AI Technologies
Anit Vallurupalli
August 23, 2024
How IBM Watson pioneered LLMs
In 2007, IBM began working on its newest revelation- a revolutionary supercomputer that could successfully rival real people on the famous quiz show Jeopardy!. Dubbed "Watson" in recognition of IBM's former CEO, Thomas J. Watson, the supercomputer was meant to be a "question-answering" machine, with the ability to answer complex, riddled questions, rather than the simplistic, straightforward ones that other computers of the time were limited to. In just a few short years, IBM's Watson went from a simple idea to a fully developed device, ready for action. By the end of 2010, Watson had already won roughly 70% of simulated games against prior Jeopardy! winners, and in 2011, Watson played its first official, televised game of Jeopardy! against Brad Rutter and Ken Jennings, current Jeopardy! champions at the time. Due to the computer's revolutionary design, Watson left the match victorious, demonstrating the rapidly developing power of AI and leaving the world ecstatic about the endless possibilities of such a computer.
The main differentiation between Watson and computers with similar functions at the time was Watson's colossal memory and the rapid speed at which it could present a correct answer. To constantly enrich Watson's knowledge bank, the supercomputer was consistently fed massive quantities of new information, often from sources such as Project Gutenberg, Wikipedia, and World Book Encyclopedia. By 2011, the computer could store the informational equivalent of roughly one million books, enabling it to compile details from a seemingly infinite variety of sources to achieve the most accurate responses. Even with an almost boundless expanse of knowledge, Watson could still produce an answer to a complex question in under three seconds, an impressive feat that paralleled human response time.
When Watson is presented with a question, a multitude of algorithms rapidly begin inspecting the question and proposing probable responses. Next, preceding algorithms establish a ranking system to order each response based on retrieved evidence for or against each answer. Each response is granted a score from 0-1, indicating the likelihood of its correctness based on the retrieved evidence. The higher the score, the more likely that the answer is correct, and the highest scoring answer will be the one Watson will respond with.Though this may seem like a hefty process, it is all done in mere seconds, underlining the vast power of the supercomputer and the advanced level of its abilities.
When Watson is given text, the computer takes it apart semantically and syntactically in order to fully understand it and analyze it from every possible angle. Then, much like how Watson ranks probable answers, it ranks sentences with "confidence scores" based on how frequently that content appears throughout the rest of its knowledge bank and what texts it appears in. Scores are also out of 1, and the score and sentence together are referred to as the "semantic frame." Sentences with higher scores are perceived to have more reliable content, and therefore are believed by Watson to lead to more trustworthy answers.
Watson has also learned to assess hypotheses through intermediate hypotheses. When given an intricate question, Watson often finds many possible answers. So, Watson often puts each of these possible answers into an intermediate hypothesis and investigates it, seeing if it will correctly answer one portion of the multifaceted question and then eliminating incorrect answers from there. For example, if one Jeopardy clue was "In cell division, mitosis splits the nucleus, and cytokinesis splits this liquid that cushions the nucleus," Watson may come up with many possible answers, such as vacuoles, organelles, or cytoplasm. To narrow these down, Watson will develop an intermediate hypothesis: that each of these three is a liquid. Then, Watson will "test" that hypothesis by calculating confidence scores that each answer is actually a liquid by sifting through related semantic frames. The answers with low confidence scores will be removed from selection, narrowing down the supercomputer's options significantly and allowing it to achieve a single, correct response.
IBM's Watson was a revolutionary device, one that superseded many expectations of what AI could accomplish and left many hopeful for lucrative success of other real-world applications of Watson and future AI inventions. Watson has inspired a new generation of AI technologies, and is often renowned as one of the significant milestones in the history of AI development.
Sources:
https://www.techtarget.com/searchenterpriseai/definition/IBM-Watson-supercomputer
https://www.nytimes.com/2021/07/16/technology/what-happened-ibm-watson.html
