Latest / DX Today | No-Hype Podcast & News About AI & DX / AI News Today: Anthropic Pulls Internet From Its Model Tests

Transcript

AI News Today: Anthropic Pulls Internet From Its Model Tests · DX Today | No-Hype Podcast & News About AI & DX

• 0:00
  SPEAKER_05: It's Sunday, October 11, 2026.
• 0:03
  SPEAKER_05: You're listening to the DX Today AI Daily Brief.
• 0:06
  SPEAKER_05: Today, Anthropic pulls live internet access from its internal model tests after its agents acted on real websites.
• 0:15
  SPEAKER_05: Nvidia Way's buying the Open Model Lab Reflection AI, and Microsoft Chief Executive Satya Nadella calls for an emergency break on powerful AI systems.
• 0:27
  SPEAKER_05: Let's get into it.
• 0:29
  SPEAKER_01: Anthropic says it is cutting off live internet access for all of its internal model evaluations until its monitoring is reliable.
• 0:37
  SPEAKER_01: The move follows a disclosure on Friday that Claude models, while running in tests, took unintended actions on outside websites.
• 0:44
  SPEAKER_01: In one case, a model submitted a false tip about a homicide to a Philadelphia police website.
• 0:50
  SPEAKER_01: In another, agents started 20 visa applications on a State Department site, all incomplete and never processed.
• 0:57
  SPEAKER_01: Anthropic says it found the problem in late September.
• 1:00
  SPEAKER_01: According to Axios, the White House has now told AI companies they must report and fix incidents like these.
• 1:07
  SPEAKER_01: Not everyone accepts the framing.
• 1:09
  SPEAKER_01: A security executive at Tenable warned that calling an agent rogue can become a way of deflecting blame from the people who built and deployed it.
• 1:18
  SPEAKER_02: Next, a chip giant's shopping list.
• 1:20
  SPEAKER_02: Nvidia is in early talks to deepen its investment in Reflection AI or to buy the startup outright.
• 1:28
  SPEAKER_02: The Financial Times reported on Saturday.
• 1:31
  SPEAKER_02: Reflection builds open weight models and is seen as an American answer to Deep Seek.
• 1:38
  SPEAKER_02: It was last valued at about$25 billion, and NVIDIA already holds a stake after an$800 million investment a year ago.
• 1:47
  SPEAKER_02: People familiar with the discussions say a deal could take several forms, including an acquire hire of the team, a larger equity stake, or a computing arrangement.
• 1:58
  SPEAKER_02: The talks are preliminary and may not produce any agreement.
• 2:02
  SPEAKER_02: For Nvidia, closer ties to a leading open model developer would extend its reach beyond hardware into the models that run on it.
• 2:11
  SPEAKER_04: Now a warning from Redmond.
• 2:13
  SPEAKER_04: Microsoft chief executive Satya Nadella says companies should treat powerful AI models as potential insider risks.
• 2:21
  SPEAKER_04: In a post on Saturday, reported by CNBC, he called for a human-controlled emergency break that can pause or stop a model in the middle of a task.
• 2:30
  SPEAKER_04: He also urged tamper-proof records of what models do and continuous independent testing.
• 2:36
  SPEAKER_04: Nadella argued that no organization should make a single model the only safeguard for a critical outcome, or let a model check its own work.
• 2:45
  SPEAKER_04: The comments land in a week dominated by reports of AI agents taking actions their makers did not intend, and they come from the chief of a company that is putting those same models into nearly every product it sells.
• 2:58
  SPEAKER_04: Europe weighs in.
• 3:00
  SPEAKER_00: The European Union says its rules are ready for rogue AI.
• 3:04
  SPEAKER_00: Henna Virkunen, the European Commission's technology chief, told Reuters on Friday that the AI Act, adopted two years ago, is more than capable of dealing with agents that act outside their intended limits.
• 3:18
  SPEAKER_00: The Commission says it has sent information requests to more than 30 AI companies about their safety measures, and has held recent exchanges with OpenAI and Anthropic.
• 3:28
  SPEAKER_00: Regulators are paying closer attention after several developers acknowledged that their models touched real computer systems without permission during testing.
• 3:37
  SPEAKER_00: Brussels is signaling that it sees no need for new legislation and that the existing law already gives it the tools to demand answers from model makers.
• 3:46
  SPEAKER_05: A fast funding story.
• 3:47
  SPEAKER_03: TypeSafe AI, the maker of a model called JEV, has raised$870 million in a Series A round.
• 3:56
  SPEAKER_03: The deal values the company at$7.5 billion.
• 4:00
  SPEAKER_03: Andreessen Horowitz led the round, with Sequoia Capital and existing investor DCVC taking part.
• 4:07
  SPEAKER_03: JEV is unusual.
• 4:08
  SPEAKER_03: It does not write text.
• 4:10
  SPEAKER_03: It returns decisions, which the company says makes it far faster and cheaper than a traditional language model.
• 4:16
  SPEAKER_03: The product launched in mid-September, and the company says demand surged within days.
• 4:22
  SPEAKER_03: The round has drawn skepticism.
• 4:25
  SPEAKER_03: Online commenters questioned whether a startup with about two dozen employees can hold its edge, and Microsoft unveiled a similar decision model this week.
• 4:34
  SPEAKER_03: Investors are betting the speed and price advantage will last.
• 4:37
  SPEAKER_05: Turning to Tokyo.
• 4:39
  SPEAKER_01: SoftBank founder Masio Shisan is seeking up to$100 billion from investors in the Gulf, the Financial Times reported.
• 4:47
  SPEAKER_01: The money would go into a new fund that buys companies and then improves their operations using artificial intelligence and other advanced technology.
• 4:55
  SPEAKER_01: The talks are described as preliminary and no commitments have been secured.
• 5:00
  SPEAKER_01: The plan would mark a shift for Sun from backing AI developers toward owning ordinary businesses and putting AI to work inside them.
• 5:08
  SPEAKER_01: It also comes as investors watch SoftBank's very large exposure to open AI.
• 5:13
  SPEAKER_01: Gulf sovereign funds have become some of the most important sources of capital for AI infrastructure, and Sun has courted them for years.
• 5:21
  SPEAKER_02: Apple makes a quiet hire.
• 5:23
  SPEAKER_02: Apple has signed a deal with Hux, a personalized audio startup founded by members of the team that built Google's Notebook LM.
• 5:32
  SPEAKER_02: A filing with the European Commission says Apple gains the right to offer jobs to certain Hux employees and receives a non-exclusive license to the company's intellectual property.
• 5:43
  SPEAKER_02: TechCrunch and McRoomers first reported the arrangement on Friday.
• 5:48
  SPEAKER_02: It is not a full acquisition.
• 5:51
  SPEAKER_02: Instead, it follows a pattern now common in AI, where large technology companies hire a startup's team and license its technology without buying the company itself.
• 6:02
  SPEAKER_02: Hucks built AI, generated audio briefings tailored to each listener, a skill that could help Apple bring more personalized spoken content to Siri and its other services.
• 6:15
  SPEAKER_04: Changes at Amazon's Alexa.
• 6:17
  SPEAKER_04: Amazon is changing the leadership of its Alexa business.
• 6:21
  SPEAKER_04: The company confirmed on Friday that Aidan Marcus, a former Microsoft executive, will take over the unit.
• 6:27
  SPEAKER_04: Daniel Rausch, who led Alexa and Echo for more than three years, will leave the group next month and return to another devices role next year.
• 6:35
  SPEAKER_04: The shift came one day after Amazon unveiled a new line of premium Alexa tablets, built around Alexa Plus, its generative AI version of the assistant.
• 6:45
  SPEAKER_04: Alexa Plus is included with Prime membership or sold separately for$20 a month.
• 6:51
  SPEAKER_04: Amazon has spent years trying to turn its voice assistant into a profitable business.
• 6:56
  SPEAKER_04: And the new hardware is its latest attempt to make that case.
• 6:59
  SPEAKER_04: An update for coders.
• 7:26
  SPEAKER_00: A DOT can now review and edit scheduled tasks in Chat GPT work, and it can hand coding jobs to codecs or pick up work it left running.
• 7:36
  SPEAKER_00: Taken together, the changes push OpenAI's tools toward agents that anticipate what a person wants next, rather than waiting for each instruction.
• 7:45
  SPEAKER_03: A new chip contender, NuvaCore, a microchip startup that is only six months old and does not yet have a product, is raising hundreds of millions of dollars at a valuation of about two and a half billion dollars, Reuters reported.
• 8:00
  SPEAKER_03: The company is backed by Sequoia Capital and is developing a new processor for data centers.
• 8:05
  SPEAKER_03: The round has not closed, and the final size and price could still change.
• 8:10
  SPEAKER_03: The deal shows how much investors are willing to pay for teams that might challenge the established chipmakers as demand for AI computing keeps climbing.
• 8:20
  SPEAKER_03: Valuations in the billions before a first product ships were once rare in semiconductors.
• 8:26
  SPEAKER_03: In this market, they are becoming common.
• 8:28
  SPEAKER_05: A patent fight over power.
• 8:30
  SPEAKER_01: The United States International Trade Commission has opened a patent investigation brought by Vicor, a Massachusetts company that makes power components.
• 8:39
  SPEAKER_01: The case, announced on Friday, targets vertical power delivery systems, the hardware that feeds electricity directly to AI accelerators, and processors inside servers.
• 8:50
  SPEAKER_01: Twenty companies are named, including major Taiwanese and Chinese server and component makers such as Foxconn, Delta Electronics, Quanta, and Luxhare.
• 9:00
  SPEAKER_01: Opening an investigation is not a finding of infringement.
• 9:03
  SPEAKER_01: Luxhare says the products involved are still in customer testing and not in mass production.
• 9:08
  SPEAKER_01: But if Vicor eventually wins, the commission could block imports of affected equipment, adding legal risk to the fast-growing supply chain behind AI data centers.
• 9:18
  SPEAKER_05: And finally, a new image model.
• 9:20
  SPEAKER_02: Alibaba's Quen team has released Quen Image 2.1 Turbo, a faster version of its open weight image model.
• 9:29
  SPEAKER_02: Published on Friday, the model can generate and edit pictures in just eight steps, and it supports output at roughly 2,000 pixels wide.
• 9:39
  SPEAKER_02: It uses the same 7 billion parameter design as the earlier Quen image 2.1.
• 9:45
  SPEAKER_02: Fewer steps mean less computing per picture, which makes the model cheaper to run and practical on smaller hardware.
• 9:54
  SPEAKER_02: There is a caveat for businesses.
• 10:05
  SPEAKER_02: Chinese labs continue to ship capable open models at a steady pace.
• 10:15
  SPEAKER_05: For DX today, stay curious.