Jensen Huang Admits to Affair and John Ternus Finally Wants a Siri Bot

Jensen Huang Admits to Affair With Suspected Chinese Spy

Nvidia CEO Jensen Huang has admitted in a message to staff to having a romantic relationship with a woman US intelligence agencies suspect of being a Chinese spy. On Sunday, the company’s board of directors announced an independent investigation. “I owe an apology to my family and to all of you,” reads a draft of the email to employees. In it, Huang writes that he was unfaithful to his wife, causing pain to her and their children, and that there is no excuse for it.

Huang states that he takes “full responsibility.” He writes that he failed to meet his obligations to employees, the board, and shareholders. Leading Nvidia is “the privilege of my life,” he writes, adding that he remains fully committed to the company and wants to regain trust.

The review will be conducted by an independent board committee supported by outside legal counsel. According to its own statement, the board has no reason to believe that national security or company trade secrets were compromised. In the same statement, it expressed its confidence in Huang as CEO.

The case is explosive because of Huang’s role in Washington. He is considered one of the key voices there advocating for an easing of export restrictions on AI chips to China. His argument: keeping American technology present in the Chinese market strengthens US competitiveness. At the same time, it discourages China from developing its own and potentially more powerful chips.

The relationship had already been a topic in government circles. Secretary of State Marco Rubio reportedly warned senior Trump administration officials about Huang’s connection to the woman, whom US intelligence officials suspected of being a Chinese agent. → Axios, CNBC

Synthszr Take: A board that announces an investigation and expresses confidence in the person being investigated in the very same statement has already written the verdict. At Nvidia, that is costlier than elsewhere. A founder-CEO controls the bottleneck of the entire AI industry, and every China deal is scrutinized by Washington. External lawyers may clear Huang. But the government and customers will only believe it once oversight visibly operates without him. So far, the board sounds like its own PR department. Nvidia CEO Jensen Huang admits to affair with suspected spy, board launches review.

Apple CEO Ternus Wants to Turn Siri Into an AI Agent Faster

Apple’s new CEO John Ternus is pressuring software chief Craig Federighi to accelerate the company’s AI plans. In a closed-door meeting in recent weeks, Ternus reportedly demanded that the software roadmap be pushed forward and go beyond the necessary improvements to Siri. According to insiders, he is frustrated by a reluctance to tackle AI with the same urgency as the competition. Since the firing of the previous AI chief last year, Federighi has also been in charge of Apple’s AI efforts. On a recent earnings call, Ternus described the capabilities of Agentic AI as “promising.”

Internally, Federighi is seen as an AI skeptic. He is said to have voiced concerns about privacy, security, societal consequences, and the impact on jobs. His leadership style is described as heavily hands-on in meetings and major decisions, leaving little room for competing ideas. As a result, parts of the software and design teams feel they have limited influence on product direction. Furthermore, Federighi plans in 12- to 18-month cycles, while hardware development operates on multi-year roadmaps. A departure is not in sight, and according to the report, succession planning is not a priority for him. What remains undisputed is that he delivered the revamped Siri in a short time after years of delays.

Two expansion stages have been announced for Siri. With iOS 27.3, potentially as early as December, the assistant is expected to control Apple’s own apps with finer granularity, such as rotating and cropping images in Photos, organizing Safari tabs, or editing messages. Developers will reportedly be able to define sequences of tasks. Integration with third-party apps is expected in the spring, such as ordering through Grubhub or booking tables via OpenTable. These capabilities remain query-based: every action still requires user initiation. With iOS 27.2 on October 26, 2026, Apple already introduced an “Excluded Apps” menu allowing individual apps to be excluded from Siri, search, and suggestions.

As the next step, Apple wants to evolve Siri into a true agent that pursues goals, determines next steps autonomously, and completes ongoing tasks without step-by-step instructions. The benchmarks are considered to be OpenAI’s Dots, Meta’s Muse, and Grok. According to observers, such an overhaul could begin as early as 2027, though no concrete timeline is known. What remains unclear is how Apple will handle permissions, erroneous actions, and access to external accounts and services outside its control.

In parallel, Apple is changing its development process for iOS 28 and iPadOS 28 (internally “Bell”) as well as macOS 28 (“Poppy”). Each of the three milestones—M1, M2, and M3—lasts eight weeks, previously divided into four weeks of coding and four weeks of integration. In the already completed M1 phase, Apple shifted the ratio to six weeks of coding and two weeks of integration, citing, among other factors, more capable coding agents. → 9to5Mac, MacRumors, The Verge, Gizmodo, AppleInsider

Synthszr Take: A new CEO needs proof that analysts can count on the next earnings call, and Siri cropping photos isn’t it. Ternus’s incentive is therefore aimed at an agent that pursues goals independently. That is precisely where the unresolved questions around permissions and erroneous actions in third-party accounts lie—the very issues Federighi has been cautious about. The acceleration shifts risk from the schedule into product liability. For Grubhub and OpenTable, this means they are building on an API whose rules are being drafted under deadline pressure.

Nous Research Tunes Alibaba’s Qwen to Western Values

Nous Research has released Arete-27B, a post-trained version of Alibaba’s Qwen3.8-27B. According to the lab, it is designed to reinforce Western values. On its in-house evaluation, the share of propaganda-laden responses dropped from 22 to 0.9 percent. Nous justifies the move citing refusals, evasive answers, and responses aligned with the Communist Party that the base model produced on politically sensitive topics. The US startup, known primarily for its Hermes models, is making the weights, evaluation dataset, and training data available on Hugging Face. After training, the factual score rose from 2.32 to 2.98 out of 5 points. The number of zero-point responses fell from 359 to 216 out of 803. The most pronounced shift was in the freedom of expression category: the factual score rose from 0.89 to 2.44, and the propaganda share dropped from 60 to 1 percent. With reasoning enabled, Arete reaches 3.07 points; this figure is currently based on two out of five runs. On the internal Hermes Bench, Arete scores 65.60 percent compared to 65.77 percent for the base model, which Nous classifies as noise. → Nous Research, implicator, Nous Research on Hugging Face, The Decoder

Synthszr Take: 52.8 million tokens, two epochs, no reinforcement learning: that’s how thin the layer was on which Beijing’s line sat in Qwen. Jailbreak communities used to take months to bypass guardrails. Here, 1,202 training steps are enough. The same logic works in every direction; the values of an open model therefore belong to whoever fine-tuned it last. For buyers, the country of origin of the weights loses significance; what matters is the chain of fine-tunes (Inkling itself learned from Kimi).

Envision Builds China’s Stargate in Inner Mongolia at One-Fifth the US Cost

Chinese energy group Envision is building “Galaxy” in Ulanqab, Inner Mongolia, a data center campus with up to 800 megawatts of IT capacity. SemiAnalysis visited the site in September and describes it as China’s counterpart to OpenAI’s Stargate campus in Abilene, Texas. According to SemiAnalysis, the initial 120-megawatt phase was completed within a year, at roughly one-fifth the US cost. All four construction phases are tailor-made for two unnamed Chinese internet conglomerates. According to a local report, nearly 30,000 construction workers were simultaneously working on 22 data center projects in the surrounding Chahar High-tech Zone. → SemiAnalysis

Synthszr Take: In AI expansion, permits are becoming scarcer than GPUs. Chips can be bought; acceptance in Texas hardly exists anymore. In the Chahar High-tech Zone, nearly 30,000 construction workers are working simultaneously on 22 projects without a single protest. Envision manufactures rotor blades for its own wind power directly across from the campus. For OpenAI and the hyperscalers, this means: every month spent disputing site locations grants a head start to two Chinese internet conglomerates, built at one-fifth the cost.

Anthropic Buys 1.7-Gigawatt Gas Generators for $3.4 Billion

According to the Financial Times, Anthropic has taken over orders for gas-fired generators with 1.7 gigawatts of capacity for $3.4 billion. The orders for heavy Caterpillar generators had previously been placed by supplier OnePWR Solutions from San Antonio. According to the FT, Anthropic took over these orders at the end of September without acquiring the supplier itself. The newspaper describes the deal as the largest known single purchase of power generation equipment by an AI lab. Anthropic originally intended to source power from OnePWR, but then opted to take over the equipment orders directly. → TheEnergyMag

Synthszr Take: Anthropic set out to buy power from OnePWR and ended up buying a spot in Caterpillar’s order book. The bottleneck thus lies in delivery slots for heavy generators, and those cannot be reordered on short notice. While Microsoft has secured long-term baseload with its nuclear deals, Anthropic is securing gas generators that can be up and running much faster. Competitors without such orders will in the future plan their data centers around someone else’s delivery schedules. Compute capacity will ultimately be decided by who ordered steel early enough.

Latent Sparse Attention Makes 10-Million-Token Contexts Up to Nine Times Faster

Researchers from Prefrontal AI and ETH Zurich have introduced Latent Sparse Attention (LSA), a method designed to significantly accelerate the processing of very long contexts in language models. AlphaSignal reports this in an article by Ben Dickson. On tasks involving ten million tokens, the time to first token dropped by a factor of 9.2 compared to a full-attention model, according to the researchers. Compared to a common Sparse Attention method, the speedup was 2.2x. The background to this is a bottleneck in previous approaches: their indexer still checks every previous key, causing the computational effort to grow quadratically with context length. For the GLM-5.3-Flash model, 98 percent of an attention layer's prefill time at ten million tokens was spent solely on selecting relevant tokens. Instead, LSA sorts the keys into clusters within a learned space, each with a representative. A query first checks only these representatives and then computes with the keys from the most promising groups. → AlphaSignal

Synthszr Take: In GLM-5.3-Flash, 98 percent of prefill time is spent searching for relevant tokens. At ten million tokens, customers are primarily paying to flip through the catalog. LSA strips out this exact line item. For agent sessions whose context grows over hours, prefill is the biggest cost driver. D-LSA can be attached to unmodified models, meaning the savings flow directly into inference providers' margins without any retraining. The long-context surcharges on their price lists thus lose their technical justification.

GitHub Upgrades Its Git Storage Architecture for 5,000 Parallel Coding Agents

On November 5, GitHub published details of a new storage architecture for Git designed specifically for large numbers of AI coding agents working in parallel. According to the company, a single repository withstood 5,000 concurrently working agents in internal tests. GitHub reported that this single repository handled up to 7,200 pushes per second. Read requests reportedly maintained consistently low latency throughout the tests. The overview on DEV emphasizes that these are internal test figures and that typical repositories do not currently experience such loads. → TechOrange 科技橘

Synthszr Take: 7,200 pushes per second will not overwhelm storage, but they will overwhelm any permission model built for humans. Branch protection and mandatory reviews assume that a human reviews every merge. With 5,000 agents, each needs its own identity with narrowly scoped permissions revocable in seconds, and every commit needs verifiable provenance. GitHub tested throughput, not this layer. Companies going into production before addressing this are handing out thousands of write keys to actors whose intent no one verifies.

Tencent’s Chief AI Scientist Warns of Bioweapons from AI Models

Yao Shunyu, Chief Scientist for AI at Tencent and former researcher at OpenAI, expects AI models to soon master biology. In his assessment, this brings significant benefits, but also serious risks. He spoke on Saturday during one of his rare public appearances. The occasion was a scientific symposium co-hosted by the Hong Kong Academy of Sciences and the Future Science Prize. According to Yao, the cyber capabilities of models—and soon their biological capabilities—are reaching a threshold beyond which harm becomes possible. He described a line that will be crossed first in the cyber domain and then in biology. He reportedly did not provide specific timelines or countermeasures. → South China Morning Post

Synthszr Take: A warning without a timeline and without countermeasures costs the one issuing it nothing. The bill goes to others: Biotech startups using frontier models for drug discovery are the first to face screening requirements, access vetting, and more expensive insurance. Health authorities have to fund surveillance for pathogens that could eventually be designed using open weights. Meanwhile, model developers continue to sell capabilities whose misuse they themselves predict, while facing almost no liability so far.

Jevons Takes Effect: US Service Providers Add 70,000 Office Jobs in October

Employers in professional and business services in the US added 70,000 jobs in October. The US Department of Labor reported this on Friday. According to the agency, it is the largest increase in this segment in several years. The sector includes lawyers, accountants, and consultants, among others, and is therefore considered a bellwether for the trajectory of traditional white-collar jobs.

A debate has been underway among economists for some time regarding whether artificial intelligence is beginning to displace tasks in white-collar professions. Recent labor market data is now being interpreted in media reports as an indication that such displacement is not yet reflected in hiring figures. For October, the data points in the opposite direction.

However, the numbers remain contentious. At least one economist described the increase as a “head fake”—a false signal that simulates a trend that may not continue. This leaves it open whether October marks a trend or an outlier. → Wall Street Journal, Washington Post

Synthszr Take: 70,000 new jobs in a single month disprove the displacement thesis just as little as a weak month would have proven it. A sector that lumps together lawyers, accountants, and consultants smooths over the exact shift in question. AI first replaces tasks, and firms continue hiring anyway as demand rises. The signal will only become resilient once hiring in exposed professions declines over several quarters. Until then, October remains a data point for headlines, and workforce planning needs the time series.

AI Mathematics Could Crack Encryption Before Quantum Computers Do

Advances by AI systems in mathematics could jeopardize the encryption of bank accounts, email inboxes, and large parts of the internet sooner than quantum computers. For decades, engineers have been preparing for powerful quantum computers. They expect well-funded attackers to exploit them to break today’s cryptographic methods. In response, new security methods are currently being rolled out under the term Post-Quanten-Kryptografie. The premise: recent mathematical breakthroughs by AI models could pull forward the moment when standard encryption becomes insecure. Concrete attacks or broken methods have not yet been demonstrated. What is at stake is the possibility that the threat timeline will shift.

Internal documents from Cisco reveal how vulnerable and politically contested the network infrastructure beneath encrypted traffic is. The corporation supplies routers, servers, and communication systems to businesses, governments, and militaries. According to the leaked documents, the Israeli military intelligence networks suffered multi-hour outages and repeated crashes about a year after the start of the Gaza war. The issues occurred during an update of the digital infrastructure, and built-in debugging mechanisms failed. A draft Cisco memo from September 2025 notes that the outages severely disrupted critical operations and damaged customer trust.

Business with military intelligence brought in around $67 million over three years, according to an internal document from 2025. The records show that Cisco sought to maintain and expand the relationship despite the criticism. Its subsidiary Splunk, a platform for analyzing large volumes of machine-generated data, applied this year to participate in Project Nimbus. This is the Israeli government’s controversial cloud contract with Google and Amazon.

The documents were leaked by employees who oppose the company’s ties to the Israeli military. The group Bridge to Humanity, made up of current and former employees, accuses Cisco of betraying its own values. Cisco did not respond to a request for comment. The Israeli military declined to comment on contracts with third parties. The US Equal Employment Opportunity Commission (EEOC) previously found reasonable cause of a hostile work environment for Palestinian, Arab, and Muslim employees. Cisco denies this assessment. → The Guardian, Wall Street Journal

Synthszr Take: Post-quantum cryptography rests on a comfortable assumption: only well-funded actors with their own quantum labs can afford the attack. A mathematical breakthrough from an AI model knows no such barrier to entry. It spreads via API or as open weights—if in doubt, to anyone with a credit card. This makes model access an infrastructure security issue. Banks should pace their migration plans according to the release cadence of AI labs; qubit roadmaps take a back seat.