As artificial intelligence continues to advance at a breathtaking pace, its dual nature of offering both remarkable benefits and potential threats becomes ever more pronounced. While acknowledging the critical role AI plays in accelerating research, enhancing decision-making, and optimizing operational efficiencies, it's also crucial to factor in the ethical considerations and challenges it presents.

With over 15 years of expertise in the AI sector, Head of Data Science Ivan Drokin has tackled a wide array of tasks, spanning from intricate research projects to complex engineering endeavors. Our dialogue ventures into the challenges and prospects that AI introduces to both researchers and industry professionals.

How did you first become interested in AI research, and what led you to pursue a career in machine learning?

I come from a background in applied mathematics, focusing on control theory and its applications. My first encounter with machine learning was in 2008 through elective courses at university, which immediately captured my interest due to its perfect mix of theory and practical application. This blend allowed me to leverage my applied mathematics skills in real-world scenarios, aligning with my preference for practical work.

Driven by this newfound interest, I embarked on a self-directed learning journey into machine learning, navigating through the limited resources and hype around the field at the time. This phase was marked by intensive self-study and hands-on experimentation with various algorithms.

My professional journey began at a small fintech startup, where I applied machine learning to analyze financial assets, aiding in decision-making processes about buying or selling these assets. This experience marked my first substantial use of machine learning in a real-world application.

Concurrently, I undertook research projects at my alma mater, focusing on optimizing the training of large models, although these endeavors have since become outdated.

Later, I ventured into the biotech industry with Biocad and then Botkin.ai, working on developing pharmaceutical drugs. Now, I serve as the Head of R&D at Prequel, a company that specializes in photo and video editing apps.

I also pursue independent research, particularly interested in blending concepts from natural language processing with computer vision, especially within the medical domain. This interdisciplinary approach keeps me engaged and motivated in my ongoing exploration of machine learning applications.

How do you evaluate the progress in the field of AI over the recent years? Do you think that everything is changing too quickly?

Over the past decade, everything has radically changed. Methods, approaches, models, and the scope of achievable tasks have all transformed. For instance, in 2022, ChatGPT sparked a massive revolution and completely altered the field of Natural Language Processing (NLP). Similarly, convolutional networks in 2012 fundamentally changed the field of computer vision. This is why everything continues to evolve today. Even solutions that are four to five years old can already seem outdated.

I feel that changes are happening very swiftly, but this isn't necessarily a bad thing. Of course, not all changes are super positive, and there are always two sides to every story. Specifically, when talking about large models, especially generative ones like Large Language models like ChatGPT or Claude, or models that generate images or videos from text prompts, such as the recently announced Sora by OpenAI, it all comes down to computation. And it's a computation that becomes a limiting factor in many respects.

Back in 2012–13, I could replicate experiments from almost any paper with the hardware I had on hand. Now, that's far from the case. Not everything can be replicated due to the need for millions of dollars in hardware investments, which is certainly a problem.

However, there are also many positives. We've gained access to a vast array of open-source models, which have enabled a multitude of startups and companies to develop various solutions, ranging from photo and video editing to automating the document flow of large companies.

What are some of your achievements that you are most proud of? You have also invented several patents; what are those?

I take pride in my involvement with robust teams that have accomplished significant feats together. One example is Botkin.ai, where our collective efforts garnered accolades from professional communities. We successfully executed over 50 intricate projects, ranging from lung cancer screening to aid in the analysis of clinical trial data. I'm also proud of my contributions to braingarden.ai. Our collaborative efforts resulted in assembling a skilled team, earning the trust of industry giants like LG and Arrival, and delivering innovative projects for our clients.

Some of the scientific papers I've contributed to have earned Best Paper Awards at international conferences like AIST. It's gratifying to know that my contributions to science are valued by the community. Being invited to join the program committee of the AIST conference also stands as a significant achievement for me.

Regarding patents I've been involved with, each represents the culmination of extensive teamwork and a rigorous R&D process. During my tenure at Botkin.ai, our research group devised and patented several solutions, such as a novel approach to lung CT analysis. This method, which conceptualizes data as a point cloud rather than an image, offers advantages like improved accuracy and reduced computational costs. 

The inspiration for this approach struck during the MIDL 2018 conference in London, where I heard a presentation on utilizing point clouds for analyzing dental conditions. Recognizing its potential, we delved into exploring this avenue, yielding successful results. Another series of patents focuses on modeling and processing patient information in medical imaging contexts. These patented technologies formed the cornerstone of our company's products.

On a broader scale, I believe in our responsibility as intelligent beings to disseminate knowledge throughout the universe. I'm grateful for the opportunities I've had to speak at conferences, deliver popular science lectures, give speeches, and teach at esteemed universities worldwide.

Do you think the rapid development of AI poses a danger to humanity, and do you believe we should perhaps slow down a bit?

In this regard, I'm a techno-optimist. I believe that technology, especially based on artificial intelligence, is something that can help humanity survive and improve. AI is indeed a technology that could allow us to solve many problems. Even the issue of unemployment isn't as dire. We humans possess one superpower: we are curious and can adapt at an incredible speed, unlike any other species on Earth. If AI were to take our jobs, albeit with some nuances, we could figure out something, like implementing a universal basic income, and just live, learn, and make everything around us better.

For example, in the field of medicine, AI can enable us to scale solutions globally. Training lab technicians is easier than training doctors, whose work is hard to scale due to decades of experience, training, and certification requirements, even as the demand for doctors continues to grow. AI could help address this issue.

Similarly, in transportation, autopilots will eventually become widespread, significantly reducing accidents and traffic congestion. This could also help address some environmental issues, presenting a massive positive impact.

Yes, AI carries real risks along with its benefits. Numerous studies show that a model good at finding medications could, with a tweak, be turned into a developer of biological weapons. Or consider the classic thought experiments about paperclips, for example. Popular culture has spent 30 years telling us about AI coming to destroy us, from HAL 9000 in Kubrick's work to the Terminator. But I believe that as humanity, we can manage these challenges, negotiate effectively, and harness all the benefits AI has to offer.

It's likely that governments will attempt to regulate this area, especially concerning large language models because they pose specific risks ranging from privacy loss to mass unemployment. So yes, regulatory efforts will be made, but I'm not among those who believe an AI apocalypse threatens us. I think we are far from such an apocalypse, and if it does happen, it won't be because AI goes rogue and tries to eliminate humanity. More likely, the threat will come from overheating our planet in our attempts to train AI. 

How can AI help with scientific research?

AI can significantly aid in research across various fields, and I'll use the medical sector as an example where I have expertise.

Generative AI models, like stable diffusion or midjourney, which convert text to image, can be utilized to create synthetic datasets for research purposes. This is particularly relevant in medicine, where access to real data is limited due to stringent privacy laws in almost every country regulating who can share medical data and how it can be done properly.

Consequently, there are not many publicly available datasets, especially when compared to the volume found in traditional computer vision. Another challenge is the complex process of data annotation, which is not only labor-intensive but also requires the time of experienced doctors, making it an expensive endeavor.

Synthetic datasets can democratize research significantly, enabling a broad spectrum of researchers worldwide to test and refine their models on synthetic data before potentially scaling them to real-world applications. This opportunity is a significant and exciting development in the field.

How do you keep up with all the trends in AI? Is it necessary to attend conferences, give presentations, and network with others?

Not necessarily, but it's highly beneficial. From my perspective, I regularly review academic publications and articles in my domain using compilations from Telegram channels and archive searches. I follow several authors on Telegram and Medium.com who share interesting ideas, with about 80% of my reading focused on academic publications.

I make it a point to read all the proceedings and accepted papers from the major conferences in the field, which number around thirty that I track. This means constant reading; I can't afford to skip publications for any period. It's a daily ritual to start my day with a cup of coffee and the latest articles.

Attending conferences and technical meetups is also incredibly useful; I highly recommend it. There's always something new to learn from someone that you weren't aware of before. For example, a recent trip to the Data Science Conference in Belgrade with a colleague provided deep insights into the practical applications of large language models and OpenAI's GPT models in various fields, from journalism to document management.

A particularly memorable experience for me was attending the MIDL conference in London in 2018, where we presented the early results of our research. Being in a place with 200+ people working in the same domain but on entirely different tasks offers a unique opportunity to converse with everyone over five days, gain a plethora of new knowledge, and view your own problems from a different perspective. This exchange of experiences and knowledge is invaluable not only for the information gathered but also for the motivation and curiosity it inspires to dig deeper.

Motivation is crucial in machine learning and academia, as you quickly learn from failures. Sometimes, the motivation to try again is exactly what's needed.