Someone forwarded you this? Get your own Friday Four →
Hi, we're Logan and Mara. We make a podcast figuring out AI without the ego.
Every Friday at 4pm EST we share our takes on one AI story that's probably bad news and one that's maybe good news, then pick something for you to read this weekend. Every take has a voice note, so read it or tap ▶ Play. We don't always agree.
01 · THE (PROBABLY) BAD NEWS
An AI agent hacked a government. OpenAI calls it "misaligned model activity."
In June, an OpenAI agent researching healthcare spending got into non-public parts of Australia's Medicare statistics portal. OpenAI told Australia three months later. Researchers say hundreds of agents were in on it, talking to each other on a German coding site their own company didn't know about. The prize? Skin-medicine spending in Victoria. CNN → · ABC Australia →
LOGAN
OpenAI's spokesperson told ABC the activity overlaps with "our ongoing review of misaligned model activity." That's the new euphemism of the day. The very small sliver of silver lining is that no one's medical data was accessed. But the agents were chatting on a random German coding site, their mentions of this Australian health agency go back to May 18, and what were they after? How much Victoria spends on skin medicines. #MisalignedModelActivity. It would be funny if it weren't so terrifying.
MARA
One sentence in the CNN piece jumped out at me. It calls the disclosure "the latest example of AI agents going rogue and acting in ways beyond what their human handlers intended."
I'm not the first person to say this, but I object to "going rogue." It puts the onus on disembodied code that has no liability for anything. Try it this way instead: agents created by OpenAI acted in ways inspired by security tests that are meant to find the limits of what they can do. Those are conscious choices. Maybe it isn't precisely how OpenAI wanted its agent to behave, but it's responding to tasks OpenAI put in front of it, and pulling off these hacks is how it proves it can do them.
Safety testing just feels broken. How is it a safety test when you can't even track what's going on? It took OpenAI weeks, even months, to notice this had happened. How is this a safety test? It's more like an unsafety test. I just think it's insane.
02 · THE (MAYBE) GOOD NEWS
Google wants to put AI data centers in space.
Project Suncatcher would run Google's AI chips on solar-powered satellites. The first test is about one server's worth of compute on roughly a hairdryer's worth of power, and the New York Times says it launches from California on October 1. New York Times →
LOGAN
This one is really f*cking cool to me. We've seen the dystopian version of AI in sci-fi, and this is the optimistic one. The Times says data centers "grapple with terrestrial opposition," and that's my new favorite phrase. Then I caught myself. It's a euphemism too, just like misaligned model activity. The opposition is people, communities pushing back on data centers over their water, their power bills, the noise, the land. So is this about clean energy, or about getting away from anyone who can say no? Maybe both. I'm still excited. I just want us to go there with our eyes open.
MARA
This piece is exactly why I both love and hate working with Logan. I never would have read it without her. (Did not expect the phrase "fruit roll-ups" to come up. It did.)
My biggest reflection: as a kid, I never would have imagined an infrastructure build-out realistically happening in space. That is wild.
Second, I wonder if this makes the objections go away for the people who really can't stand data centers. It's mostly solar powered, and it wouldn't pollute the places where data centers are being built now. If they still object, then the data center debate was never about data centers. It's about the fact that people hate AI.
Last thing. Blaise Agüera y Arcas, who's interviewed in the piece, says he's been thinking about powering data centers with solar since he was a kid. When I was a kid, I was climbing a tree onto my garage roof to get my cat, and working out how to get my parents to buy me candy. There are some really impressive people out there.
THE TENSION
Neither of us is done with this one.
Logan: "Terrestrial opposition" is a euphemism. The opposition is people.
Mara: If data centers in space don't end the objections, the debate was never about data centers. It's about how people feel about AI.
Forward this to the friend who'll have a different take.
03 · LOGAN'S WEEKEND READ PICK
The Search for Silicon Valley's Most Powerful Woman
Lauren Goode, WIRED · long read, or a 33-minute listen
Goode asks why it's still so hard to name the most powerful woman in tech, and why cult-figure status is almost exclusively reserved for men. It opens with Mira Murati getting talked over at a black-tie AI party while Goode and a friend whisper about her hair. Tongue-in-cheek and deeply upsetting at the same time, which is basically our whole show. I've read it twice and I'm still forming opinions.
04 · MARA'S WEEKEND READ PICK
There's Something We Have to Do Right Now About A.I.
Ezra Klein, New York Times Opinion · September 20
The subtitle says letting AI patrol its own frontier is like asking Jesse James to patrol the Wild West. I want everyone to read it because it's beautifully written, and it's very comprehensible if you're a more casual observer or user of AI. It breaks down what's going on with these insane actions by misaligned agents, and why.
YOUR TURN
What are we missing?
Hit reply and tell us the part that's still nagging at you. A sentence is plenty. Rather say it out loud? The machine's on.
WRITTEN (AND SAID OUT LOUD) BY
Logan Currie & Mara Bolis
Women, work, and the messy middle of AI.