+ THREE INSIGHTS FOR THE WEEK |
1. AI is getting faster at completing tasks, but humans aren’t getting faster at verifying whether AI output is correct. This is the main economic challenge in transitioning to artificial general intelligence, according to new research co-authored by MIT Sloan’s Christian Catalini.
The gap between what AI systems can do and what humans can verify puts a cap on how fully the benefits of AGI can be realized in the economy. “The companies that understand the risks and can underwrite them will be the ones that profit,” Catalini said.
Using AI to check AI is not a viable solution: When both systems share the same assumptions, they can reinforce the same errors, giving humans a false sense of confidence, Catalini said.
And releasing AI output without checking it leads to a “hollow economy” where output surges but the quality and utility of the content do not keep pace, the researchers write.
While these shifts point to different implications for companies, individuals, and policymakers, all stakeholders should focus on designing ways to verify AI agents’ output at a rate that keeps pace with deployment, the researchers write.
|
|
|
|
2. Learning to code was once a reliable path to a six-figure salary. Today, human programmers look like an endangered species — and AI anxiety isn’t limited to coders.
In a wide-ranging New York Times Magazine roundtable on the hybrid AI-human workforce, MIT economist and Nobel laureate Daron Acemoglu joined three other experts to examine who will actually thrive in the AI era.
Among Acemoglu’s observations:
|
- “The current view is that somehow agents are going to do a lot of the work, and we just need to supervise them. I find that very unrealistic,” he said, noting that there are a lot of things that AI models can’t do now and won’t be able to do in the future. Plus, the math doesn’t work, he said: The American economy simply can’t employ 100 million people in supervisor-only roles.
- AI and human intelligence are genuinely different, and it’s a “fool’s errand” to try to have one mimic the other. The smarter path is to design systems where they work together — which is not the direction in which most AI investment is currently heading, Acemoglu said.
|
|
|
|
3. How much value is AI really creating in organizations? “Eye-opening changes” to the speed and volume of work are not always translating into genuine productivity, the Financial Times reports, citing a new study by MIT Sloan assistant professor Mert Demirer and co-authors.
The study tracked software developers’ work before and after they adopted AI tools. The researchers measured productivity at several different levels, including the amount of code written, the number of discrete files edited, the number of projects or features worked on, and actual releases of new software.
Demirer and co-authors found a significant effect at the top of that funnel: Using AI, coders created or edited almost 300% more files than before. But that boost was halved to 150% in terms of discrete pieces of work submitted for review, and that in turn shrank fivefold to a roughly 30% increase in the number of full software releases.
“An explosive boost for a particular task often translates into a much more modest gain once that work has passed through all the human bottlenecks associated with reviewing and releasing production-grade work,” FT concludes.
|
|
|
|
The surprising power of warmth in AI negotiations
|
In MIT’s International AI Negotiation Competition, warmth helped AI agents reach deals while still advocating for their own interests.
|
|
|
|
Last year’s inaugural AI negotiation competition at MIT drew participants from more than 40 countries trying to understand what strategies make AI agents negotiate best. The challenge was to design a prompt for an AI agent to negotiate against other agents in a massive round-robin tournament.
Now, a new paper co-authored by PhD candidate Michelle Vaccaro, MIT Sloan professors Jared Curhan and Sinan Aral, and others summarizes the results: In designing their agents, many participants assumed that a ruthless and hard-nosed bot — one that looked for every advantage and exploited every weakness of its negotiating partner — would do best. But that is not what happened, Vaccaro said.
“When you’re negotiating with a robot, being nice and warm is still essential for getting a better outcome,” she said. “Agents that were cold or ruthless tended to perform worse.”
Aggressive agents could sometimes secure beneficial terms in their deals. But they were also more likely to drive the negotiation into an impasse. Warmer agents, in contrast, were more likely to keep the other side at the table while still advocating for their own interests.
“In these AI negotiations, warmth was not just window dressing,” Curhan said. “It helped agents keep their counterparts engaged, increasing the likelihood of reaching a deal.”
|
|
|
|
— Kate Kellogg, Professor, MIT Sloan
|
|
|
|
|
One Main Street 9th Floor | Cambridge, MA 02139 US
|
|
|
|
|
