Nvidia on Monday launched the Open Agent Safety Platform, a sandbox it says can quarantine rogue AI agents within "milliseconds", as the FTC reportedly opens an investigation into OpenAI and Anthropic and Beijing tightens its grip on the open-model hubs Chinese developers rely on.
Nvidia released the Open Agent Safety Platform on 28 September, a reference architecture that quarantines misbehaving AI agents "within milliseconds", according to The Verge. The same week, Rest of World reported that Chinese platforms ModelScope and MoArk are scaling domestic alternatives to Hugging Face.
On 28 September, a US court filing quoted an OpenAI engineer saying users "won't click" links, according to Search Engine Journal. It was the same week researchers and vendors published fresh numbers on how much of the search box has stopped behaving like a search box.
Chinese open-model platforms MoArk and ModelScope are racing to replace Hugging Face as Beijing weighs the risk of Washington banning Chinese models, Rest of World reported on 29 September.
On 27 September, PlanetScale published a technical deep dive into how full-text search engines work. The machinery that decides which pages publishers get traffic from is being rebuilt from the index up. The post landed the same day two smaller search projects were still drawing attention on Hacker News. One of them had been submitted just 7.5 hours earlier.
A behind-the-scenes clip from a Palestinian short film, reposted as proof that Israel stages casualties, drew about 2m views on X within days of the 7 October Hamas attack. It was one of the first viral fakes of a war that would keep producing them, and it set the pattern for how online disinformation networks now operate.
A behind-the-scenes clip from a Palestinian short film was viewed about 2m times on X with the caption "Video showing Israel attempting to create fake footage of deaths," according to the Guardian. The clip was real. The claim was not.
At least 27 online information operations have been partially or wholly attributed to PR or marketing firms since 2011, with 19 occurring in 2019 alone, according to a BuzzFeed News review. This shift from state troll farms to commercial contractors marks a quiet structural change in a decade of research on online disinformation.
More than 85% of people worldwide worry about the impact of online disinformation, and 87% say it has already harmed their country's politics, according to a Unesco-commissioned Ipsos survey of 8,000 people in 16 countries.
A behind-the-scenes clip from a Palestinian short film, recaptioned as proof that Israel fakes casualties, drew about 2m views on X. It was not an isolated hoax: it was one node in an industrialised system for manufacturing falsehood at speed.
PlanetScale published a deep technical explainer on Postgres full-text search on 27 September, the newest item in a week of search infrastructure releases that lands as publishers report steep falls in search referrals.
TikTok dropped its appeal against a £12.7m UK fine for failing to keep under-13s off the platform on 24 September. Three days later, Bill Gates told NBC that AI regulation has to be mandatory, not self-imposed.
Search referrals to publishers have dropped by a third globally in a year, and some media bosses now expect another 43% cut over three years, according to a Reuters Institute report. The data behind the collapse is patchier than the headlines suggest.
PlanetScale published a technical anatomy of Postgres full-text search engines on 27 September, the same weekend a Digiday recap described publishers rebuilding for a post-search era, as search referrals continue to slide.
TikTok has dropped its appeal and will pay a £12.7m fine to the UK's Information Commissioner's Office, the Guardian reported on 24 September, while Bill Gates used a Sunday interview to demand mandatory AI regulation.
Harvard psychologist Steven Pinker declined a public debate on AI's existential risks in an open letter published on Quillette on 28 September, arguing that doomsday rhetoric distracts from bioterrorism, cyberattacks and unchecked AI agents.
Search referrals to publishers have fallen by a third globally in a year, and new size breakdowns show small sites losing 60% over two years, nearly three times the decline at large publishers.
Facebook's moderators were told to leave up posts urging violence against women but to remove a threat to shoot Donald Trump, according to leaked training manuals published by the Guardian and covered by Ars Technica.