Podcast Speaker / June 10, 2024
Ep 2 - Future of AI isn't scary but ideal place to be | Solopreneurs, Freelancers & Governments
Podcast with Varu, hosted by Varu. Watch the episode, or read the transcript below.
Transcript
This transcript was cleaned up post-production for readability. Repeated words, filler, and caption errors are removed. Words in [brackets] were reconstructed where the captions were incomplete or garbled.
Varu 00:00
[In] business it functions in a way where there are different types of workers — like doers, or something of the sort. There’s the gig economy: freelancers, consultants, or agencies, [and so on]. And there’s someone who’s internally working with a company. Especially in [a] solopreneur [setup], let’s [keep] that [as] just the CEO or the founder. Now the founder goes to a freelancer or someone to basically get things done. So my question is: is it going to cost less for a solopreneur when he or she is going for agentic AI? How expensive is that? And B: what we are actually saying — that the gig economy is growing — is [that it is] actually going to also see a sort of decline.
Slava 00:49
I think, Varu, you touched on great points there. You’ve mentioned what could be the future where you have the possibility of using AI agents in your workforce: what will be the future of the gig economy, what will be the future of entrepreneurship. And I do think we are in a transition where we are no longer the executives of work. We can decide on how we perform work, and that creates what I would call a new dimension: if we are outsourcing work, does it really matter who is performing that work?
Gig economy is really based on this idea that work has to be, at a fundamental level, at its most basic primitive. In a gig economy you’re not outsourcing a job, you’re outsourcing a task. The traditional corporate world is outsourcing based on responsibility, based on title, based on rank. Gig economy says you outsource based on task: I want this food delivered to this address, I want this blog post to be written in a specific way. These are tasks. Agents are actually ideally designed for that.
So the first type of automation that we will see — and it’s already happening — is automation of tasks. This is why gig economy will be the first to be impacted by agentic AI. We already start to see examples of that. For those who are curious what that looks like, check out a platform like Fiverr. Fiverr has been one of the leaders of the gig economy space. They’ve allowed anyone to really simplify the value creation to one service, where you’re giving one service in the process of exchange of value. In that process you’re promising to do an SEO review, to write a blog post, to create leads, or any other specific [task]. If you, as the creator of your Fiverr profile, are offering one task at a time, then this is actually where AI, and specifically agentic AI, really shines. Agentic AI can perform one task very well, and that task is really nothing more than an AI solution that can not only perform the task but can also learn based on the feedback of those who pay for the gig, or based on any other feedback you give it.
This is essentially the reason why gig economy will be the first to embrace agentic AI: because AI is very well suited to solve one task at a time.
Now the second industry that will be impacted, of course, is the traditional corporate world: the enterprises, the companies that are on the mission to use responsibilities [and] titles to solve problems. If you are, for example, OpenAI, and you want to create the next evolution of AI solutions, you’re not hiring a person to solve a specific problem. You’re hiring a person based on their role or their title. You’re hiring an AI researcher, a language modeling researcher. Once you’ve hired for that position, that person will help you solve the problems you want, and that person will be responsible to solve many problems, many tasks. So the traditional world, the corporate world, will also benefit from agentic AI, but not from the perspective of solving one task at a time — from the perspective of solving many tasks, so that it’s able to save people time.
I’ll give you a simple example. Language models, the technology that is powering agentic AI, are able to understand what you want not just based on your intention, not just based on what you say in the prompt. The task level of prompts looks like: give me a blog post that I want to promote my company. Agentic AI is not limited to that. [Agentic] AI can also define the role, the persona, or the responsibility of that language model. I’ll give you [an] example: with agentic AI you can say you are a blog copywriter, you are responsible for creating the world’s best content around marketing, you have access to the following resources. So the role, the responsibility, the tools you have access to, you can now define as part of that prompt. You’ve shifted from telling the language model to generate a piece of content for you — generate the blog — to telling the language model what it’s responsible for, what it can and cannot do. In that process you’ve delegated what would have been a task for a person to an agent. That agent now has [the] ability to execute for you, but in the process of execution it’s not solving one task at a time, because you’ve defined the roles, the responsibilities, the personas. It’s able to solve multiple tasks for you, and the complexity of [the] task [it] is able to solve well depends on the complexity of the agent that you [built].
Varu 07:14
Got it. And so to sum it up, what I understand is agentic AI is going to push individuals to become leaders, [with] subject matter [expertise], so that you don’t have to worry more about the operations and execution, because you have bigger things to focus on. Got it. So I had this one more thing which suddenly hit me. When you take the entire economy of the world, it is dependent on a lot of things: the [gross] income, the per capita income of a person. The entire world is dependent on that. The moment agentic AI [is adopted] — the adoption of AI is going to increase, it’s not just agentic AI I’m talking about — there is going to be an evolution. A lot of people are going to get AI into their systems very strongly. How do you think all the big [powers], or the countries, will be interested to push this aggressively, because it’s going to shake the economy for sure?
Slava 08:16
Yes. I do think that the first step here is for both industry and countries to use language models. We haven’t really talked about this, but this is important to understand as a foundation for agentic AI. Language models are models that we use to chat with AI technology. These are conversational primitives that we can now leverage to have a conversation with a [chat], to have a conversation with a company, to have a conversation with a country.
At this moment the conversational elements of language models feel like it’s a chatbot: you type something, you get a response, and it feels conversational, it feels interactive. However, we don’t really have good examples of that on a country level, so we primarily see that on a consumer level. We’re chatting with ChatGPT, we get back a response, it’s a consumer product. However there are countries around the world who are pushing this level of technology to the country level, meaning the national level, the international level. Examples of that include the [Emirati] nations, the countries of [the] Middle East. These are the types of countries that are using technology like large language models, and not only are they using them, they’re developing their own language models. Why? Because what they’re able to do is offer that technology to all of their citizens, to all of their population, and allow the population to have a conversation with either themselves or the government, or to have a conversation in multiple [languages]. For example, if you only speak Arabic but you want to have a conversation in [another] language, country-level language models allow that.
Where this is going is: as soon as countries, not just companies, embrace language models, you will be able to have a conversation, a natural conversation, with not just the government but also with other citizens of the same country, and other countries as well. At this moment in time I speak English and Russian, I’m learning Spanish, I’m learning Portuguese. Language [models] allow me to have a conversation with other cultures that I wouldn’t have had access to. So the first unlock here is [the] ability for people to have conversations with other domains that they’re not familiar with.
Once you have a conversation with the power of [the] language model, now you can start to take action on that conversation. This is where agentic AI will start to shine. Agentic AI uses language models behind the scenes to create a conversational experience with you, and then it uses the elements of decision-making, the elements of planning, the elements of taking action, to then decide on what to do for you.
A simple example of what that may look like: if you speak, for example, Arabic, but you want to have a natural conversation with the rest of the world, and let’s say that your ideal customer or your ideal conversation is in English, then agentic AI can help you to have those conversations. You speak in Arabic, the language model will understand Arabic, it’ll translate everything you say to English, and agentic AI will have the ability for that conversation to happen at scale. You’ll be able to talk to your ideal customer at scale in English. You’ll be able to have any other types of interactions, but it will not force you to learn English. The language model will have the ability to translate for you. The agentic AI will have the ability for you not to perform the work.
So the evolution of the space is: use language models to interact with the world, take action with the world in a way that feels natural to you.
Varu 12:51
So as for this, what I see is that probably it’s going to improve the globalization part of it, and also for the businesses to expand easily to the non-English speaking countries, and same for the non-English speaking countries with the English speaking countries. That’s definitely amazing, Slava. I did not even imagine that this is going to be the future.
Slava 13:14
Yeah, it’s really incredible, because the level of automation of what’s possible is not something we’ve seen before. At best, automation was at a point where if you hired somebody who was a [technologist] or developer, they would help you implement what would be considered, from a programming perspective, basic automation: if one condition happens, then do this. It’s the basics of if-then conditions. This was the level of automation that companies were using for a long time.
However, as you can imagine, in a circumstance where things are not as expected, that level of automation will quickly fail. When you have the dynamic reality of the changing world, when you have changing consumer demands, changing user demands, you’re not able to write every single scenario as an if-then condition. At some point the complexity of what you need to do to create automation changes too quickly for you to keep up. Even if you hire more developers to try to [solve] this problem, even if you try to really outsource the technology as much as possible to handle automation for you, you will not be able to keep up with the demand of the changing world.
Agentic AI is able to keep up with that demand, and for that reason it’s able to create a new dynamic, a new reality, a new circumstance where companies are able to adapt in real time to the demands that they see. Automation by itself — traditional automation — is not able to keep up with that demand. This is why people sometimes are on the receiving [side], and they sometimes get frustrated.
A simple example for those who are curious to understand what I’m referring to: today if you want to have a good conversation with your bank, and let’s say you want to understand why a transaction failed, why something did not work the way you expect, [or] you want to have even a simple status update on where your current checking account is, [what] your balance is — to have a conversational experience with a bot today is extremely frustrating. Why? Because chatbots were built traditionally on that automation where at best they would decide to answer one question at a time. And if your question is different than what the developers had set out before the conversation started, then the chatbot can quickly become frustrating. If the chatbot is not able to adapt in real time, it will answer you in a very simple way. It will either give you not what you expect, or it will tell you ‘I don’t understand,’ or it will fall back on ‘please ask a human being’ or [a] human agent to jump into that conversation flow.
This is a frustration that many people have had for a long time with traditional chatbots. They were not able to intelligently route us to what we want, to answer our questions, and they would often fall back to human agents to try to solve that problem.
Now with agentic AI you have a new opportunity to [serve] customer service better, to create better sales, [to] create better marketing. This is a level of insight that is important to understand in a new dimension where agentic AI is a reality: customer service improves, sales improves, marketing improves. All the traditional domains of creating value are so much better than if you just use traditional [automation].
Why even myself, as a professional in the space who decided to focus on agentic AI — this for myself was the real unlock — is when I noticed that my clients kept asking me for more complex edge cases, they kept asking me for more complex use cases to solve their problems. I knew that generative AI was not able to keep up with the demands. This led me to start to embrace agentic AI. This led me to start to develop AI agents as part of my agency. It led me to start creating solutions for my clients. And this led me to start curating resources for the community, to educate people on the advantage of this technology.
I’m a strong believer of the technology. I went all in, as [in] the poker analogy: I put all my chips on [the] deck to really go all in on this technology. Why? Because it was the only technology possible to solve problems of complexity that generative AI was not able to solve.
Varu 18:34
Basically agentic AI is going to be a solution for all the edge-case scenarios. I have this one last question, because you’ve given me a full download of what it looks like, and I’ll have to give a surety, [an] assurance that I’ve understood it, which means that all the other non-tech people will also understand what this is all about. If you have to integrate this agentic AI technology, which is the large language models or the language models that you’re talking about, what are the challenges or limitations that we should be careful [about] — that businesses should be careful about, if there are any — and what capability should they have so that it sits perfectly into the technology that they have built their product [on], or [something]?
Slava 19:20
Absolutely. So the challenges here really lie on the ability for this technology to give you what you expect. At this time the biggest challenge to getting what you expect is what the industry calls hallucination. This is not a term that oftentimes is talked about in media. It’s sometimes talked about from the health domain: people have been using certain levels of stimulant to hallucinate. But in the context of agentic AI, hallucination refers to the unexpected ability of agents or language models to make up things that you didn’t expect. Whether it makes up something that is not what you understand, or makes up something that looks like it’s a fact or statement but it’s not truthful, it’s not factual. So [hallucination] is a challenge to using agentic AI. Sometimes language models hallucinate and give you not what you expect.
That could be as simple as: hey, tell me what you found in your research about a specific person. If the agentic AI is not grounded in its technology to prevent hallucination, [the agent] may tell you something that is not truthful. This is, I would say, one of the bigger challenges in this space: preventing AI technology from making up facts or creating statements that appear truthful when [they’re] not.
The second biggest challenge I see in the space is to make sure that [an] agentic AI solution is not going off [the] rail. The first was creating a safety guardrail. The second is creating a technology guardrail. This is where we’re aiming to prevent AI agents from performing work that you didn’t expect that agent to perform. Now this of course is on us as well: the clearer we are in our prompts, the better we guide the technology to give us what we want. But sometimes if people are not clear with their intention, that creates the circumstance where [the] AI agent starts to perform the work but it’s not what we expect. If we, for example, want a blog post written in a specific way, and we’re not clear about that, [agents] will start to generate the blog post, but when it’s finished it’ll look like not exactly what you are wanting to receive as [a] response.
So [the] first big challenge of using agents is ensuring that there are safety guard[rails] to prevent hallucination. The second challenge that I see in the space is ensuring that AI agents are performing the tasks that you want, and they’re not going off course based on what you’re expecting. And there are ways to constrain that. There are ways to use prompt engineering techniques to enable that.
The element I’ll mention here briefly is, I would say, the third big challenge here: to ensure that if you are sending data to [a] language model, if you are using [an] agentic AI solution, [you] prevent that AI agent from disclosing your private data to those you don’t want. If you are, as I mentioned before, in [a] domain that [has] sensitive data — perhaps you’re in [the] healthcare domain, you’re dealing with sensitive customer or client data — you don’t want that data to be exposed. You want to protect that data. You want to make sure that data stays within your cloud environment. In that circumstance you need to ensure that whenever you do send that data to a language model, whenever an AI agent has access to your data, it’s protected.
So at this moment the challenges lie primarily around guardrails and safety.
Varu 23:46
Got it. Yeah, it makes a lot of sense, and I really like the way you put hallucination here. It’s interesting. So you mentioned something about [how] you write about agentic AI. Is there a newsletter that me and other viewers or listeners would be able to subscribe to, or is there a resource link or something that you can tell me, so I’ll definitely share it in my YouTube description?
Slava 24:13
Absolutely, yes. I do have two resources that are very useful for those who are curious to learn more about agentic AI, or maybe want to start going down the rabbit hole of AI agents. The first is the curation of 100 different resources that I’ve been able to curate over the past three months, and this is a free resource that I share for anyone who’s interested. This resource is found on GitHub, and I will send that link to you, and you will have access to that as well. This is a resource that allows you to click, learn more about AI agents, about different frameworks, about different ways that you can start using AI agents today, whether you want to use an off-the-shelf solution or you want to develop the solution yourself. The second resource I’ll share is the newsletter that I’m currently managing, that I’m currently creating content for. The newsletter is called Agentic AI. The free resource on GitHub is called Awesome AI Agents [list].
Varu 25:25
Okay, thank you. I will definitely mention these resources, and yeah, please do share the link. I’ll share them in the YouTube description.
Slava 25:32
Absolutely. And I do invite everyone here who is listening: AI agents are a new term. Agentic AI is something that you will hear more and more as you start to pay attention to the AI space. But you don’t need to learn everything today. You do need to get yourself up to speed, and the two [resources I] shared with you will help you to make sure you’re informed, will help you to make sure you stay up to date. I encourage you to subscribe and do follow both myself and Varu to make sure you stay educated, to make sure you stay informed. However this is not something that is very easy for you to understand unless of course you are conscious, unless you spend time, you spend energy. So don’t feel overwhelmed. Do feel informed. Do feel like you have tools at your disposal. There are resources. There are people like myself and Varu who will make sure you get up to date as quickly as possible.
Now start with using one AI agent. Start by watching the Devin demo that I mentioned by Cognition AI. Start by using ChatGPT in a way that you’re giving ChatGPT a persona, you’re asking it to perform different tasks for you. As you start to use the technology you will start to appreciate its power. You will start to better understand what’s possible. And then at that point you will decide: do you want to use an agent that you can purchase and you can essentially pay for? Do you want to develop your own? Do you want to work with an agency like myself and Produvia that allows you to then implement something, not just use it on a high level?
And so this is a call to action for anyone in the space. The technology is available. The technology is ready for use. This is the perfect time to use it. The only limitation we have as [a] society, as people, is what we apply it for. We’re not limited by its availability. We’re not limited by its potential. We’re only limited by our ability to apply. The beauty of language models is that as soon as we’re clear with what we want, and we give that prompt to our language model, the rest can be solved for. We don’t need to know how to solve the problem, but we do need to know what problem to [solve].