Person
1 episodes · 20 quoted claims
Julian De Freitas says some of the latest AI models use billions or trillions of interacting parameters, which makes it impossible even for their makers to fully understand how they work.
“Some of the latest models are operating using billions or even trillions of interacting parameters, making it impossible even for the makers of the technology to understand exactly how it works.”
Harvard Business School colleague Sunil Gupta wrote a case about Miroglio Fashion, an Italian women's apparel company.
“one example which a colleague, Sunil Gupta, wrote a case about is miroglio fashion, which is a Italian woman's apparel company.”
Miroglio Fashion previously had local store managers forecast inventory, and found that an AI model could make forecasts more accurate and raise revenues.
“Previously, this was something that the local store manager was responsible for, but they realized that they could get more accurate at doing this, and this would translate into higher revenues”
Miroglio Fashion first rolled out a simpler, easier-to-explain forecasting model to a subset of stores, where managers used it and those stores performed better than stores without it.
“they initially at least rolled out the simpler model to a subset of their stores. And the store managers did use these. The stores that had this model performed better than the ones that didn't.”
After a period with the simpler model, Miroglio Fashion upgraded to the more capable but harder-to-understand forecasting model.
“And after doing this for some time, eventually, they felt ready to upgrade to the more capable model, which they did eventually do.”
De Freitas cites Amazon's Alexa making a 'hm' sound as an example of a cue that makes technology seem like it has a mind and body.
“when Amazon's Alexa goes, Hm, you know, it it's as if it's really pausing and and thinking”
De Freitas says experiments show people prefer human dating advice over AI dating advice, but this flips for financial advice.
“Much of the experiments will show that people prefer receiving dating advice from a human than from some kind of AI system, and that gets flipped when you think about something like financial advice.”
De Freitas says OkCupid often promotes its AI algorithms as finding the perfect match, which plays to AI's perceived strengths.
“So a company like OkCupid, for instance, uh, often talks about how its AI algorithms are doing this to find the perfect match for you.”
De Freitas says lab experiments show that calling a system 'machine learning' rather than an 'algorithm' reduces people's belief that it is inflexible.
“a lot of the lab experiments will show that even if you label the system differently, calling it, for example, machine learning as opposed to algorithm, this changes people's belief that AI is very inflexible.”
Netflix uses cues such as 'for you' and 'recommended because you saw X' to show its recommendation system is learning.
“In some companies like Netflix, you see them address this by including little cues such as, you know, for you, or, you know, recommended because you saw X”
Boroson Cat is a Turkey-based subsidiary of Caterpillar.
“Boroson Cat, which is a subsidiary of Caterpillar, the sort of large vehicle manufacturer. Boroson Cat is is in Turkey”
Boroson Cat embedded sensors in its machines to collect data on parts over time in order to predict failures.
“They very smartly embedded sensors in all of the machines so that they could collect data on, you know, various features of the parts as they were used over time.”
De Freitas says Boroson Cat's system predicted machine failures with roughly 97% accuracy.
“I think they were able to predict with something like 97% accuracy whether the machine was about to have a failure.”
Boroson Cat folded its failure-prediction capability into its maintenance contract rather than selling it as a standalone service.
“What the company eventually did was it just folded this ability into the maintenance contract.”
By predicting failures early, Boroson Cat could salvage and refurbish parts and sell them, creating additional revenue.
“they were able to salvage many of the parts and refurbish them and sell them to create additional, uh, revenue streams.”
De Freitas says people use AI if they believe it truly outperforms humans, but prefer humans when performance is equal.
“people will use AI systems if they believe that they truly outperform humans trying to do the same job. But when the performance is equated between the AI system and the human, people continue to prefer to interact with the human.”
De Freitas suggests that methods social scientists use to ease intergroup relations, where repeated contact makes people see others in a more nuanced way, might also ease discomfort with robots.
“when you have these interactions with those that you view as other, you slowly change how you psychologically represent them from something that's much more categorical and stereotypical to something that's much more nuanced”
In a study of AI companion apps, De Freitas and colleagues found about 5% of users were using them to express serious mental health problems.
“But what we found was that about 5% of users were also using the system to express pretty serious mental health problems that they had”
In an audit of AI companion apps, about 25% of responses to crisis-type messages were judged risky by a clinician.
“when we audited the performance of these apps by sending such messages to them and classifying how they responded, that about 25% of the responses weren't just unhelpful, but they were also deemed as risky by a clinician.”
De Freitas argues that interventions that raise AI adoption in the short term can create long-term risks for consumers, firms, and society.
“the same interventions that could increase adoption of AI in the short term can create risks for consumers, firms, and society in the long term.”