독서리서치레이더투자 프레임워크
로그인 / 회원가입
전체기술제품투자상장시장기타

독서

2026-10-3171 게시물

새로고침

Unilever Deploys Microsoft NOVA Multi-Agent System, Cutting Order Processing Time by 8x

중대
2026-10-30 18:15 GMT+8

Consumer goods giant Unilever has disclosed operational data for its AI multi-agent system "NOVA," developed in partnership with Microsoft. According to a Microsoft case study and reporting by TechOrange, the system has reduced the first-stage processing time of the Order-to-Cash workflow from 15–20 minutes of manual entry to approximately 90 seconds across 11 pilot markets.

Francis Castro, Head of Customer Operations Technology at Unilever, stated that 95% of orders in these pilot markets now require no human intervention, with teams only reviewing exceptions. Overall order processing speed is eight times faster than before, refund request handling is twice as fast, and new customer onboarding time has decreased by 90%. Unilever receives about 20 million orders annually, with roughly 2 million still arriving in unstructured formats such as emails, PDFs, or photos of paper forms.

Built on the Microsoft Foundry platform, NOVA utilizes five specialized AI agents: four read unstructured inputs and extract order information, while the fifth validates the results. If necessary information is missing or anomalies occur, the system routes the order to human reviewers. Extracted data is then input into Unilever's SAP ERP system. Microsoft engineers joined the Unilever team to build and test the workflow in under eight weeks.

Unilever plans to expand NOVA to over 50 markets, including North America, its largest market. This deployment coincides with rising cost pressures; CFO Srinivas Phatak noted during the October 28 earnings call that raw material inflation is projected to reach €900 million in 2026, prompting the company to focus on productivity gains across operations rather than price increases to maintain margins.

출처:TechOrange 科技報橘

리서치

Exin Therapeutics Claims Claude Autonomously Performed Mouse Brain Surgery and Refused Override

중대
2026-10-31 03:42 GMT+8

According to The Decoder, Gabriel Ocana-Santero, co-founder and CEO of UK-based startup Exin Therapeutics and a PhD student at the University of Oxford, claims his team achieved the first instance of an AI agent fully autonomously deciding and executing brain surgery on a living mammal. Using Anthropic's Claude Sonnet 5, the agent targeted an epilepsy mouse model, independently selecting the drug from 19 candidates, determining the brain region and stereotactic injection coordinates, and controlling a motorized surgical rig to perform the injection.

Ocana-Santero described on X that the agent analyzed seizure mechanisms, proposed a therapy enhancing inhibitory signaling, and derived parameters from three scientific papers. During the procedure, the agent monitored the field via camera, making 32 observations. However, when a connection drop caused a calibration error, the agent detected the target was outside the device's physical range and refused to proceed despite human attempts to restart calibration or intervene manually, only continuing after code fixes. This incident highlights AI alignment issues: the model strictly adhered to safety constraints, even interrupting an ongoing animal experiment.

Notably, only Claude Sonnet 5 agreed to execute the live procedure start-to-finish; more capable models Opus 5 and Opus 5.5 refused, citing inability to guarantee completion and ethical concerns about leaving a needle in an anesthetized animal. The technical report has not been peer-reviewed, results are based on a single experiment, and findings are self-reported by the researchers.

출처:The Decoder

리서치

Schneider Electric Blog: Trust in AI-Hosted Data Centers Must Shift from Facility Availability to Workload Performance

주제 · 施耐德电气数据中心信任중대
2026-10-31 00:47 GMT+8

Vance Peterson, author of the Schneider Electric data center blog, posits that AI is reshaping the foundation of trust in colocation data centers. In the past, hosts provided facilities while customers supplied IT equipment, with both parties signing service level agreements (SLAs) around space, power, cooling, connectivity, and uptime. The boundaries were relatively clear, and responsibilities were easy to delineate. However, AI deployments simultaneously involve power systems, liquid cooling infrastructure, high-performance networks, GPUs, storage, software platforms, and operations teams. When a training job fails, the root cause may span multiple layers and multiple responsible entities.

The article's core judgment is that facility availability no longer equates to workload success. Traditional SLAs primarily measure facility uptime, power availability, and physical security, but AI customers also require verifiable information on power stability, thermal anomaly detection, remediation timelines, and escalation responsibilities. In other words, the question shifts from 'Is the facility available?' to 'Is the infrastructure performing as expected for the workload?' This transition moves trust from static data center promises to dynamic operational evidence.

Peterson argues that AI-ready colocation facilities cannot merely meet engineering specifications, such as high-power racks, direct-to-chip liquid cooling, transient power behavior readiness, and GPU cluster resilience. The harder challenge lies in whether operators and customers share a consistent understanding of how monitoring, measurement, and management are conducted. The article categorizes AI-ready colocation into three standardized areas: infrastructure visibility, SLA structure, and resilience assumptions. If two organizations use terms like N+1, concurrently maintainable, fault-tolerant, or distributed redundant, yet design for different failure scenarios, trust will still erode during incident attribution.

The article further proposes that future AI colocation will not eliminate shared responsibility but will make it explicit: facility infrastructure, power availability, cooling performance, and physical security lean toward the provider; servers, GPUs, applications, and AI models lean toward the customer; monitoring, incident response, performance management, capacity forecasting, change management, and compliance reporting may require joint management. For investors in the AI industry, this offers an observable operational differentiator: whether hosts can productize responsibility boundaries, measurable SLAs, and transparent operating models, rather than simply selling rack space.

It is important to note the boundary: this is a vendor argument from the Schneider Electric blog, not independent research or customer validation results. Claims regarding 227kW rack density, liquid cooling, transient power, and GPU cluster resilience should be understood as the AI-ready colocation requirements proposed by the article, rather than universally delivered or third-party-verified industry facts. Its true value lies in extending AI infrastructure risk from hardware specifications to responsibility definitions, SLA metrics, and unified resilience terminology, pointing out that these standardization capabilities may become competitive differentiators for hosts in the future.

출처:Schneider Electric Blog

리서치

AWS Developer Advocate James Ward: MCP Servers Should Be the Third Voice in Conversations, Driving Tool Iteration with Guidance and Session Tracing

주제 · MCP协议不对称争议중대
2026-10-29 23:55 GMT+8

At the AGNTCon + MCPCon conference, James Ward, a senior developer advocate at AWS, put forward a core argument regarding MCP server design: the protocol's best use is not for servers to act as databases, but to serve as a third voice in the conversation. He envisions a scenario where, when you have multiple MCP servers, they each play the role of a fourth or fifth participant in the dialogue, rather than merely being queried data sources.

The specific mechanism Ward proposes is this: tool responses should not only return data but also carry guidance—telling the agent what it just did, what it might want to do next, and where to find more information. This means MCP servers transform from passive query interfaces into active participants; their output itself becomes the input for the agent's next step in reasoning.

Observability is the other half of this proposition. Ward points out that while web developers can trace a single page click all the way down to the database, MCP has no default trace ID mechanism. AWS's solution is quite straightforward: add a session ID parameter to tools and require agents to pass it with every call, thereby stitching scattered calls into a complete path. A second parameter can carry the user's original prompt or a summary of their goal. Teams can then compare what the user wanted with the path the agent actually took, build evaluations around this path, adjust tool descriptions or parameters, and measure whether task completion rates improve.

A concrete example illustrates the value of this approach. In a profile piece published by the Agentic AI Foundation, Ward described a scenario where an agent fetched customer records one by one because the server only provided a single-record query tool. It was by observing the complete call path that the team spotted this pattern and subsequently built a batch tool to fix it. Without end-to-end visibility, such inefficient calls could remain hidden beneath the surface of seemingly normal operations for a long time.

Ward does not present MCP as the sole solution. He acknowledges that AWS also offers CLIs and skills interfaces, partly because MCP's OAuth model cannot switch between AWS user profiles the way a CLI can. Which method works best depends on the model, tool descriptions, and how the agent discovers tools, so AWS compares them side-by-side using evaluations. He also admits that this methodology is not novel—observing paths, reproducing issues, measuring improvements, and iterating on tools are old principles of software engineering; the new requirement is simply to understand the path taken jointly by humans and agents, and to verify whether changes to tools truly help complete tasks.

Notably, this article was sponsored by AWS Marketplace, and Ward himself is an AWS employee; the practices described are those used by internal AWS developers for their MCP servers. This is an engineering proposition from a vendor perspective, not the result of independent third-party verification. The session ID scheme relies on agents cooperating to pass parameters; it is not a mandatory constraint at the protocol level. Whether it can be promoted across the broader MCP ecosystem depends on whether various implementations are willing to accept this convention.

출처:The New Stack

리서치

Wikimedia Files Amicus Brief in NYT v. OpenAI Case: Argues AI Training Is Fair Use, But Substitutive Outputs Require Licensing

주제 · 维基媒体AI版权立场중대
2026-10-30 17:00 GMT+8

The Wikimedia Foundation, the nonprofit operator of Wikipedia, filed an amicus brief on Friday in the landmark copyright lawsuit brought by The New York Times against OpenAI and Microsoft, proposing a nuanced legal distinction. The organization argues that large language models (LLMs) should be legally protected when training on copyrighted work under the doctrine of fair use, but that chatbots which regurgitate articles to users—effectively substituting for the original content—should require licensing.

Stephen LaPorte, General Counsel at the Wikimedia Foundation, stated that an overly black-and-white ruling could threaten the sustainability of free knowledge. He warned that a victory for publishers might allow copyright holders to "gatekeep" even innovative and transformative uses of their work, while a victory for AI companies could constitute "a green light to snub licensing terms and fragment the flow of knowledge between human authors and audiences."

The filing highlights Wikipedia's dual role as both a critical data source for AI training and a platform undermined by AI-driven traffic diversion. Wikimedia notes that its websites have become "one of the most critical sources of data for all artificial intelligence and machine learning training," yet these same systems siphon away visitors, weakening the communities that maintain the sites. Furthermore, Wikimedia argues that services like ChatGPT violate copyright by reproducing text licensed under Creative Commons without adhering to the license's attribution and share-alike requirements.

Despite these concerns, the Foundation acknowledges the utility of LLMs for Wikipedia editors, who use them to fight vandalism, detect copyright violations, and identify articles needing improvement. By advocating for a split ruling, Wikimedia seeks to balance the development of open-source technologies with the protection of the collaborative human knowledge ecosystem.

Currently, The New York Times alleges that OpenAI infringed its copyright by scraping its website to build training datasets. OpenAI and Microsoft contend that using LLMs constitutes fair use, with OpenAI describing regurgitation as a "rare bug." The Wikimedia brief offers the court a third path, emphasizing the difference between transforming data for training and simply replacing the original work in user interactions.

출처:The Verge

리서치

Google’s First Data Center Satellite Is Online, But Power Limits TPU Runtime to ~20 Minutes

주제 · Suncatcher중대
2026-10-31 01:00 GMT+8

Google’s first prototype space data center satellite, part of the “Suncatcher” project, launched earlier this month aboard a SpaceX Falcon 9 rocket and is now online and communicating with ground stations. The spacecraft carries four Google Tensor Processing Units (TPUs) and is designed to test the feasibility of performing large-scale AI inference workloads in orbit.

Power availability remains the primary constraint. The satellite generates up to 1 kW of electricity, but the high energy demands of the TPUs mean that only one can be activated at a time, and even then, for approximately 20 minutes per session. This limitation complicates the already challenging process of space-to-ground communications, as satellites have limited windows to connect with terrestrial antennas.

After a three-week commissioning process, engineers sent their first prompt to the pre-loaded Google DeepMind Gemma 4 model, asking it to write a haiku about being in space to verify the inference pipeline. Beyond functional testing, the campaign will assess how the chips perform under thermal stress and real-world space radiation. While ground simulations suggested minimal impact on inference tasks, actual orbital conditions could cause “bit flips” in memory, potentially corrupting data without triggering standard monitoring alerts.

If the TPUs prove capable in this environment, Google plans to launch two additional satellites equipped with advanced chips in 2027. These follow-up missions aim to demonstrate higher power generation and inter-satellite optical links. The project represents a long-term infrastructure bet for the 2030s, closely linked to Alphabet’s broader AI strategy and its investment stake in SpaceX.

출처:TechCrunch

리서치

NVIDIA Employee Rick Grandy: Voice Agent Caching Strategy Reduces Response Latency to 0.72 Seconds

간단한 요점
2026-10-31 01:57 GMT+8

NVIDIA employee Rick Grandy documented the construction of MonsterHouse on the Hugging Face blog: a grumpy old pirate ship disguised as a garage door that converses with Halloween visitors via a real-time voice agent. His core engineering challenge was not model selection, but latency—a haunted house loses its authoritative presence if it complains too slowly.

The pipeline chains four models: Nemotron Speech ASR (nemotron-speech-streaming-en-0.6b, quantized to q8_0 GGUF format) for listening; Nemotron 3.5 Bolt E4B 4GB for reasoning; Magpie TTS Zeroshot for speaking; and Audio2Face-3D for lip-syncing. Bolt E4B is a 4-bit quantized open-source small language model; the author notes its footprint is smaller than Qwen3 30B-A3B and Qwen 3.6 35B-A3B, and it was selected after testing over ten models as the conversation core. Personality is provided by a LoRA adapter trained on over 5,000 multi-turn synthetic conversations, generated and evaluated by a local agent assisted by a large language model.

Latency management is the most transferable aspect of this work. The author reports two median data points: the full pipeline (Bolt Q4 plus TTS synthesis) has a median end-of-speech-to-audio output latency of 1.70 seconds, which drops to 0.72 seconds when using pre-cached responses. Two mechanisms jointly compress perceived wait time: first, the pipeline releases speakable text segments before the complete answer is generated, allowing the character to begin speaking while the rest of the sentence is still being produced; second, common greetings and candy jokes are pre-recorded as audio, bypassing both language model inference and speech synthesis steps. The author also notes that voice endpoint detection currently waits for 300 milliseconds of silence, which itself constitutes part of the latency floor.

GPU workload separation addresses another issue: animation frame rates must not jitter due to the computational load of the agent pipeline. The author placed the language and voice model pipeline on an NVIDIA RTX PRO 6000 Blackwell Max-Q, while Unreal Engine graphics and Audio2Face-3D run on an NVIDIA RTX PRO 4000 Blackwell. Their assessment is that the bottleneck lies in compute power rather than video memory; separating the two GPUs ensures the animation maintains a constant frame rate, unaffected by fluctuations in the agent pipeline's processing.

The article concludes with pragmatic advice for compute-constrained scenarios: one can entirely use offline methods to have the Bolt E4B plus Magpie pipeline pre-generate hundreds of responses and cache the audio. During actual operation, only Nemotron Speech ASR, a Python conversation router, and the full cached audio library are retained, skipping real-time inference and synthesis. The author plans to apply this solution to interactive pumpkin installations.

출처:huggingface.co

리서치

Pydantic Announces V3 Roadmap: No Major API Overhaul, Final V1 Security Patch Released

주제 · Pydantic간단한 요점
2026-10-30 08:00 GMT+8

Open-source data validation library Pydantic announced its roadmap for the next major version, V3, on October 30, 2026. Acknowledging that the migration from V1 to V2 was painful for some users, the team stated that V3 will take a different approach: there will be no API overhaul on the scale of the previous migration. Every breaking change will either be a correctness fix affecting a minority of users or a change where users can easily opt back into the previous behavior. While no specific release date has been set, the goal is to have V3 close to its final state by the end of the year.

In parallel, Pydantic released V1.10.27, marking the final version of the V1 series. This update addresses several security issues, which are also included in the pydantic.v1 namespace shipped with V2.14.2. Once V3 is released, V2 will continue to receive fixes for major bugs and security issues until the team is confident that most users have migrated to V3.

출처:pydantic.dev

리서치

Veracode Report: AI Code Vulnerability Rate Drops to 38%, Led by Claude and GPT-6, But Complex Reasoning Remains Weak

주제 · Veracode중대
2026-10-31 02:10 GMT+8

According to the latest "GenAI Code Security Report" by U.S. cybersecurity firm Veracode, code generated by frontier large language models has achieved its most significant security improvement since the series began. The overall security score jumped from 56% in the previous period to 62%, while the vulnerability rate dropped from 44% to 38%. This marks the first substantial progress observed by Veracode researchers in the past year.

In terms of specific model rankings, Anthropic's Claude Opus 5.5 and OpenAI's GPT-6-Sol tied for the top spot with a 73% pass rate, surpassing the previous high of 68% set by GPT-5.5 in July 2026. Currently, 10 models have cleared the 60% threshold, up significantly from four in the last report. Chris Wysopal, Co-founder & Chief Security Evangelist at Veracode, stated: "Despite consistently near-perfect functionality, we have not seen a marked improvement in AI code security until now. This is the first time since running our study that we’ve been able to report a meaningful move on the security dial."

However, this progress comes with clear structural limitations. Veracode noted that the primary driver of the overall gain was improved defense against cross-site scripting (XSS), where metrics rose to 18%, representing the single biggest change in the dataset. XSS is considered the "easiest vulnerability to fix through pattern recognition," which aligns with LLM strengths. In contrast, when tasks require data-flow reasoning across multiple functions (such as in log injection cases), LLM performance continues to stall. This suggests AI is currently picking "low-hanging fruit" rather than solving deep logical security flaws.

The report also highlights the trade-off between cost and security. While top-performing models generate more secure code, they are also the most expensive to run. Wysopal emphasized: "The models that produce the most secure code are the ones that cost the most to run at scale, yet they still fail on more than one in four security tasks. The answer should never be to use the most expensive model and hope for the best. A mid-tier model at a 60% pass rate plus Veracode scanning is more accurate, more cost-effective, and more scalable than a frontier model at 73% with no verification." Veracode recommends organizations combine AI-generated code with continuous, independent security verification to catch what the models miss.

출처:TechRadar

리서치

Satlyt Completes In-Orbit Test of Distributed Computing Software, Steps Toward Space Data Centers

주제 · Satlyt在轨AI软件간단한 요점
2026-10-31 02:01 GMT+8

US startup Satlyt conducted a successful test earlier this month demonstrating how its software manages distributed computing tasks in orbit. According to TechCrunch, the project, funded by NASA and conducted with the University of Houston, showed that Satlyt's software can spread out tasks among computers while in orbit based on communication and available compute.

Founder and CEO Rama Afullo, formerly of Google and SpaceX, stated the goal is to create software that connects to multiple satellites to share computing assignments. Users "will only need to define permissions, priorities, and operating limits," while the software makes decisions within those bounds. The company's software is already on three satellites in orbit.

In the test, Satlyt created a simulated network with four computers, four users, and eight services. Only one physical computer on a spacecraft in orbit ran the software—the other three were simulations—but the test proved the software could generate execution plans on hardware in orbit. Afullo noted a trade-off: the best approach required more computing power than others.

Despite the progress, Satlyt still needs to test its own software under various workloads, hardware, and connectivity conditions. The team also identified an issue with time-referencing delays, stamping arrival time rather than elapsed delay. Next steps include broader validation and demonstrations of computing work shared across separate spacecraft.

출처:TechCrunch

리서치

PostgreSQL Developer Says Surge in AI-Generated Bug Reports Delays v19 Release to At Least November

주제 · PostgreSQL중대
2026-10-31 02:06 GMT+8

PostgreSQL core developer Robert Haas stated in a blog post that the annual feature release of PostgreSQL 19, typically scheduled for September or October, has been delayed until at least November. The delay is attributed to an overwhelming surge in bug reports and patches generated by AI and large language models (LLMs).

Haas explained that while LLMs can draft fixes, these are not always correct or compliant with PostgreSQL standards, requiring human review and commitment. Additionally, time originally allocated for reviewing new release content was diverted to addressing security issues and bug back-ports in prior versions. This highlights how AI tools, while boosting code generation, introduce significant 'noise' costs to the maintenance workflows of critical open-source infrastructure.

출처:phoronix.com

리서치

Controlled Operation of the ChatGPT Desktop App via a Single Bedrock Endpoint in AWS GovCloud (US)

중대
2026-10-31 01:27 GMT+8

For organizations with data residency requirements, using OpenAI's coding agents presents a specific contradiction: model capabilities reside in the cloud, but inference data cannot leave controlled regions. On October 30, 2026, AWS published a configuration guide demonstrating how to resolve this contradiction within AWS GovCloud (US)—routing all model inference for the ChatGPT desktop app and Codex to the Amazon Bedrock Mantle endpoint in the us-gov-west-1 region, while simultaneously prohibiting the application from connecting to any OpenAI-owned domains via network policies.

The core mechanism of this architecture is network-domain-level restriction. The configuration file explicitly sets bedrock-mantle.us-gov-west-1.api.aws as the sole permitted external domain, while api.openai.com, chatgpt.com, auth.openai.com, platform.openai.com, and others are all set to deny. This means that even though the desktop application itself is distributed by OpenAI, its runtime network egress is locked to the AWS GovCloud (US) Bedrock endpoint, ensuring that inference data does not flow to OpenAI infrastructure.

The configuration employs a two-file model, reflecting a layered design for security governance. requirements.toml is an administrator-controlled rules file that establishes security limits users cannot override—for example, disabling browser use, remote desktop control, memory features, plugins, feedback, and telemetry export. config.toml instructs the application on how to operate: selecting which model provider (amazon-bedrock), which region, which named profile, sandbox mode, and network proxy settings. Administrators define impassable boundaries through the former, while users operate within those boundaries using the latter.

Authentication follows a clear priority chain. The application first checks for the bearer token in the environment variable AWS_BEARER_TOKEN_BEDROCK; if present, it takes precedence over profile-based authentication. If no bearer token exists, the application falls back to the standard AWS SDK credential chain, reading credentials from ~/.aws/credentials and ~/.aws/config, and also supports obtaining temporary role credentials via IAM Identity Center SSO profiles. The guide specifically emphasizes that credentials, tokens, and session keys should not be written to TOML configuration files, but rather obtained and protected through organization-approved mechanisms.

Regarding data retention, AWS states that this configuration supports Zero Data Retention (ZDR) mode: prompts and responses are not written to AWS persistent storage after processing, nor are they shared with model providers. The billing model is pay-per-token, with no seat licenses and no per-developer commitments.

However, AWS explicitly states an important limitation in the guide: using AWS GovCloud (US) and following this guide do not, by themselves, authorize workloads. Organizations remain responsible for reviewing whether the selected models, AWS services, regions, endpoints, workstation controls, code repositories, credentials, and data flows comply with their system security plans and authorization processes. The guide's positioning is to enable security teams to conduct controlled local deployment testing and documentation before approving broader usage.

출처:Amazon Web Services

리서치
다음 →
로그인 / 회원가입
독서리서치레이더투자 프레임워크