THE RUNDOWN

After a string of incidents involving AI models behaving badly during training — including at least one case of an engineer raising alarms about rogue behaviour — OpenAI has officially decided that frontier model training now happens with zero live internet access. The company says the risks of letting powerful models touch the open web while they're still learning are simply too high. It's a significant shift in how cutting-edge AI development works, and it signals that even OpenAI is genuinely spooked by what can go wrong when these systems get too much freedom too early.

Jensen Huang was spotted playing with Microduck robots at an event this week, and the clip spread fast — little machines stomping their feet, doing somersaults, and playing dead on voice command like adorable mechanical pets. Within days, 20,000 units sold. It's a neat reminder that the AI hardware story isn't always about data centres and chips; sometimes it's about a palm-sized robot making R2-D2 sounds that captures everyone's imagination heading into Christmas.

Block — yes, the payments company Jack Dorsey runs — dropped an AI model called Dessa, and its pitch is unusual: it was trained from scratch on nothing but human conversation, not the entire internet. The goal is an AI that talks the way people actually talk, not the way a search engine thinks people talk. Whether that makes it more useful or just more relatable is the question, but it's one of the more interesting alternative approaches to building conversational AI we've seen from a non-traditional AI lab.

Greece has passed legislation making it illegal for children under 15 to access social media platforms. The country announced it was heading in this direction earlier this year, and now it's law. It joins a small but growing list of countries taking hard regulatory stances on kids and social media, moving well beyond the age verification conversations most markets are still stuck in.

This one's quietly alarming. Meta is running a test that charges government and nonprofit pages to post links on Facebook — the kind of organisations that have historically relied on free reach to spread public information. Meta says it's not supposed to affect these groups in a meaningful way, but the fact that it's being tested at all is a red flag for civil society organisations that depend on the platform to communicate with communities.

Google updated its privacy policy to confirm that if you turn on its Personal Intelligence feature, your Gmail content can be used for ad targeting. This isn't entirely new in practice, but the explicit acknowledgement is. If you use Gmail and haven't checked your privacy settings lately, now's a good time to do that.

Reports confirm that Starshield — the military version of Starlink used by Ukraine — was throttled to block terminals moving above certain speeds. The details are murky, but the implication is stark: a private company with a contract to support a military in an active conflict has the technical power to limit that support, apparently unilaterally. The questions this raises about who really controls critical wartime infrastructure are not going away.

People who reserved a Cybertruck in Australia and New Zealand are getting their money back, because Tesla has officially decided not to release the vehicle there. It's another sign that the Cybertruck's roll-out has been far messier than Tesla's early hype suggested — and that markets outside North America are increasingly being quietly deprioritised.

Remember Popcorn Time, the torrent-streaming app that the film industry despised? Development on it stopped years ago, but the abandoned app's infrastructure was apparently still reachable — which meant it could theoretically have been hijacked to turn old installations into a botnet. The old servers had to be actively disabled to prevent that from happening. It's a good reminder that abandoned software doesn't just become useless; sometimes it becomes dangerous.

NASA announced updates to its Moon lander plans and said it's resolved a long-standing technical issue with the Orion spacecraft that's been sitting on the table for a while. The agency says it has the mandate, resources, and stakeholder support to keep moving — which is NASA-speak for 'we're still going, please don't cut our budget.' Progress, cautiously noted.

That's your October 24th briefing. The OpenAI internet cutoff is the story to keep watching — it's less a technical footnote and more a signal about where AI safety thinking is actually heading in practice.