독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-11-0925 게시물

새로고침

Tencent Chief AI Scientist Yao Shunyu Warns of AI Bioweapon Risks, Advocates Training Data Filtering to Counter Open Model Vulnerabilities

주제 · 姚顺雨AI生物安全警告중대
2026-11-09 23:44 GMT+8

According to a report by the South China Morning Post, Vinces Yao Shunyu, Tencent's Chief AI Scientist and former OpenAI researcher, recently issued a warning that artificial intelligence will soon be capable of developing bioweapons. In the debate over open versus closed AI models, Yao argues for a middle ground, stating that neither purely open-source nor purely closed-source approaches can solve this security problem on their own. He notes that since the gap between Chinese and US models is not large, the entire industry must proceed carefully.

Yao cites the hacking incident at Hugging Face as an example of the defensive advantages of open models: when closed US models like Anthropic's Claude refused to analyze attack commands due to safety filters, the platform used the open Chinese model GLM-5.2 to parse the logs instead. However, the Hugging Face report also highlights the limits of this approach: GLM-5.2 was only used for forensic analysis, and the actual attack was stopped by patching vulnerabilities and rebuilding servers. Hugging Face stressed that the incident is not an argument against safety filters on hosted models.

Meanwhile, safety filters on open models are extremely fragile. A technique called 'Abliteration' allows anyone with access to model weights to surgically remove trained refusal behaviors. Thousands of these modified variants currently sit on Hugging Face, and commercial services selling unfiltered versions have emerged. In contrast, closed models often filter too aggressively. Since Claude Fable 5, Claude's filters route suspicious queries about cyber and biology topics to the weaker Opus 4.8 model, causing even MRI scan analyses to be flagged as bioterrorism, a limitation Anthropic has acknowledged.

To address this dilemma, Yao proposes a more robust approach: keeping dangerous knowledge out of the model during training in the first place. Tencent plans to release its new model Hy4 as an open-weight model with an accompanying safety report in the coming days. The company states that Hy4's biological capabilities fall below the high-risk threshold and that it removed the model's hacking capabilities by filtering the training data. Researchers say this approach survives abliteration as well. This strategy reflects China's latest moves in AI safety standard-setting, attempting to solve security alignment problems at the architectural level while maintaining model openness.

출처:The Decoder

리서치

Databricks releases UCDataset beta to stream Unity Catalog tables directly to GPUs, claiming 10x speedup on simulated loads

주제 · UCDataset간단한 요점
2026-11-10 03:00 GMT+8

Databricks announced the beta release of UCDataset (Unity Catalog Dataset) on November 9, 2026, available on its AI Runtime platform. This native PyTorch data loader is designed to stream table data directly from Unity Catalog to GPUs during distributed training jobs, eliminating the need for pre-exporting data to intermediate formats like Parquet or writing custom loader code.

The tool addresses a common bottleneck in large model training: expensive GPUs sitting idle while waiting for data. Traditional workflows often require exporting tables to volumes before each run, which incurs storage costs, leads to stale data, and breaks governance lineage. UCDataset resolves this by reading table files directly from cloud storage and automatically sharding rows across multiple GPUs and workers.

In a blog post, Databricks stated that on simulated production workloads, UCDataset reduced average training time by 10x compared to loading data from an exported Parquet copy. This performance metric is based on internal benchmarks and has not been independently verified. Currently, UCDataset supports Delta tables registered in Unity Catalog but does not yet handle row filters or column masks in the beta version.

출처:Databricks

리서치

Anthropic Spends $3.4 Billion to Take Over OnePWR Orders, Securing 1.7GW of Gas Power Capacity

중대
2026-11-09 21:45 GMT+8

According to Reuters, Anthropic has agreed to spend $3.4 billion to secure 1.7 gigawatts (GW) of gas-fired generators for its AI data centers. Rather than purchasing new equipment directly, the company took over existing generator orders previously placed with Caterpillar by Texas-based supplier OnePWR Solutions. The agreement, reached in late September 2026, also covers the engineering and construction work needed to install the equipment.

This move highlights the core bottleneck facing current AI infrastructure expansion: power supply. Global gas turbine orders have reportedly exceeded 100GW, while manufacturing capacity stands at an estimated 60GW to 70GW. Average delivery times in the U.S. have reached six years, with some prospective orders extending into 2030 or 2031. By taking over OnePWR's orders, which were placed roughly two years ago but required additional financing, Anthropic avoided starting from scratch and significantly shortened its timeline for acquiring power.

Bypassing traditional grid connections and manufacturer waitlists is becoming common across the AI industry. Elon Musk's xAI recently installed eight converted General Electric jet turbines at a data center in Mississippi, yielding a potential combined output of 400MW. OpenAI's first Stargate data center followed a similar route, with builder Crusoe acquiring 29 GE Vernova gas turbines. However, Anthropic is taking a different path by stepping into existing supply chain backlogs rather than buying new equipment or relying on retrofitted retired aircraft engines.

While this strategy rapidly solves the power issue for computing facilities, it has drawn environmental criticism. Shakeel Hashim, founder of ReadTransformer, noted that these kinds of generators are terrible for the environment. This points to an ongoing issue as AI companies expand their power infrastructure: securing electricity quickly does not remove the environmental costs of generating it.

출처:Tom's Hardware

리서치

US NHTSA Investigates Waymo Robotaxis for Continuing School Bus Violations After Software Fix

중대
2026-11-09 23:24 GMT+8

The US National Highway Traffic Safety Administration (NHTSA) filed documents on November 6 announcing a new "recall query" investigation into Alphabet-owned Waymo's robotaxis. The probe aims to assess whether the software update deployed by Waymo last November was adequate to resolve issues with vehicles illegally passing stopped school buses.

According to TechCrunch, NHTSA noted that a majority of the reported incidents actually occurred after Waymo completed the software update for all affected vehicles. This finding prompted the agency's Office of Defects Investigation (ODI) to open proceedings to evaluate the remedy's effectiveness.

Previously, Waymo faced an initial investigation in October 2025 after footage showed its vehicles maneuvering around stopped school buses with flashing lights in Atlanta and Austin. Despite completing the software update on November 17, 2025, regulatory data suggests the behavior persisted. Waymo has not yet publicly responded, and the NHTSA investigation remains ongoing.

출처:TechCrunch

리서치

OpenAI Announces Community-Friendly Data Center Policy: Peak Throttling and Grid Feedback

주제 · OpenAI중대
2026-11-09 22:00 GMT+8

According to a report by heise online, OpenAI has announced changes to its data center operating model to reduce negative impacts on surrounding communities. This move is seen as a direct response to the anti-data-center sentiment observed during recent US midterm elections, which contributed to a sharp decline in semiconductor stocks.

Under the new policy, OpenAI commits to ensuring that local residents do not pay for the electricity consumed by its facilities. When power demand peaks in neighboring municipalities, OpenAI intends to throttle its own consumption and feed energy back into the grid. Additionally, the company plans to disclose the specific environmental footprint of each data center, including power usage, noise levels, and water requirements.

Chris Lehane, OpenAI's head of global affairs, stated that these rules had been under development for months and were not timed to coincide with the election results. However, the announcement highlights that the expansion of AI infrastructure now depends not only on technology and capital but also on local community acceptance.

출처:heise online

리서치

Diagrid Releases Dapr 1.19: Optimizing AI Agent Workflows and Simplifying Control Plane

중대
2026-11-09 22:00 GMT+8

Agent infrastructure startup Diagrid released Dapr 1.19 on November 9, 2026, introducing changes that cut the overhead of durable workflows behind long-running artificial intelligence agents and simplify its control-plane architecture.

Prior to version 1.19, Dapr’s workflow engine shipped the entire execution history to the worker on every step. As jobs ran for hours or fanned out into dozens of parallel tasks, the payload kept growing, pushing up latency and memory use. Now, a worker keeps what it already holds and receives only new events. This makes the data sent per step independent of how long a workflow has been running. The optimization currently applies to workers built on Dapr’s Java and Python software development kits.

A second workflow change, an opt-in preview feature called WorkflowsFastPath, stops pushing every workflow event through a durable scheduler reminder. Wake-ups and local activities are handled in place. In Diagrid’s internal engine benchmarks, median step latency fell from 553 milliseconds to 153 milliseconds with the feature enabled.

Regarding the control plane, Dapr 1.19 introduces a new mode that retires the Placement service, which previously distributed the actor placement table. The surviving Scheduler service takes over actor placement, along with jobs, reminders, and workflow scheduling. This spares operators from deploying a stateful service and the Raft consensus cluster behind it. Scaling events also become less disruptive: previously, a host joining or leaving paused placement across the entire namespace, but in the new mode, only affected actor types are locked and updated.

New alpha building blocks aimed at AI application developers include a Vector building block for storing and querying embeddings (with Meilisearch as the first supported component) and a Binary Store for streaming large files like model outputs. For .NET developers, the new Actors.Next SDK generates proxies and registration code at compile time, cutting runtime overhead and enabling Native AOT compilation.

출처:SiliconANGLE

리서치

Broadcom Open-Sources Its Kubernetes VM Manager to CNCF, Leads New Conformance Initiative

주제 · 博通VM API开源중대
2026-11-09 22:00 GMT+8

U.S.-based semiconductor and infrastructure software company Broadcom Inc. has announced that it is contributing its Virtual Machine Application Programming Interface (VM API) project to the Cloud Native Computing Foundation (CNCF). The project has been accepted into the CNCF Sandbox, an entry point for early-stage cloud-native projects, aiming to simplify operations across diverse enterprise computing environments.

The VM API provides a vendor-neutral framework for provisioning and managing virtual machines using the same APIs that Kubernetes employs to orchestrate modular and portable software containers. Prashanth Shenoy, Broadcom’s Chief Marketing Officer and Vice President of Marketing for Cloud Platform, Infrastructure and Solutions, stated that the goal is to make virtual machines "first-class citizens" of a Kubernetes environment without sacrificing capabilities such as hardware isolation and specialized input/output performance.

In addition to the open-source contribution, Broadcom is co-chairing a newly established Workload Conformance Working Group. This group aims to develop common expectations for how applications operate across different Kubernetes distributions. Existing conformance programs verify that distributions support standard APIs, but do not guarantee identical application behavior due to differences in infrastructure capabilities, configuration, and operational practices. "Just because two Kubernetes environments are conformant doesn’t necessarily mean that the application will behave consistently across both," Shenoy said.

A Linux Foundation survey of 492 practitioners, scheduled for release, found that VMs account for 53% of production applications, with most respondents expecting this proportion to remain stable. Among organizations expanding their VM deployments, 34% cited AI and agentic workloads as contributing factors. These moves reflect the growing need for infrastructure that can support traditional applications alongside modern containerized and artificial intelligence workloads.

출처:SiliconANGLE

리서치

HOPE: Incorporating Expert Pairwise Interactions into Quadratic Programming to Maintain MoE Cooperative Groups at 50% Pruning Rate

주제 · MoE专家协作剪枝중대
2026-11-09 20:34 GMT+8

The vast majority of parameters in Mixture-of-Experts (MoE) models remain idle—of the approximately 250 GB of VRAM occupied by Qwen3.5-122B, 95% is taken up by the MLPs of experts across all layers, yet only 8 experts are actually activated per token. Directly deleting inactive experts is the most concise path to compressing this memory bottleneck: it requires no training, no dynamic loading, and can be combined with quantization. The difficulty lies in deciding whom to delete. Existing pruning methods—such as REAP, MAN, and EAN—assign a scalar importance score to each expert independently and then remove those with the lowest scores. Author Alex M. Tseng (Amazon) points out that this approach assumes each expert's contribution is independent of others, but experimental observations indicate that experts in MoEs form stable cooperative groups that are consistently co-activated. Independent scoring completely ignores this: removing expert A may seem harmless on its own, but if A frequently cooperates with high-importance expert B, a single deletion could dismantle an entire cooperative group.

The core of HOPE is an E×E interaction matrix F. The diagonal of the matrix measures the independent importance of individual experts—which correlates highly with REAP rankings, with a Spearman correlation coefficient of 0.988—while the off-diagonal elements capture the average gate-weighted output product of experts i and j on tokens where they are co-activated. With this matrix, selecting the pruning set becomes a quadratic programming problem: minimizing the total interaction strength among pruned experts. The solution employs continuous relaxation followed by rounding, taking 1 to 2 seconds per layer, which is negligible compared to the total overhead of pre-calibration forward propagation.

The advantage of this mechanism is most significant at a 50% pruning rate. Taking Qwen3.5-35B-A3B as an example, after HOPE pruning, the model scored 42.3 on SWE-Bench Pro, matching the unpruned model's 41.8; whereas REAP dropped to 36.2 under the same conditions, representing a regression of about 13%. The authors attribute this to the fact that at high pruning rates, many cooperative links face the risk of being severed, while HOPE's quadratic programming automatically sacrifices some experts with high independent scores but weak cooperative node status to preserve the overall cooperative structure.

However, the authors also acknowledge the boundaries of HOPE: at low pruning rates of 10% to 20%, first-order methods are sometimes equally competitive, narrowing HOPE's advantage; the current implementation prunes each layer independently without considering cross-layer interactions; the modeled relationships are limited to pairs, making triplet and higher-order interactions combinatorially difficult to estimate. All performance figures are self-tested by the author's team and have not yet been independently reproduced.

출처:huggingface.co

리서치

German ifo Institute Releases AI Early Indicator: Entry-Level Job Decline Associated with AI, Youth Employment Down 14% in High-Exposure Roles

주제 · ifo AI劳动力指标중대
2026-11-09 21:30 GMT+8

The ifo Institute in Munich published a new AI early indicator (KI-Frühindikator) on November 9, 2026, designed to measure the impact of widespread AI adoption on the labor market. The study compared employment and new hires across 140 occupations, relating them to the substitutability of human work by AI.

According to the data, full-time employment among 20- to 24-year-olds in the most AI-exposed occupations has fallen by 14% since 2022, dropping from approximately 279,000 to 239,000. These high-exposure roles include data entry and text processing, software developers, accountants, database and network technicians, as well as authors and journalists. "We do not yet see broad job cuts due to AI," said Clemens Fuest, President of the ifo Institute. "But those starting their careers feel the change first."

However, the impact of AI on the labor market is not uniformly negative. The study notes that while entry-level positions have decreased, total employment in AI-exposed occupations actually grew by 1.8% since 2022, whereas employment in non-exposed occupations declined by 6.4%. This suggests that AI is altering the nature of work through 'augmentation' rather than just 'automation,' enhancing the productivity of existing employees. Co-author Oliver Falck views the loss of jobs for beginners as an early warning sign for the labor market.

It is important to note that researchers emphasize this indicator does not provide a monocausal explanation. Demographic changes and longer training periods may also contribute to the observed employment trends. Additionally, even executives face risks, as AI can handle significant portions of management tasks; conversely, occupations relying primarily on manual labor, such as construction workers, carpenters, cleaners, and farmers, are least affected by AI.

출처:t3n.de

리서치

Open-Source Project Darling Adopts AI-Assisted Contributions, Runs League of Legends on Linux

주제 · Darling간단한 요점
2026-11-09 21:02 GMT+8

The macOS compatibility layer project Darling has officially adopted rules for AI-assisted code contributions. Designed to allow Linux users to run macOS binaries, the project had progressed slowly for over a decade due to limited developer resources. The team states that large language models (LLMs) have changed the equation, making previously unsolvable problems financially and practically realistic.

According to Phoronix, the Darling team used AI-generated patches to successfully run the macOS version of League of Legends on Linux. This breakthrough demonstrates that for small open-source teams, AI tools can significantly lower the barrier to complex compatibility development, allowing them to compete with Apple's ecosystem on closer to level ground.

The policy explicitly views AI as a powerful tool but emphasizes that its use must be bound by human oversight. Currently, several Darling forks embracing AI-generated code are working to merge these patches into the upstream repository.

출처:phoronix.com

리서치

Google Launches Two New TPU Variants as External Customer Count Nears 100

중대
2026-11-09 20:59 GMT+8

On November 9, 2026, Google announced the addition of two new systems to its eighth-generation AI accelerator lineup: the TPU 8tx for confidential computing and the TPU 8i Max for memory-intensive workloads. This marks the first time Google has shipped four TPU systems within a single year, signaling a strategic shift from treating TPUs as an internal tool to marketing them as a competitive platform for external customers.

The TPU 8tx targets enterprises and government agencies working with confidential data. While data is typically encrypted during storage and transmission but decrypted in memory during processing, Google states that the TPU 8tx keeps data and model weights protected and isolated even during active use. The system uses the low-latency Boardfly network and delivers up to twice the latency reduction compared to the TPU 8t.

The TPU 8i Max addresses the memory bottleneck common in AI agent applications handling large codebases or video. By placing two TPU 8i chips on a single package, Google achieves double the compute power, double the memory, and double the memory bandwidth within the same rack format, significantly increasing performance per square meter of data center space.

Beyond hardware, Google released TorchTPU as open source to allow PyTorch code to run on TPUs with minimal changes, alongside support for inference frameworks like vLLM and SGLang. With nearly 100 external TPU customers, Google is increasingly positioning its in-house chips as an alternative to Nvidia GPUs, particularly for day-to-day model inference.

출처:The Decoder

리서치

Score Centering: Eliminating Gradient Drift in RL Training with Additive Correction

중대
2026-11-09 18:33 GMT+8

Reinforcement learning (RL) training for large language models (LLMs) typically employs a disaggregated architecture: an inference engine is responsible for generating rollouts, while a training engine computes policy gradients. These components run on different hardware, utilizing distinct precisions, kernels, and batch shapes; consequently, they produce different probabilities for the same token—a phenomenon known as Training-Inference Mismatch (TIM). Asynchronous training further exacerbates this issue, as policy updates may be based on rollouts that are already stale.

The core contribution of Marek and Ryabinin’s paper lies not merely in proposing a correction method, but in providing a deep analysis of why TIM causes instability. They demonstrate that under a constant reward setting, the theoretical policy gradient should be zero (since all actions receive the same reward, there is no learning signal). However, when TIM is introduced, the probabilities of the sampler q and the trainer p no longer cancel out, causing the expected gradient to become non-zero—the trainer begins to mimic the distributional differences of the sampler regardless of whether the reward provides meaningful information. This phenomenon is referred to as gradient drift.

More generally, by decomposing the policy gradient into covariance terms, two parts can be separated: the drift term equals the expected reward multiplied by the expected score of the ratio of sampler to trainer probabilities; the signal term is the covariance between the reward and the score. The drift term does not distinguish between successful and unsuccessful rollouts, relying solely on the mean of the rewards. In online RL, where the sampler is continuously refreshed from the trainer, drift forms a feedback loop: TIM leads to biased updates, which propagate to the sampler, which then produces new biased rollouts—errors accumulate gradually until training collapses.

Score Centering works by subtracting the expected score under the sampler from the token-level scores calculated by the trainer at each prefix position. This subtraction ensures that the centered scores have an expectation of zero under the sampler distribution, thereby completely eliminating the drift term in the covariance decomposition while leaving the signal term unchanged (since subtracting a constant does not affect covariance). Unlike importance sampling, this is an additive correction that avoids high-variance multiplicative weights and requires no clipping or masking to limit extreme values.

In practical implementation, precisely calculating the expected score over the entire vocabulary for the sampler is prohibitively expensive (LLM vocabularies typically exceed 150,000 tokens). The paper employs a top-k approximation: retaining only the 128 tokens with the highest sampler probabilities and approximating the remainder using scaled tail probabilities from the trainer. A key derivation utilizes the property that the expected score is zero under the trainer’s own distribution, allowing the correction term to be computed using only top-k tokens. Experiments show that k=128, and even k=32, achieve effects comparable to full-vocabulary approximation, though the guarantee of drift elimination is only approximate in these implementations.

출처:cameronrwolfe.substack.com

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크

Under mild TIM caused by quantization, Score Centering alone is sufficient to stabilize training. However, under severe staleness (such as in asynchronous training where the sampler synchronizes only once every 64 steps), Score Centering alone is insufficient—because the learning signal is still measured under the sampler distribution, the deviation from the trainer distribution requires correction via importance sampling. The paper suggests combining Score Centering with truncated or masked importance sampling when significant policy differences may arise between the trainer and the sampler.

Cameron Wolfe is a Deep Learning Researcher at Netflix holding a Ph.D., known for his communications that deconstruct AI research papers. He cites Nathan Lambert, noting that many frontier RL laboratories are already using this method, albeit previously unpublished.