Evgeny Smirnov obtained his M.Sc. Degree in Astronomy, Mathematics, and Computer Science at Saint Petersburg State University in 2009, closely followed by his PhD in Astronomy & Mathematics in Pulkovo Observatory (2016). Furthermore, he acquired another M.Sc. Diplomas:
1) Philosophy, Science, and Religion – The University of Edinburgh (2021)
2) Personality Psychology – Saint Petersburg State University (2022)
In his professional life in the private sector, he held software engineering roles in StegSoft and Vstretcher. After that, he founded 4xxi in the project management and consultancy field, Newtonew in agile project management, social media management, sales, and Denovo on IT consultancy and AI research. Considering this wide range of experiences and expertise, the interview contained almost all of them in the questions around large language models (LLMs). The questions were mainly on our motivation to use LLMs, their potential benefit on project management tasks, using LLMs in scientific research, conducting literature review with them, and the place of LLMs and even more potentially advanced future AI versions in scientific discoveries.
Very shortly, he explained that he had both extrinsic motivation to use LLMs that allowed him to be around 10 times more efficient compared to three years before, to the point that his two AI agents were realizing different research studies while we were doing the interview. Intrinsically, he stated that LLMs can help us to understand who we are better, and LLMs also have a strong mathematical roots that’s why he had an interest in comprehending that deeply.
In project and product management which are not fundamentally different than each other, LLMs has roles to fill but not roles to replace, with an example from User Story creation, some part of the documentation and direction/guidance to developers might be accelerated, yet there are lots of decision making and fine tuning using the information and data not loaded on LLMs, on the actual work it does not make user story creation by humans obsolete currently. Making story points and similar metrics more accurate has its deep challenges as well, specifically coming from the problems with the metric itself.
Moving on to the scientific research, LLMs and AI has obvious advantages in summarizing, literature review, number crunching, and similar, but these are dangerous when the users can’t do these successfully by themselves and don’t validate the outputs. In the case of literature review, too, losing the context of a scientific field and new developments is among the dangers of poor utilization of LLMs. It is much better to use LLMs to optimize tasks you can already do yourself and validate outputs so that you save time/resources. One example research study he published while using automatic classification in some parts of the study is this.
The last question was the potential of LLMs to make discoveries. The way we craft knowledge and discovery is built up from texts, and what LLMs generate is not significantly different, though they also contain biases we introduce to them. This is good as it shows us who we are, but the actual breakthrough will be when artificial general and super intelligence will be realized.
As his final remarks, Evgeny Smirnov cautioned us that even though the technology of LLMs, like that of atomic bombs or similar things, has its dangers, humans have a passion to discover, get interested, and learn in their roots, so our fear should not stagnate growth.
The interview can be watched here.