Jeopardy host Ken Jennings gives RI talk on future of AI

SMITHFIELD – Jeopardy host Ken Jennings is a self-proclaimed expert in what it’s like to lose a job to artificial intelligence.

The takeover happened in 2011, years after Jennings had won 74 consecutive Jeopardy games to clinch the all-time record, when he was brought back to the show to play against IBM’s Watson computer.

“I said yes, immediately, for three reasons,” Jennings told an audience at Bryant University on Wednesday. “First of all, you don’t get many chances to play Jeopardy. Second of all, this seemed like the future to me. The third reason I wanted to go back on Jeopardy was I was 100% sure I was going to win.”

Spoiler alert: Watson steamrolled Jennings and the other human contestant, Brad Rutter, winning by over $50,000.

“It was disconcerting and a little depressing to see my life’s work – whatever I thought my valuable skill was – to be appropriated by a computer,” Jennings said, remembering the final round of that famous game.

Jennings said his confidence in his ability to beat Watson came from his certainty that computers would get tripped up trying to query for the answer to the riddle-like, dual-meaning, punny clues often appearing on Jeopardy.

But Watson was the product of years of IBM’s research on natural language processing, much of which was in pursuit of the very specific goal of winning a Jeopardy game. Jennings recalled being sent a technical article explaining how Watson worked before going up against it on the show. The article included a graph showing the amount of questions answered versus the percent answered correctly, with Watson’s iterations plotted as trend lines and human Jeopardy winners plotted as points.

The plot of human winners created a “winner cloud” that showed they were answering about half of the questions and getting them right about 90% of the time. With each iteration, Watson’s trend line inched higher toward the winner cloud on the graph.

“This is what it looks like when the future comes for you,” Jennings said, showing a slide of the graph. “It’s just a line on a chart, gradually but inexorably moving closer to human performance.”

As Jennings explained, Watson never quite closed the gap in question accuracy (getting answers right) between itself and human performance. But the advantage it had turned out to be on the buzzer.

“The Jeopardy buzzer is a very tricky part of the show,” Jennings said. “You can’t press the button as soon as you know the question.”

The buzzer activates when the host finishes reading the clue – signified by lights visible only to the contestants. Buzz too early and a contestant gets locked out for a brief moment when the buzzers activate. Jeopardy’s producers insisted that Watson buzz in like other contestants. To do so, engineers built Watson a “thumb” by adding a plunger to a coil that received a signal when the lights turned on, generating an electric current that pushed the plunger down.

“It turned out to be very precise on the buzzer,” Jennings said. “Human reflexes turned out to be no match for an electric field that says turn on the buzzer now.”

Watson’s final score at the end of three rounds ended up being $77,147. Jennings, with extra time to write in his Final Jeopardy answer, added in parentheses “I for one welcome our new computer overlords” under his correct answer. The whole game, he recalled, felt like “an away game for humanity,” because it was played on a mini Jeopardy set constructed at an IBM research facility, where the audience was almost entirely IBM employees, board members and shareholders.

But what was most interesting about the Watson game, Jennings said, “beyond its domination,” was how Watson made mistakes.

“Its mistakes were almost more revealing than its successes,” he said. “When artificial intelligences screw up, they’re not going to screw up in predictable ways.”

At its most basic level, Watson worked by analyzing the clue in different ways, such as pulling out keywords, and then sorting through its enormous database of learned content (called a “corpus”) to come up with many different plausible answers. Algorithms then rank those different answers based on bits of evidence that either support or refute the answer.

Many of Watson’s modern successors, including ChatGPT, have transformed this idea into a model called a “generative pre-training transformer.” Similar to how Watson analyzed a Jeopardy clue and ranked plausible answers, these artificial intelligence programs basically analyze snippets of text and rank plausible options for the next word in a sequence based on probabilities gathered from appearances of those word sequences in its corpus. By doing so, they are able to produce text writing that “feels” like it was written by a human – but that may contain factual errors.

Jennings said the questions he fielded from reporters at a press conference about the Watson match were mostly comparing Watson to the artificial intelligences of science fiction staples, such as HAL 9000 from “2001: A Space Odyssey.”

“Everyone has these preconceptions of AI that are based on popular culture,” Jennings said. “Generally people have this great built-in suspicion. It’s just the fear of obsolescence.”

In the face of that obsolescence, however, Jennings makes the case for learning facts and trivia.

“It still does matter, even in the post-Watson world, what we know, what we have learned,” he said. “You don’t even notice when the facts in your life pay off.”

As an example, Jennings told the story of Tillie Smith, a then-10-year-old girl who is credited with saving 100 people in Thailand after she recognized the early warning signs of a tsunami while on a family vacation there in 2004. Not everyone will have a story like Smith, but at the very least, a shared fact can often lead to a more engaging, unexpected social connection, Jennings said.

“That one fact in the right place changed everything,” he said.

The secret to learning and remembering those facts, he said, is to be naturally curious and engaged about everything, or “omnivorous about learning.”

“The stuff in our heads is important,” he said. “We are the sum of all the things we’ve learned.”

Follow Stella Lorence on Twitter @slorence3.