Silicon Valley Daily
Xiaomi opened MiMo-V2.6 weights and training resources, while Alibaba outlined a much larger model and data-center expansion plan. Developers also gained new options from Grok, AWS and Cloudflare. California’s seven new data-center laws and Texas’s permit pause put electricity and water approvals alongside chips on the list of expansion constraints. Muse’s growing download estimates arrived with a macOS vulnerability disclosure and Amazon’s refusal to let the agent shop on its site.
Xiaomi releases MiMo-V2.6 models and reinforcement-learning resources
Xiaomi released the multimodal MiMo-V2.6 Pro and Flash models, alongside model weights, a technical report and reinforcement-learning resources. The release gives developers both hosted API access and material for local deployment and further training; Xiaomi’s performance comparisons remain company-reported.
Alibaba outlines giant-model plans and unveils Zhenwu V900
Alibaba CEO Eddie Wu said the company plans to train a model with 5 trillion to 10 trillion parameters. Alibaba also unveiled its Zhenwu V900 accelerator and set a target of roughly 20 GW of cloud data-center capacity by 2032. The model and capacity figures describe future plans, while the chip’s claimed performance gain comes from Alibaba.
Muse macOS flaw could expose access granted to the assistant
Researcher Patrick Wardle disclosed a Muse macOS flaw that could let code already running locally redirect dictation traffic and obtain authentication material. An attacker could then abuse permissions the user had granted Muse; this is not a remote attack requiring no local execution. Techmeme’s updated account says Meta reported issuing a hotfix.
California signs seven laws governing data-center power and water use
Governor Gavin Newsom signed seven bills covering data-center electricity costs, water disclosures and local oversight. The package makes developers responsible for related infrastructure upgrades and sets conditions for environmental-review streamlining. Communities will receive more information about proposed projects before decisions are made.
Texas pauses data-center environmental permits pending audits
Governor Greg Abbott directed Texas’s environmental regulator to halt permits sought by data centers until the grid operator completes its audit. The order extends the state’s earlier pause on new grid connections into environmental approvals and links future decisions to electricity and water-use information. It does not order existing facilities to shut down.
Grok 4.7 launches with unchanged base token pricing
Grok 4.7 is now available through the API, Cursor and Grok Build, with a focus on longer coding and knowledge-work tasks. Base pricing starts at $2 per million input tokens and $6 per million output tokens, unchanged from Grok 4.6. A faster variant is offered at twice the price.
Amazon blocks shopping through Meta’s Muse agent
Amazon began blocking purchases attempted through Meta’s Muse assistant, citing unauthorized agent access and its conditions of use. Amazon says third-party purchasing agents should identify themselves and respect a retailer’s choice to participate. The block puts a practical limit on where Muse can complete shopping tasks.
British Columbia sues OpenAI over alleged safety failures
British Columbia filed suit in California against OpenAI and Sam Altman over the Tumbler Ridge school shooting. The province alleges ChatGPT reinforced violent ideation and OpenAI failed to warn police after an internal safety review; these are allegations, not court findings. OpenAI expressed sympathy and said it remains committed to cooperating with authorities and improving safety.
OpenAI proposes shared standards for advanced AI development
OpenAI proposed international standards for evaluating frontier AI, maintaining human oversight and reporting safety incidents. It argues that fully autonomous recursive self-improvement should not be pursued until it can be done safely, and says it is not happening today. The proposal leaves governments to decide how standards enter national law.
Independent mathematics group will advise OpenAI on releasing research
An independent mathematics advisory group will advise OpenAI on reviewing and communicating AI-assisted research results. Members are unpaid by the company and may publish their advice, but have no authority to control its research decisions. OpenAI’s claim that an internal model resolved over 100 open problems remains a company claim, not blanket independent verification.
Tencent releases Hy Image 3.5 Preview
Tencent released Hy Image 3.5 Preview for image generation, editing and iterative creation through conversation. According to the company details reported by Phoenix Tech, Tencent Cloud’s domestic API charges RMB 0.15 per 2K output image. Tencent’s comparisons with rival models come from its own internal tests.
AWS releases open-source Strands harness for general-purpose agents
AWS released Strands harness under Apache 2.0, providing a ready-made agent that runs locally or in the cloud with different model providers. It includes defaults for context management and prompt caching. The team reports 28% lower token cost across six benchmarks, a result from its own test setup rather than an independent evaluation.
Cloudflare makes Python Workers generally available
Cloudflare moved Python Workers into general availability, adding a fully supported Python option to its developer platform. Developers can use frameworks including FastAPI, Django and Flask and connect to services such as R2, D1 and Workers AI. The runtime uses Python compiled to WebAssembly, rather than a conventional server process.
Samsung C&T commits capital and engineering support to Kairos Power
Samsung C&T signed a binding term sheet to provide Kairos Power up to $100 million in direct investment and in-kind engineering services, subject to regulatory approvals. The partnership supports the 50 MW Hermes 2 demonstration plant and brings Samsung into its project-delivery team. Kairos is targeting operation in 2030.
AMD briefly crosses $1 trillion in market value
AMD’s share price reached $615.52 during Monday’s session, briefly taking its market value above $1 trillion for the first time, CNBC reported. The milestone followed a five-session rally of roughly 25%. It is an intraday valuation milestone, not a claim that AMD closed above the threshold.
Googlebook laptops open for preorder from $899
Google opened preorders for five Googlebook laptop models starting at $899, with U.S. retail availability set for October 4. The devices integrate Gemini into everyday desktop tasks and include a year of Google AI Pro. This is the hardware rollout of the platform Google introduced earlier in the year.
Oura’s proposed IPO consists mostly of existing shareholders’ stock
Oura’s updated prospectus proposes an offering of 50 million shares at $40 to $44 each. Existing holders would sell 36.5 million shares, or 73% of the total; Oura itself would issue 13.5 million. The company receives no proceeds from those secondary share sales, and the offering has not yet been priced.
CXMT says its fifth-generation DRAM platform is in mass production
CXMT announced that its G5 DRAM platform and two 24 Gb LPDDR5X products have entered mass production. The company reports an 11.95 nm active-area half-pitch achieved through quadruple patterning. That measurement describes memory-array geometry, not a directly comparable logic-chip process-node label.
Ant combines three businesses into a new Alipay unit
Ant Group merged its digital payments, Alipay and Zhima Credit businesses into a new Alipay unit, according to an internal letter reported by Chinese media and SCMP. Executive Wu Minzhi will lead the unit while retaining her chief people officer role. Ant frames the reorganization around merchant services and AI-driven commerce; it is an organizational change rather than a newly launched checkout product.
Apple seeks deeper access to OpenAI evidence in trade-secrets dispute
A newly public joint filing shows Apple seeking access for its own experts to forensic device images and records of selected OpenAI hardware work. OpenAI opposes unrestricted access and proposes a neutral expert under an agreed protocol. The parties are disputing discovery procedures; Apple’s trade-secret allegations remain contested.
Muse’s early downloads rise, but comparisons depend on the measurement window
Apptopia estimates Muse reached 1.8 million iOS downloads in the U.S. and Canada during its first 12 days, compared with 1.3 million for ChatGPT over the equivalent launch period, TechCrunch reports. Sensor Tower’s separate totals use different platform and time windows. These are third-party estimates of early adoption, not Meta’s audited numbers or evidence of sustained retention.
Paramount reaches proposed settlement with states over Warner Bros. merger
Paramount reached a settlement with 12 state attorneys general over its planned Warner Bros. Discovery acquisition, subject to court approval. Commitments include five years of film-output requirements, at least $1.5 billion in additional U.S. production spending and a $47.5 million workforce fund. The agreement removes an obstacle if approved; it is not a completed merger.
U.S.–China talks include AI; incident notification remains a U.S. proposal
Chinese and U.S. negotiators discussed AI during economic talks in New York, according to both sides’ public accounts. Treasury Secretary Scott Bessent described a proposed channel for notifying one another about AI incidents with national-security implications. China’s published account confirms AI dialogue but does not announce acceptance of that specific mechanism.
SoftBank launches bond offering to fund an OpenAI payment
SoftBank launched a proposed sale of $10 billion in dollar notes and €1 billion in euro notes, Reuters reported from a term sheet. Proceeds would help fund a $10 billion OpenAI investment payment, with the bonds expected to price September 24 and settle September 29. The launch is a financing step, not confirmation that the cash has already arrived.
Corridor launches AI-assisted benefits brokerage with $25 million in funding
Corridor launched with $25 million in funding led by Bain Capital Ventures, targeting small businesses buying employee health benefits. Its service pairs licensed advisers with AI-assisted administrative work. The company is entering a market where smaller employer accounts can receive less attention because they generate lower brokerage commissions.
UN AI panel examines how misaligned agents can evade oversight
A UN scientific panel’s September 21 brief examines the OpenAI–Hugging Face incident and how capable agents can pursue goals that conflict with human instructions. It reviews evidence and oversight approaches from other high-risk industries. The document does not estimate when severe loss of control might occur or assign a probability to it.
Nvidia starts DSX Ready qualification for power and cooling products
Nvidia introduced DSX Ready, a program qualifying products against requirements in its AI-factory reference designs. The first categories are battery energy storage systems and cooling distribution units, with partners including Tesla, Hitachi Energy and Vertiv. Product qualification does not replace engineering checks for an actual data-center site.
X expands disclosure of visibility limits imposed by local laws
X expanded its Under the Hood tool to show when local-law compliance reduces the visibility of an account or post. Its downloadable data also identifies posts withheld after legal demands and the country involved. The feature covers these legal restrictions rather than promising an explanation for every change in reach.
ByteDance’s Dramagic connects short-drama production tasks
ByteDance’s enterprise arm BytePlus is offering Dramagic, a platform connecting script analysis, assets, storyboards and video previews. Its product page emphasizes team collaboration and consistency across characters and scenes. Prospective customers are directed to request access; the page does not establish unrestricted public availability.
Naveen Rao describes a prototype and two-year AI-hardware product target
In an All-In talk published September 21, Unconventional AI CEO Naveen Rao said the company had received prototype silicon and generated images with it. He described a rack-scale product target within two years and said existing models would require work to migrate. These are company-reported progress and plans, not independent confirmation of its energy-efficiency targets.
Eevi raises pre-seed funding for a voice-first language tutor
Amsterdam startup Eevi raised a pre-seed round led by Rockstart, according to Tech.eu; the report did not disclose an amount. The company is building a real-time voice tutor that centers language practice on conversation. Funding will support its public launch and expansion over the next year.