読書リサーチレーダー投資フレームワーク
ログイン / 新規登録
すべて技術製品投資株式市場その他

読書

2026-11-1392 投稿

更新

c't 3003 Editor Lukas Rumpler: The Security Key for AI-Controlled Smart Homes Lies Not in Cloud vs. Local, but in the Tools You Grant

トピック · MCP智能家居安全重大
2026-11-13 23:02 GMT+8

In his latest video, c't 3003 editor Lukas Rumpler demonstrates a controversial practice: connecting large language models (LLMs) to Home Assistant via the Model Context Protocol (MCP), granting them administrative permissions over smart home systems. His core conclusion is that security does not depend on whether the model runs in the cloud or locally, but rather on which tool permissions you actually grant.

MCP is an open standard that allows Claude, ChatGPT, Gemini, or local programs like LM Studio to issue operational commands to smart home systems through a unified interface. Home Assistant offers two integration paths: an official built-in MCP server and a community project called ha-mcp installed via HACS. The difference in permissions is vast—the official integration can only query status and toggle devices, whereas ha-mcp possesses nearly 90 tools capable of creating automations, writing dashboards, reading logs, making backups, and even installing HACS extensions. Rumpler likens the former to a remote control and the latter to an administrator login.

He used ha-mcp and Claude to build several items to demonstrate their practical value. The most intuitive example is the dashboard: he described wanting an interface for a kitchen iPad that looks like home decor rather than a technical panel. Claude generated a visualization using Amsterdam canal houses as a metaphor—each house represents a room; windows glow when lights are on, doors open when someone is in the room, and the washing machine is depicted as a small boat with a rotating drum during operation. A second dashboard, inspired by Van Gogh's 'The Yellow House,' uses SVG graphics and animations to present the status of various rooms. He admits he could never have written this code by hand.

More practical scenarios involve complex automation and troubleshooting. He describes bedroom lighting logic involving multiple conditions: lights turn off if no one is present for five minutes, restore previous settings if someone returns within two hours, keep specific lamps on between 8 PM and 2 AM, and partially illuminate upon entry while dimming within two minutes. Writing these branch conditions manually is tedious; Claude directly generates the required automations, scenes, and scripts. Troubleshooting is also effective: when lights unexpectedly turned green, Claude checked the logs and found it was caused by his girlfriend accidentally pressing the wrong switch; when a voice file failed to play, Claude identified that German umlauts in the filename were unsupported by the system.

However, Rumpler emphasizes that this setup can cause serious problems. He lists three categories of risk: First, the secret address key for ha-mcp acts as a password; if leaked, an attacker can install arbitrary code on your home network. Second, if the AI can read your calendar, a meeting invitation containing a malicious instruction could be triggered weeks later—security researchers demonstrated last year using Gemini in a real apartment to execute such prompt injection attacks, resulting in blinds opening and water heaters turning on. Third, all household data processed through Claude passes through Anthropic servers.

ソース:c't Magazin

リサーチ

AI Chip Firm Cerebras Reports Record Q3 Revenue of $210.8 Million, Raises Full-Year Outlook

トピック · Cerebras业绩重大
検証済み 2026-11-13 23:32 GMT+8

US-based AI infrastructure company Cerebras Systems (Nasdaq: CBRS) released its financial results for the third quarter ended September 30, 2026, on November 12. The report shows record GAAP total revenue of $210.8 million, up 55% year-over-year. Cloud and other services revenue reached $141.9 million, a 265% increase from the same period last year. Core revenue, which excludes non-cash items like customer warrant amortization, stood at $230.9 million, growing 70% year-over-year.

Despite strong top-line growth, Cerebras remained deeply unprofitable on a GAAP basis, reporting a net loss of $161.2 million and an operating margin of -106%. However, the company maintains a robust liquidity position with $7.8 billion in cash, cash equivalents, restricted cash, and short-term investments, which management states will support its aggressive expansion plans.

On the business front, Cerebras announced commercial wins with two new hyperscalers, including an integration deal with Oracle Cloud Infrastructure expected to go live in 2027. The company has also secured contracts for 700 megawatts of data center capacity for delivery by the end of 2027, positioning itself as one of the largest non-hyperscaler AI clouds.

Based on current pipeline strength and capacity expansion, Cerebras raised its full-year 2026 financial outlook. The company now expects core revenue between $940 million and $960 million, with gross margins in the 42%-44% range. CEO Andrew Feldman projected that revenue would triple in 2027, supported by the production ramp-up of its next-generation CS-4 systems and the development of the upcoming CS-5.

ソース:sec.gov

リサーチ

Former OpenAI Researcher Claims to Find Blackwell Chip Signatures in Six Chinese AI Models

重大
2026-11-13 23:21 GMT+8

AI researcher Will DePue claims that a forensic analysis he conducted has turned up the first technical evidence that leading Chinese AI labs, including DeepSeek, Moonshot AI, MiniMax, and Ant Group, are using Nvidia's export-controlled Blackwell GPUs. DePue played a key role at OpenAI in building models like Sora, GPT-4o, and GPT-5 before founding the nonprofit research group Unemployment Labs. He states that senior researchers at OpenAI and Anthropic have already independently checked and confirmed his method.

According to DePue's analysis, six Chinese models released since April 2026 carry the distinct signature of Nvidia's Blackwell generation, while earlier models such as DeepSeek-V3 and Kimi K2.5 show traces of Nvidia's older Hopper generation. The evidence comes directly from the published language models. Before training starts, the software framework PyTorch fills empty slots in model weights with random numbers generated from a seed value. Because different chip models have different core counts, this filling process produces a characteristic pattern: older Hopper chips use 132 compute cores, while newer Blackwell chips use exactly 148 or 152 cores depending on the model. Although training adjusts these values millions of times, some slots for unused words or symbols keep their initial random values, leaving a faint statistical echo.

DePue's report estimates that the five training runs he examined required about 12,000 Blackwell GPUs combined, equivalent to the computing power of roughly 30,000 older H100 chips. At a unit price of about $70,000, each cluster would cost $200 million to $300 million, totaling more than half a billion dollars for all five systems. While the method strictly shows only which hardware initialized the weights and does not reveal where the GPUs are located or who owns them, it confirms that Chinese companies can access Nvidia's top hardware by renting computing capacity at data centers outside China. This does not necessarily prove physical smuggling or a violation of US export law, but highlights loopholes in current regulations regarding cloud-based compute leasing.

ソース:The Decoder

リサーチ

Lambda et al. Release TurboPairFormer: 1.73x Faster OpenFold3 Training with Reproducible Gradients

重大
2026-11-13 23:00 GMT+8

Lambda, Stevens Institute of Technology, and the OpenFold team released TurboPairFormer on November 13, 2026, an open-source GPU kernel library designed to optimize the training efficiency of the biomolecular structure prediction model OpenFold3. The tool restructures the computationally expensive "triangle attention" operation, achieving up to 1.73x faster training on NVIDIA H100 and H200 GPUs.

According to the official blog post, TurboPairFormer improves performance through three key techniques: first, it recomputes attention probabilities only once during the backward pass, reducing redundant overhead; second, it employs output-residual compensation to capture rounding errors when converting from FP32 to BF16, lowering mean gradient error by 28% to 47%; third, it uses a fixed accumulation order to ensure gradients are bitwise identical across repeated runs, addressing non-determinism caused by floating-point atomic operations.

In tests on 16 H100 GPUs, TurboPairFormer demonstrated the lowest GPU computation time per optimizer step at all tested crop sizes. For instance, at a 768-token crop size, it was 1.73x faster than the original OpenFold3 Triton backend and 1.13x faster than NVIDIA cuEquivariance. Additionally, all gradients were fully reproducible across 600 fixed-input test cases, contributing to more stable large-scale training processes.

ソース:lambda.com

リサーチ

Canadian AI Firm Cohere Completes Acquisition of Cybersecurity Startup AISLE and Integrates Cubic Team

トピック · Cohere重大
2026-11-13 22:00 GMT+8

Canadian artificial intelligence company Cohere announced on November 13, 2026, that it has completed the acquisition of cybersecurity startup AISLE and formally welcomed the team behind AI-powered code review platform cubic. This move marks a significant expansion of Cohere's "sovereign AI" strategy—systems that allow enterprises to deploy AI in private clouds, on-premises, or air-gapped environments while maintaining full control over their data—by integrating autonomous vulnerability discovery and secure software development capabilities to meet the demands of governments and critical infrastructure sectors.

With the closure of the AISLE deal, Cohere has completed four strategic mergers and acquisitions this year, following earlier acquisitions of Aleph Alpha and Reliant AI. AISLE, which specializes in using AI to automatically discover and remediate software vulnerabilities, will form the core of the newly established "Cohere Cyber" business unit. Simultaneously, the cubic team will lead Cohere's secure coding vertical, offering enterprise-grade AI code review and automated remediation services. Together, these additions aim to address the security and compliance challenges organizations face when embedding AI into mission-critical workflows.

To support this new architecture, Cohere has implemented key executive appointments. AISLE co-founder Ondrej Vlcek has been named Senior Vice President of Cohere Cyber, Jaya Baloo has been appointed as the company's first Chief Information Security Officer (CISO), and Stanislav Fort will serve as Vice President of AI Security, leading security efforts within model research. Paul Sanglé-Ferrière, founder and CEO of cubic, will lead his team in accelerating the integration of enterprise-grade coding capabilities into the Cohere platform. Additionally, Cohere plans to make substantial long-term investments in Prague, Czech Republic, establishing the city as its global center for AI cybersecurity and autonomous security systems.

ソース:Cohere

リサーチ

Voice AI Firm ElevenLabs Launches Experimental Team App ElevenAI Combining Notes and Dictation

トピック · ElevenLabs重大
2026-11-13 22:44 GMT+8

Voice AI startup ElevenLabs has launched an experimental application called ElevenAI, which combines meeting note-taking, dictation, and agent integration. Unveiled during its 11/11 Summit in New York, the app is currently in preview and requires users to sign up for early access rather than downloading directly.

According to TechCrunch, ElevenLabs, valued at $22 billion, positions ElevenAI as a product that lives within a team, distinct from productivity tools designed for individuals. The app features standard capabilities such as cross-meeting notes, recording notifications, historical data queries, and live chat assistance. Companies can also integrate ElevenAgent to perform follow-up actions like drafting emails based on meeting notes. Its dictation feature supports over 90 languages and understands industry-specific terminology.

This move places ElevenLabs in a competitive market where Fireflies, Wispr Flow, and Granola have recently released similar dictation or meeting note capabilities. ElevenLabs' strategy leverages its own models to offer enterprise teams an alternative to purchasing separate licenses for notetaker software.

ソース:TechCrunch

リサーチ

vLLM Announces Day-0 Support for Thinking Machines' Inkling 2 Small Multimodal MoE Model, Achieving 630 Tokens/s Per User on GB200

重大
2026-11-12

Open-source inference engine vLLM announced Day-0 support for Thinking Machines' newly released Inkling 2 Small model on November 12, 2026. Inkling 2 Small is a multimodal (text and vision) Mixture-of-Experts (MoE) model with 594 billion total parameters and 17 billion active parameters, supporting a context window of up to 1 million tokens. The model also comes with a native DSpark speculator.

The vLLM team performed deep optimizations targeting NVIDIA Blackwell architecture GPUs (such as the GB200). Through techniques including kernel fusion, programmatic dependent launch (PDL), highly tuned Gluon attention kernels, and sequence parallelism, vLLM achieved significant performance improvements. According to official tests, single-user decoding speed reaches approximately 630 tokens/s at TP4 with an acceptance length of 3.5. Additionally, sequence parallelism optimization reduces the 8K-token prefill step by 22% at TP4 and 26% at TP8, freeing 1.3–1.9 GiB per GPU for KV cache.

Inkling 2 Small adopts new architectural features, including block attention residuals, gated attention, and mixed-precision MoE (NVFP4 for the first 40 layers, MXFP8 for the last 8). vLLM maintains the reference implementation's numerical precision in its integration and verified that accuracy remains consistent with non-speculative runs on benchmarks like NIAH, MMMU-Pro, and HLE when DSpark is enabled. Currently, vLLM also supports LoRA fine-tuning and reinforcement learning training for this model.

ソース:vllm.ai

リサーチ

Google brings standalone Gemini app to Apple CarPlay with voice-only limits

クイックテイク
2026-11-13 22:22 GMT+8

Google has introduced its artificial intelligence assistant, Gemini, as a standalone application within Apple's CarPlay system. Previously, Gemini was only accessible as an embedded feature within the Google Maps app on CarPlay. This new capability is included in version 1.2026.4470202 of the iPhone Gemini app, which activates the feature upon update.

Due to strict Apple regulations regarding driver distraction, the Gemini app on CarPlay operates primarily as a voice assistant, suppressing colorful images and long text outputs. Users can start conversations on their iPhones and continue them in the car via the sidebar. The app supports natural language interaction, allowing users to interrupt, pause, or resume dialogues to manage driving conditions.

According to German tech outlet Mac & i, Google states that the design aims to make Gemini a "conversation-ready companion while driving," minimizing distractions. Beyond basic queries and schedule checks, authorized users can allow Gemini to access data from Gmail and Google Calendar. The rollout is currently gradual, so some users may not yet see the option.

ソース:Mac & i

リサーチ

Tomasz Tunguz: Nvidia's Three AI Capital Channels and the 112% Cash Flow Overdraft in the Last Quarter

トピック · 英伟达AI资本管道構造的
2026-11-12 08:00 GMT+8

Tomasz Tunguz, Partner at Theory Ventures and former Google product manager, published a structured industry analysis on the day Jensen Huang resigned. His core argument is that over the past four years, Nvidia has acted as the 'Central Bank of AI,' and Huang's departure exposes this role to systemic tightening risks. This analogy is not mere rhetorical exaggeration; he provides specific figures and compositions for three capital channels, along with a triggering concern: last quarter, Nvidia deployed 112% of its operating cash flow into equity and M&A.

The scale of the three channels is as follows: $42 billion in M&A and technology acquisitions over the past two years (including Groq's $20 billion LPU technology license and engineering acqui-hire, Hugging Face's $12.93 billion acquisition, Poolside's $6 billion Model Factory license, etc.); $61 billion in cash-cost-based equity investments covering 229 private companies and 13 public companies (with disclosed stakes in OpenAI reaching $30 billion); and $164.5 billion in lease guarantees and cloud capacity backstop commitments (including up to $105 billion in guarantees for OpenAI’s Ohio data center leases and $6.3 billion in capacity guarantees for CoreWeave). Additionally, there are $63.1 billion in customer accounts receivable, with payment terms extended to as long as one full year.

Tunguz points out that in the last quarter (FY27 Q2), Nvidia's operating cash flow was $24.08 billion, while the cash deployed into equity and M&A was $26.98 billion, a ratio of 112%. He traces data across eight quarters: rising gradually from 4% in FY25 Q3 to 72% in FY26 Q4, and reaching 112% in FY27 Q2. Over the past four quarters combined, Nvidia generated $134.4 billion in operating cash flow but invested $75.8 billion into equity and M&A. This means the deployment pace has exceeded the organic cash generation capacity of a single quarter, requiring the consumption of existing cash reserves or debt financing to maintain. Under Huang's leadership, this unconventional deployment was viewed as a strategic bet driven by personal conviction and relationships; however, Tunguz believes that facing stricter board scrutiny, a more conservative successor might compress these expenditures.

This is precisely where the 'Central Bank' analogy lies in its mechanism. Tunguz writes that if Nvidia tightens equity checks, payment terms, or lease guarantees, AI construction will slow down, just as the broader economy slows when the Federal Reserve tightens credit. He specifically emphasizes that these deals run on Huang’s personal conviction and relationships—Nvidia simultaneously acts as chip supplier, lead venture investor, anchor cloud customer, and lease guarantor across more than 200 startups. What sustains these multiple roles is not contract terms, but trust among founders. A successor must maintain these founder partnerships without the co-founder who built them.

Tunguz’s argument has a notable boundary. The 112% figure is a single-quarter peak rather than a sustained rhythm—the average deployment rate over four quarters is approximately 56%, and of Groq's $20 billion, only $2.94 billion fell in Q2 FY27 (the remaining $13 billion was paid in the 72% quarter of Q4 FY26). More importantly, Tunguz himself lists two alternative paths: First, custom AI ASICs already accounted for 28% of AI server shipments and 15% to 20% of accelerator revenue in 2026, with Google TPUs holding 65% of the custom silicon market, and 'leadership turbulence at a single-source supplier accelerates this shift.' Second, Meta, Google, and at least eight US startups that have raised over $100 million continue to train open-weight models, and they will now have to carry more of the American open-source mantle themselves. Furthermore, the $164.5 billion in guarantees represents 'contingent liabilities that do not appear on the balance sheet unless triggered,' but require maintaining substantial liquidity reserves. Therefore, a more accurate phrasing of the 'Central Bank tightening' narrative is: Nvidia’s capital channels serve as a liquidity supplement mechanism for the AI ecosystem; tightening them will raise financing costs for mid-sized operators dependent on its credit endorsement and slow project progress, but it will not eliminate the capital expenditure needs of hyperscale cloud service providers themselves.

ソース:Tomasz Tunguz

リサーチ

Project Claims No Reverse Engineering, Only Adherence to Public Standards: AI Agents Release 16 Software Clones in Weeks

トピック · AI清洁室软件克隆重大
2026-11-13 14:00 GMT+8

By late September 2026, developer Brandon Thomas had released open-source clones of several Adobe Creative Cloud applications in a very short time, including alternatives to Photoshop, Lightroom, and Premiere Pro. A few days later, clones of Microsoft Word, Excel, and PowerPoint also appeared. By mid-November, the project included 16 programs, all written in Rust, all marked as incomplete, yet many already functional.

Ulrich Wolf of iX Magazine reported on this phenomenon, highlighting the development method claimed by the project: no reverse engineering whatsoever. All clones position themselves as clean-room reimplementations—starting from observable application behavior and comparing results against publicly available specifications, such as ISO 32000 (PDF format), ECMA-376 (Office file formats), or Adobe's PSD file description.

This methodological choice is critical for AI-assisted development. Reverse engineering requires understanding a massive, undocumented codebase, which is nearly impossible for current AI agents. The clean-room approach reframes the problem as specification compliance: given an input, what should the output behavior be? Given a file format, what should the parsing result be? These are testable, decomposable tasks with clear success criteria—exactly the strengths of AI agents and automated testing pipelines.

The article's subtitle mentions that the synergy between AI agents and CI pipelines is the source of this speed. Although the full technical architecture explanation is behind a paywall, public information allows us to infer its working mode: AI agents generate implementation code for specific functional modules, while the CI pipeline automatically runs tests against public specifications; failure results are fed back to the agents for correction. This loop shifts the human developer's role from writing code to defining tests and reviewing results.

Boundaries that must be preserved: the legal safety of the clean-room method is not automatically guaranteed. The project states it does not perform reverse engineering, but external observers cannot confirm this solely from the code surface. Additionally, all 16 programs are marked as incomplete; actual production usability, long-term maintenance costs, and the ability to synchronize with upstream specification changes have not yet been independently verified.

A transferable insight for developers: when target behavior can be precisely described by public specifications, the combination of AI agents and automated testing can compress development cycles to a matter of weeks. This may not apply to all software—areas lacking clear specifications or where specifications themselves are ambiguous remain difficult—but it provides a replicable path for file processing, format conversion, and functional modules with public standards.

ソース:iX Magazin

リサーチ

Noah Smith and Brad DeLong: The U.S. Needs Land and Loans, Not Workers, for Reindustrialization

トピック · 美国再工业化瓶颈重大
2026-11-13 15:10 GMT+8

Economist Noah Smith and University of California, Berkeley economics professor Brad DeLong discussed whether and how the United States should advance reindustrialization in the latest episode of the Hexapodia podcast. Their conclusion defied most expectations: the bottleneck lies not with workers, but with two institutional constraints—land use permitting and bank financing.

They first dismissed the path of economic equality among three common justifications for reindustrialization. Noah Smith and Brad DeLong argued that revitalizing the American middle class by creating high-paying blue-collar manufacturing jobs is unworkable because AI-driven robotics will make most manufacturing highly automated; competing with robot hands using human hands in factories is no longer viable. This means the era of mass semi-skilled blue-collar factory employment has definitively ended.

However, this does not mean manufacturing cannot create jobs. They pointed out that income generated by automated factories creates a multiplier effect locally, driving local service sector employment. This requires that the operators of these factories actually reside near the factories so they can spend their money locally.

With the path of economic equality ruled out, geopolitics and military security become the primary drivers of reindustrialization. They cited an example that is not necessarily optimistic but is significant: Ukraine is the only known country to have successfully achieved reindustrialization, having built a strong drone and robotics industry out of necessity to resist the Russian invasion.

Noah Smith and Brad DeLong ultimately attributed the lack of reindustrialization in the United States to two missing elements: land use permits and bank loans. Regarding land, a permitting reform bill currently moving through the U.S. Congress attempts to address this issue. Regarding banking, they consider this an area where the United States truly lags behind other countries.

Noah Smith proposed a specific plan: utilize America’s sluggish megabanks as pools of capital for reindustrialization by granting them selective exemptions from the Dodd-Frank restrictions enacted after the 2008 financial crisis. He stated that bank financing appears important for manufacturing, but promised to elaborate on the specific mechanisms in subsequent articles.

It is worth noting that Noah Smith’s proposal is currently a suggestion rather than implemented policy, and he personally acknowledged that detailed arguments regarding bank financing mechanisms are incomplete. The logical chain linking Dodd-Frank exemptions to manufacturing financing, along with the potential systemic risks involved, still requires more comprehensive argumentation to support it.

ソース:noahpinion.blog

リサーチ

Amazon Deploys Over 1 Million AI Robots: Severe Injuries Down 40%, But Repetitive Strain Up 77%

トピック · 亚马逊仓库机器人安全争议重大
2026-11-13 10:11 GMT+8

Global logistics giant Amazon currently deploys over 1 million robots in its automated warehouses, coordinated by systems like Sequoia and the DeepFleet AI model. This scale aims to boost package processing efficiency and relieve human workers from the most physically demanding tasks. However, independent research and company statements present a complex picture regarding the specific impact of this human-machine collaboration on worker safety.

According to a study published in the ILR Review in April 2025, severe injury cases in Amazon's robotic warehouses decreased by 40% compared to traditional logistics centers. This suggests that assigning high-risk tasks, such as retrieving items from high shelves or handling heavy loads, to robotic arms like Vulcan has indeed yielded safety benefits. Yet, the same study found that minor physical injuries caused by "repetitive motions" among human employees increased by 77% in automated warehouses. Experts attribute this rise in minor injuries to the fast-paced work rhythm driven by robots, which forces employees to perform quick, repetitive actions to keep up with the system.

In response, an Amazon spokesperson emphasized that since 2019, the company's "global recordable incident rate" has improved by 43%, while the "global lost-time incident rate," which covers the most severe injuries requiring time off, has improved by 70%. Amazon rebutted reports from labor groups and the Strategic Organizing Center (SOC), arguing they contain fallacies and rely on old data from before the deployment of new robot types. Joan Moriarty, SOC's Director of Strategic Research and Campaigns, noted that Amazon tends to highlight its latest robotics technology while avoiding discussion of the tens of thousands of human employees injured due to the company's obsession with maximizing speed and minimizing costs.

Furthermore, a 2024 investigation by the U.S. Senate Committee on Health, Education, Labor, and Pensions (HELP) stated directly that Amazon's internal data shows a correlation between increased robot productivity and heightened stress levels for human employees, leading to higher injury rates. Although advanced robotic fulfillment centers typically require 30% more staff for maintenance and engineering roles with higher pay, The New York Times previously reported that as automation expands, Amazon expects its demand for talent in the U.S. to decrease by over 160,000 people by 2027 and by 600,000 by 2033.

ソース:TechOrange 科技報橘

リサーチ
次へ →
ログイン / 新規登録
読書リサーチレーダー投資フレームワーク

His proposed layered defense scheme is as follows: Use the official integration's Assist address for daily use (exposing only devices authorized for the voice assistant, unable to create automations or install extensions); temporarily enable ha-mcp only when construction is needed and immediately disable access afterward; disable dangerous tools unnecessary for the task (installing HACS packages, restarting, deleting) in the ha-mcp panel; set all changes in the Claude connector to require step-by-step approval, allowing each click only once rather than permanently; and simultaneously disable Claude's web search and email functions to prevent the AI from accessing external content and private data simultaneously.

For local model alternatives, Rumpler performed a direct comparison on the same task: instructing both Claude and Qwen3.8-27b (running on an M2 Max MacBook with 64GB RAM) to build an automation to turn off a spherical bedroom light after two hours. Claude completed the task in under a minute; Qwen took 30 minutes across three attempts, with single reasoning pauses exceeding four minutes, struggling for minutes solely over whether the duration should be formatted as 02:00:00 or hours: 2. Later, he asked Qwen to troubleshoot a light that wouldn't turn on. The model did find a logical error Rumpler admitted he would never have discovered independently (the automation checked if the light was off before it had actually turned off, so the restoration condition never triggered). However, its proposed fix introduced new issues—the light would jump to maximum brightness every time it restored, ignoring previous dimming states.

Rumpler explains that local models are slow because Home Assistant returns massive amounts of data at every step, which the model must read before proceeding; Mac inference chips are far slower than dedicated GPUs when processing large context inputs. His conclusion is: use cloud models for building complex content, and local models for simple operations and privacy-sensitive scenarios, but regardless of the choice, human review cannot be skipped. AI handled the writing for him, but not the verification.