Viorel Spinu interviewed by StirileProTV, August 2024

"AI never tires. It relentlessly persists, trying over and over until it achieves the goal."

[Marian Andrei] Hello, everyone! Welcome to the show "I Like AI," a show made by humans, because artificial intelligence constantly asks us, every five minutes, if we're satisfied with its services.

Our guest today is Viorel Spinu,  AI specialist. And he's not a specialist just since 2023, when AI became trendy. You have a lot of experience.

[Viorel Spinu] I could say I've been working in this AI field for about 12 years, since I was training models on a GPU on my personal computer.

[Marian Andrei] So, almost an era if we look at it. We've discussed AI, especially since the emergence of LLMs like Gemini or CCPT. How should we view or relate to the past year, since this technology has reached regular consumers?

And how do we avoid feeling overwhelmed and a bit fearful at work?

[Viorel Spinu] I understand what you're asking, and it's very natural to feel this fear. Things are very, very new and overwhelming us at an incredible speed. Let me share something from my job.

I was working on a system, and in the evening when I went to bed, the system wasn’t working too well. I was thinking about how to improve the prompting and programming to make things better. The next morning, I woke up and saw that Lama had released Lama 3.1. I was using Lama 3.0. And I said, "Let’s upgrade the version before making any other changes to the code." I switched from 3.0 to 3.1, and it worked wonderfully. Overnight, without making any other modifications. So yes, I understand this fear because things move very quickly.

In just one night, the level changes without you doing anything.

What I think we should do is take a step back and look at the bigger picture. Such changes have happened many times in history.

[Marian Andrei] That’s true, but when you look at people who aren’t as passionate or interested in technology as we are, those people need empathy. You need to take them by the shoulder and show them how it works, what it looks like.

[Viorel Spinu] Yes, yes. My advice to them would be to just start. Because, you know, the hardest step in a journey of a million steps is the first one.

After you’ve taken the first step, the second and third follow. And before you know it, you’ve already taken the first 1,000 steps. It becomes automatic.

When you’re outside the phenomenon and view AI as something opaque, unknown, it’s difficult. Fears and internal anxieties keep you away. But it’s all just a mental block.

[Marian Andrei] You offer some free online courses. What’s the feedback? I saw that the starting level is like learning the first letter of the alphabet.

[Viorel Spinu] Yes, that’s true. I started this course precisely to address the fear I noticed in the people around me. I could sense it.

And I said, "Hey, I could do something to change things." So, I created a free online course about AI, starting from the basics.

The feedback is fantastic. I started two months ago. It’s interesting that you receive the first lesson as soon as you sign up.

It’s not about a clear start and finish. It’s a continuous process. And I receive emails.

People ask me: "I tried this, I tried that, but how do you do this?" And I respond because I really like helping. I really enjoy seeing through others' eyes how they perceive this phenomenon.

Being inside it, I sometimes lose perspective. And this helps me see where the real value is.

[Marian Andrei] We had a very smart guest, just as smart as you, who said something very interesting. After the mouse and keyboard, we now have another way to communicate with the computer: verbally.

[Viorel Spinu] Yes, that’s exactly right. The latest demonstration from OpenAI, GPT 4.0, was absolutely fantastic. The way it counts.

Exactly. And how it emulates feelings in its voice, how it adds emotion to its voice. It makes you feel something.

I read somewhere that it even had moments where it mimicked your voice.

[Marian Andrei] Which is a bit scary, indeed. But we see solutions like this being developed both by OpenAI and Google, which are starting to understand our world. If we open the phone’s camera, they can fully understand what’s happening around us.

And again, that’s a scary chapter, isn’t it?

[Viorel Spinu] Yes, I wouldn’t use the word "understand." Because when you say they understand, your mind assumes that the system becomes your counterpart. It becomes human.

And it’s not human. There’s nothing magical there. An LLM, a large language model, is nothing more than a huge collection of numbers.

We’re talking, in OpenAI’s case, about trillions of numbers, and in Lama 3.1’s case, about 450 billion numbers. Just numbers and nothing more. So, it’s indeed fascinating when you talk to it.

It’s very easy to fall into the trap of believing you’re talking to a human counterpart. Because those numbers have been trained to do just that.

[Marian Andrei] So it’s not intelligence, it’s not an emulation of the human brain, but just…

[Viorel Spinu] Just numbers that have learned to predict the next word. That’s what it’s about. An LLM is very good at predicting the next word in a sentence.

You give it a sentence, and you ask what word fits best here.

[Marian Andrei] Let’s break that down because this is where I wanted to get to, how to write a proper prompt. Let’s break it down and describe what you see here. In this setting, where I’m wearing an orange shirt, how would…

[Viorel Spinu] Look, before getting to prompts, I’d like to go through how an LLM is trained. Because that will help a lot in understanding what a prompt is. When I train an LLM, I take a lot of text from the internet.

Text from websites, news, books… Wikipedia, anything. And I give it these texts.

Initially, when the LLM starts training, it’s like a blank slate. It’s a collection of random numbers. And I give it pieces of this text and hide a word.

I ask it what word fits here. And those numbers organize into structures, so the model becomes better and better at predicting that word. Now, let me go back a bit.

I said I give it text from millions of sources. Books, the internet, forums where people sometimes curse. Can you imagine that this LLM is like having a million different personalities inside it?

It’s a scientist because it’s seen things in the scientific field. But at the same time, it can be very vulgar. So, when I ask it to solve a math problem, the question is: which personality is responding? There’s a good chance it’s intelligent enough to respond with the scientist’s personality.

But it’s not guaranteed. And so, the prompt does exactly that. The prompt sets the LLM into the personality I want to interact with.

And then, when I ask it to solve a programming problem, it’s helpful to start with: "You are a senior Python programmer, very detail-oriented." And only then do I present the problem.

In this way, I’ve set it to focus on the area I’m interested in. And now, coming back to your question. Sure, there are multiple aspects.

If I want it to describe what’s here, the first problem I’d like to solve is to take the image. I send the model the image and then set the context.

I say: "You are an interior designer." Maybe I’m interested in having it do an architectural analysis of this room. How does it look architecturally?

And then I say: "You are a very good interior designer. I’ll give you a picture. Please suggest how I could improve the atmosphere in this room."

And it might say: "Hang a painting on the wall." For example. Because I’ve generated that specific personality in it.

And still on the topic of prompting. Look, I saw an example a few days ago that I found absolutely fantastic. You can try it too.

Anyone can. You ask GPT a very simple problem: "My biological mother was in Rome, my father was in Paris. When was I born? In which city was I born?"

If you just ask this question, chances are it will respond: "I don’t know, I don’t have enough information." Or it might give a vague answer.

But if you add a simple phrase at the end of the question: "Think step by step," it will reason: "Your biological mother was in Rome, the child is born where the biological mother is, so you were born in Rome."

That simple. Just by adding "think step by step." The technique is called "chain of thought."

Chain of reasoning. And it’s not very different from what a teacher does with a student: they say, "Think step by step to solve the problem." I do the same thing with the LLM. And at that moment, you see, if you think about what I was saying that an LLM predicts the next word and nothing more, at the moment I make it think step by step, it starts with a very simple problem.

If I don’t tell it this, it might start solving a very complex problem and get lost. It hallucinates. But if it thinks step by step, it starts with a simple problem.

And the correct answer comes when it goes through the steps sequentially. It becomes easier for it to manage.

[Marian Andrei] So there’s no intelligence in all this?

[Viorel Spinu] In my opinion, no. Intelligence has other characteristics.

[Marian Andrei] But when you add into this context the development of agents, how do things evolve?

[Viorel Spinu] Yes, that’s a very interesting area.

[Marian Andrei] Let’s explain what an agent means.

[Viorel Spinu] Absolutely. When you talk to a large language model, I mentioned that a prompt sets it into a certain personality.

Correct. You have a single interlocutor you talk to. You talk, it responds, everything is fine.

Things work very well because what happens there is a kind of brainstorming. The large language model says something, you correct it, it corrects itself, and so on.

But what if… I’m a lazy guy. I wouldn’t want to do this brainstorming myself.

I’d prefer another large language model to do the brainstorming with my model. I withdraw because I’m lazy. And you let two such entities talk to each other. It’s much better than if there were only one. Because one says something, and the other, looking at it differently, might say: "No, you’re wrong, look, here’s a problem."

And that corrects it, yes, that’s right. What do you think? These are agents. It’s very spectacular how they work.

For example, I’ve done some interesting experiments in code generation. A model writes code as well as I do.

But if you just tell it to write code, sure, it helps, and it’s very okay. But it helps even more if you have multiple checkpoints. If you want to develop an entire system using an LLM, you need to move into the agent zone.

You need to try to replicate what happens in a software company. You need to have an agent who is a programmer, who knows how to write code. You need an agent who is an analyst, who knows how to analyze the problem and break it down into specifications.

You need an agent who is a project manager and who understands… Knows how to delegate and organize. Exactly, and what I did there was an entire flow with about six agents.

We had tester agents, architect agents, and each of them had a point of intervention. Initially, the analyst would read the business requirements and translate them into specifications. I’m not sure how well it does that.

[Marian Andrei] Very well, very well. You have to think that even if it doesn’t do it right the first time, the advantage of AI over me is that it doesn’t get tired.

[Viorel Spinu] Correct, but then why do you still get a salary?

[Viorel Spinu] Because I know how to put AI to work.

[Marian Andrei] Okay, I get it. And this is something I really want to discuss because I’ve encountered this question frequently in my courses: "Is AI going to take my job, what do I do?" My conclusion is that it’s not.

It’s still not good enough for that.

[Viorel Spinu] I would say that AI will take a job, but it will give you another one. And not just one, but two or three.

I believe that for every job that AI eliminates now, it will add two or three new jobs to the ecosystem that we can’t anticipate now. And look, I propose we do an exercise. It’s a very interesting exercise that I often talk about.

Let’s take John. John lived in 1960. At that time, there were no software compilers.

People were writing code in machine language. It was very complicated to write code back then. John worked for six years to develop an ERP, an information system for a company.

And he was very proud of it. Things were going well.

Now, let’s bring John to our days, here, in the studio, and show him that I can do the same thing he did in six years in a week. With new low-code, cloud technologies, etc. What would John say?

"This guy from 2024 did my six years of work in a week." He might think that there’s no need for programmers in 2024.

What John doesn’t realize is that, yes, I am a thousand times more efficient than he was back then, but business has also exploded.

In 1960, programming was only used in banks and rockets. Now, it’s used even in my microphone here.

[Marian Andrei] Yes, but maybe that’s where people’s fear of technology comes from, especially when you look at how fast it evolves. And that’s something I’d like to discuss further: how much can AI evolve in the coming months? We’re eagerly waiting, for example, for video generation. We already have image generation, which is very good, but video generation through Sora is next.

I’ve seen some very interesting Chinese solutions.

[Viorel Spinu] I think we’re seeing some predictions now, we’re making some assumptions, but it might be that things advance slowly enough for people to adapt. Look, I’m looking at what’s happened in the last two years, compared to 2022. When GPT first appeared, I was really worried.

"I won’t have anything to do anymore. I’m a programmer, what do I do now?" But after two years, GPT can write code, but it doesn’t write code instead of me. I still have to guide it. So I don’t think things will move as fast as they seem.

It’s very spectacular to see a prototype, but from there to implementation in production and systemic change, there’s still a long way to go. And in my opinion, a human being has a few years to adapt.

[Marian Andrei] I think it helps small companies a lot where you usually have one person in ten positions who has to do a bit of everything. I’d also like to point out something: I know you work for a very large company that develops its own LLMs. True.

What do companies use these internal LLMs for?

[Viorel Spinu] I’d start with simple things that all companies have. An example is information extraction. Imagine a huge corporation that has accumulated, over 10-20-30 years, a ton of internal documentation. From employment contracts to procedures, no one knows exactly where everything is.

Using an LLM, you can answer questions and find relevant information quickly. You talk to a language model just like you would talk to a colleague. You ask: "I want to take a vacation. What do I need to do?" And the LLM knows exactly where to find the information and tells you the steps you need to follow.

These aren’t science fiction things. They’re happening now.

And even more, AI can do certain things instead of me. For example, I can say: "I want a vacation from tomorrow for 5 days. Make a vacation request for me." And the AI knows what systems to call and with what data to do that.

[Marian Andrei] So, there are employees who request vacation now just by talking to an AI?

[Viorel Spinu] Yes, exactly. And it works. This is just one area of applicability.

Then, there’s the whole area of Generative AI. It’s called Generative because it generates. For example, it generates code.

As I said, I’m not at the point where I implement a system that generates complete code in production, but I’ve seen systems that do code review. Programmers write code, and the AI looks at it and says: "Look, here you could have done it this way," and it does it very well. Because, you know, it doesn’t get tired and has incredible memory.

In my mind, AI is like a work partner with fantastic memory and knowledge. At this point, it doesn’t surpass me in reasoning, but it completely outmatches me in knowledge.

[Marian Andrei] What do you think about the CrowStrike code? Would it have been different if they had used such a system?

[Viorel Spinu] I don’t know what happened there.

[Marian Andrei] Viorel, how can people sign up for your course?

[Viorel Spinu] curs.viorelspinu.com is the course’s website. You just give your email address, and that’s it.

[Marian Andrei] You don’t do anything with it?

[Viorel Spinu] No, absolutely nothing. I encrypt it in about 33 levels.

[Marian Andrei] So, even you don’t know what email address it is, just your system?

[Viorel Spinu] Yes, and I’ve automated the process. As soon as you sign up, you get the first lesson, and then, as time passes, you receive another lesson each week.

[Marian Andrei] So if I sign up now, I get lesson 1, and someone who signed up three months ago is up to date with lesson 9 or 8.

[Viorel Spinu] Exactly.

[Marian Andrei] Okay, very interesting. Viorel, thank you very much. So, search for Viorel Spănu on Google, and you’ll quickly find his course. I can’t wait to talk again in 6 months to review the latest updates.

[Viorel Spinu] With great pleasure. Sometimes it’s hard to keep up, but yes, sometimes I take a step back and look with some fear at what’s coming.

That’s right. But I look with optimism because now I work less. And I like that.

[Marian Andrei] Thank you again. You can always find more information in the "I Like AI" section, where we have many very interesting interviews.