak. M

if you can read and write then you can create workflows in our ecosystem because our workflow engine is mainly

23:08about telling the computer what to do. So the nodes are mainly navigation nodes. You know if you want the AI to open the browser, click on the URL bar, type in twitter.com and then go and click post new tweet,

you would literally explain that step by step inside of our workflow engine. you know, step one, open Chrome, step two, click the URL bar. You know, you're

23:34speaking to it in natural language, and that is what basically helps the AI be flexible, but provides it with the flex provides it with the capability to be able to do almost anything.

Yeah, it's very cool. You mentioned earlier about how, you know, there's thousands of other open- source AI models that people are experimenting with, but they really don't have the

24:01processing power and the libraries that, you know, the fang of AI have. is how will action model you know is the community-based efforts enough to try and keep up with the growth of you know we're seeing every month like there's a new Gemini model there's a new claude and like they're getting past PhD level how can open- source and decentralized AIS compete with that

24:28

so LLM's I mean you know I think it will be it will be a tighter difference between top models.

Mh.

The main difference might end up just being time to first token, you know, and the speed of the result maybe at the actual cost, you know, because those companies have such scale. I mean, Google has their own 100% infrastructure, you know, they own their

24:55own chips.

and so their their economies of scale are much better as well. With regards to what we're building, there is a significant difference. You know, if you try and use a traditional LLM to navigate the internet and use the interface of a browser, it struggles to do that for a number of different reasons.

One, the training data of an LLM is

25:21about a year old, right? So, if there's been any changes to the UI, to the platform, then it doesn't even know the correct sequence of events. It doesn't know the functionality that's available,

right? And secondly, it's not trained on the UI of a website. So it doesn't know where the buttons are. It doesn't know what the buttons look like. So it makes

25:42it hallucinate much more when it comes to actions than it does with normal text. And hallucination with text is already around 20 to 30%. Depending on topic, right? M

so if you want an AI as a business to do the work for you, it needs to be close to 100% success rate, right?

And that's where the action model really prospers because we get live data. We

26:09have already tens of thousands of users within a couple of days that have downloaded the browser extension and that browser extension is tracking and learning from all the different platforms that our community is using. That means the AI is learning live on all of these platforms exactly how they work, what is the functionality, what is the sequence of steps required to

26:34complete a specific workflow, what do the buttons look like, the mouse coordinates, all of this information is live and accurate, not one year old, not guessing it, right? And so that drastically increases the accuracy, which is exactly what businesses need. I agree. I'm excited to check it out. I can leave a link to the Chrome extension in the show notes below and

26:59and what's the best way Cena to follow along with this onboarding, find out the workflows as more come in and just the progress of action model and to just dive into it.

Well, depends [clears throat] on how involved you want to be. If you want to follow us and the journey, you can go on X action model AI, follow us there. If you want to get involved and you know

27:22actively contribute our discord channel but everyone can go and sign up go to action model.com and you know go and search for a referral code. I'm sure you'll publish your one within this video as well.

Sounds great. I will definitely I can add all that in the show notes. And I appreciate your insights into where AI

27:46is heading and you know automation is exciting. It's just a matter of getting it in the right format and not being thrown out of your job for potentially a billion people by the automation. So I think there's a lot of people that they understand what's coming. You and I both know and getting involved in something like this will

28:09help yourself and others as well. So I appreciate the initiative from action model and I'm looking forward to following along. I would love to follow up again as you know it's only been two days since the launch and it's already considered pretty successful but there's a long way to go to get to save a billion people. Looking forward to following up in the near

28:32future again with more updates and as AI continues to grow quickly.

Thanks for having me Ashton. It's a pleasure.