 |
|
|
|||||
It has been an odd few weeks for generative AI systems, with ChatGPT suddenly turning sycophantic, and Grok, xAIs chatbot, becoming obsessed with South Africa. Fast Company spoke to Steven Adler, a former research scientist for OpenAI who until November 2024 led safety-related research and programs for first-time product launches and more-speculative long-term AI systems about bothand what he thinks might have gone wrong. The interview has been edited for length and clarity. What do you make of these two incidents in recent weeksChatGPTs sudden sycophancy and Groks South Africa obsessionof AI models going haywire? The high-level thing I make of it is that AI companies are still really struggling with getting AI systems to behave how they want, and that there is a wide gap between the ways that people try to go about this todaywhether it’s to give a really precise instruction in the system prompt or feed a model training data or fine-tuning data that you think surely demonstrate the behavior you want thereand reliably getting models to do the things you want and to not do the things you want to avoid. Can they ever get to that point of certainty? I’m not sure. There are some methods that I feel optimistic aboutif companies took their time and were not under pressure to really speed through testing. One idea is this paradigm called control, as opposed to alignment. So the idea being, even if your AI wants different things than you want, or has different goals than you want, maybe you can recognize that somehow and just stop it from taking certain actions or saying or doing certain things. But that paradigm is not widely adopted at the moment, and so at the moment, I’m pretty pessimistic. Whats stopping it being adopted? Companies are competing on a bunch of dimensions, including user experience, and people want responses faster. There’s the gratifying thing of seeing the AI start to compose its response right away. Theres some real user cost of safety mitigations that go against that. Another aspect is, Ive written a piece about why it’s so important for AI companies to be really careful about the ways that their leading AI systems are used within the company. If you have engineers using the latest GPT model to write code to improve the company’s security, if a model turns out to be misaligned and wants to break out of the company or do some other thing that undermines security, it now has pretty direct access. So part of the issue today is AI companies, even though they’re using AI in all these sensitive ways, haven’t invested in actually monitoring and understanding how their own employees are using these AI systems, because it adds more friction to their researchers being able to use them for other productive uses. I guess weve seen a lower-stakes version of that with Anthropic [where a data scientist working for the company used AI to support their evidence in a court case, which included a hallucinatory reference to an academic article]. I obviously don’t know the specifics. Its surprising to me that an AI expert would submit testimony or evidence that included hallucinated court cases without having checked it. It isnt surprising to me that an AI system would hallucinate things like that. These problems are definitely far from solved, which I think points to a reason that its important to check them very carefully. You wrote a multi-thousand-word piece on ChatGPTs sycophancy and what happened. What did happen? I would separate what went wrong initially versus what I found in terms of what still is going wrong. Initially, it seems that OpenAI started using new signals for what direction to push its AI intoor broadly, when users had given the chatbot a thumbs-up, they used this data to make the chatbot behave more in that direction, and it was penalized for thumb-down. And it happens to be that some people really like flattery. In small doses, thats fine enough. But in aggregate this produced an initial chatbot that was really inclined to blow smoke. The issue with how it became deployed is that OpenAIs governance around what passes, what evaluations it runs, is not good enough. And in this case, even though they had a goal for their models to not be sycophanticthis is written in the company’s foremost documentation about how their models should behavethey did not actually have any tests for this. What I then found is that even this version that is fixed still behaves in all sorts of weird, unexpected ways. Sometimes it still has these behavioral issues. This is what’s been called sycophancy. Other times it’s now extremely contrarian. It’s gone the other way. What I make of this is its really hard to predict what an AI system is going to do. And so for me, the lesson is how important it is to do careful, thorough empirical testing. And what about the Grok incident? The type of thing I would want to understand to assess that is what sources of user feedback Grok collects, and how, if at all, those are used as part of the training process. And in particular, in the case of the South African white-genocide-type statements, are these being put forth by users and the model is agreeing with them? Or to what extent is the model blurting them out on its own, without having been touched? It seems these small changes can escalate and amplify. I think the problems today are real and important. I do think they are going to get even harder as AI starts to get used in more and more important domains. So, you know, it’s troubling. If you read the accounts of people having their delusions reinforced by this version of ChatGPT, those are real people. This can be actually quite harmful for them. And ChatGPT is widely used by a lot of people.
Category:
E-Commerce
When Nicholas Bloom, the William Eberle Professor of Economics at Stanford University in California, started studying working from home in 2004, it was hard to get anyone engaged, he says. Even in 2018, no one had any interest whatsoever. In 2025, thats hard to fathom. Between the pandemic and technological advancements, WFH has become a norm among white collar workers. Not only has it normalized; its also destigmatized. The act that used to generate memes of Homer Simpson on the couch, prodding a distant computer with a stick has gained positive connotations, says Bloom. Working from home is seen as a privilege. Its also here to stay. For their latest study, Working from Home in 2025, Bloom and his collaborators analyzed responses from 16,000 college graduates across 40 countries and discovered that WFH levels appear to have stabilized as of 2025, but its embrace hasnt been universal. WFH rates vary by location: highest in English speaking regionsthe U.S., UK, Australia, Canada, New Zealandthe rate dips a little across continental Europe, then dips a lot across Africa and Central and South Americas. WFH is least prevalent in Asia. To be clear, when Bloom says WFH, hes mostly talking about those on hybrid work schedules. Sixty percent of people work fully in-person, 30% are hybrid, and 10% are fully remote, he says of those countries where the policy has stuck. Hybrid typically means Tuesday through Thursdays in the officea schedule Blooms values at about 8% more paybecause it saves two to three hours a week of commuting [and] enables people to live further away from their offices, often to where real estate is cheaper. Companies also benefit from hybrid policies, Blooms study found, since fewer employees tend to quit. With all these advantages, youd think bosses would have embraced WFH worldwide. Why on earth does, say, Japan have a third the work from home rates of the U.S.? Bloom says. After looking at factors including development (Japan is about as developed as the U.S.), population density, industrial structure, and connectivity (no big differences there), it left Bloom and fellow researchers with one notable variable. The big factor is cultural, he says, and it’s around individualism. In conversation with Fast Company, which has been edited for length and clarity, Bloom elaborated on how individualism drives working from home, how much the pandemic really increased at-home work rates, and why people still tend to think were returning to the office even though the data says otherwise. Fast Company: What inspired you to look globally for your latest study? Nicholas Bloom: If you look at the data, there was clearly a return to office movement from summer 2020 onwards after the lockdown in the U.S. But from Spring 2023 onwards, the return to office seems to slow down. People seem surprised by that. They’re like, Isn’t the media full of stories of Zoom canceling [WFH], Amazon canceling [WFH]? Yes, there are a bunch of high-profile firms canceling or reducing work from home. Turns out there are just as many on the other side, because their leases expire. If youre Goliath National Bank and your lease expires, it’s a perfect opportunity to reduce days in the office and save a chunk of money. What we’ve seen over the last couple of years in the U.S. is like a war, and it’s been fought to a standstill. That sparked the big question for us: What on earth does this look like globally? We last collected global data in 2023, so I really didn’t know. It turns out, globally, work from home has also stalled out. There has been no change since 2023. Globally, we’re in a new norm. Folks saying when we return to the office at this point are dreaming. This is the future. One of your findings I found particularly interesting was that WFH rates are higher in individualistic societies than in collectivist ones. Can you unpack that? In individualistic societies, managers typically aren’t micromanaging their employees. The U.S. setup is: A manager tells an employee what to do and gives them strong incentives, like performance evaluations and bonuses. In Japan, theres much more micromanaging, because there’s much less hiring, firing, and bonuses. Managers want to see employees there. In Japan, you can’t leave the office until the boss has left. This long-hours culture exists for everyone. When the boss leaves, their junior leaves, then their junior leaves, etcetera. That is very problematic for work from home. If you talk to folks working for American firms in Japan, they’re typically on a hybrid setup. If they work for Japanese firms in Japan, often doing the same job, they’re required to come into the office every day. Culture seems to have a huge explanation for this difference across countries. To what extent do you think this comes down to bosses trusting their workers, or not? It is kind of trust, although in the U.S., it’s trust but verify. Bosses don’t just trust workersthey trust them, but then they monitor. Should companies without a WFH policy reconsider? The big selling point is that it’s profitable. In my paper in Nature in June 2024, we did a massive, randomized control trial at a big company called Trip.com. They’re a publicly listed company worth about $40 billion. They randomized whether you got to work from home two days a week or come in all five daysthe former if your birthday fell on an odd day, the latter if it fell on an even day. For 24 months, we tracked 1,600 employees working in finance, marketing, computer engineeringprofessionals with college degrees. There was no effect on performance. However, quit rates fell by 35% for people allowed to work from home two days a week. For Trip.com, every person that quits costs about $50,000. If someone quits, you have to advertise, re-interview, re-recruit, get them up to speed, and take managers off activity to train them. By reducing quit rates by 35% with no effect on productivity, that’s increasing business profits by like $20 million a year. That is ultimately why work from home has stuck. On the flipside, an Economist article that mentions your study cited JP Morgan CEO Jamie Dimons worry that the young generation is being damaged by increased working from home. To what extent do you agree or disagree with that statement, and why? I advise my Stanford undergrads, particularly in their first five years of work, that it’s a good idea to go into the office four days a week, because Jamie Dimon is exactly right. It is easier to mentor, learn, and build connections in person. Typically, when I poll students, that’s what they wantthey want to socialize, be mentored, and they don’t have a lot of space at home. As people get to their 30s and 40s, they’ve moved up that learning curve, but they still benefit from coming in, maybe three even two days a week. Another interesting data point from your study was the similar WFH rates for men and women across regions. What do you think accounts for that? They want to. You see a slightly higher preference for women to work from home. The main decider in the U.S. is: Do you have kid? A man with children under the age of 12 has a higher preference to work from home than a woman without kids, for example. Having a disability is also a huge driver, but gender doesn’t matter that much. What you see in countries like India is gender matters a lot more, because for women, there’s assault risk and massive sexism in the workplace. In lower income countries, the gender gap grows. What was the most surprising takeaway from your study? Working from home has stabilized globally. I did an online presentation for Australia last week, and people there are under the same view as in the U.S., that big companies were banning it. We just don’t see that in any data set. Fact and opinion are about as divergent as people’s views on crimethey always think crime is rising. On average, its tending to decline. Everyone thinks work from home is ending, but you don’t see it globally.
Category:
E-Commerce
As a lifelong pop culture aficionado, I have a tendency to connect my favorite media to whatever Im currently doing. When I purchased a secondhand easy chair a few weeks ago, my husband and I spent a sweaty 30 minutes struggling to get it up the stairs. With the chair still wedged at an impossible angle, we paused to catch our breath and I said, Youre gonna need a bigger boat. Similarly, anytime I use up the last of the milk or take the last cookie from a package, my brain always bellows FINISH HIM! But the pop culture in my head is more than just a running commentary on mundane moments. My favorite entertainment has also been an excellent teacher. In particular, the pop culture of my childhood taught me a number of financial lessons that Ive never forgottenincluding instruction on how to be an investor. Here are the timeless investing lessons I learned from 1980s pop culture. Lemonade Stand: the risk of playing it safe While the majority of my fellow late Gen Xers have deep and visceral memories of dying of dysentery on the Oregon Trail, my early childhood gaming trauma stemmed from the lesser-known Apple II game Lemonade Stand. This simple game teaches the basics of business planning by simulating a childs lemonade stand. The player receives a weather report for the day and has to decide what to spend on lemonade ingredients and advertising as well as determine the price for each glass of lemonade. As a frugal and business-minded 7-year-old, I invested heavily in lemons and sugar when the game predicted a hot summer day on my first turn. I also set a reasonable per-glass lemonade price, knowing that it was folly to overcharge my customers. Though it seemed unnecessary on such a beautiful, 8-bit sunny day, I also splurged on a single advertising poster. Once my preparations were complete, I leaned back in my chair, ready for profits to rain down on me. To my shock, I only had two customers all day. I didnt even make back the money I spent on cups. Pop culture lesson: know where to invest To little Emily, it made sense to spend money on ingredients, since you cant sell lemonade without them. But I balked at the expense of advertising, which seemed unnecessary compared to lemons and sugar. I couldnt predict or measure advertisings return on investment, so I assumed it was a waste of money. (Unfortunately, I continued to make this mistake into adulthood. When I first started freelancing, I only owned a desktop computer. Investing in a laptop seemed like an unnecessary expense with no potential upside for my fledgling writing careerexcept that I traveled at least once a month and had to move heaven and earth to either work ahead or find a computer at my destination every single time.) The shock of losing my Lemonade Stand money taught me that playing it safe cant protect you from loss. There is a risk to investingwhether youre investing in advertising, a new laptop, or in the stock marketbut theres also a risk to playing it safe. You could lose your business because no one knows about it, lose your time (and your mind) because you dont have the equipment you need, or your uninvested money could lose buying power over time because of inflation. There is no such thing as a risk-free financial decision, and playing Lemonade Stand in second grade taught me that better than anything else. The Westing Game: invest independently Ellen Raskin may as well have written her 1978 Newbery Medal winning book The Westing Game specifically to appeal to me. The novel begins after Westing Paper Products tycoon Sam Westing is found dead. Westings lawyer invites his 16 heirswho all happen to be the only tenants of the newly constructed Sunset Towersto the reading of the will. Once there, the heirs are paired off and given $10,000 and an envelope of mysterious clues written on paper towel scraps. They are invited to figure out who has taken Sam Westings life, and the winner will receive his $200 million estate. As much as that set up is more than enough to get my attention, it was the character of Turtle Wexler that really established this book as one of the pop culture giants of my childhood. This 13-year-old budding entrepreneur and investing genius captured my heart by being smart, funny, and financially confident beyond her years. Turtle and her partner, a 60-year-old dressmaker named Flora Baumbach, receive the incomprehensible clues SEA, MOUNTAIN, AM, and O, which the teen girl believes to be stock symbols. Since Westing was known to be a business wizard, Turtle thinks the stocks indicated by the clues must be clear winners. The pair invests the $10,000 in the clue stocks and in Westing Paper Products (stock symbol WPP). The clue stocks dont perform as well as Turtle had hoped. Her daily perusal of The Wall Street Journal indicates that the Westing Paper Products stock is likely to go up, so she dumps the clue stocks and puts all their money in WPP. By the end of the game several weeks later, Turtle and Floras $10,000 stake has grown to $11,587.50. Pop culture lesson: lean into your knowledge Turtle taught me the importance of investing based on my own knowledge, expertise, and instincts, rather than following someone elses lead. She starts her investing journey with the knowledge that Westing was a remarkably astute investor. She assumes the clue stocks must have been handpicked by Westing. But when the clue stocks dont do well, she pulls her money from them and invests in something she has direct understanding of, rather than doubling down on her assumption that Westing must have known better. She changes her investing tactics once she has new information. Turtle also shrugs off Flora repeatedly asking if she is sure about her investing choices. She doesnt let the concerns of her 60-year-old partner sway her, because she knows Flora doesnt understand the stock market as well as she does, even though she is much older. All together, Turtles example made it clear to me that successful investing requires knowledge and a willingness to trust yourself. Its helped me avoid following the crowd into decisions that dont fit my investing strategies. Trading Places: anatomy of an investing scheme I loved the 1983 film Trading Places for Eddie Murphys brilliant comedic timing, but I was even more fascinated by the movies portrayal of revenge via short sale. It took me several rewatches to fully understand how the investing scheme resulted in financial doom for the films villains, the Duke brothers. To exact their retribution, Murphys character Valentine and Dan Aykroyds inthorpe show up to the New York Commodities Exchange ready to trade. Their goal: sell as many orange juice concentrate futures as they can before the U.S. Department of Agriculture report on the nations orange crop. Meanwhile, the Dukes are buying as many OJ concentrate futures as they can before the report, essentially trying to corner the market. Between the heroes feverishly selling and the Dukes feverishly buying, OJ future prices skyrocket until the moment trading pauses for the crop reportwhich reveals the orange harvest will be strong. In the aftermath of the announcement, the Dukes are stuck with all the futures they purchased at inflated prices. To fulfill the margin call, they must pay $394 million. At the same time, Valentine and Winthorpe busily purchase futures from everyone but the Dukes at a greatly reduced price. This allows them to fulfill the orders they sold before the report droppedand make a ridonculous profit. This scene fascinated me as a kid, but it also confused me. I knew that successful investing was about buying low and selling high. But I couldnt understand how the characters could sell high then buy low. How could you sell something before you bought it? Pop culture lesson: stock sales aren’t always linear After many years of catching the movie on TBS reruns, I finally grasped that stock and commodities sales dont have to follow a linear progression of cause then effect. Its possible to buy low after selling high, provided you plan your investment strategy carefully. Thats because you dont have to own something you sell. You can borrow a stock (or an OJ future, for that matter) for a fee. As long as you return it or an identical asset before the margin call, you can sell the borrowed asset even though you dont own it. This is what Valentine and Winthorpe did to ruin the Dukes. They took their pooled money to pay the borrow fees of the futures, sold those futures at inflated prices before the crop report, then bought back the futures at the rock-bottom price afterwards so they could return the borrowed assets. While short sales like the one in Trading Places are unlikely to ever be part of my own investment strategy, understanding the fluid nature of ownership in stock and commodities trading has made me a better investor. It broke me out of the rigid cause-and-effect thinking that limited my investing creativity. Learning through story Despite being a lifelong money nerd, stories are my first love. So its no wonder the most enduring lessons I learned about finance come from the pop culture I loved as a child. Playing Lemonade Stand in my elementary school computer lab disrupted the story Id told myself that it was possible to make a profit without risking an investment. Reading The Westing Game gave me the story of a confident financial heroine to remember when Im tempted to follow the crowd. And Trading Places folded a satisfying revenge story into a creative investing scheme, which helped me feel smart and savvy when I finally wrapped my head around the details.
Category:
E-Commerce
All news |
||||||||||||||||||
|
||||||||||||||||||