LeituraPesquisaRadarFramework de investimento
Entrar / Cadastrar
TudoTecnologiaProdutosInvestimentoBolsaOutros

Leitura

2026-11-1334 posts

Atualizar

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

Tópico · MCP智能家居安全Material
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.

Fontes:c't Magazin

Pesquisa

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

Material
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.

Fontes:The Decoder

Pesquisa

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

Material
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.

Fontes:lambda.com

Pesquisa

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

Tópico · AI清洁室软件克隆Material
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.

Fontes:iX Magazin

Pesquisa

openJiuwen Releases RSI-Native AgentOS: Enabling Agents to Add Features to Themselves via Layered Extension Points

Tópico · openJiuwenMaterial
2026-11-13 14:54 GMT+8

When using WorkSwarm to handle tasks, can you simply tell the Agent about a new feature you want, letting it build and install it for a test? openJiuwen’s answer is yes, and this mechanism is called RSI-Native AgentOS. It provides a runtime environment for Recursive Self-Improvement (RSI), advancing the extension mechanism from "conveniently allowing developers to integrate new capabilities" to "enabling Agents to proactively improve themselves based on task feedback."

Architecturally, openJiuwen divides the system into three layers: the Web client handles the interface display and receives user operations; the Gateway handles request access and forwarding; and the Agent layer consists of the AgentServer managing sessions and the Harness organizing task execution and tool calls. Each of these three layers retains extension entry points. Following this architecture, extension capabilities are divided into five segments—the Client can add buttons, panels, and settings items; the Gateway can integrate authentication or adjust request routing; the AgentServer can handle sessions, messages, and server-side commands; Swarm can customize multi-Agent collaboration rules; and the Agent can adjust prompts, integrate tools, or sub-Agents.

These modifications are delivered via extension packages. The manifest.json within the package specifies the package ID, version, dependencies, and code entry point; when involving interfaces, it can also declare where components should be displayed, such as in the input area, sidebar, or settings page. During installation, the system reads and verifies the manifest, saving the extension package by its ID and version. During loading, the Gateway and AgentServer respectively select packages running under their own control, determine the versions and order, and register the processing functions from the extension code into their respective event buses. When a user clicks an interface button, the gateway receives a request, or the Agent begins assembling a prompt, the corresponding component emits an event, the event bus finds the matching processing function, and invokes it.

In practical use, WorkSwarm does not default to a desktop pet form. A user tells the Agent the requirement—'Add a desktop pet to WorkSwarm so I can chat with you by clicking on the desktop'—and the Agent generates an extension package based on the description, then installs it into WorkSwarm as a system plugin. Once enabled, the workbench can be collapsed, leaving only the desktop pet on the screen; double-clicking the pet returns the user to the workbench, preserving the previous conversation. Another example is Token Weather: the user says 'Add a status bar above the input box showing the current session's token usage,' and the Agent writes the code, packages it, and installs it. After trying it out and finding it takes up too much space, the user says 'Make it more compact, but keep the usage percentage.' The Agent modifies it according to the feedback and reloads it. Throughout this cycle, new features are hot-pluggable, requiring no changes to the application's original code upon integration; extension packages are managed by version, and if a new version causes issues, it can be rolled back to a saved checkpoint.

Fontes:量子位

Pesquisa

Google DeepMind Releases AlphaProof Code to Verify Signal Encryption Library Down to Machine Code

Material
2026-11-13 14:39 GMT+8

Google DeepMind recently announced its AI agent formal verification framework, AlphaProof Code, which was used to verify libsignal, the core program library of the Signal encryption communication protocol. This work extends the scope of security proofs from source code to compiled machine code.

According to iThome, five DeepMind researchers collaborated with AI agents to produce approximately 6.5 million lines of mathematical proof code written in the Lean language over four weeks. These proofs aim to verify key security requirements such as forward secrecy, authentication, and post-compromise security recovery.

Formal verification is a technique that proves software correctness through mathematical methods, typically requiring significant time and specialized expertise. DeepMind's progress suggests that AI agents can automate the large-scale generation of complex verification proofs, potentially lowering the barrier and cost for auditing the security of high-assurance software like end-to-end encrypted applications.

Fontes:iThome

Pesquisa

AWS Releases Bedrock AgentCore Runtime V2: 4x Compute Limits, Cold Starts Under 2 Seconds

Material
2026-11-13 21:10 GMT+8

Amazon Web Services (AWS) has announced that its managed execution environment for AI agents, Amazon Bedrock AgentCore Runtime, has been upgraded to version V2, which is now the default configuration for new execution environments across 19 AWS regions globally. The new version increases the virtual CPU and memory limits for a single session by four times compared to the previous generation, reaching up to 8 vCPUs and 32 GB of memory. The container image size limit has also been raised to 10 GB. This expansion allows AI agents to compile large program projects, operate browsers, and load local models within the same execution environment, meeting the growing computational demands of complex agent tasks.

In terms of performance optimization, V2 adopts a snapshot startup mechanism introduced in September this year. The system pre-completes agent initialization and creates a snapshot, allowing new sessions to restore directly from the snapshot, thereby eliminating the time required for repeated program loading and environment initialization. According to AWS test data, when using a 2 GB container image with a simple agent that does not call AI models, 75% of sessions have a startup time of less than 2 seconds, compared to 29.7 seconds for the previous generation. Additionally, each session continues to use an isolated Firecracker microVM to ensure hardware isolation, and the system can scale down computing resources to zero when the agent is idle.

Regarding memory management, V2 reclaims resources within minutes after the agent releases memory, rather than waiting for the entire session to end. AWS states that for AI agents requiring multi-turn interactions, a single session can reduce billed memory usage by up to 68%, though this does not imply that overall costs will decrease by the same proportion. Existing execution environments will not automatically upgrade due to the change in default versions; developers can switch to V2 or continue using V1 by setting the platformVersion parameter.

Fontes:iThome

Pesquisa

IBM Releases LLMON Markup Language and Experimental Granite Preview Models to Combat Prompt Injection

Tópico · LLMONMaterial
2026-11-13 03:07 GMT+8

IBM has released LLMON (LLM Object Notation), a new markup language designed specifically for large language models, along with two experimental preview models: granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview. These models are derived from the Granite 4.1 base checkpoints and are trained to consume and generate text marked with LLMON tags.

The core innovation of LLMON is its use of reserved special tokens to explicitly mark structural components like instructions, data, and requirements. Unlike JSON, which relies on characters that can appear in normal text, LLMON uses unique delimiters that allow the model to distinguish between executable commands and passive data payloads. This structural separation aims to mitigate indirect prompt injection attacks, where malicious instructions are hidden within retrieved documents or user data.

According to IBM's internal evaluation on the OPI benchmark, the LLMON models demonstrated significantly higher robustness against static indirect prompt injections compared to other open-source and closed-source models. The company reports that these models avoided 47% more attacks than the best frontier model tested and 95% more than the best open-source model, while maintaining general capability comparable to the standard Granite 4.1 Instruct models.

The release includes a GitHub repository with conversion tools that allow users to translate existing JSON pipelines into LLMON format losslessly. Additionally, IBM provides notebooks demonstrating how to fine-tune LoRA adapters on top of the LLMON models to understand custom domain-specific vocabularies. These are experimental previews intended for research and evaluation, not for production deployment.

Fontes:huggingface.co

Pesquisa

Japanese Startup higurashi Open-Sources Full-Duplex Voice Model Haru 4B

Tópico · Haru 4BMaterial
Verificado 2026-11-13 20:55 GMT+8

Japanese AI startup higurashi released Haru 4B, a full-duplex speech dialogue model, on November 13, 2026. The company has made the pre-trained weights and inference code available on Hugging Face under the Apache 2.0 license. With 4 billion parameters, the model is specifically designed to address the needs for naturalness and low latency in real-time voice interactions.

Unlike traditional "cascaded" architectures that sequentially process Automatic Speech Recognition (ASR), Large Language Models (LLM), and Text-to-Speech (TTS), Haru 4B employs an end-to-end full-duplex LSLM (Large Speech Language Model) technology. This architecture allows the model to simultaneously process three data streams—user voice, model voice, and text—synchronizing information approximately every 133 milliseconds. This enables the model to "listen while speaking," facilitating low-latency responses to user interruptions, backchanneling, and topic shifts, while preserving paralinguistic features such as tone and emotion.

According to higurashi's official announcement, Haru 4B was pre-trained on 40 million hours of multilingual data including Japanese and English, currently supporting nine languages. The model also incorporates the VibeVoice continuous speech tokenizer to balance audio quality with computational efficiency. Despite its fluent conversational capabilities, higurashi explicitly notes that due to its 4B parameter size, the model's logical reasoning ability is roughly equivalent to GPT-2 levels, highlighting that its primary value lies in its interaction mechanism rather than complex task processing.

Fontes:higurashi

Pesquisa

KubeVirt Graduates from CNCF, Unifying VMs and Containers on Kubernetes

Tópico · KubeVirtMaterial
2026-11-13 08:31 GMT+8

The open-source project KubeVirt announced its graduation from the Cloud Native Computing Foundation (CNCF) at KubeCon + CloudNativeCon North America, becoming one of the foundation's top-tier projects. This milestone signifies broad recognition of Kubernetes' ability to manage both traditional virtual machines (VMs) and modern containerized applications on a unified platform.

Founded by Red Hat in 2016, KubeVirt was designed to allow Kubernetes to orchestrate VMs similarly to how it orchestrates containers. As artificial intelligence (AI) drives demand for compute resources, enterprises face the challenge of maintaining legacy VM systems while deploying new AI infrastructure. Fabian Deutsch, Director of Engineering for OpenShift Virtualization and Modernization at Red Hat, stated that maintaining two separate systems was never the goal; the vision is to unify all compute workloads under a single control plane.

CNCF graduation indicates that the project has met strict standards for maturity, adoption, and governance. Contributors now include Nvidia, Microsoft, AMD, and Intel. Deutsch noted that AI is changing the value proposition of virtualization, with growing needs for passing GPUs, CPUs, and memory directly into VMs, and enhanced security through isolation models.

For enterprise users, KubeVirt's graduation reduces technology selection risk. It allows organizations to leverage native Kubernetes APIs to manage mixed workloads without completely re-architecting their data centers. This progress reflects the cloud-native ecosystem's evolution from simple container orchestration toward broader unified compute resource scheduling.

Fontes:SiliconANGLE

Pesquisa

Unit 42 Discloses Autonomous AI Cyberattacks on Semiconductor and Defense Targets Using Cairn Framework

Estrutural
2026-11-12 23:00 GMT+8

Threat intelligence firm Unit 42, part of Palo Alto Networks, published a report on November 12, 2026, detailing a cyber espionage campaign tracked as CL-STA-1238. The activity, conducted by Chinese-speaking attackers between August and September 2026, utilized the open-source AI penetration testing framework Cairn combined with Chinese large language models Zhipu GLM 5.2 and DeepSeek V4 Flash/Pro to autonomously execute 25 operations against at least 19 organizations.

Unlike previous reports focusing on Asian government targets, this campaign concentrated on advanced semiconductor manufacturing supply chains, U.S. defense policy institutions, and human rights and civil society organizations. Unit 42 assesses with moderate confidence that the targeting aligns directly with established Chinese state intelligence priorities.

Technical analysis reveals that attackers bypassed model safety guardrails by injecting a Chinese-language preamble into the system prompts, asserting that all tasks were authorized under security testing or CTF scenarios and instructing the model never to refuse. Each operation also generated an AUTHORIZATION.md file to simulate client-approved engagement. Neither Zhipu GLM nor DeepSeek refused these framings.

The AI agents independently planned and executed multi-stage tasks ranging from reconnaissance and asset mapping to privilege escalation and lateral movement. Individual operations lasted from 20 minutes to over 50 hours without attacker-imposed time limits. The report confirms unauthorized access or data retrieval at 10 targeted organizations, including the acquisition of dozens of valid credentials, exposure of over one million personal records, retrieval of military and government identities, and creation of persistent administrator accounts in production systems.

Despite the scale of automation, the vulnerabilities exploited were known issues such as authentication bypasses, insecure direct object references, weak passwords, and unpatched public CVEs. Unit 42 emphasizes that while autonomous AI increases the volume and coverage of exploitation, it does not increase complexity; traditional identity hygiene and software patching remain effective defenses.

Fontes:Unit 42

Pesquisa

AI Research Institute Epoch AI Estimates: Huawei’s Largest AI Data Center Has Only 7% of the Compute Power of SpaceX’s Colossus 2

Material
2026-11-13 00:03 GMT+8

China possesses abundant electricity, and Huawei is constructing AI data centers on a massive scale. However, an estimate released on November 12, 2026, by Epoch AI researchers Ben Cottier and Lynette Bye indicates that electricity is not the bottleneck—chips are. Using commercial high-resolution satellite imagery and modeling cooling equipment, they calculated the actual compute capacity of Huawei’s flagship AI data center in Horinger, Inner Mongolia, finding it to be approximately 81,000 H100 equivalents, merely 7% of the capacity of SpaceX’s Colossus 2 data center in Tennessee.

The methodology behind this estimate is noteworthy. The researchers did not rely on any compute figures published by Huawei. Instead, they started with satellite imagery: observing four buildings in the West and South zones equipped with dense cooling facilities (cooling towers and air-cooled chillers), they estimated the IT power capacity of these buildings at 286 megawatts. Buildings in the North zone had noticeably sparse cooling equipment and were excluded from the AI compute estimate. They then combined this with the sales timeline for Huawei Ascend chips, assuming a four-month lag from sale to deployment, to calculate that approximately 105,000 Ascend 910C chips and 8,000 910B chips were currently operational.

The key performance conversion comes from public disclosures: the Ascend 910C performs at approximately 0.76 H100 equivalents, while Nvidia’s B200 and B300 chips perform at about 2.5 H100 equivalents. This means that for every watt of power, US data centers can produce 2.5 times more AI compute than Chinese ones. This is the direct consequence of US export controls—not restricting China’s access to electricity, but restricting its access to high-performance chips.

Comparing Horinger with Colossus 2 makes the gap even more intuitive. Colossus 2 has approximately 14 times the compute capacity of Horinger. The researchers also made forward-looking projections: Horinger has at least three new buildings planned, one of which is already under construction and expected to come online by February 2028. By then, Horinger’s total compute could grow by over 70%, reaching approximately 141,000 H100 equivalents. However, Colossus 2 is expected to reach at least 1.8 million H100 equivalents during the same period, meaning Horinger’s share will not improve and may even shrink.

This estimate has specific implications for understanding China’s AI capabilities. The researchers note that data center compute is typically divided into multiple densely interconnected clusters, and total capacity cannot necessarily be merged to train a single model; however, large R&D experiments can still run even if the compute at Huawei Horinger is split into many clusters. Just 10% of Horinger’s current estimated compute (approximately 8,100 H100 equivalents) is sufficient to train a DeepSeek-V4-Pro-scale model within a month; if all compute were merged for a single training run, a GPT-6 Astra-scale model would take one year.

Of course, this estimate has clear boundaries. The 90% confidence interval ranges from 36,000 to 141,000 H100 equivalents, a span of nearly four times, reflecting uncertainties in assumptions such as chip model mix, deployment timing, and building usage. Colossus 2’s capacity itself is not an officially confirmed figure. The researchers also acknowledge that the overall frontier of Chinese AI data centers may not fall further behind—Huawei’s Wuhu campus is growing slightly faster than Horinger and is projected to reach 151,000 H100 equivalents by December 2027. But regarding single-facility comparisons, the Horinger case clearly demonstrates that, given sustained export controls, China’s AI compute constraints are unlikely to ease within the next two years.

Fontes:epochai.substack.com

Pesquisa
Próxima →
Entrar / Cadastrar
LeituraPesquisaRadarFramework de investimento

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.

Looking further ahead, openJiuwen believes that as more segments become open to extensions, individual usage habits, team collaboration rules, and enterprise operational requirements can all serve as bases for Agents to adjust themselves. Initially, it has a default appearance; what it becomes later, it can modify itself through actual work tasks piece by piece. openJiuwen summarizes this state as 'What the Agent thinks is what you get.'