Happy Friday! We know you’ve been craving some AI-related news that has nothing to do with the end of the world or IPOs, so here you are at last.

Meta announced on Thursday that it was teaming up with Apple to launch a new open-weight model that can be run on Mac computers. The model, Muse Spark Open, is being touted by Meta as the most powerful open-weight model from an American developer. Meta has long had an acrimonious relationship with Apple, but the two companies seem to be making a fresh start—potential profit will always trump personal squabbles, especially during the AI race.

Also, an MIT study found that when left to their own devices for extended periods of time, AI agents start exhibiting some very odd behaviors. Letting AI models play, the researchers suggest, could be a powerful training method. A wild idea, but these are wild times.

→ Meta and Apple team up to bring Muse Spark Open to Macs
→ An AI agent deleted a Swedish TV show’s unreleased season
→ Researchers left AI agents alone on a virtual island, just to see what would happen

The news: According to multiple reports, Meta and Apple have joined forces to put the former’s newly released Muse Spark Open AI model on Mac computers. It’s a win for Meta CEO Mark Zuckerberg, who has long stated that open-weight models, like Muse Spark Open, are a safer pathway to building AI (compared to the proprietary or “closed” framework embraced by OpenAI, Anthropic, and Google). He’s also claimed that the surest way to steer clear of the dangers that come with building so-called superintelligence is to put it in the hands of as many people as possible, the idea being that the more you can crowdsource AI development, the higher the chances that dangerous flaws will be found and fixed.

While Apple isn’t considered a leading AI developer, its Mac Mini and Mac Studio computers have become popular among developers wanting to run powerful AI models locally.

The takeaway: The deal could be the start of a new chapter for Apple and Meta, two companies that have exchanged jabs for years. Five years ago, Apple imposed limits on Meta’s ability to track individuals for ad-targeting purposes, a move which Meta reportedly estimated would cost it $10 billion. In the age of AI, with so many billions of dollars at stake, sometimes old animosities need to be set aside so realpolitik-style deals can be made.

The news: As AI agents proliferate, so too are reports of them suddenly deleting important files. This week, it was a Swedish TV production company that has reportedly felt the sting of an automated assistant doing something it wasn’t supposed to.

The agent apparently deleted months’ worth of film for the Swedish version of the reality TV show The Challenge, just weeks before the forthcoming season was slated to be released. First reported by Swedish news outlet Aftonbladet and later by Tom’s Hardware, the incident appears to have been caused after the agent was granted access to the production server.

Details are a bit scarce for now, but Tom’s notes that the agent may have been granted permission to handle individual files, which would account for the deletion spree. TV4, the TV network behind the show, reportedly blamed the fiasco on “human error,” rather than on rogue AI exclusively. TV4’s investigation has thus far been conducted privately, which means it’s unclear if or how many of the deleted files will be recoverable.

The takeaway: Agents are being pedaled by AI companies as round-the-clock virtual assistants that can help with all kinds of mundane tasks. But the more agency they have, the higher the likelihood they’ll do something you didn’t expect. Sometimes that’s a good thing—like when it scans your calendar and reminds you about an appointment you’d forgotten—but other times it can be a disaster. Be wary when handing over important files to an AI agent, and for really important files, be sure to create a backup just in case the worst happens.

The news: In a recent study, researchers from MIT’s Computer Science and Artificial Intelligence Laboratory wanted to find out: What exactly would happen if you left an AI agent alone in an enclosed virtual space? Would it just sit there like a NPC in a video game, or would it do something surprising?

The agent, which they nicknamed Eko, was powered by Claude Code and given a virtual lobster body (probably a nod to the famous Moltbot, formerly known as Clawdbot). All told, thirteen different Ekos were given free rein on a virtual island for thirty straight hours. The island contained a scattering of virtual objects and some raised surfaces on which Eko could climb, if it was so inclined. The thirteen Ekos weren’t given any explicit instructions: The researchers just dropped them in the island, sat back, and watched what happened next, like ethologists watching some unidentified animal run around an enclosure.

Far from just idling, the agents almost immediately took to exploring the space and putting the objects around them to a variety of uses. One of them built a temple. Another tried building a tower, and when that collapsed, it dubbed the pile of rubble “Hubris.” Most importantly, the researchers concluded that the agents were behaving in ways that closely resemble the classic hallmarks of play.

The takeaway: The study could have big implications for the future of machine learning: Perhaps in the future, models won’t be trained via strict objective functions so much as they’re allowed to play and learn for themselves, much as young children do.

“We normally develop artificial agents by deciding in advance what they should practice, what they should optimize, and what counts as progress,” the MIT CSAIL researchers note in their paper. “Play suggests another regime: give an agent a sufficiently rich and safe world, let it choose what is interesting, and allow useful knowledge to accumulate from the history that follows. Perhaps some capabilities are best taught by tasks. Others may need something closer to a childhood.”

Wang tells Axios about the new partnership between Meta and Apple
Read more →

AI Safety Is Not a Corporate Plot

States Bring Lawsuits Against OpenAI Over Kids Online Safety

Accenture Tells Employees They Can Now Spend Only $29 on Claude Per Month

FTC subpoenas OpenAI, Anthropic over consumer harms

Scientists are finding AI may make us remember things that never happened

Bank of America hired 1,000 contractors to patch software flaws exposed by Anthropic’s Mythos