AI News

New Model Additions

This release adds support for several prominent model families. Inkling2 is a multi-modal model supporting text and image inputs with a large parameter scale. Liquid AI's LFM3 and LFM3-VL introduce hybrid architectures combining gated short convolutions and grouped-query attention. Other notable additions include IQuestCoder for code generation, Shanghai AI Laboratory's InternLM3 instruction-tuned model, and experimental decision heads like JEV.

Exporters and Backend Runtimes

The update brings a major overhaul to model exports. Every export now returns an ExportedModel, which shares the same API as the original model and can be pushed to the Hub and reloaded. It supports Dynamo, ONNX Runtime, OpenVINO, and ExecuTorch, with built-in runtime fusions that make decoding significantly faster than standard PyTorch eager execution.

Core Bugfixes and Optimizations

Beyond new models, version 5.23.0 addresses a wide range of issues across vision pipelines, distributed training, attention mechanisms, caching, and audio processing. These include fixing cache correctness in various models, optimizing expert parallelism, and improving overall stability for multi-GPU setups.

背景与事件起因

据美联社等多家媒体报道，英伟达联合创始人兼CEO黄仁勋于周四辞职，并由首席财务官科莱特·克雷斯担任临时CEO。此前有报道称，黄仁勋卷入多起婚外情，其中包括与涉嫌中国间谍的艾丽斯·黄（Alice Huang）的关系。随着美国情报部门的相关报告在特朗普政府高层流传，舆论风暴迅速升级。

商务部长的压力

批评矛头直指美国商务部长霍华德·卢特尼克。多位国会议员指出，商务部在将与艾丽斯·黄有关联、且曾购买数十亿美元英伟达芯片的新加坡公司Megaspeed列入出口管制黑名单一事上存在拖延行为。白宫官员目前正在讨论是否应将卢特尼克撤职，而他本人已同意在未来几周内出席国会听证会。

行业与政策影响

这起丑闻正值美国收紧对华人工智能芯片出口的关键时刻。多位国会领袖呼吁彻底调查并堵住出口管制漏洞，而中国AI公司正试图在海外寻找替代芯片产能，北京方面也在加速推进以华为为代表的国产芯片替代进程。同时，黄仁勋在中国的巨大影响力和个人崇拜现象也在社交媒体上引发了广泛关注和讨论。

Following the resignation of Nvidia's long-time leader after reports of extramarital affairs, US lawmakers have moved quickly to demand answers. Republican Senator Tim Scott requested that the former CEO and Commerce Secretary Howard Lutnick testify before the Senate Banking Committee in December, while three House committees also called for a closed-door interview.

Reports allege that one of the individuals involved was an asset of China's Ministry of State Security, raising serious concerns regarding the protection of sensitive US technology. The unfolding situation has wiped almost half a trillion dollars off Nvidia's market valuation, thrown the company into leadership transition with CFO Colette Kress serving as interim CEO, and left the broader AI industry navigating uncharted political and security waters.

Background and Motivation

Large language models rely heavily on context to generate tokens, but traditional formats like JSON were designed for parsers rather than causal models. Because transformers struggle with long spans and statistical quotation matching, models can become brittle or vulnerable to attacks such as prompt injections. To fix this, developers created LLMON (LLM Object Notation), which uses an XML-like structure with reserved special tokens to cleanly separate instructions, data, and requirements.

Implementation and Capabilities

Today, experimental preview models derived from Granite 4.1 base models, specifically granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, have been released alongside a GitHub toolkit. Unlike regular chat templates that flatten inputs into a single string, LLMON wraps semantically coherent components in meaningful tags and uses an execution block to bind specific instructions to their exact data payloads. Evaluated on the OPI benchmark, these models virtually eliminate static indirect prompt injections, showing significantly higher robustness compared to standard open-source and closed-source models while matching general capabilities.

Background and Positioning

Google has introduced Gemini 4 Hydrogen as the third model in the Gemini 4 series, following Argon and Carbon. Google positions Hydrogen as its mainstream model for daily agent tasks, emphasizing speed and efficiency.

Performance and Comparisons

According to Artificial Analysis, Hydrogen scores 51 points on standard reasoning, outperforming Argon. Google reports it is about 2.5 times faster than Gemini 3.8 Flash. In Google's benchmark comparisons against Anthropic's Claude Sonnet 5.5 and OpenAI's GPT-6.1 Terra, Hydrogen led in 7 out of 10 items, including the HLE-Diamond reasoning test at 58.8%, while trailing in terminal operations and computer use.

Pricing and Availability

Hydrogen features a 1-million-token context window. API pricing is set at $1.50 per million input tokens and $7.50 per output token. It is available immediately across multiple developer platforms, Google Workspace tools, and the free tier of the Gemini app.

The Hidden Cost of Cheaper Oversight

Large companies traditionally relied on a human chain of supervisors and managers to enforce rules and check compliance, a process naturally limited by time and labor costs. However, recent deployments of AI tools in chains like Xushengji, Yiguotang, and Bestore allow companies to automate daily store audits, analyze thousands of operational data points continuously, and catch deviations cheaply.

The Shift from Hierarchy to Micro-Rule

This technological leap creates a paradox of organizational flattening. While fewer managers and shorter approval chains make corporate charts look cleaner, AI bypasses traditional layers to enforce standards directly onto every minute action. What used to be periodic weekly checks turn into relentless daily tasks. Front-line workers must spend their time explaining minor anomalies, leaving less room for local adaptability and proactive problem-solving.

Balancing Control and Exploration

Organizational research shows that while algorithm-driven monitoring can improve baseline execution, heavy-handed digital control often reduces employee initiative and critical feedback. As companies become hyper-efficient at enforcing existing standards, they risk penalizing non-standard yet innovative attempts, ultimately sacrificing long-term adaptability for short-term compliance.

Policy and Community Pressures on Infrastructure

As US AI infrastructure investments scale up, data center projects are encountering unexpected hurdles from local zoning boards, state politics, and community opposition. Industry experts note that local residents in regions like Texas and California are increasingly pushing back against massive power and water consumption, raising concerns about utility costs and long-term community benefits. These local friction points, ranging from environmental permits to grid interconnection bottlenecks, threaten to increase project costs by 10% to 20% and cause widespread construction delays.

Market Dynamics and Supply Chain Vulnerabilities

Wall Street's recent sell-off in AI-related stocks reflects growing anxiety over the sustainability of massive capital expenditures and circular financing structures. Panelists discussed the potential mismatch between hardware supply and actual software demand, raising questions about whether data center slowdowns could create shortages in specific components while leaving upstream chip capacity vulnerable if token economics or model efficiency shift.

The Rise of AI Authors

Some self-published authors are leveraging generative AI to rapidly produce books, navigating platforms like Kindle Unlimited where high output is rewarded. Studies estimate that a significant portion of recent self-published books contain AI-generated text, though writers note that understanding structure and craft remains essential for success.

Scandals and Accusations

Mainstream publishing is facing a period of intense turbulence, with several high-profile book deals cancelled after authors were accused of using AI. Writers of color have disproportionately faced these investigative accusations, while innocent authors live in fear of false positives from detection tools or being penalized for trying to sound non-AI.

The Copyright and Future Dilemma

US copyright law currently requires books to be human-authored to receive full protection, leaving publishers hesitant to embrace AI fully. While traditional publishers refuse to promise a complete ban, many in the industry remain committed to protecting human artists against a backdrop of growing market anxiety.

Background and Rollout

OpenAI is expanding its Dots AI agents platform, introduced earlier this year, by adding a dedicated payment tool called ChatGPT Wallet. The feature is rolling out to ChatGPT Pro subscribers in the US who are 18 and older, requiring a US billing address and eligible major credit cards from American Express, Chase, Mastercard, and Visa.

Security and Privacy

OpenAI reports that users must add a card to the wallet, after which Dots agents can initiate purchases based on user prompts. However, the agent must check in with the user before finalizing any transaction. Every purchase requires a security check using biometric verification, such as Face ID or a fingerprint scan. Similar to Apple Pay or Google Wallet, the system replaces actual credit card details with a payment token, meaning card numbers are hidden from both the AI agent and the merchant.

Future Plans and Limitations

Looking ahead, OpenAI expects users to eventually fully delegate purchases to their agents by defining parameters like spending limits, approved categories, or specific timeframes. However, practical hurdles remain. Some retailers restrict or block AI agents entirely, similar to how Amazon blocked Meta's Muse agents earlier this year. OpenAI notes that online stores must update their websites, payment gateways, and risk systems to accommodate legitimate agent purchases and prevent accidental blocking.

According to LangChain Blog:

LangChain announced that it has raised $300 million in Series C funding at a valuation of $3 billion, led by CapitalG with participation from GIC, DFJ Growth, IVP, and Sapphire Ventures. The company believes that AI agents will transform the economy, and every enterprise will need to control its own intelligence, data, and infrastructure to build competitive advantages.

To capture this market, LangChain highlights its LangSmith platform, which serves over 7,000 organizations—including Fortune 500 enterprises like NVIDIA and Toyota as well as startups like Harvey. LangSmith is structured around three core pillars: Agent Runtime (infrastructure and architecture for running agents), Agent Improvement (testing, monitoring, and automated optimization), and Agent Governance (centralized control over models, tools, and sensitive data). The company also announced an upcoming release event scheduled for Friday, November 20th, to introduce new platform capabilities.

According to Cohere Blog:

Background and Acquisitions

Cohere states that organizations increasingly require AI systems that are powerful, secure, and under their control, a concept the company refers to as sovereign AI. To advance this goal, Cohere has completed its acquisition of AISLE, a builder of autonomous systems for vulnerability discovery and remediation, and welcomed the team behind cubic, an AI-powered code review platform. These transactions mark Cohere's fourth and fifth strategic M&A deals this year, following Aleph Alpha and Reliant AI.

Cybersecurity and Prague Hub

Following the acquisition, AISLE co-founders are taking key leadership roles: Ondrej Vlcek becomes Senior Vice President of Cohere Cyber, Jaya Baloo becomes Chief Information Security Officer, and Stanislav Fort becomes Vice President of AI Security. Cohere plans a long-term investment in Prague, establishing the city as its global center for AI cybersecurity and autonomous security systems, leveraging AISLE's expertise to build cyber-resilient AI capabilities.

Secure Coding and Enterprise Scale

Cubic's team, led by founder and CEO Paul Sanglé-Ferrière, brings expertise in AI, security, and developer tools. Their review agents support environments such as on-premises and air-gapped deployments. According to Cohere, this team will establish the company's secure coding vertical, accelerating the integration of enterprise-grade coding and vulnerability remediation capabilities into its platform.

According to NVIDIA Generative AI:

How Synapsis AI Works

Hospitals accumulate vast amounts of patient history in free-text formats like physician notes and operative reports. Dyania Health developed Synapsis AI to read across these scattered documents, connecting findings from years apart to answer complex clinical questions and surface actionable signals for care teams.

Security and Local Processing

To protect patient privacy, Synapsis AI operates entirely on NVIDIA-accelerated computing inside the health system's firewall. The platform is trained using NVIDIA NeMo libraries and optimized via TensorRT, allowing hospitals to process records onsite without sharing sensitive data externally.

Real-World Application at Cleveland Clinic

In a study involving over 7,800 aortic valve replacement patients between 2005 and 2021, Cleveland Clinic utilized Synapsis AI to review historical charts and build predictive models for valve deterioration and reintervention risks. According to the health system, this approach completes manual chart reviews that would otherwise take years, with clinicians retaining final decision-making authority.

Project Details

According to a Reuters report, multiple regions in China have incorporated restricted Nvidia processors into their AI infrastructure plans. Projects using H100 and H200 chips span across Fujian, Shaanxi, Jilin, Inner Mongolia, and Gansu provinces, with investments ranging from billions of yuan. Some facilities, such as a computing cluster in Wuxi using H800 processors, are already operational, while institutions like Shenzhen University and Northwestern Polytechnical University have also made smaller chip purchases.

Potential Impact

The heavy reliance on restricted American technology exposes Chinese developers to significant supply disruptions, construction delays, and rising costs as Washington tightens export controls. Furthermore, developers might eventually be forced to redesign their facilities to accommodate domestically produced alternatives.

The authors report in this preprint: Digitizing handwritten chess scoresheets has traditionally been framed as a handwriting-recognition task combined with move legality checks. However, the authors argue that this misses silent errors where a misread yields a legal but incorrect move, corrupting the remainder of the game. To address this, the authors present Zugwise, which repurposes computer-chess machinery—such as quiescence search, static exchange evaluation, and pin detection—to evaluate whether a human could have plausibly written a given sequence. It incorporates an absurdity detector, backward search for tracing error symptoms, and a Needleman-Wunsch merge combining evidence from both players' sheets. Evaluated on tournament sections in an oracle replay, the system reduced silent errors from 16 to 1 and from 50 to 13. Two learned signals (CTC re-scoring and a Maia-style human-move prior) raised the correct fix's top-rank rate from 58.8% to 78.0%. The software runs entirely in the browser and is free, private, and offline-capable, accompanied by a three-way adjudication protocol.

The authors report in this preprint: This paper investigates the proposal by Kalai et al. to use explicit probability thresholds in prompts as scoring rules, directing language models to answer only when their confidence exceeds threshold t. The authors probe 12 checkpoints across three post-training stages from four open 3-8B model recipes, alongside exploratory models up to 32B. The empirical tests reveal that while a strict abstention format line shifts mean abstention from 20% to 91% at t = 0.75, varying t from 0.5 to 0.9 within this format changes abstention by a mere -0.8 percentage points. Without formatting prompts, higher thresholds can paradoxically lower abstention, with up to 40% of items receiving decisions incompatible with threshold-independent beliefs. Furthermore, external thresholds on recalibrated P(True) showed brittle generalization across datasets like TriviaQA and PopQA, and models trained on threshold-conditioned targets failed to track unseen thresholds consistently. The study concludes that high abstention rates do not imply true rule compliance and outlines rigorous evaluation criteria for stated scoring rules.

The Shift in Recording Hardware

Standalone AI recording cards like Plaud have enjoyed massive success since 2023, but the core function of recording and transcribing is easily matched by phones and apps. To survive, newer devices are shifting focus from professional meetings to capturing a user's entire day to build personal AI contexts.

The Phone Maker Advantage

Tech giants are quickly closing this gap. Apple added live dialogue rewind and daily recaps to the Apple Watch, while domestic brands like OPPO and Alibaba are launching devices that instantly discard raw audio after cloud transcription, keeping only the text. However, third-party hardware still faces friction because major apps block direct AI control, making standalone or wearable mics a necessary workaround to capture offline conversations.

The Race for the Agent Gateway

Ultimately, these collected contexts are read by AI agents rather than humans. As phone makers race to integrate system-level memory engines by mid-2027, standalone hardware faces a tight six-month window to position itself as a practical peripheral before users rely entirely on built-in phone capabilities.

According to Google AI:

Google reports that Gemini 4 Hydrogen is now available globally for all users in Google Search, representing its largest model upgrade for the search engine so far.

Key Improvements

According to Google, the upgrade features better understanding of intent, allowing Search to handle multi-part questions, typos, and specific requests. It also uses advanced reasoning to improve query fan-out and linking, simultaneously searching subtopics to surface a wider range of relevant websites.

Visual and Language Experience

Users will see more visual responses, including images, structured layouts, comparison tables, and custom tools. Additionally, the system provides richer language understanding across all supported AI Mode languages.

Background and Survey Data

Survey data indicate that generative AI users in China will exceed 700 million by mid-2026, with about a third using the tools for work. Research shows that more than a third of respondents pay for large-model access, and nearly three in ten buy dedicated AI office products. Outside a few tech giants, corporate reimbursement remains rare, leaving individuals, students, and early-career researchers to cover the shortfalls.

Spending and Use Cases

Workers use paid AI tools across four main clusters: drafting and slides, meeting notes and translation, data cleaning and light coding, and specialized work like contract review. Costs vary by user needs, ranging from a few hundred yuan a month for journalists to around 3,000 yuan for fund managers who rely on AI for research code. E-commerce platforms and telecom operators now offer token recharges alongside ordinary consumer goods, turning AI usage into a regular household bill.

Cost Factors and Future Outlook

Pricing design opacity and token-based metering drive up monthly expenses, as users often stack multiple subscriptions and burn credits on prompt tuning. Until employers treat AI access as standard infrastructure or vendors make pricing more transparent, workers will continue to subsidize the efficiency expected by their employers.

According to AWS Machine Learning:

Background

Amazon announced that SGLang-Neuron is now publicly available through an AWS Deep Learning Container. Built by RadixArk in collaboration with Annapurna Labs, the tool brings SGLang's popular serving framework to AWS Trainium, Amazon's proprietary AI chip portfolio designed for cost-effective training and inference at scale.

Technical Details

SGLang-Neuron operates as an out-of-tree plugin that connects SGLang with AWS Trainium. The core SGLang framework handles the API server, request management, and RadixAttention, while the plugin manages Trainium-specific models, attention mechanisms, Mixture of Experts (MoE), and graph execution using native PyTorch and the Neuron Kernel Interface (NKI). It also supports PD disaggregation, allowing prefill and decode phases to run on separate servers and transfer KV caches via Elastic Fabric Adapter (EFA).

Practical Impact

Developers can now deploy models on Trainium using the exact same SGLang APIs and server interfaces they already use. This lowers the barrier to adopting AWS hardware for large-scale model serving and reinforcement learning workloads without requiring a complete rewrite of their serving pipelines.

According to Anthropic Research:

Background and Challenges

Safety evaluations guide critical decisions regarding AI model deployment and safeguards. However, researchers have found that AI evals are easily gamed, poorly documented, and sensitive to minor formatting changes. Drawing on psychological research reforms, Anthropic outlines best practices for building trustworthy safety evaluations.

Design and Execution Best Practices

Building strong evaluations requires matching the design to the core question, whether observational or causal, and measuring the right construct, such as appropriate response rates instead of simple refusal rates in sensitive contexts like self-harm or dual-use biology. Evals should use multi-turn simulations to capture nuanced, escalating user behavior and run in secure, sandboxed environments to prevent unintended actions.

Grader Validation and Reporting

Since modern evaluations rely on AI graders, teams must calibrate them rigorously, ensure atomic and binary criteria, and validate them against human labels. Furthermore, researchers should preregister analysis plans, report full statistics with confidence intervals, and disclose model configurations transparently to ensure results are reproducible.

Jensen Huang’s resignation has caused anxiety in Israel, where Nvidia employs around 6,000 people and has deep roots following its 2019 acquisition of Mellanox. Local tech figures note that Huang personally drove the company’s heavy investment in the region, raising questions about whether his successor will maintain the same level of commitment. Interim CEO Colette Kress has deep ties to the region, and Israeli executives like Apple’s Johny Srouji and Nvidia’s Michael Kagan have been floated as potential successors.

Despite the local uncertainty, venture capitalists in Israel view the transition as a potential turning point for the broader AI industry. According to Vertex Ventures Israel General Partner Yanai Oron, slower decisions at Nvidia could give rivals an opening, though a broader slowdown in AI spending remains a risk. Selah Ventures founder Adi Levanon notes that while Huang's departure does not dismantle Nvidia's ecosystem, it might prevent the company from becoming a full-on monopoly, giving alternative chip and inference software developers a chance to prove their economic value.

After three and a half years of growth, the llama.cpp project has become a foundational tool for running multi-modal large language models locally on everything from consumer computers to enterprise servers. With the release of version 1.0.0 for both llama.cpp and its tensor library ggml, the team aims to provide a more reliable foundation for the expanding ecosystem of local AI applications.

To support this growth, the project is establishing strict semantic versioning rules. Any breaking changes to the public C application programming interface will now trigger major version bumps, making interface boundaries clearer for developers who integrate the software into their products. Additionally, the project continues to expand its distribution channels and hardware support, utilizing infrastructure from Hugging Face alongside hardware partnerships with major chipmakers to ensure code quality and broad platform compatibility.

Jensen Huang, the 63-year-old founder and CEO of Nvidia, has stepped down from his role after a series of personal revelations. On Sunday, November 8, Huang admitted to having a relationship with Alice Huang, a Chinese entrepreneur suspected by US intelligence agencies of spying for Beijing. Shortly after, on Wednesday, November 11, news emerged of a second romantic relationship with a subordinate who is an executive at Nvidia.

In response to these developments, Nvidia's board of directors convened urgently on Wednesday evening and announced his resignation on Thursday. Huang, who led the company since its founding in 1993, had previously shifted its focus toward artificial intelligence and ruled the organization with direct management over roughly 60 reports. He had not established a succession plan, leaving the tech giant suddenly without its long-time leader.