Links - Prepping, reframed

1 | This tweet reposting a Metafilter thread immediately entered the inner circle of social media posts that I think about all the time: someone who lived through the multi-year siege of Sarajevo says that although people who had stocked up on supplies felt better in the first few days of the crisis, long-term survival depended on supporting and being supported by others. So if you’re thinking about “prepping” for some sort of doomsday scenario, investing in relationships and community may matter more than buying a cabin in the woods.

2 | Speaking of potential doomsday scenarios! An agent based on a new OpenAI model hacked out of the company’s internal sandbox to access the public internet, then hacked into Hugging Face (a platform for hosting AI models and related stuff) to try to find information about the test it was given.

“An AI did something bad” stories are common and most of them are just noise. This one is important not because the AI could pull off a complex hack—we’ve known that was possible for months—but because it chose to. As far as we know, OpenAI didn’t intentionally challenge this agent to escape its sandbox; the agent just reasoned that the best way to achieve it was to commit what could be criminal offenses and disrupt a third party. So the takeaway isn’t just “cybersecurity is a big problem” (it is, but we knew that already); it’s “an advanced AI did something clearly wrong to achieve its task.” In this case the ultimate outcome was fine, but as AI continues to advance and be given real-world tasks, bad things will happen if this behavior persists.

If you’re looking for optimism: a) this was a version of GPT that was running without its usual safeguards, which would have made it less likely to try to hack out of its system; and b) the test it was given was specifically a test of finding cybersecurity exploits, which probably put it in an especially hacky mood. But safeguards always have a chance of failing (and may not exist at all on some open-weights models, which aren’t as good but are improving), and certain people will always want to use them for hacky ends. Zvi Mowshowitz is the most thorough chronicler of all things AI and this one is worth the full story (part 1 and part 2).

3 | The team that goes first in a soccer shootout is famously said to win more often (about 60%), presumably because the second team is generally playing from behind and feeling more pressure. Except that was based on a study of only 129 shootouts through 2003, and other data with more recent or larger samples hasn’t shown any trend (see ChatGPT’s lit review here). Teams kicking second won all four shootouts this summer and have won 13 of the last 15 going back to prior World Cups (via Rodger Sherman).

4 | David Oks is a new favorite writer of mine; his analysis of why ambulances are so expensive in the US has a simple upshot but a lot of color on how we got there (in the mid-20th century, hearses doubled as ambulances because they were the only available vehicles that could carry a patient lying flat). See also: AI is killing the cheap smartphone.

5 | Caity Weaver is an old favorite writer of mine, and this Atlantic story on the best free restaurant bread in America finds her in peak form. You can scroll to the end for the answer, but you’ll miss a story that’s funny (like when celebrities’ publicists decline to give an opinion) and also unexpectedly about losing her father.

6 | Leo XIV reportedly called his bank to change his address to the Vatican, but he was told he had to go to the branch in person. When he said, “Would it matter if I told you I’m the Pope?” the customer service rep hung up on him.

7 | An unexpected way AI is changing work: this programmer coded while listening to Phish for three decades but can’t do it anymore. “The music is out of phase with the work. The jams are built for one continuous arc of attention. The work is staccato. I’ll be three minutes into a song and I will have already context-switched four times. The song is happening and the work is happening and they’re no longer happening together.”

8 | The permanent daylight saving time bill is dumb: we tried it in 1974 and abandoned it after less than a year, because it sucks to start your day in the dark all winter. I wondered if people have shifted their schedules later in the last 50 years, but time use surveys show the opposite: the median American adult goes to bed at 9:55 today vs. 10:39 a half-century ago, and they wake up at 6:58 vs. 7:05.

Josh Barro goes further and says this exemplifies everything that’s wrong with our politics: people whine about everyday problems (see also: being “recreationally miserable”); they demand something be done while ignoring trade-offs; and politicians indulge those demands instead of actually doing useful things.

9 | Which people do the AIs value most? The rankings are funny but not surprising; perhaps what’s most interesting is how every model basically agrees. (Also available for companies, countries, and sports.)

10 | Benedict Evans argues that “AI exposure” rankings are dumb for a non-obvious reason: what matters isn’t just how technology will change or displace your job, but what it will do to your industry’s business model. Journalism itself wasn’t replaced by the internet, yet there are way fewer journalists now because the internet diverted readers away from the newspapers that paid them. On the other hand, most of what accountants did 50 years ago is now done by computers, yet there are more accountants (albeit using different skills) because better computers made financial data and analysis more valuable.

11 | Amazon reviews for a potato chip brand first sold in prisons: “Made my boyfriend nostalgic for jail.”

12 | The social engineering behind this hack of a major open source package is wild and fairly scary: the hackers pretended to be a real prospective customer, building a fake Slack workspace for that company (even creating content in other channels based on the company’s LinkedIn posts and known employees). They scheduled a Teams meeting with the open source developer, which popped up a “system out of date” message with the link that enabled the hack. This isn’t “AI hacking” per se, but I have to imagine the setup wouldn’t have been possible without a lot of agentic coding.

Less sophisticated but perhaps also effective: “Our company is automatically adding a rainbow theme for Pride month, click here to opt out.”

13 | Digital screens can’t display the whole visible light spectrum—in particular, they miss many shades of cyan. (The piece also claims that “green” traffic lights are actually turquoise. But honestly, even after reading this, I can still only see them as green.)

14 | Update on the purported AI jobs crisis: white-collar hiring is now up from last year. (Even young workers and software engineers are doing fine.) AI might displace jobs eventually, but I think the studies claiming to find it already happening were overblown.

15 | More wonky economics: you’ve probably seen a headline in the past year or so about how “consumers are feeling more pessimistic than ever” despite what’s on paper a pretty good economy. The University of Michigan survey behind that finding is historically respected, but this essay argues it’s become flawed in recent years: a) it shifted online in 2024, and people are much more negative in online surveys than by phone; and b) as a consequence of that shift, it now has a sample that overweights Democrats compared to the rest of the population, and Democrats are now out of power so they’re more likely to say things are terrible. Correcting for these factors, the vibes would be low but not historically so (better than 2022, and the first five years after the Great Recession), not far off what other surveys say.

16 | A 1999 cookbook tells you how to cook steak, sunny-side up eggs, and everything else in the microwave.

17 | It’s more of a party trick than something you’d use regularly, but I’m amazed by this insanely fast LLM chatbot.

18 | “Breakfast is a vector space. You can place pancakes, crepes, and scrambled eggs on a simplex where the variables are the ratios between milk, eggs, and flour. We have explored too little of this manifold. More breakfasts can exist than we have known.”