LesenForschungRadarAnlageframework
Anmelden / Registrieren
AlleTechnikProdukteInvestierenBörseSonstiges

Lesen

2026-11-0542 Beiträge

Aktualisieren

Dropbox Engineering Team Upgrades AI Coding Agents to Task-Level Delegation, Completing 53 Legacy Test Cleanups in One Week

Thema · DropboxWesentlich
2026-11-05 22:00 GMT+8

The Dropbox engineering team faced specific time pressure: they needed to launch a new experience that had been brewing for three years in November, while complex integration work between the desktop and web versions remained far from complete. Previously, AI tools mainly helped engineers handle individual coding steps—completing a snippet of code or explaining an error message. For this project, they advanced their usage by handing over clearly defined tasks entirely to coding agents, with engineers reviewing the results.

This shift did not happen gradually. During the May Hackathon, the team used AI to assist in building a prototype, exploring two paths for bringing the web experience to the desktop: one involved embedding existing web pages into desktop windows using the Electron framework, and the other involved importing selected web components directly into the desktop interface. AI helped them track dependencies, identify platform-specific assumptions, and test integration schemes. The prototype gave product and design teams something tangible to try, rather than leaving it at paper discussions.

What truly demonstrated the effect of task-level delegation was a technical debt item shelved for two quarters: automated tests for the legacy file list required substantial engineering time to update and had been left untouched. The team broke this cleanup work down into 53 tasks and handed them to coding agents. The agents made changes according to a written plan agreed upon with engineers beforehand, ran relevant builds and tests locally, recorded results, and then submitted them for engineers to review whether to accept the code. All 53 tasks were completed within one week.

The key constraint of this process is that engineers remain responsible for the results. Before each task begins, the engineer and agent jointly confirm a written plan; after the agent completes its work, the engineer reviews the code and verifies that behavior matches expectations. Instruction files stored in the codebase explain the rules agents must follow, providing engineers with a repeatable way to delegate work rather than letting agents decide on their own.

The same pattern was applied to accessibility compliance work. During the launch preparation phase, the team delivered 22 accessibility features assisted by AI development, covering the Home page, All Files, file previews, and Photos modules. These included announcements for screen reader users to track carousel changes, fixes for how keyboard focus moves between controls, and layout adjustments for pages at 200% to 400% zoom. Coding agents also took on writing automated tests for five core modules of the web application, including the file browser users use to browse content.

A notable detail is the agents' ability to proactively discover problems during development. In one instance, an agent found that deleted items in chat history remained visible until the page was refreshed during a trial run; engineers used this to fix the underlying synchronization issue and added tests. This allowed human reviewers to spend more time judging whether the experience was reasonable, while functional issues were covered by automated checks.

Quellen:dropbox.tech

Forschung

YC Startup Discovered Materials Puts $1M Behind Five Unsolved Materials Science Problems to Test AI

Wesentlich
2026-11-05 21:00 GMT+8

Discovered Materials, a San Francisco-based startup that uses AI models to discover new substances for advanced chips, announced on November 5 a $1 million prize challenge for AI models capable of solving five critical unsolved problems in materials science. The company, part of Y Combinator’s spring 2026 batch, aims to address the gap between AI's success in mathematics and its limited progress in complex physical systems.

While frontier AI models have solved outstanding math problems and made strides in life sciences, they have yet to truly plumb the complex physics underlying advanced materials. “It’s more difficult to specify problems clearly in materials science because it’s not like math, where you have a very specific definition and can get to a yes or no,” said Advaith Sridhar, co-founder of Discovered Materials. To define these challenges, the company recruited an advisory group of academics from leading global universities, including Stanford professor Felipe Jornada.

The five problems include developing a theory to explain cuprate superconductivity, creating numerical methods to predict silicon-silicon dioxide interface structures, mapping grain structure to thermal properties, predicting how light changes material behavior, and improving Density Functional Theory (DFT). Sridhar stated that the goal is to establish these problems as “goalposts” for the field, regardless of who solves them—whether it be AI, professors, or other researchers.

Before announcing the challenge, Discovered Materials tested frontier lab models against these problems, giving AI agents internet access and resources to work autonomously for days. The models spent several days on each problem without making much headway. However, Sridhar noted that tracking efforts to solve these problems will show “whether the models are climbing in this domain,” providing a clearer metric for AI's evolving capabilities in scientific discovery.

Quellen:TechCrunch

Forschung

AI Cloud Provider Lambda Announces Plans to Build Data Center Campus in Alabama, Investing $33.1 Billion

Thema · LambdaWesentlich
2026-11-04 23:02 GMT+8

US-based AI cloud company Lambda announced on November 4, 2026, plans to build a large-scale data center campus named "Red Clay Technology Center Campus" in Lowndes County, Alabama. The project, readied for development by Cloverleaf Infrastructure, represents an estimated $33.1 billion investment.

According to Lambda's announcement, the data center campus is expected to create more than 1,000 construction jobs during development and support at least 225 full-time roles upon operation. The project is projected to generate significant economic benefits for local schools and communities, with more than $2.8 billion in local tax revenue through 2049, including over $1.6 billion for Lowndes County Public Schools. Combined state and local tax revenue is projected to reach $3.4 billion.

To address community concerns about infrastructure burdens, Lambda committed to paying the full cost of the power and grid infrastructure required to serve the facility. Through upfront financial commitments, these costs will not be shifted to Alabama Power's existing customers even if project plans change. Both Lambda and Alabama Power have signed the White House Ratepayer Protection Pledge. Additionally, the facility will deploy closed-loop cooling technology to reduce water use and plans to partner with N-Drip to support local farmers in conserving water.

The project is currently in the planning stage, with construction expected to begin in the first quarter of 2027 and take up to five years. Power supply remains pending regulatory approval. Stephen Balaban, Co-founder and CTO at Lambda, stated that the company is committed to building strong relationships with the local community and ensuring resident perspectives inform their approach.

Quellen:lambda.com

Forschung

Bosch Sensortec CEO Stefan Finkbeiner: Edge AI Keeps Raw Data on Device, Privacy Ensured by Architecture

Thema · 博世边缘AI隐私架构Wesentlich
2026-11-05 17:00 GMT+8

In a Leaders Talk interview with eeNews Europe, Bosch Sensortec CEO Stefan Finkbeiner presented a core assessment regarding edge AI and privacy: the advantage of privacy protection stems from architectural design rather than transmitting data to cloud-based processing.

His argument is that when AI algorithms run directly on sensor silicon, physical signals are processed locally and discarded; the device outputs only the specific conclusion requested by the user—for example, "air quality is moderate" or "step count updated"—without raw data ever leaving the device. This contrasts sharply with traditional architectures: in conventional designs, raw sensor data is continuously transmitted to external clouds for algorithmic processing, a process that inherently creates privacy vulnerabilities. Finkbeiner argues that edge AI returns data sovereignty to the user, making on-chip processing the strongest technical safeguard for privacy.

The physical prerequisites supporting this architecture are breakthroughs in two sensing technologies. TMR (Tunnel Magnetoresistance) technology provides ultra-high precision, extremely low noise, and ultra-low power magnetic sensing, enabling indoor navigation in GPS-denied scenarios and head tracking in AR/VR headsets. ASICAP capacitive sensing pushes miniaturization and signal sensitivity to new boundaries, allowing wearable devices to detect micro-gestures and acoustic details with minimal energy consumption. Finkbeiner concludes that when these physical breakthroughs are combined with on-sensor AI processing, real-time response, near-zero latency, battery life, and privacy can be simultaneously guaranteed. He states that all new products feature microcontrollers equipped with edge AI capabilities.

This narrative originates from a manufacturer's CEO describing their own product roadmap, constituting an attributed argument rather than independent verification. Specific performance metrics for TMR and ASICAP, the quantitative definition of "near-zero latency," and the coverage scope of all new products have not been confirmed by third-party testing. While Finkbeiner’s privacy architecture logic—local processing, discarding raw data, and outputting only conclusions—is technically clear, whether actual products strictly adhere to this design depends on specific implementation rather than visionary statements.

However, behind this architectural choice lies a notable design philosophy. Finkbeiner characterizes Bosch Sensortec’s innovation culture as "constructive naivety": engineers are encouraged to try new things and take calculated risks, rather than merely manufacturing components according to customer specification sheets. He believes that in consumer electronics, doing only what is listed in specifications means becoming a commodity supplier, which is already too late. This culture provides fertile ground for the development of edge AI sensors—not waiting for customers to define needs, but proactively demonstrating what can be achieved by combining physical sensing with on-chip intelligence.

Quellen:eeNews Europe

Forschung

GitHub Announces General Availability of Dependabot Support for Agent Package Manager

Thema · GitHub代理包管理器Kurzfassung
2026-11-04 04:24 GMT+8

On November 3, 2026, GitHub announced that its dependency management tool, Dependabot, now officially supports Agent Package Manager (APM). This update allows developers to use Dependabot to automatically create pull requests (PRs) for updating AI agent components managed by APM, including skills, prompts, and instructions.

APM is a tool designed to manage dependencies for AI agents within projects, allowing developers to specify sources and versions for required skills in an apm.yml file. With Dependabot's APM support enabled, the system checks Git dependencies declared in apm.yml and generates update PRs when new versions are available. If an apm.lock.yaml file exists in the project, Dependabot also regenerates this lock file to ensure consistent versions of AI agent components across teams.

The feature is currently generally available on github.com for all plans and is expected to be included in GitHub Enterprise Server 3.24. Note that this support covers version updates only and does not include security vulnerability alerts or fixes.

Quellen:The GitHub Blog

Forschung

Baidu Open-Sources ERNIE 5.5 Flash Base: Claims SOTA Among Same-Size Models with New Sparse Attention

Wesentlich
2026-11-05 15:15 GMT+8

On November 5, Baidu officially open-sourced its latest base model, ERNIE 5.5 Flash Base. The model utilizes a Mixture-of-Experts (MoE) architecture with 280 billion total parameters and 14 billion active parameters per token, trained on 33.5 trillion tokens. It supports context lengths up to 1 million tokens during fine-tuning and inference.

To address high computational costs in long-document analysis and agent tasks, the model introduces Latent-Gated Sparse Attention, combining hybrid sparse attention with low-rank gated attention to reduce gating projection weights by 75%. It also employs Multi-Router Latent MoE to improve the correlation between routing scores and expert utility.

Baidu claims the model achieves state-of-the-art (SOTA) performance among same-size models across benchmarks for knowledge, reasoning, math, code, multilingual capabilities, and long context, surpassing DeepSeek-V4-Flash-Base, Hy3-preview-Base, and MiMo-V2.5-Base. Model weights are available on Hugging Face.

Quellen:量子位

Forschung

Moonshot AI Releases Open Weights for Kimi K3.1 Pro with $20M Commercial Threshold

Thema · 月之暗面Kimi模型系列Wesentlich
2026-11-05 16:34 GMT+8

Chinese AI company Moonshot AI officially released the open weights for its large language model, Kimi K3.1 Pro, on the Hugging Face platform on November 5. Previously, the company had launched the model via API on October 23, promising to release the weights by early November. This release fulfills that promise, allowing developers to locally deploy this Mixture-of-Experts (MoE) model with 2.8 trillion total parameters.

Kimi K3.1 Pro is based on the Kimi K3 architecture released in July, further trained with larger computational resources. The model activates 104 billion parameters and supports a context window of 1 million tokens, handling both text and image inputs. Architecturally, it replaces full attention layers with Gated DeepSeek Sparse Attention (Gated DSA) to optimize long-context processing. Moonshot recommends vLLM, SGLang, and TokenSpeed as inference engines.

According to vendor self-reported data, the model scored 65.7% on the Terminal-Bench 4.0 coding agent test, slightly below Claude Opus 5.5 (66.4%) but above GPT-6 Astra (58.2%). In 3D production tasks, it achieved 92.7% accuracy on the BenchCAD benchmark when using tools, also trailing Opus 5.5 and GPT-6 Astra. Notably, to reduce security risks, Moonshot intentionally removed cyber-attack-related data during training, resulting in significantly lower offensive code generation capabilities compared to unfiltered checkpoints.

The release uses a custom license named "Kimi K3.1-Pro License." This license generally permits free use, reproduction, modification, redistribution, sale, and fine-tuning. However, if enterprises provide the model as a service (Model as a Service) to third parties, and their combined associated companies' total revenue exceeds $20 million in any consecutive 12-month period, they must sign a separate commercial contract with Moonshot AI. Additionally, commercial products with more than 100 million monthly active users or over $20 million in monthly revenue must prominently display "Kimi K3.1-Pro" on their user interface.

Quellen:PC Watch

Forschung

spaCy Author Honnibal Critiques MIT Astra Study: Conflating 'Reading Weights' with 'Modifying Weights' is Misleading

Wesentlich
2026-11-05 15:10 GMT+8

Recently, a study by MIT researchers regarding the Astra model caused a sensation in the AI community. The research claimed that Astra could identify tasks performed by small neural networks simply by observing raw weight numbers, and even modify the weights of the Qwen3.5-0.8B model without using gradients, boosting its math score from 7.9% to 48.1%. Many netizens interpreted this as 'AI starting to read another AI's brain,' implying a unified, deep capability for model comprehension.

However, Matthew Honnibal, author of spaCy and co-founder of AI company Explosion, publicly questioned the presentation of this research. He argued that the study's experimental design resembles a stage magic 'misdirection': magicians may use different techniques to perform the same trick, but audiences often assume there is only one mysterious method. Honnibal pointed out that 'reading weights' and 'modifying weights' in the Astra study are actually two completely different technical paths, and bundling them together easily leads to reader misunderstanding.

After carefully reviewing eight sets of experiments in the paper, Honnibal found that the experiments strictly adhering to the condition of 'observing raw weights only' were limited to networks with very small parameter sizes (between 17K and 172K). In these experiments, Astra indeed did not run Python code or execute the target network, but instead placed the weight floating-point numbers directly into the context for analysis. While this demonstrates impressive pattern recognition, it remains far from truly understanding modern large language models.

In contrast, the most attention-grabbing experiment—fixing Qwen's math capabilities—was conducted entirely differently. When handling the 800-million-parameter Qwen model, Astra did not directly read the raw weights. Instead, it was authorized to use Python programming and allowed up to 50 forward passes. Each pass involved processing batches of practice problems, analyzing and modifying weights based on observed performance. Honnibal emphasized that this looks more like an 'off-field coach' performing engineering debugging by writing code, running programs, and observing feedback, rather than a mystical 'brain-reading' technique.

Despite pointing out the misleading presentation, Honnibal did not dismiss the value of the research itself. He admitted that prompting an AI model to design and execute tasks typically done by human ML engineers—writing code snippets to analyze or even directly modify network weights—is a huge step forward. AI models from a year or two ago would have failed at all these tasks. This indicates that today's top-tier large models possess complex tool-use and reasoning capabilities, allowing them to diagnose and fix other models as 'AI engineers.'

The real issue lies in how the results were packaged. Blurring two distinct capabilities into a single 'superpower' capable of reading and writing AI brains not only exaggerates the actual boundaries of current technology but also reflects the widespread hype in the AI industry. For readers, distinguishing between 'pattern recognition based on raw data' and 'engineering intervention based on code execution' is crucial for accurately assessing the real progress of AI in model interpretability and automated engineering.

Quellen:量子位

Forschung

Google Provides Credit Support for TeraWulf's Justified Data Center Project to Back Anthropic Lease Financing

Wesentlich
2026-11-05 13:00 GMT+8

According to a report by Zhidongxi, US data center developer TeraWulf announced on November 4 that it has signed a credit support agreement with Google. Under the agreement, Google will provide credit support for raising debt financing for TeraWulf's Justified data center project. This move aims to improve project financing conditions and help TeraWulf secure competitive long-term debt funding to advance the construction of three data center buildings in the park.

The Justified data center project is a key vehicle for the massive lease between TeraWulf and AI unicorn Anthropic. In July this year, TeraWulf subsidiary Raylan Data signed a 20-year lease with Anthropic, agreeing to provide approximately 401 megawatts of IT load capacity, expected to generate about $19 billion in contract revenue during the initial term. Since data center construction requires significant upfront investment while rental income begins only after facility delivery, financing capability has become a bottleneck for project implementation.

The three credit support agreements signed correspond to the three buildings in the Justified park respectively. Google's obligation to support the performance of the lease for each building takes effect after the last machine room in that building starts paying rent. If Anthropic defaults on payments or goes bankrupt, Google can pay due rents and replace Anthropic as the lessee according to the agreement, or pay a termination fee and acquire equity under specific conditions. This structure converts future rental flows into current financing credit collateral.

This is not the first time Google has supported TeraWulf's projects. As early as August 2025, Google provided support for $1.8 billion in lease obligations for TeraWulf's collaboration with Fluidstack and received warrants to purchase TeraWulf common stock. Analysts believe that Google's successive guarantees for TeraWulf's customer leases may be intended to secure more compute orders for its self-developed TPU chips, reflecting how tech giants are deeply intervening in supply chain financing through financial tools in AI infrastructure construction.

Quellen:zhidx.com

Forschung

Lai Jie, the First Employee of Tencent's Robotics X Lab: This AI Wave Is an Intellectual Revolution, Not a Robot Revolution

Thema · 星尘智能融资上市Wesentlich
2026-11-05 14:29 GMT+8

When Lai Jie left Tencent in late 2022 to found Astribot, ChatGPT had just been released days earlier. However, his technical foresight went even further back: while conducting data-driven research at Tencent’s Robotics X laboratory, he observed that Transformers were evolving from language to multimodal processing. He reasoned that if Token design was reasonable, all modalities could be aligned in the same space—a crucial step for data-driven robotics.

His core judgment now is that this wave of AI is more of an intellectual revolution than a robot revolution. Companies like Unitree have already demonstrated that robotic physical capabilities can be made sufficiently strong; the next challenge is enhancing intelligence. This implies that robot bodies should be repositioned as learning tools for AI—serving simultaneously as data collectors and application terminals.

This positioning rests on both evolutionary reasoning and experimental evidence. Lai Jie derives design principles from biological evolution: humans have five fingers, and eagles have four toes (two front, two back), primarily because they grasp different things, and what they grasp is determined by their brains. In other words, intelligence determines bodily morphology, not the other way around. A Brief History of Intelligence notes that the emergence of the opposable thumb was a key physical node in human intelligent evolution; only with this change could humans adapt to various tools.

Another piece of evidence comes from his experience at the Tencent lab. They trained industrial robots like KUKA and UR using bartender data, but later purchased an ABB robotic arm separately—not for its industrial reliability, but because its structure resembled a human’s, with degrees of freedom and joints close to those of the human body. The result was that models trained on the ABB arm performed better. Lai Jie concluded that when a robot’s structure approximates the human form, the transfer gap from human data is minimized, leading to better model performance.

Astribot names this philosophy DFAI (Design for AI). Regarding hardware selection, they adopted a cable-driven solution—using high-strength ropes to transmit power, moving motors from the joints to the base. Lai Jie explained that the initial choice of cable drive was not to mimic human tendons, but to find a transmission method with the best force transparency.

What is force transparency? Lai Jie uses a daily example: when a person shakes hands with a robot, either you are forcing it or it is forcing you, with no sense of mutual responsiveness. Many robots, when placing a coffee cup down, release their grip as soon as they judge the cup is near the table, causing it to drop with a clatter. Humans do not act this way; we release our grip only when we feel the table supporting the cup and our hands no longer need to exert force. Cable drive allows robots to perceive changes in end-effector forces, enabling them to learn this tactile-feedback-based intelligent relationship rather than relying on absolute position judgments.

Quellen:爱范儿

Forschung

Japanese Ceramic Giant TOTO Announces ¥60 Billion Investment in New Semiconductor Equipment Component Plant

Thema · TOTO半导体陶瓷扩产Wesentlich
2026-11-05 22:30 GMT+8

Japanese ceramic manufacturer TOTO announced on October 30, 2026, that it will construct a new production facility within its TOTO Fine Ceramics Buzen plant in Buzen City, Fukuoka Prefecture. The facility is dedicated to mass-producing next-generation electrostatic chucks and Aerosol Deposition (AD) components. The project involves an investment of approximately ¥60 billion, with production scheduled to begin in the first half of fiscal year 2029.

Electrostatic chucks are critical components used to secure silicon wafers in semiconductor manufacturing equipment, while AD components are installed in etching process tools. TOTO positions these ceramic materials for semiconductor equipment as a growth segment within its "Ceramics Business." The company states that its plasma-resistant electrostatic chucks and AD components contribute to the multilayering of memory and the miniaturization of logic semiconductors, trends driven by surging demand for AI chips.

The new production building is a three-story steel structure with a construction area of approximately 12,000 square meters and a total floor area of about 36,000 square meters, making it the largest production facility in the company's ceramics business to date. Additionally, TOTO is constructing a sintering building at the same plant to increase electrostatic chuck production, scheduled for completion in January 2027. This ¥60 billion investment is separate from the approximately ¥30 billion capital expenditure plan announced in April 2026 for the Chigasaki, Nakatsu, and Buzen plants, and is described as a "large-scale investment plan for further production enhancement."

Quellen:EE Times Japan

Forschung

webAI Releases TwIL-8B Reasoning Model, Self-Tests Score 98.2 on MATH-500

Thema · TwIL-8BKurzfassung
2026-11-05 06:25 GMT+8

webAI released TwIL-8B on November 4, 2026, a dense decoder-only reasoning model with 8.79 billion parameters. The model is post-trained from IBM's Granite 4.2 8B, aiming to deliver efficient capabilities in mathematics, code generation, and instruction following at a smaller scale.

According to the model card published by webAI, TwIL-8B achieves a score of 98.2 on the MATH-500 benchmark and 81.6 on LiveCodeBench (verbose variant). However, it performs relatively poorly on knowledge-intensive tasks: its GPQA Diamond score is 70.2, which is 11.5 points lower than Qwen3.5-9B, and its MMLU-Pro score is 76.5, also trailing behind the competitor.

The model comes in balanced and verbose variants, supporting a context window of 131,072 tokens. A quantized GGUF version is approximately 4.98 GiB, suitable for CPU or modest-VRAM GPU inference. Notably, all weights are released under a non-commercial license, and the performance metrics are self-reported by webAI without independent third-party verification.

Quellen:MarkTechPost

Forschung
Weiter →
Anmelden / Registrieren
LesenForschungRadarAnlageframework

Integration of cross-platform shared components also benefited from this approach. Web and desktop each use different build tools, login mechanisms, and data retrieval methods. Engineers need to determine which code can be reused and which parts must be handled separately on each end. AI helped them track connections between components, extract reusable parts, write adapter code (connecting shared features to each app's services), and corresponding tests. Engineers reviewed the final code to confirm it worked as intended.

The team also emphasized that customer feedback and engineering review together determined what could be rolled out to more users. All eight initial design partners chose to keep the new experience enabled. In broader gray-scale testing, they tracked usage of familiar features such as upload, preview, search, and sharing, as well as whether AI actions succeeded and whether users turned off the new experience. These signals helped determine priorities, while the coding agent workflow gave engineers a consistent way to implement and verify fixes.

This directly connects to the current data bottleneck in embodied AI. Lai Jie believes that native multimodal large models (which no longer separate LLMs and ViTs) have solved robots’ abstract understanding of the physical environment, but concrete understanding remains missing—internet data and existing collected data lack concepts of size and spatial awareness, possessing only visual awareness. The solution is to provide large models with complete spatial-concrete information through the robot itself, forming a data flywheel.

Lai Jie predicts this flywheel has started turning this year, as GPT also begins training first-person perspective (EGO) data and robot data. He explicitly stated that it is very likely robots will enter households next year.

Regarding the relationship between embodied models and foundational large models, Lai Jie uses a cooking metaphor: You video call your mother, and she guides you to add oil, but you must know how to perform the action yourself, because she cannot pour the oil for you. The embodied model is that part which 'knows how to pour the oil'—it perceives physical parameters (weight, force, grip stability) and uses natural language to ask GPT for solutions when uncertain. He opposes training all tactile information into the foundational large model, arguing it is cost-inefficient and would create communication bandwidth issues.

Lai Jie firmly opposes robots entering factories to screw bolts. His reason is that a robot’s greatest advantage lies in its broad capability, which cannot be confined to a single domain. Just as large models can answer questions on medicine, management, history, and science, robots entering homes should not be limited to single tasks. On industrial cost considerations, he personally prefers using fixed programs for repetitive tasks, noting that bosses might not be willing to pay token fees for every single action.

However, Lai Jie admits that Astribot is far from reaching the level of the iPhone 4. He maps their S1 model to the Apple I, and the T1 to the Apple II, stating they have not yet reached the Macintosh stage. But he believes industry development speed exceeds expectations, and blockbuster robot products at the Macintosh or iPhone 4 level will appear next year.

It must be emphasized that all of the above represents strategic judgments and first-party narratives from Lai Jie as the founder of Astribot. The cost advantages of cable drive, the intelligence gains from force transparency, and the prediction of household entry next year have not been independently verified. His DFAI concept offers a route for embodied intelligence distinct from 'build a strong body first, then install AI,' but whether this route can be realized in engineering and commerce depends on product iterations and market feedback that are not yet public.