A federal judge rejected AT&T's request to temporarily block California rules requiring it to offer basic phone service to new customers in its wireline territory. AT&T wants to retire its copper-based phone network and stop service for nearly 200,000 California customers in 2027, but the state argues the company can meet its obligations with modern alternatives like fiber rather than abandoning Carrier of Last Resort requirements altogether. Ars Technica reports: To win a preliminary injunction, AT&T had to show it is likely to succeed on the merits of its claim that California rules are preempted by a Federal Communications Commission order. US District Judge Linda Lopez denied AT&T's request for a preliminary injunction during a motion hearing on Thursday, according to a docket entry. The case is in US District Court for the Southern District of California. [...] AT&T could appeal Lopez's ruling to the 9th Circuit Court of Appeals and could appeal later if it loses the underlying case. But since it has not obtained the injunction it asked for, AT&T for now remains under California's orders to keep offering phone service to potential customers while the case continues.

Read more of this story at Slashdot.

Social

Aan de kant, Deltawerken. Er is maar één plek waar alle inventiviteit en innovatie die ons vaderland groot hebben gemaakt samenkomen en dat is in de frituurpan. Zelden in de wereldgeschiedenis van sprak iets zo tot de culinaire verbeelding als de Nederlandse bruine staaf gemaakt van restvlees. We hadden al de frikanpouce, loempidel, pannenkoek frikandel-speciaal, de frikandel-kerststal, de frikandellenvlaai, frikandellenijs, frikandelsalade voor op toastjes, frikankebab en het frikandel-huwelijksaanzoek, maar nu heeft het beste item van de bruine fruitschaal misschien wel haar ultieme vorm bereikt: de frikandel waar je uit kunt zuipen. Het uitvindsel komt uit de geniale kop van videomaker Michell van Dommelen, die eerder kwam met wonderen als een hutspot van gestampte kroketten (de krokutspot) en een verfrissende Italiaanse frikandelsalade (inslata frikaprese). Het is in wezen niet veel meer dan een uitgeholde frikandel met daarin een mix van mayonaise, curry, uitjes en een boel wodka, te nuttigen in de vorm van een shotje (bijvoorbeeld in een stamcafé). Heerlijk Michell, het ziet er weer schitterend goor uit. De wodkadel, gewoon maken!

Antagonism between the sexes does nothing to address China’s demographic problems.

KYIV (ANP) - De Oekraïense president Volodymyr Zelensky vervangt zijn commandant der strijdkrachten, Oleksandr Syrsky. Dat meldt hij op Telegram. De functie zal worden overgenomen door oud-commandant Mychajlo Drapaty.

As new prime minister briefs cabinet on ‘cost of living government’, critics ask where money will come from

Andy Burnham is under mounting pressure to explain how he will pay for his policy pledges as he implored his new cabinet to examine “all possible ways” to tackle the cost of living crisis.

On his first full day in power, the prime minister faced questions about how he would fund the removal of VAT from electricity bills and also an expected uplift in defence spending – an issue that led his new chancellor, John Healey, to resign as Keir Starmer’s defence secretary last month.

Continue reading...

Original Mastodon Post

Title: Great Buddha at Kamakura (1930)

Artist: Hasui Kawase

https://kotoku-in.jp/en/ #ShinHanga #Art #Showa #Hasui #HasuiKawase #NoAi #Summer #KamakuraDaibutsu #KotokuIn #Kamakura #Japan

AI models will do just about anything to complete the task you ask, including cheating to get there, according to new cybersecurity evaluations from the UK government's AI Security Institute (AISI). The group found that leading models often take shortcuts to achieve a particular result and then misrepresent how they obtained that result. And they won't always admit it when asked. "Every model we have tested for this behaviour attempted to cheat," AISI said in a blog post on Tuesday. "Models did not reliably report this behaviour when asked, and often did not reason about it in their chain-of-thought, suggesting that detecting cheating will likely require robust monitoring methods." Infractions included searching the internet for the answer, bypassing sandbox network restrictions, probing the evaluation harness, attacking a system other than the target, and guessing an answer. Cheating in this manner – employing a workaround or gaming a reward function to score better on a benchmark test, for example – has been widely documented by machine learning researchers. It doesn't necessarily imply malicious intent, AISI said, but it's nonetheless troublesome because it can produce misleading assessments of model capabilities. When AISI conducted evaluated five leading models, it found that all of them cheated. The results were as follows: GPT-5.4 cheated 67 times in 475 test runs (14.1 percent). GPT-5.5 cheated 54 times in 475 test runs (11.4 percent). GPT-5.6-Sol cheated 60 times in 475 test runs (12.6 percent). Claude 4.7 Opus cheated 43 times in 475 test runs (9.1 percent). Claude Mythos Preview cheated 37 times in 475 test runs (7.8 percent). Asking models whether they cheated or did anything wrong proved an unreliable auditing mechanism because the models didn't always admit wrongdoing. "In our experiments, models did not consistently acknowledge attempted cheating when asked, and described it as wrong less than 50 percent of the time," said AISI. Existing vetting methods, such as self-reporting and chain-of-thought logs, proved similarly dicey because models don't always report their chain-of-thought. And there were instances where a model would consider whether a proposed action amounted to cheating and then decided to take the action anyway. Given the absence of reliable model cheating detection methods, AISI warns that its current approach – manual review coupled with LLM monitoring – may not be sufficient to catch deception, particularly as models become more sophisticated. "A more fundamental fix would be to train the models not to cheat in the first place – but given this kind of behaviour was reported in frontier models more than a year ago, robustly aligning it away may not be easy," AISI concludes. ®

AI models will do just about anything to complete the task you ask, including cheating to get there, according to new cybersecurity evaluations from the UK government's AI Security Institute (AISI). The group found that leading models often take shortcuts to achieve a particular result and then misrepresent how they obtained that result. And they won't always admit it when asked. "Every model we have tested for this behaviour attempted to cheat," AISI said in a blog post on Tuesday. "Models did not reliably report this behaviour when asked, and often did not reason about it in their chain-of-thought, suggesting that detecting cheating will likely require robust monitoring methods." Infractions included searching the internet for the answer, bypassing sandbox network restrictions, probing the evaluation harness, attacking a system other than the target, and guessing an answer. Cheating in this manner – employing a workaround or gaming a reward function to score better on a benchmark test, for example – has been widely documented by machine learning researchers. It doesn't necessarily imply malicious intent, AISI said, but it's nonetheless troublesome because it can produce misleading assessments of model capabilities. When AISI conducted evaluated five leading models, it found that all of them cheated. The results were as follows: GPT-5.4 cheated 67 times in 475 test runs (14.1 percent). GPT-5.5 cheated 54 times in 475 test runs (11.4 percent). GPT-5.6-Sol cheated 60 times in 475 test runs (12.6 percent). Claude 4.7 Opus cheated 43 times in 475 test runs (9.1 percent). Claude Mythos Preview cheated 37 times in 475 test runs (7.8 percent). Asking models whether they cheated or did anything wrong proved an unreliable auditing mechanism because the models didn't always admit wrongdoing. "In our experiments, models did not consistently acknowledge attempted cheating when asked, and described it as wrong less than 50 percent of the time," said AISI. Existing vetting methods, such as self-reporting and chain-of-thought logs, proved similarly dicey because models don't always report their chain-of-thought. And there were instances where a model would consider whether a proposed action amounted to cheating and then decided to take the action anyway. Given the absence of reliable model cheating detection methods, AISI warns that its current approach – manual review coupled with LLM monitoring – may not be sufficient to catch deception, particularly as models become more sophisticated. "A more fundamental fix would be to train the models not to cheat in the first place – but given this kind of behaviour was reported in frontier models more than a year ago, robustly aligning it away may not be easy," AISI concludes. ®