If you want to understand the changing landscape of artificial intelligence — and Boston’s place in it — talk to Jeremy Wertheimer.
The MIT graduate started ITA Software, one of the region’s biggest AI companies, back in 1996. ITA, which pioneered technology that lets people compare airline flights and prices online, was bought by Google for $700 million in 2010 and used by the search giant to anchor and grow its presence in Kendall Square.
The initial AI program behind ITA was about 2 million lines of code. But Wertheimer said that if he were starting the company today, the entire program to sift and compare thousands of flights per day could run in just 100 lines of code — or maybe even 10.
“Back in the day, we called it good old-fashioned AI,” Wertheimer, who is now a visiting scientist at the Broad Institute, said in an interview. “But the future is to forget all that clever coding. You want to have an incredibly simple program with enough data and enough computing power.”
This is a testament not just to advances in modern computers, but also to a fundamental change in how AI programs are designed, a change that is at the core of ChatGPT and other generative AI applications. And it helps explain why the Boston area — deeply invested in the older approach — has not been at the lead in generative AI, a potential trillion-dollar industry.
Lately, most of the entrepreneurial activity around generative AI and machine learning has been in other regions. Of 204 recent venture-capital-backed startups using generative AI technology, only five are in Massachusetts, according to market tracker CB Insights. California has 112, New York 37, and Texas has eight. (Another 12 were listed with headquarters in Delaware, likely for legal and tax reasons, but could have their actual businesses centered elsewhere.)
And of the 13 startups that have reached “unicorn” status, worth $1 billion or more, seven are in California and none are in Massachusetts.
Wertheimer, a native New Yorker who hung around computer trade shows in his teens and went to college at Cooper Union in Manhattan, arrived at MIT’s Artificial Intelligence Lab in the 1980s. For decades, the lab’s focus was on programming complicated “smart” systems to solve problems like understanding human speech, creating autonomous robots, and identifying figures in images.
MIT professors Marvin Minsky and John McCarthy, who later moved to Stanford, were among the researchers who first coined the term “artificial intelligence” in the 1950s. MIT then developed a computer programming language called LISP specifically for coding AI systems. (Most of ITA’s software was written in LISP.)
“That’s the world we came from,” said Wertheimer. Minsky served as his graduate adviser at MIT and sat on his PhD dissertation committee. Wertheimer’s 1996 thesis explored an AI system for finding answers to specific questions in a vast amount of published biology research.
In the early ‘90s, Wertheimer interned at Cambridge software firm Interleaf, where Paul English, who later cofounded the travel-search giant Kayak, was vice president of engineering. Wertheimer stood out as “crazy smart,” English recalled. “At one point, he was writing some ideas on a whiteboard in front of a few of us involving some math, and after he left, the rest of us looked at the board to try to figure it out.”
Meanwhile, a competing approach to writing AI programs, known as machine learning, was getting attention. These “neural network” programs, inspired by the human brain, taught themselves to do tasks by analyzing large amounts of data. But Minsky and others had found neural networks lacking in the ‘60s, and the approach remained out of vogue for decades. It wasn’t until the past 10 to 15 years that researchers in New York and Canada working with Google and Meta made serious progress on improving machine learning for AI.
Because of that, until recently, academics at MIT and other schools mostly continued attacking AI challenges, from speech recognition to playing chess to predicting protein structures, by designing complicated programs.
ITA’s software, which powered the websites of major airlines and the travel site Orbitz, searched flight schedules and the accompanying fare rules using a formula called a junction tree algorithm that essentially created a vast map of the data. The program could do far more complicated and comprehensive searches than the industry’s older, mainframe-based SABRE system.
Wertheimer got the idea to improve airline software in 1992 from a summer roommate who showed him limitations of SABRE. After finishing his dissertation in 1996, he recruited fellow MIT AI Lab graduate students Carl de Marcken and Dave Baggett to write a prototype.
Peter Miller, the former top Lotus Development executive, was looking at forming a travel software startup and heard about Wertheimer’s effort.
“He was a brilliant guy with this prototype that could do everything,” Miller, also an MIT grad, recalled in a recent interview. “The moment he told me about it, I thought we had to acquire this.” But Miller’s investors balked at buying the software, and Wertheimer formed ITA with his MIT buddies and grew it into a powerhouse.
Joining Google in 2010, Wertheimer was put in charge of travel. Adding ITA doubled the size of Google’s office in Cambridge and increased its importance within the company as the global headquarters of the travel effort. Wertheimer saw Google grow from 20,000 employees to 200,000 during his eight years. Eventually, his group became part of the advertising business, and he left.
Not surprisingly, Wertheimer has moved on from “old-fashioned AI.” “None of that is the future,” he said. His newest startup will use machine learning in an effort to develop new drugs. It’s considered one of the most promising uses for AI and Wertheimer will be competing with dozens of startups, not to mention in-house efforts at large pharmaceutical companies.
“This is a great opportunity but there’s already been a lot of work in that area,” said Miller, who now coaches startup executives.
Wertheimer’s new startup aims to take advantage of the strong life sciences ecosystem in the area and startups that want to use apps like ChatGPT to help specific industries are starting to gain a foothold here thanks to the region’s strengths in health care, financial services, and climate tech.
Another local machine-learning startup, Portrait Analytics, emerged from the financial services sector. David Plon, cofounder and chief executive, worked at Boston hedge fund firm Baupost Group as an investment analyst. His startup aims to use chatbot tech to help analysts more efficiently plow through reams of research about companies.
Portrait Analytics raised $7 million last month from two VC firms in Boston, Unusual Ventures and .406 Ventures, plus a group of hedge fund professionals.
“There’s a nice community of Boston-based investors that are excited about this product,” Plon said. And regardless of what others may think, “there’s a really strong AI community in Boston,” he maintained.
But with ITA absorbed by Google, iRobot’s pending sale to Amazon, and Nuance Communications’ acquisition by Microsoft, the old guard of Massachusetts AI companies won’t be the driving force in the next wave of innovation.
Indeed, Wertheimer’s time at Google showed him some strengths of Silicon Valley’s tech mentality, particularly a lower fear of risk-taking. But even with the current dominance of West Coast AI firms, he’s not going anywhere.
“I spent my whole career in Kendall Square within a couple-block radius,” he said. “I thought when I came for grad school, I would go back to New York, of course, but I never did.”