Generative AI Development
August 6, 2024
10 min read
Today, when the AI revolution seems to be gaining momentum, for businesses it’s very important not to miss their chance to join it, or maybe even to head this transformation. At Sigli, we want to help you gain a competitive advantage by explaining how you can leverage the power of this technology.
Check out the full Innovantage episode with Vasil Simanionak here:
In the second episode of the Innovantage podcast, Max Golikov talked to Vasil, the Chief Delivery Officer at Sigli, a person who was captivated by AI long before it became available to a wide audience heard about. This sphere looked completely different from run-of-the-mil computing which made it extremely interesting for him. Being inspired by such films as Terminator and Star Trek, Vasil chose AI as his major.
In a dialog with Max, Vasil shared his vision of the past, present, and future of Artificial intelligence and named the task that he will never delegate to AI.
In our article, we’ve gathered the most interesting ideas from this discussion and we hope that you will find them quite insightful.
It would be completely wrong to say that AI appeared together with ChatGPT or 1–2 years earlier. In reality, some products powered by AI of this or that kind were developed quite long ago.
The first expert systems were delivered around 50 years ago and they already represented an example of a very narrowed AI. Of course, their capabilities, as well as use cases, were rather limited.
For example, such systems could have been used by a lawyer in some specific cases. Lawyers often need to ask standard questions to their clients, like the place of birth, the date of birth, the place of residence, etc. Based on the answers to these questions, an expert system can prepare a document that will be further submitted to some authorities or used for other purposes.
So what are expert systems? They can be defined as early forms of AI that rely on a set of rules provided by human experts to make decisions or solve problems within a specific domain.
The development of these solutions is related to usual coding stuff because such things are based on conditions like “If something — Then do something”. The main task and challenge in this case is to define the right rules. This means that human experts who work on these rules should deeply understand the specificity of all the related processes.
The next stage of AI development is something that is considered to be AI in our modern understanding.
While expert systems were difficult to understand for the general public and they had only specific narrow use, with ChatGPT-like models everything is different. They have gained enormous public attention and they are available to everyone. These solutions allow users to input queries and get clear results.
While talking about that kind of system, in the majority of cases namely ChatGPT will be mentioned and that’s an example of excellent marketing and branding.
The majority of people definitely consider ChatGPT to be AI. But is it true? While talking about that Vasil highlighted that the correct answer depends on our perspective and exact understanding of artificial intelligence.
On one hand, large language models (LLMs) do not have common sense but they can process data. They are built on neural networks that mimic the human brain.
A neuron has, for example, two inputs and a single output. If the first input is triggered, an output will be triggered. If the second — an output won’t be triggered. In networks, neurons are put in millions of layers. Users need to make an input and wait for an output. That’s how they work.
When it comes to deep learning with LLMs, we do not define the underlying model to process this data. We just define a kind of infrastructure with the neural network where we have a lot of neurons and they are interconnected at different layers.
We throw data and expect the result. But even a creator of this model has no idea how an LLM will answer.
Due to the huge media influence, today these ChatGPT-like solutions are widely believed to be true AI despite some limitations in their capabilities.
AI is a huge set of everything related to something that machines can do quite similar to what humans can do. Of course, people can calculate but a calculator is not an AI solution. So we can say that in the context of AI, machines should do something as well as humans can or maybe even better.
Despite all aspirations around AI, it is still a tool, not a different species or something like that.
Today, we can define several models (or levels) of AI. They differ from each other not only in their functionality but also in how they deal with data. Let’s briefly summarize them.
As described above, expert systems do not actually work with data. These systems are nice straightforward tools but they do not provide you with the impression that you deal with intelligence.
ML systems work with some data but there are no strict rules. Engineers and analysts define the model of how this data should be gathered and processed. So we have control over how the solution will work with our data. We throw this data into this model and we check how to use it.
A good example here is an ML-powered app for the real estate market. You can input different parameters like the size of an apartment and its location, while an app calculates the price depending on the parameters.
Text models are the simplest ones of this kind. They operate on the text input and can convert this text into a new one. Here, their work can be compared with the work of programmers who need to convert requirements into code.
When an output offered by the model is not good enough, a user can provide feedback. In such a way, a model can be trained to ensure better outputs.
Moreover, there are a lot of talks about the quality of data used for training and their origin, such as, whether they were obtained and used legally or illegally. However, there is still no single opinion on that.
That’s one of the questions that may sound really controversial and sometimes even a little bit naive but it’s really interesting how AI experts answer it. Vasil provided a quite worrying reply. He said that everything depends on our behavior. Nevertheless, it’s not a reason to look for ways to be as good to AI as possible in order to survive. It’s just a reason to study this aspect a little bit deeper.
According to Vasil, there is a possibility that AI will exterminate humanity and there is also a possibility to see a dinosaur outside. But still, it is just a possibility.
Our future, and our chances to stay alive:), will depend on how AI-powered solutions, including LLMs, will be designed and how we will use them.
If we let any ChatGPT-like solution interact with the internet, it will be able to perform rather complex tasks. For example, it will be able to start a website, buy a domain (if you give it some money), and create a no-code or low-code platform.
Even an LLM can interact with the real world and actual AI can greatly mimic a human not only in text conversations but also in live streams. If you have ever seen videos with lifelike talking faces generated by Microsoft’s VASA, you know that they can be very convincing.
So can AI overtake the world? Theoretically yes. But only if a human lets it do this.
In the conversation with Max, Vasil named several examples of widely adopted business use cases of AI.
AI is always willing to help and can reduce the time to answer to a client. However, in this context, it’s important not to omit a financial factor. If you want to get almost real client support that will function practically without human participation, this will turn out to be more expensive than hiring human specialists.
The cost of such projects can greatly vary based on various factors and parameters. For example, the basic infrastructure for models like ChatGPT represents a huge number of graphics processing units or GPUs. This specialized hardware is essential for processing complex computations, as well as training and running AI models.
That’s why it will be necessary to calculate the cost of GPU rental services provided by Nvidia or Microsoft, for example. They have different subscription models that can address different needs.
Moreover, you can opt for on-premises infrastructure and locate all the required software and hardware resources within your physical premises. This model will be also associated with some additional expenses.
If we turn to the use of AI models, here, we will also have various scenarios.
Vasil noted that in the case of using a commercial model when you do not need to train it, the cost of one query will be a couple of cents. However, when you need to train and finetune your solution, it will be a completely different story. The price will be significantly higher and it’s very challenging to define it.
It’s also crucial to bear in mind that with LLMs, you can’t expect a 100% correct result for every query. That’s why to get the desired outcome, several interactions may be required.
In any case, the principle of quality-ratio principle works here quite well. The bigger your investment is, the better result you can expect. However, you should admit the fact that it won’t be a human result. Given this, businesses should find a balance between the amount that they are ready to pay and the quality that they will accept.
While talking about the future both Max and Vasil agreed that technologies are changing too quickly. It’s very hard to make any predictions for more than 5 years.
However, according to Vasil, in the near future, ChatGPT and similar solutions can become great personal assistants. The use of such assistants can go much beyond purely business applications. For example, they will be able to check the health of users, send them reminders, and fulfill a lot of other tasks that will make people’s lives better,
Another interesting and highly promising sphere of AI use is communication which is highly important in business.
Let’s admit that even when speaking the same languages, we all have different understandings of some things. AI-powered personal assistants can make sure that our thoughts can be perceived by others in a good way.
ChatGPT-like systems will be able to translate our ideas into bigger definitions that are more comprehensive for others. They will serve as bridges between people as they can translate not just words one by one. They can translate what is really said.
That is a positive side of their implementation. Nevertheless, there is a negative one as well: some translators can lose their jobs.
One of the key issues about AI highlighted by Vasil is that you can’t always check whether ChatGPT offers you something that is true or not. That’s why according to him, it’s definitely not the best idea to rely on AI in explaining something to children. Here, a human is an undisputable leader (especially, when it comes to your own child).
Of course, there are solutions like Google’s Gemini. In this case, answers are googleable and you can see the source of information. Nevertheless, AI can’t fully understand the context in which a child can pose this or a question. Moreover, human interaction is something that we all need.
During their discussion, Max and Vasil also touched on a very important topic about the skills that are required today.
Earlier, teachers and books were the sources of truth for the young generation. Then, the internet joined this list. Now, everything is quite unclear.
What sources can be trusted? Whom can we believe?
That’s why for a new generation, it is very important to develop an ability to check the source of data and understand whether it is trustworthy. A human can be good at some things but can be completely wrong in others. Given this, it’s crucial to have critical thinking and see whom and when we can trust.
While talking about the value of AI, Max and Vasil also highlighted the importance of human connection and personal touch in communication. These are something that we should preserve even in the era of AI and significant digital transformations.
If you want to learn more about AI, its current role for businesses, and its future prospects, do not miss our next episodes of the Innovantage podcast hosted by Max Golikov.
.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
.svg)
.png)
.png)
