❯ OpenAI expands ChatGPT and Codex into physical engineering with Mazda collaboration
A prototype designed in a weekOpenAI disclosed physical-engineering projects on November 3. Two employees used GPT-6 Astra on a side project to design a device for talking to Codex in under a week, covering its enclosure, circuit board, firmware and software integrations. A manufacturer produced a small prototype batch within a month. OpenAI says manufacturing preparation still required several reviews and revisions because the model inconsistently handled tolerances, assembly details and production constraints.
Turning scans into engineering filesThe company is expanding work with Chip Ganassi Racing to convert three-dimensional part scans into editable computer-aided design files and manufacturing drawings. Engineers can modify dimensions and structures in these files, making the workflow useful when original designs are unavailable. Racing-team engineers review accuracy and manufacturability.
Vehicle simulation joins the effortA new Mazda research collaboration will use airflow and structural simulations to explore vehicle designs that balance styling, cost, aerodynamics and crash safety. OpenAI’s hardware teams already use ChatGPT and Codex for modeling, drawings and simulations; their token usage has increased 23-fold since February. These remain exploratory engineering workflows with continuing review and feedback from collaborators.
▮ SIGNALAI can give engineers more time to compare designs, while production still requires checks against dimensions, materials and assembly constraints.
❯ Apple Intelligence reaches China’s iOS 27.3 developer beta with Qwen integration
Developer testing starts todayApple released the iOS 27.3 developer beta on November 4. ITHome’s testing found the Apple Intelligence and Siri settings available on mainland-China devices. Apple’s announcement schedules testing from today and a public release in December. This update excludes the next-generation Siri AI announced at WWDC 2026.
Qwen inside familiar workflowsUsers can invoke Alibaba’s Qwen for text and image understanding in selected Siri, Writing Tools and Visual Intelligence experiences. Access is available through a Qwen account or Sign in with Apple. Initial features also include rewriting and summaries, photo cleanup, natural-language photo search, and translation in Messages, Phone and FaceTime.
Hardware and language requirements applySupported phones include the iPhone 15 Pro family and compatible iPhone 16 and later models, with device and Siri languages meeting Apple’s requirements. Face-to-face translation needs compatible AirPods, current firmware and an Apple Intelligence-enabled iPhone. Supported earbuds include AirPods 4 with active noise cancellation and AirPods Pro 2 or later. System integration brings assistance into existing apps and reduces switching between standalone AI services.
▮ SIGNALQwen gains an operating-system entry point as Apple connects local model services to frequent writing, image-understanding and translation tasks.
❯ Anthropic uses Claude in custom-chip design as its 15-person team pitches efficiency gains
Claude joins chip engineeringCailian Press’s account of the reporting says Anthropic has told prospective IPO investors that its new 15-person chip team uses Claude for custom-chip engineering and design. The company claims shorter development times than traditional workflows. The Information first reported the investor discussions on November 3; the efficiency claims come from Anthropic’s presentation to investors.
An established hardware effortAnthropic publicly confirmed its custom-chip team in August, aiming to co-design hardware and Claude for faster, more efficient operation. It continues a multi-chip strategy using Amazon, Google, Nvidia and AMD hardware. The new reporting also describes approaches to electronic design automation companies for specialized design data, which could help models understand complex engineering constraints.
Internal infrastructure comes firstAnthropic has announced neither a commercial offering for Claude’s chip-design capabilities nor a production schedule for its custom silicon. Faster iteration could help it obtain hardware suited to its workloads sooner. Serving other chipmakers would also require addressing confidentiality around proprietary designs and data.
▮ SIGNALModel capabilities are reaching into compute development itself: faster custom-chip design could improve the operating conditions for future models.
❯ Anthropic says future models may soon cross its AI R&D acceleration threshold
A new forecast in the IPO filingIn its November 2 amended S-1, Anthropic says internal estimates and expectations for future models indicate it may soon meet its Responsible Scaling Policy threshold for AI-driven acceleration of development. It has not designated any current model as exceeding that threshold and continues to assess the data.
AI helps develop its successorsThe discussion concerns recursive self-improvement: models assist research, coding and engineering, speeding the development of successors that could accelerate research further. Anthropic says successive Claude versions already help build later ones. Its policy explanation distinguishes acceleration in aggregate AI capabilities from an increase in an individual researcher’s productivity.
Measurement can lag progressThe filing says Anthropic has previously recognized significant development acceleration only after it happened because some indicators lag. Even if technically possible, the company does not expect to safely let Claude autonomously design successive frontier-model generations without meaningful human supervision and understanding. It is considering policy changes to keep people in control of development.
▮ SIGNALWhen AI is both a research tool and its subject, capability assessments and safeguards must keep pace with experimental iteration.
❯ DeepSeek showcases Harness Creator Mode for building reusable plugins through chat
Requests become working toolsDeepSeek Harness showcased Creator Mode on November 2 and released seven sets of demonstration prompts in Chinese and English. The Paper reports examples including design workspaces, sandbox games and three-dimensional model visualizers. Users describe their requirements in chat to create tools and interfaces inside Harness.
Plugins persist beyond a conversationDeepSeek’s product page describes Harness as an open-source agent application in worldwide public preview, built around an everything-is-a-plugin architecture. Users can install existing plugins or ask the agent to write, install and inspect new ones in Creator Mode. A floating Pomodoro timer demonstrates building a tool and then adjusting its interface.
Workflows become customizableHarness offers a desktop application and a web interface launched from code. It handles files, data analysis, documents, slides, coding and background tasks. Plugins turn a conversational request into a reusable feature and extend tools, skills and interfaces. The product and its core plugins remain under active preview development.
▮ SIGNALGenerated results can persist as reusable tools, allowing users to accumulate agent capabilities suited to their own work.
❯ DeepSeek V5 plans Ascend pretraining as its Huawei collaboration extends deeper into infrastructure
A pretraining plan from supply-chain sourcesSTAR Market Daily reported on November 3 that multiple supply-chain sources say DeepSeek plans to pretrain V5 on Huawei Ascend chips; The Paper specifically identifies Ascend 950. Ascend previously mainly ran trained DeepSeek models to answer requests. Pretraining develops foundational capabilities from large datasets. DeepSeek has not publicly confirmed V5’s training arrangements or release schedule.
Cooperation reaches the software stackThe collaboration already covers the V4 family. The report describes DeepSeek’s September 30 release of Ascend-oriented infrastructure, including TileLang compilation tools, high-performance computing libraries and distributed communication libraries, alongside Huawei’s supernode networking and low-level programming support. These components map model operations onto chips and coordinate data exchange between them.
Training requires system-wide coordinationLarge pretraining runs keep many chips working together for extended periods, increasing requirements for operation support, communication efficiency and stability. Previously disclosed full-parameter V4-Pro post-training on Ascend further optimized an existing foundation model. The V5 plan concerns foundational pretraining, extending cooperation into the model’s initial development if implemented.
▮ SIGNALAscend adaptation now reaches compilation, computation and communication; training capability depends on sustained coordination across hardware and cluster software.
❯ AMD revenue rises 45% as its data-center business nearly doubles
Data centers drive the increaseAMD reported third-quarter revenue of about $13.38 billion on November 3, up 45% year over year. Data-center revenue rose 97% to approximately $8.5 billion. Its earnings release attributes growth to EPYC server processors and the launch of Instinct MI450 GPUs. Data centers contributed more than 60% of quarterly revenue.
Accelerators and complete systems scale togetherAMD sells GPUs for AI training and inference alongside CPUs for general server computing. It says Anthropic, Meta, Microsoft, OpenAI and Oracle Cloud Infrastructure are running models and workloads on initial Helios rack-scale systems, which combine GPUs, CPUs and interconnects for large model clusters.
A larger delivery quarter aheadAMD forecasts fourth-quarter revenue of $17 billion, plus or minus 3%, with the midpoint about 27% above the third quarter. Adjusted gross margin is expected at roughly 55%, versus 56% in the third quarter. Client and Gaming revenue fell 8%, underscoring the concentration of growth in data centers. Larger AI-system deliveries will test both production scaling and product-mix economics.
▮ SIGNALAI data centers have become AMD’s main growth engine, extending competition into rack-scale delivery and system deployment.
❯ Supermicro revenue jumps 298% and annual sales guidance rises to $72 billion–$82 billion
AI-server demand converts into salesSupermicro reported fiscal first-quarter revenue of $19.95 billion for the period ended September 30, up approximately 298%. Reuters reports that sales exceeded the $16.39 billion LSEG consensus, while adjusted earnings of $1.81 a share also beat estimates. Shares were up more than 9% after hours in the updated report.
Deployable infrastructure is the productSupermicro combines chips, servers, networking, racks and services for customers building AI data centers. Its release reports gross margin of 12.7%, above 9.3% a year earlier but below the previous quarter’s 17.5%. Operating cash outflow was $434 million, showing that large system deliveries continue to absorb working capital despite rapid revenue growth.
Annual guidance moves higherThe company forecasts second-quarter revenue of $19.5 billion–$21.5 billion, above the $17.14 billion estimate cited by Reuters. Annual sales guidance rises from $65 billion–$72 billion to $72 billion–$82 billion. The financial information is preliminary and unaudited, with quarterly closing and review procedures still incomplete.
▮ SIGNALAI infrastructure demand is becoming system deliveries, while expansion also increases server makers’ working-capital requirements.
❯ Fal raises $1.3 billion at a $16 billion valuation to expand generative-media inference
Valuation doubles in six monthsFal’s latest financing totals $1.3 billion, led by Bessemer Venture Partners with participation from Nvidia and others. Its valuation rises from April’s $8 billion to $16 billion. ChainCatcher’s financing report cites sources saying annualized revenue reached $800 million earlier this year. That extrapolates revenue at a particular pace rather than reporting completed full-year sales.
Running models for creative softwareFal hosts and runs image, video and audio models for creators, developers and businesses. Studios, advertising agencies, e-commerce brands and game developers can integrate these models into production pipelines without separately building model-serving and compute-scheduling infrastructure. Inference executes generation requests and turns trained models into continuously available services.
Capacity and research receive fundingFal’s announcement confirms the simultaneous disclosure of Series E and F: Sequoia led the earlier E round, while Bessemer led F with continued Nvidia investment. The rounds total $1.65 billion, including $1.3 billion in the latest F round. Funding supports inference capacity, staffing, research and creative-tool development for 2027.
▮ SIGNALAs image and video models enter production pipelines, inference platforms gain recurring demand for reliable generation capacity.
❯ Salesforce agrees to acquire Cresta to connect voice agents with customer-service workflows
A definitive agreement is signedSalesforce’s November 2 announcement confirms an agreement to acquire Cresta, with closing expected in the first quarter of fiscal 2028, subject to customary conditions including regulatory approvals. Terms were not disclosed. The Stack, citing The Information, reports a $2.7 billion price.
From coaching to handling callsCresta supplies a voice-first customer-experience platform. Its AI conducts real-time conversations, handles interruptions and resolves complex calls. The same platform coaches human representatives, analyzes interactions and tests voice agents. Customers identified by Salesforce include Marriott, Comcast and Alaska Airlines.
Connecting conversations to customer recordsThe integration targets the gap between voice conversations and enterprise customer information. Service Cloud provides customer history and systems for taking action, while Cresta handles conversation understanding, representative assistance and voice tasks. Outcomes can then connect to issue resolution and customer retention. Salesforce says Cresta will continue supporting customers’ existing major contact-center and telephony platforms after closing.
▮ SIGNALVoice-agent value increasingly depends on customer data and the authority to act, bringing conversation and execution into a single service workflow.
❯ Shopify revenue grows 32% as AI commerce adds compute and infrastructure demands
Merchant sales keep expandingShopify’s November 3 results show third-quarter revenue of $3.742 billion, up 32%. Gross merchandise volume rose 29% to $119.145 billion, net income increased 75% to $462 million, and adjusted earnings were $0.49 a share. GMV measures orders facilitated for merchants; Shopify earns revenue from subscriptions and merchant services.
Growth must cover service costsSubscription Solutions generated $836 million and Merchant Solutions $2.906 billion. Free cash flow reached $816 million, a 22% margin. The Wall Street Journal reports that results and holiday guidance also reflect cost pressure from AI compute and infrastructure construction. Revenue was slightly below the $3.75 billion expectation it cited.
Commerce reaches additional interfacesPresident Harley Finkelstein says Shopify is extending commerce across more surfaces and agents. Fourth-quarter guidance calls for low-thirties revenue growth and high-twenties gross-profit growth. As shopping discovery and transactions enter AI conversations, the platform must adapt product data, payments and merchant services to new interfaces while financing the infrastructure behind them.
▮ SIGNALAI shopping interfaces broaden merchant distribution while requiring commerce infrastructure to balance transaction growth with service costs.
❯ Joe Tsai says Qwen exceeds 470 open models and four billion downloads as Alibaba backs openness
Ecosystem figures from WuzhenAlibaba co-founder and chairman Joe Tsai said on November 3 that Qwen now includes more than 470 open models and has passed four billion cumulative downloads worldwide. Yicai’s on-site report quotes him outlining the openness strategy Alibaba adopted three years ago and arguing that open models support wider AI adoption.
Developers build derivative modelsChina Youth Daily’s report also cites more than 330,000 Qwen-derived models and 33 million ModelScope users across 239 countries and regions. Downloads, derivatives and community membership describe distribution, further development and participation respectively. This ecosystem lets developers build regional and industry applications on existing models.
Reducing the starting costTsai argues that openness lowers R&D costs for smaller companies and is particularly useful for less widely spoken languages and emerging markets. Developers can adapt models to local languages, specialist information and workflows, reducing the investment needed to build foundations from scratch. Alibaba’s approach therefore depends on community infrastructure, tools and continued third-party development as well as model releases.
▮ SIGNALDerivative development extends open models’ reach, while the ease of local adaptation shapes how quickly applications spread.
❯ Unitree targets robot intelligence as Wang Xingxing seeks global collaboration
Intelligence becomes the next priorityUnitree founder Wang Xingxing said before the Wuzhen summit opened on November 3 that the company will focus on robot-brain technology and seek global collaboration. DoNews, citing Caiwen, reports that he sees promise in physical AI and intends to pursue technical advances. No specific new model or partnership was announced.
Hardware, models and compute develop togetherCCTime’s republication of NetEase’s report describes three priorities Wang outlined on October 22: expanding robot manufacturing, combining internal and collaborative development of physical-AI models suited to Unitree hardware, and building large compute clusters. His renewed focus on robot intelligence follows that coordinated strategy.
From movement to task understandingRobot intelligence must turn visual, language and other sensor inputs into task decisions and physical control so the same hardware can address more environments and jobs. Larger hardware deployments also increase demand for model adaptation and improvement. Unitree’s stated collaboration approach leaves room for outside algorithm, model and engineering teams to contribute.
▮ SIGNALAs robot hardware scales, software that understands environments and completes tasks determines how many settings the same machine can serve.
❯ Apple’s home robot has largely finished hardware, but software challenges threaten its 2027 target
Next year remains a targetThe Next Web, citing The Information, reports that Apple targets 2027 for its J595 tabletop home robot, with a possible delay to 2028. Much of the hardware was finished about two years ago, but the software is not ready. Apple has announced neither the device nor its release date.
A screen that turns toward peopleThe reported design places an approximately nine-inch screen on a rotating arm. Cameras and sensors would support awareness of surroundings, following users during video calls and home tasks such as turning on lights. The roughly $1,000 price and HomeView Ultra name remain reported possibilities.
Autonomy increases software demandsThe report describes roughly 300 software staff but only about ten robotics specialists, alongside discussions about reducing first-generation features. Compared with a fixed screen, sensing surroundings, deciding when to move and controlling physical components require tighter coordination, making software readiness central to delivery.
▮ SIGNALHome robots bind interaction software to physical action, leaving environmental understanding and behavior control to shape release timing even after hardware is finalized.
❯ DeepSeek restores normal service after yesterday’s web and API disruption
Three services were affectedThe Paper reported on November 3 that DeepSeek experienced partial web and API disruption at around 4:58 p.m. Beijing time. Affected components included V4.1 Flash API, V4 Pro API and chat services. The company subsequently implemented a fix and monitored results while the components still showed degraded performance.
The current status is normalAs of the morning of November 4, DeepSeek’s official status page shows all systems operating normally. Monitoring and degradation were earlier stages of the incident response. The disruption involved both the consumer chat interface and APIs used by developers to integrate models into applications.
Reliability is part of deliveryChat users can retry requests, while applications making continuous calls also depend on timely responses. Separate status reporting for the website, APIs and model components helps users identify affected dependencies and decide whether their workloads require waiting, retrying or another service.
▮ SIGNALAs models enter continuously running applications, availability and response reliability shape the experience alongside reasoning capability.
❯ Binance details the $6.34 billion attack chain involving fake recruitment and compromised internal tools
A prolonged intrusion preceded the theftBinance published initial findings on November 3 about its October 29 security incident. The Block’s account of the report says attackers infected an employee’s computer through fake recruitment and a malicious coding assignment, then used an internal AI tool to reach wallet-related systems. Approximately $6.34 billion in digital assets was affected.
Several entry points reached the walletsThe report describes a months-long chain: an intrusion attempt involving tampered developer tools in June, fraudulent recruiting in September, and access to more engineers’ devices through a compromised computer and internal tools on October 20. Attackers obtained deployment credentials and later accessed wallet-signing services to initiate unauthorized transfers. BleepingComputer also details the connections between devices, credentials and container-deployment privileges.
Rebuilding systems and reopening withdrawalsBinance says it closed the entry points, rebuilt or replaced affected systems, strengthened device controls and commissioned an independent security review. It began reopening withdrawals for additional assets in phases on November 3. The investigation remains preliminary, with the findings showing how links between endpoints, internal tools and privileged production systems widened the consequences of compromise.
▮ SIGNALWhen internal AI tools connect to privileged environments, separation between identities, devices and deployment rights helps determine the reach of an endpoint breach.
❯ SK hynix’s prospective New York plants face equipment-approval delays affecting its investment plans
New York remains under evaluationReuters, published by The Korea Times, reports discussions about two NAND plants in Marcy, New York, through SK hynix subsidiary Solidigm, with approximately $45 billion in initial investment. Solidigm’s public statement says it is evaluating locations worldwide and has made no decision on a specific course.
Existing production supports future financingSources say the company seeks higher Dalian output to generate revenue for U.S. expansion, but licenses for American equipment needed to expand the plant remain unapproved and negotiations continue. NAND retains data without power and supplies SSDs and data-center storage, a different function from AI-processing GPUs or high-bandwidth working memory.
Several conditions shape expansionThe New York scale, location and schedule remain under discussion. A potential eventual investment of approximately $100 billion is also a planning scenario. Revenue and upgrades at existing plants connect with financing for new facilities, making additional storage capacity dependent on equipment, funding and final investment decisions.
▮ SIGNALStorage expansion links existing cash flow, equipment delivery and new investment, with several execution steps between a proposed budget and operating capacity.