WEBVTT

1
00:00:10.000 --> 00:00:24.485
Welcome to the Global Advisors A.I Voices podcast. I'm Maya.
Today is Tuesday, the tenth of November 2026.
This is your morning catch-up on A.I. Please like and follow the show for your daily A.I update.

2
00:00:26.534 --> 00:00:50.193
Nvidia's board has launched an independent investigation into C E O Jensen Huang after he admitted to an affair with Alice Huang, an executive whom United States intelligence officials believe was recruited by Chinese spies. The investigation, conducted by independent directors and outside counsel, comes as Congress demands answers about the relationship.

3
00:00:50.373 --> 00:00:57.477
Huang apologized in a memo to the company, acknowledging the personal breach, but the professional stakes are far higher.

4
00:00:57.827 --> 00:01:22.403
The core concern is national security. United States officials suspect Alice Huang was tasked with gathering information and influencing chip export restrictions, a critical lever for Nvidia's business. TechCrunch's Rebecca Bellan argues that these allegations could taint Nvidia's assurances regarding preventing sensitive technology from reaching unauthorized customers.

5
00:01:22.583 --> 00:01:30.306
Secretary of State Marco Rubio reportedly warned officials about the relationship earlier this year, adding weight to the scrutiny.

6
00:01:30.656 --> 00:01:52.394
While the investigation is ongoing, the question of Huang's future remains open. Yale's Jeffrey Sonnenfeld estimates it is roughly a coin flip whether Huang stays on as C E O. The outcome will likely depend on how the board weighs the personal misconduct against the potential national security risks and the company's need for stability in a volatile market.

7
00:01:54.794 --> 00:02:16.575
The fallout from the Nvidia investigation has hit the markets hard. Nvidia's stock fell more than 4 percent on Monday following the board's announcement of the probe into C E O Jensen Huang's relationship with Alice Huang. The drop reflects investor anxiety over the governance crisis and the potential legal ramifications for the chip giant.

8
00:02:16.925 --> 00:02:41.565
Legal experts are already circling. Several law firms, including Schall, Brown and Schwartz and Levi and Korsinsky, have announced investigations into whether Nvidia made false or misleading statements to investors. Stanford Law professor Robert Bartlett expects a securities lawsuit to be filed, noting that plaintiffs' lawyers have strong incentives to be first in line under federal securities law.

9
00:02:41.915 --> 00:03:06.086
However, Bartlett cautions that these early announcements are not necessarily signs of a strong case. He points out that plaintiffs' firms often publish notices quickly when they see an opportunity, meaning the volume of investigations does not automatically equate to legal merit. The real test will be whether Nvidia's earlier insistence that its operations were 'normal' can withstand scrutiny.

10
00:03:08.485 --> 00:03:19.919
SoftBank is expected to face scrutiny of its A.I strategy at its second-quarter earnings following a sell-off. The tech sell-off raises risks for SoftBank's massive A.I bet.

11
00:03:22.319 --> 00:03:43.460
Baidu has released ERNIE Evolve, an automated research system designed to improve A.I training and inference code by picking research directions, editing code, and running experiments. The company describes this as a practical step toward recursive self-improvement, where A.I systems enhance the software that builds and serves A.I.

12
00:03:43.810 --> 00:04:08.279
The system uses a multi-agent architecture, including a Director to pick research directions and a Selector to choose candidates from code trees. To prevent reward hacking, an Evaluator and Reviewer are employed, with the Evaluator being a fixed program outside the repository that agents cannot see. This ensures that code changes are accepted only if they meet strict, measurable objectives.

13
00:04:08.629 --> 00:04:33.205
In tests, ERNIE Evolve reported 33.7 percent higher throughput in one deterministic inference test and achieved better results than public comparisons on several metrics. Baidu states that recursive self-improvement must remain human-controlled and interruptible, with continuous safety evaluation. The company plans to apply this loop to training-system optimization in the future.

14
00:04:35.604 --> 00:04:55.679
Nathan Benaich, speaking on The MAD Podcast, estimates that all 27 European countries possess roughly 80,000 H100-equivalent cards. This figure is lower than any single top-five GPU class in the United States, highlighting a significant infrastructure gap between the two regions.

15
00:04:56.029 --> 00:05:23.357
The disparity is driven by capital expenditure and energy costs. Benaich notes that United States hyperscalers are projected to spend approximately 730 billion dollars in 2026, while European commitments remain in the single-digit billions. He describes energy costs in Europe as 'really bad' and significantly more expensive than in the United States, with the United Kingdom noted as having one of the most expensive energy prices.

16
00:05:23.707 --> 00:05:45.851
These barriers have tangible consequences. Benaich cites a recent deal between Microsoft and Nscale that was impacted because the data center provider couldn't get the energy connection in time. This suggests that Europe's struggle to match United States compute capacity is not just a funding issue, but a structural challenge tied to energy infrastructure.

17
00:05:48.250 --> 00:06:06.341
Meta A.I Research has announced the rollout of specialized versions of models behind Muse and Meta A.I, trained using Muse Wonder. The current Muse model scores 47 percent on Meta's Personal Superintelligence Index, indicating significant progress in general capability.

18
00:06:06.691 --> 00:06:30.520
A key improvement is in memory retention. Meta reports a 2.4x reduction in forgetting information critical to users' goals. The Active Memory Eval score for correct context usage has risen to 90 percent. Active Memory allows assistants to retain important information and retrieve it when relevant, distinguishing between temporary and permanent data.

19
00:06:30.870 --> 00:06:48.641
Meta has also introduced Active Focus, a new feature intended to keep tasks moving within the Muse application. This update aims to enhance the assistant's ability to maintain context and drive productivity, addressing common challenges in long-running A.I interactions.

20
00:06:51.040 --> 00:07:08.299
Physical Superintelligence reports that an internal research system running a non-public Anthropic model has proved the Aaronson-Ambainis conjecture. The lab says the system produced a complete proof on the third day of its run, which was accepted by the Lean proof assistant.

21
00:07:08.649 --> 00:07:34.803
Stanford University researchers Guy Blanc, Jordan Docter, Carmen Strassle and Li-Yang Tan described the result as huge, noting it resolves a decades-old conjecture with implications for quantum speedups. The manuscript and Lean 4 formalization are publicly available, though experts must still confirm that the encoded statement matches the original conjecture and review parts of the paper outside the Lean verification.

22
00:07:37.203 --> 00:07:56.211
Nate Herk, an A.I automation expert, suggests that personalization significantly boosts the perceived value of A.I tools. He cites a braces calculator that achieved a 45 percent conversion rate, attributing this success to the IKEA effect, where users value things more when they help make them.

23
00:07:56.561 --> 00:08:19.643
Herk argues that allowing users to input their own information makes the result feel like their own decision, increasing its worth. However, he cautions that these tools must be short and easy to ensure users finish them. He points to examples like Wise's currency converter, Zillow's home estimate, and Sherwin-Williams' paint visualizer as models of this value-add approach.

24
00:08:22.042 --> 00:08:46.597
Anthropic has disclosed in its revised S-1/A filing that its A.I systems may be approaching recursive self-improvement, a threshold that could accelerate A.I development while creating new challenges for human oversight. Anthropic warns that using A.I to help build more capable A.I could pose a considerable safety risk, particularly if it sharply accelerates the pace of progress.

25
00:08:46.947 --> 00:09:06.403
A report by METR indicates that Anthropic's models are improving roughly twice as quickly as a year ago, with little indication of slowing down. This disclosure comes as Anthropic prepares to go public, even as its C E O, Dario Amodei, has publicly called for labs to slow down A.I development.

26
00:09:08.802 --> 00:09:28.536
Anthropic has published an engineering post describing 'dreaming' as a process where Claude reviews session logs to correct faulty memory notes and build wiki-style knowledge. In one internal test, performance on evaluation questions grew from 13 percent to 48 percent after the dreaming process was applied.

27
00:09:28.886 --> 00:09:48.896
On a separate merger-document test, Claude scored 87 percent when consulting the wiki compared to using search alone. Early testing with legal A.I company Harvey reportedly showed task completion rates jumping sixfold after dreaming was applied, though these results come from a single early tester in a research preview.

28
00:09:49.246 --> 00:10:07.038
Anthropic notes that some comparisons rest on single runs, so agent builders should treat them as design guidance and test selective memory on their own tasks. Dreaming remains behind a request-for-access gate, while other complementary features have moved into broader public beta.

29
00:10:07.388 --> 00:10:23.836
That's A.I Voices for today. I'm Maya. Thanks for listening.
Follow the show on Spotify, YouTube, Apple Podcasts, Amazon Music, or your favourite podcast platform, and explore the original voices and sources in the episode notes. See you tomorrow.