News

AWS lost customer data in Bahrain for good - three zones turned out to be one address

Amazon admitted on September 15 that after the Iranian strikes it cannot recover data held only in its Bahrain region and in one zone in the UAE. According to accounts on Hacker News, some customers had no copy abroad because the law forbade them from moving data out of the country.

Read →

Salesforce has its own model, Koa - it matches the labs in the press release, not in the paper

Salesforce and Nvidia unveiled Koa on Tuesday, Salesforce's first in-house reasoning model for CRM work, built on Nvidia's open Nemotron with no customer data. The press release says it matches leading models with three times fewer errors. The company's own research paper, submitted a day earlier, says this: better than GPT-4.1, weaker than GPT-5.5 and Claude Opus 4.8. The phrase "3x" does not appear in it.

Read →

iOS 27 is out with the new Siri - a waitlist, English only, and nothing at all in Europe

Apple released iOS 27 and macOS 27 Golden Gate on Monday. The new Siri AI arrived more than two years after it was announced, but in beta, in English only, behind a waitlist and with daily limits. On the iPhone in the European Union it is not available at all, and Apple gives no date. Everything else in iOS 27 lands on every iPhone from the 11 up.

Read →

Amodei wants to slow AI down - Altman, Musk and Nadella agree, Trump and the White House do not

Dario Amodei wrote on Saturday that AI companies must slow down, and committed Anthropic to letting embedded evaluators in, with desks and badges. By Monday everyone had answered. Altman: "we will do the same." Musk: "Dario is right." Trump: "whoever wins AI, wins." And a White House adviser: slow down on your own, without a cartel. That same weekend Anthropic told investors its gross margins are above 80%.

Read →

The A20 Pro in the iPhone 18 Pro - single-core faster than a desktop Ryzen and a Core i9

The A20 Pro from the iPhone 18 Pro scored 4,006 points in the Geekbench 7 single-core test, 26 percent more than the desktop Ryzen 9 9950X3D and 32 percent more than the Core i9-14900KS. In the multi-core test the desktop still has almost three times as many points.

Read →

Anthropic discloses how Claude was misused - biological weapons, missiles and surveillance of 25 million SIM cards

154 pages and seven areas of misuse over eight months: an interception system for Mali's intelligence service, a weapons cell in Yemen writing missile software in Claude Code, five biological research cases and seven Chinese labs extracting Claude's reasoning. Banning the account in Mali switched nothing off - the finished system runs on local servers.

Read →

25 Fields Medalists against AI companies - famous problems are not a benchmark

After a rumor that turned out to be false, OpenAI pointed 10,000 agents at one of the Millennium Prize Problems and had a proof 88 hours later. Three days after that, 25 Fields medalists, more than a third of everyone who has ever received the medal, signed a declaration against treating mathematics as a benchmark.

Read →

DeepSeek releases V4.1 Flash - a million-token context, images, and prices dozens of times lower than Anthropic's

552 billion parameters, of which 8-16 billion work per token, a million-token context, image support and open weights under the MIT license. Off-peak, $0.15 per million input tokens - dozens of times cheaper than Claude Fable 5. First independent tests from X and Hacker News: close to Opus 5 for a fraction of the price, but "better than OpenAI and Anthropic" exists, for now, only on DeepSeek's own charts.

Read →

Apple unveils its first foldable iPhone - iPhone Duo from $1,999, plus iPhone 18 Pro, new watches and AirPods 5

John Ternus's first keynote as CEO: the foldable iPhone Duo (from $1,999, on sale October 23), iPhone 18 Pro and Pro Max (from $1,199), Apple Watch Series 12 and Ultra 4, AirPods 5 with noise cancellation in both versions. Siri AI launches September 14, but not on iPhones in the EU. At the end: what the rumors said, and what was missing.

Read →

Researcher quits Anthropic with a warning - and the head of safety says he is right

Jacob Coxon left on 8 September, writing that AI labs are gambling with our lives. Evan Hubinger, who leads the Alignment Science team at Anthropic, replied that he puts the risk of human extinction at over 10% within a decade and that the company has no plan to prevent it. Two days earlier OpenAI's chief scientist had written almost the same thing. Coxon's post and Hubinger's reply have millions of views, and under both posts the same question keeps coming back: how, exactly?

Read →