April 4, 2023 | By Aayush Thapa (SF22)

The release of Open AI’s ChatGPT last fall came with a flurry of headlines about its potential for good and evil.  Articles about the capacity of artificial intelligence (AI) to harness and organize knowledge or help us cheat our way through an English essay—as well as the program’s propensity for error and bias —raised many questions. Perhaps most fundamentally: How do computers understand what we’re asking them and what are their responses based on?

ChatGPT is an example of a model that learns to imitate language to the point where it can perform relatively high-level tasks. ChatGPT is the implementation of a type of neural network known as a transformer, which itself is a type of deep learning algorithm commonly used in the field of natural language processing (NLP).

Seek AI, a startup based in New York, applies natural language processing to AI-based data storage, management, and analysis.

Interestingly, Seek AI is populated with Johnnies: Joseph Lee (A22) is a software engineer, Ece Tuglu (A21) is a social media manager, and Raz Besaleli (A21) is the founding head of artificial intelligence at the company.

Following a recent $7.5 million investment round, Seek AI will channel AI innovations into concrete solutions for data-driven companies and institutions. The funding came from several private sources, including Conviction Partners and Battery Ventures; Bob Muglia, former CEO of Snowflake Computing; Mustafa Suleyman, cofounder of both DeepMind and Inflection AI; and Tristan Handy, founder and CEO of dbt Labs.

Currently, electronic data management requires special skills, including knowledge of programming languages like MySQL. Data specialists are often costly, and the process of querying databases to retrieve data sets is time consuming. “This kind of specialized know-how leads to inaccessibility and inefficiency [regarding] a company’s data,” Besaleli said.

Besaleli and Seek hope to alleviate the need for such expertise by using the same principles behind successful AI platforms like ChatGPT and DALL-E. “Just as you can ask any question of ChatGPT in plain English and get an immediate response, Besaleli said “Anyone in a company would be able to ask questions about their data naturally. The program can then quickly generate code based on these questions, search the database, and return answers in a comprehensible way.”

This means data would be easily accessible to employees in nontechnical, business-facing roles, increasing a company’s efficiency and improving its financial bottom line.

Besaleli began thinking about and working on NLP soon after graduating from St. John’s, embarking on graduate studies at Montclair State University and joining Seek AI. They came to St. John’s with a passion for linguistics, diving into ancient Greek in the freshman and sophomore language tutorials. They were surprised to find comfort in the precision of the mathematics tutorial, which covers the development of mathematics from ancient geometry to modern mathematical systems, such as Goedel’s incompleteness theorem—crucial to early computing systems.

“After freshman year, I started to make all my papers about math,” said Besaleli. But language tutorials were provocative because it was there that they realized how little we actually know about language.

“We make many assumptions. We have intuition about what people mean in saying something.” But all these assumptions jump starkly into the light when our plain language must be made clear to a computer.

Besaleli combined the comfort of mathematics with the unknowable depths of language by pursuing graduate studies in computational linguistics—a discipline in which languages are analyzed and understood from a computational perspective. At Montclair, Besaleli continues to identify questions about language. “My first project was tracking the lexical changes in modern Hebrew using transformer-based language models,” a deep-learning model that processes sequential data like natural language. “This is all relevant to the work I am doing at Seek. For example, we are still asking what the best way is to represent a simple sentence in a structured way.”

Besaleli explained that one of the biggest conceptual challenges AI developers face is figuring out how to create a probabilistic understanding of data that describes a set of rules instead of attempting to prescribe those rules using deterministic methods. “AI models cannot read our minds. They don’t have the same intuition about language. We have to train them to be able to fill in these missing contexts when we ask simple questions.”

For example, the word ‘bank’ can be a financial institution, or it can be the land along a river or lake, or it can mean to save money (or something less concrete), for a future need. How do we get a computer to understand the meaning of the entire question, rather than as individual words with discreet and unrelated definitions? Essentially, Besaleli says, it’s a philosophical problem. “How do we know what people mean when they ask a question?”