Nvidia beat expectations and issued strong guidance while tech executives called for former CEO Jensen Huang’s return. Elon Musk and others circulated a petition, and Huang said he was prepared to come back. Nvidia’s public stance rejects that possibility, with current executives saying, “We’ve got this.”
Interim CEO Colette Kress said Nvidia will no longer sell H200 chips in China, while the Trump administration moves to close remote-access workarounds through data centers in Malaysia and elsewhere.
Critics estimate China-linked demand represents 20% of Nvidia’s business, though it does not appear directly in the company’s financials. That makes the export decision a potential earnings risk even as Nvidia’s near-term results remain strong.
Why it matters: Nvidia’s board is trying to preserve control while export restrictions threaten demand that critics peg at 20% of the business, giving investors a potential earnings hit to weigh alongside calls for Huang’s return.
Google integrated Duplex into Tasks in Gemini, allowing it to call businesses for reservations, hair appointments and other errands when online APIs or booking forms aren't available. It can also adapt when a receptionist proposes a different time.
Duplex began as a narrow system trained on task-specific phone conversations, then used LaMDA in live calls in 2022. Google's current stack uses Gemini for multimodal understanding, planning and dialogue generation, including requests such as scheduling car maintenance around calendar conflicts.
The launch brings Duplex into Gemini alongside its existing roles in Search and Maps. Google said in 2023 that information gathered through Duplex calls had appeared in Search and Maps more than one trillion times.
Why it matters: People whose errands still require a phone call can now give Gemini a free-text request, while businesses without booking APIs become reachable through Google's existing Search and Maps calling infrastructure.
Standard Bank Group deployed an automated assessment that scores generative-AI use cases against international standards and the bank’s own policies. Built with AWS Professional Services, the system gathers evidence, tests applications, and produces reports for the risk committee, automating evidence collection and testing that previously depended on manual review.
The pipeline runs in the bank’s AWS account. A pre-assessment agent scans code and documentation against roughly 40 risk questions, while AWS Step Functions orchestrates evidence checks and live application tests. Those tests include jailbreak attempts, off-topic prompts, repeated or reworded questions, and response-consistency checks.
Reviewers still make the final approval decision; automation standardizes the evidence they receive rather than removing human governance.
Why it matters: Standard Bank Group can give its risk committee consistent evidence as models and tooling change every three to six months, while keeping approval with human reviewers instead of an automated gate.
Shortical is turning Aki Avni’s likeness into a digital avatar. The microdrama startup’s wholly AI series, “Inevitable,” marks its first real-person face-licensing deal. Avni, an Israeli actor who appeared in Fox’s “24,” plays a version of himself who licenses his face to a film studio.
Shortical says lower AI production costs let smaller teams make stories involving kingdoms, dragons, superheroes and werewolves. AI programs on the company’s website appear to have millions of views, while its subscription model gives viewers 7–10 free episodes before charging for the rest.
Shortical is soliciting more faces, but its offering has produced no additional signed contracts. The company will not disclose Avni’s fee; royalty terms, CEO Guy Shimoni says, will vary case by case.
Why it matters: For young actors, Shortical’s AI-first production can remove the live-action sets where they gain experience, while creators considering face licenses still face royalty terms negotiated case by case.
AWS launched managed cluster extensions for Amazon SageMaker HyperPod Slurm, adding API- and console-based software management for observability and user management. The extensions apply settings to every node, including nodes added during scale-up or replacement, and report per-node health.
Previously, teams relied on lifecycle scripts for changing node software. Each update could require editing scripts, replacing nodes, or rerunning them across the fleet; configuration errors surfaced only when scripts executed, leaving drift invisible. HyperPod now validates extension settings against a schema before applying them and keeps the fleet consistent as it scales and heals.
The first extensions cover node, GPU, EFA, Slurm job and cluster metrics, plus directory-integrated user management.
Why it matters: For HyperPod platform teams, the payoff is earlier failure detection: HyperPod rejects invalid settings before deployment and reports whether each node’s extension is healthy.
Anthropic mathematician Levent Alpöge says Claude helped produce a Lean-compilable proof of the Riemann Hypothesis, the prime-number problem proposed by Bernhard Riemann in 1859. MIT mathematician Andrew Sutherland and others have confirmed that the code compiles, but experts still must assess its correctness and whether Alpöge translated the original problem into Lean correctly.
Hours later, Stanford professor Ravi Vakil posted Alessandro Masullo’s all-human Generalized Riemann draft, which Stanford number theorist Brian Conrad had reviewed for months. Vakil later said mathematicians using other AI models flagged problems within eight hours. Neither proof is accepted: Alpöge’s formalization spans 14 data packages and would exceed 250,000 printed pages, while Masullo’s draft is not Lean-formalized.
Why it matters: Anthropic has a proof that software can compile but not yet one mathematicians accept; Stanford has months of human review but must address AI-flagged problems and formalize Masullo’s draft in Lean.