Robots need people, too

DEEP Robotics' X30 robot dog walks down steps in Hangzhou, China, on Sept. 22. (Philip Fong/AFP/Getty Images)
DEEP Robotics' X30 robot dog walks down steps in Hangzhou, China, on Sept. 22. (Philip Fong/AFP/Getty Images)

Manuela Veloso is one of the world’s most renowned roboticists, artificial-intelligence researchers and hand talkers. Born and raised in Portugal, Veloso has spent most of the past 40 years lecturing to packed halls. When she’s really feeling one of her greatest hits — the creation of her robot soccer team, the eternal mysteries of AI — she puts her whole body into it, like the Tina Turner of computer science.

Today, Veloso is telling a eureka story. In 2010, she and her fellow researchers were working on robots designed to navigate their way across the campus of Carnegie Mellon University. “Literally, there was always a problem”, says Veloso. “A door would be closed. There was water on the floor. Stupid people would stand in front of the robot. I mean, it was just a nightmare”.

After a year passed with little progress, she woke one night with a jolt. “At 3 a.m. in the middle of Pittsburgh, I realized: This is never going to happen”. The robots were constantly demonstrating what they couldn’t do, but, lying in bed, Veloso was struck by a simple thing they could do. “It cannot press the elevator button but it could say, ‘Human, can you press the elevator button for me?’ Or ‘Human, can you get out of the way?’ This was the revelation!”

The robots were renamed CoBots. They’ve since covered thousands of miles crossing Carnegie Mellon’s campus and escorted hundreds of visitors to Veloso’s office. Thanks to Veloso, the concept of human and robot cooperation is now called symbiotic autonomy — and if you have a white-collar job, you might want to throw that phrase on a Post-it, or a tattoo. Because we’ve reached the moment when it’s not the robots who need to learn how to ask for help.

History might record it differently, but the first three words uttered by most white-collar workers who try generative AI are “Wow! … Oh, s---”. While concern naturally races ahead to the extinction of entire professions, the Bureau of Labor Statistics projects that even the most vulnerable white-collar job categories — financial adviser, interpreter, paralegal — will keep growing through at least 2029. “The jobs thing isn’t gonna change next year”, says Reid Hoffman, co-founder of LinkedIn and Inflection AI, which makes an emotionally supportive personal assistant bot called Pi. “These things happen slowly, then quickly. You’re gonna be able to see them coming. And the thing everyone forgets with AI is that it can be part of the solution. If I have someone who is outmoded in their work, well, I can use AI to help them figure out other work. And I can use AI to help them be better at the work they already have”.

Social scientists have just begun to explore the interaction of humans and AI in the workplace, and management consultants are some of the first guinea pigs. “Consultants are people who are highly skilled, highly paid, very well trained, and that previous waves of automation typically spared”, says Fabrizio Dell’Acqua, a postdoctoral research and teaching fellow at Harvard Business School and the Laboratory for Innovation Science at Harvard. “But [large language models] are very good at a lot of the same tasks required by consultancy, and it only charges a $20 subscription per month”.

Dell’Acqua and his collaborators worked with Boston Consulting Group to recruit 758 consultants for an experiment. Some were given access to ChatGPT-4. The rest weren’t allowed to use any AI tools. Then the whole group was assigned 18 consultant-y tasks for a made-up shoe company — ranging from idea generation and names for new sneakers to writing a rah-rah memo to inspire employees.

The result was a John Henry-esque contest, minus the death and the folk songs. In a new paper called “Navigating the Jagged Technological Frontier”, Dell’Acqua and his co-authors reveal that the consultants using AI did almost everything better. They were 12 percent more productive and 25 percent faster, and their work was rated 40 percent higher-quality by a human jury. Every consultant who used AI benefited, but to Hoffman’s point, the ones who gained the most were those previously identified as the worst performers. The losers leveled up. Symbiotic autonomy worked in reverse.

The experiment also showed what happens when collaboration veers into “AI do my job for me”. The researchers created a question outside of ChatGPT’s knowledge base — which the program went ahead and answered anyway, like the confident dumb guy it sometimes is. Consultants who should have known better passed along GPT’s answer without scrutiny; it was the only task in which those working without AI outperformed their peers. This tracks with earlier research Dell’Acqua did into HR recruiters who used high-quality AI tools to sort résumés. “They became lazier”, says Dell’Acqua. “They fell asleep at the wheel and missed out on good candidates. They underperformed”.

AI still messes up spectacularly sometimes. It hallucinates and can be weirdly bad at math, which feeds a temptation toward professional abstinence. But waiting to use AI until the technology is perfect presumes that AI will ever be perfect — and that you’ll still have a job when it is. Better to dive in before your human competition beats you to it. A simple set of guidelines, blatantly plagiarized from Michael Pollan’s famous advice about food, can help: Use AI. Not too much. Mostly for support.Which brings us back to the lab at Carnegie Mellon — and Veloso’s punchline. When she told her team that the solution to their years-long problem was essentially robotic humility, the humans went nuts. “They were all furious!” says Veloso, screaming with laughter. “They said I was cheating! Just buy us this $20,000 robotic arm and we can fix it”. Veloso recognized that the outrage wasn’t because the robots had failed to be autonomous but because their human designers had to confront their own failure to make them so. “If you deny the fact that you can use other resources, you are assuming that you are autonomous yourself. And that’s wrong”, says Veloso. “I mean, not wrong. But you are doomed to failure. Because no matter how smart you are, you will still find yourself in situations for which you need help”.

Josh Tyrangiel is The Post’s columnist covering artificial intelligence. He was previously the editor of Bloomberg Businessweek and chief content officer for Bloomberg Media.

Deja una respuesta

Tu dirección de correo electrónico no será publicada. Los campos obligatorios están marcados con *