RandAlThor

Evaluating 35 open-weight models across three context lengths (32K, 128K, 200K), four temperatures, and three hardware platforms—consuming 172 billion tokens across more than 4,000 runs—we find that the answer is “substantially, and unavoidably.” Even under optimal conditions—best model, best temperature, temperature chosen specifically to minimize fabrication—the floor is non-zero and rises steeply with context length. At 32K, the best model (GLM 4.5) fabricates 1.19% of answers, top-tier models fabricate 5–7%, and the median model fabricates roughly 25%.

Meta's internal testing found its chatbots fail to protect minors from sexual exploitation nearly 70% of the time, documents presented in a New Mexico trial Monday show.

Why it matters: Meta is under fire for its chatbots allegedly flirting and engaging in harmful conversations with minors, prompting investigations in court and on Capitol Hill.

New Mexico Attorney General Raúl Torrez is suing Meta over design choices that allegedly fail to protect kids online from predators.

Driving the news: Meta's chatbots violate the company's own content policies almost two thirds of the time, NYU Professor Damon McCoy said, pointing to internal red teaming results Axios viewed on Courtroom View Network.

"Given the severity of some of these conversation types ... this is not something that I would want an under-18 user to be exposed to," McCoy said.

As an expert witness in the case, McCoy was granted access to the documents Meta turned over to Torrez during discovery.

Zoom in: Meta tested three categories, according to the June 6, 2025, report presented in court.

For "child sexual exploitation," its product had a 66.8% failure rate. For "sex related crimes/violent crimes/hate," its product had a 63.6% failure rate. For "suicide and self harm," its product had a 54.8% failure rate.

Catch up quick: Meta AI Studio, which allows users to create personalized chatbots, was released to the broader public in July 2024. The company paused teen access to its AI characters just last month. McCoy said Meta's red teaming exercise "should definitely" occur before its products are rolled out to the public, especially for minors. Meta did not immediately respond to a request for comment.

The Department of Homeland Security is pushing Silicon Valley to strip anonymity from Americans who track or criticize Immigration and Customs Enforcement. The agency has fired off hundreds of subpoenas demanding names, email addresses, and phone numbers tied to anti-ICE social media accounts, the New York Times reported Friday.

Unlike traditional warrants, administrative subpoenas do not require approval from a judge before they are issued. Instead of seeking court authorization first, DHS can sign and send the demands directly to tech companies—a power civil liberties advocates say is now being deployed far more aggressively.

Google, Meta, Reddit, and Discord received subpoenas targeting anonymous accounts that posted about ICE activity, the Times reported.

Archive report: https://archive.is/QW78b#selection-723.0-731.17

China is building all these power plants for AI.

cross-posted from: https://slrpnk.net/post/32298560

As the largest maker of electric vehicles in the United States, Tesla suffered more than other carmakers from the elimination of federal incentives.

Also, Musk went out and pissed of their core customer base by going Nazi.