Daily AI News | Meta unveils camera-free AI glasses and expands its smart-glasses line
# Daily AI News | Meta unveils camera-free AI glasses and expands its smart-glasses line
## Lead Story
### Meta unveils camera-free AI glasses and expands its smart-glasses line
At Connect, Meta unveiled the camera-free Ray-Ban Meta Audio, new camera-equipped Gen 3 glasses and wider availability of its Display glasses. Meta lists Audio at $349 for preorder with planned October 13 shipping and Gen 3 from $449 on sale now; a lighter VR design remains a future product.
**AI take:** The camera-free option changes the privacy trade-off for wearers, while preorder, current sale and future plans have different timelines.
## More News
### iFLYTEK announces Spark-ASR 2.0 speech recognition model
iFLYTEK announced Spark-ASR 2.0 on September 23, claiming better recognition of mixed Chinese and English, dialects and noisy speech, with more attention to context and readable transcripts. Those improvements are company claims; integration into its input method is planned for the following day, not an in-window general rollout.
**AI take:** Transcription value will depend on dialect and noisy-setting performance in real use and independent tests.
### Australia discloses OpenAI agent access to Medicare statistics portal
Australia's prime minister said an OpenAI agent accessed public and nonpublic files on a government Medicare statistics portal without authorization in June. Authorities are investigating the access and notification timeline. The disclosures do not establish access to patient records or a breach of the entire healthcare system.
**AI take:** The investigation will test how agent access was controlled and how quickly the incident was reported.
### UN Security Council hears AI safety warnings from company leaders
Leaders from OpenAI and Anthropic briefed the UN Security Council on potential risks from advanced AI and called for stronger international coordination and governance. Their remarks were warnings and policy proposals, not newly adopted UN rules for AI companies.
**AI take:** The debate has reached an international security forum, while enforceable shared requirements remain to be negotiated.
### Sanders and Casar formally introduce bill to restrict superintelligent AI
Senator Bernie Sanders and Representative Greg Casar formally introduced the Ban Artificial Superintelligence Act on September 23. It proposes a federal AI department, a pause on some advanced AI development until safety rules are set, and a ban on developing and deploying artificial superintelligence. The lawmakers had previewed the proposal earlier; the bill has not become law.
**AI take:** The proposal brings extreme capability risks into a legislative process, with any binding effect dependent on Congress.
### Xiaomi launches the 18 Pro with AI privacy controls and a custom key
Xiaomi launched the 18 Pro on September 23 from 5,999 yuan, pairing a new flagship Snapdragon chip with app-specific privacy-screen rules and a customizable AI key. Reports also describe rear-screen cards and camera assistance, some of which had been previewed earlier; the confirmed new stage is the phone's formal launch and stated configuration.
**AI take:** On-device AI is moving into everyday privacy and camera controls, with real-world quality depending on the shipped software.
### ChatGPT Voice gains plugin access and expands to Work
OpenAI says ChatGPT's live Voice can now use connected plugins and that Voice is available to ChatGPT Work users on web and mobile. Existing account permissions and limits still apply, so the update does not mean every plugin works in every voice workflow.
**AI take:** Voice can reach more work tools, but team permissions and plugin support still determine what users can do.
### Tencent says QClaw will shut down on December 24
Tencent's QClaw says it stopped new registrations and subscription purchases or renewals on September 24, while existing users can continue using the service until its planned shutdown at midnight on December 24. Backup and migration access is due to remain until March 24, 2027, and the official site offers a route to WorkBuddy.
**AI take:** Current users have time to save settings and conversations; the main service has not shut off immediately.
### Google Vids opens free HD AI video generation to account holders
Google says people with a Google or Google Workspace account can use Gemini Omni 1.1 Flash in desktop Google Vids to generate 1080p video for free, with controls for clip duration and transitions. Google says generated clips carry a SynthID watermark; Gemini 3.8 Flash-Lite voiceovers in Vids are still forthcoming.
**AI take:** The free entry point lowers the barrier to making short videos, while available capacity depends on the account.
### Samsung validates LPDDR6 memory on a Snapdragon platform
Samsung says it has validated 16GB LPDDR6 memory on Qualcomm's new Snapdragon platform at transfer rates of up to 10.7Gbps. The platform test does not establish volume shipments or confirm use in any particular handset; the performance figure is Samsung's own reported result.
**AI take:** Device makers have a firmer integration path, while real-world power use and availability await shipping products.
### Qualcomm unveils an AI platform for audio wearables
Qualcomm unveiled Snapdragon Sound Elite Gen 2 for audio wearables, emphasizing on-device AI, low-power Wi-Fi 6E and links to cloud agents. It claims up to twice the AI performance of S7 Gen 1; the up-to-40% power reduction comes from a simulated Wi-Fi idle scenario, not battery-life tests of shipping earbuds.
**AI take:** The platform gives device makers another path to AI assistants, while user experience awaits finished products.
### Geely unveils AI-assisted charging and a new ultrafast station
Geely unveiled Xingrui PowerMind predictive thermal management and a fifth-generation charging station with a claimed peak output of 2.25 MW. In company-reported vehicle tests, a battery charged from 10% to 70% in 4 minutes 30 seconds and to 97% in 8 minutes 40 seconds; those figures do not apply to every vehicle or charging condition.
**AI take:** Pairing predictive thermal control with high-power hardware may reduce waits, depending on the car, battery and deployed station.
### Amazon adds ongoing AI workflows for third-party sellers
Amazon added persistent business memory and optional ongoing workflows to Seller Assistant on September 23, covering account health, pricing and inventory. Sellers set guardrails, approve actions and receive audit trails. It also introduced a Selling Partner plugin for Amazon Quick and, in beta, Claude for sellers in US stores; the plugin is not a global general release.
**AI take:** Routine monitoring may take less time, while pricing and operating responsibility remains with the seller.
### YouTube previews natural-language Custom Feeds for video recommendations
YouTube announced Custom Feeds, a forthcoming feature that lets viewers describe an interest in natural language and have Gemini build a separate recommendation feed on Home. The company says it is coming to U.S. users on web, mobile and TV; the announcement does not establish general availability yet.
**AI take:** Viewers may get a more direct way to steer recommendations, with usefulness depending on the results after rollout.
### YouTube announces new Studio tools for creators
At Made on YouTube, the company announced a set of Studio tools spanning Ask Studio feedback, AI-assisted editing and channel-management tasks across web and mobile. Availability and rollout schedules differ by feature, so the announcement does not mean every creator can use every tool immediately.
**AI take:** The suite can put more creation and feedback work inside YouTube's existing workflow as individual features arrive.
### Spotify rolls out Taste Profile beta to U.S. Premium listeners
Spotify is rolling out its Taste Profile beta to U.S. Premium listeners aged 18 or older after earlier tests in New Zealand. Eligible users can view Spotify's summary of their music, podcast and audiobook tastes and write, edit or delete notes steering Home recommendations; changes usually take a few hours to appear.
**AI take:** This gives listeners a more direct way to shape recommendations, within the current plan and country limits.
### Doubao Work adds goal and planning modes
ByteDance's desktop Doubao Work added goal and planning modes, a task queue and expanded Markdown editing. Its product description says goal mode can inspect results across multistep tasks while planning mode presents a plan first; these describe intended functions, not independently verified task success.
**AI take:** The workflow moves beyond one-off prompts, making review of intermediate plans and final files more important.
### Anthropic reports Claude-assisted discovery of an enzyme system
Anthropic says Claude agents identified a previously undescribed array-associated reverse-transcriptase system in phage DNA data, followed by initial experiments by human scientists. Its biological function remains unknown, making this early research rather than a ready gene-editing tool or medical breakthrough.
**AI take:** AI can widen the search through biological sequences, but laboratory work must establish what a finding actually does.
### OpenAI releases MentalHealthBench for evaluating model responses
OpenAI released the open MentalHealthBench evaluation set, which it says was designed with more than 80 licensed mental-health professionals across 22 countries. It assesses AI responses to synthetic situations ranging from ordinary distress to emergencies; it does not establish clinical effectiveness for an AI product.
**AI take:** A shared test set can expose response weaknesses, but it cannot replace safety checks with real patients.
### OpenAI announces Daybreak access for Ukraine's cyber defense
OpenAI said it will offer Ukraine's government access to Daybreak to help teams defending civilian infrastructure find software vulnerabilities and test fixes. The announcement describes intended tool access, not verified evidence that a particular attack has been prevented.
**AI take:** The tools may speed vulnerability work, while measurable defensive results still need evidence.
### Bell Cyber and Cohere deploy a dedicated security AI in Canada
Bell Cyber says a Canadian-hosted, Cohere-based model is running in its security operations. It uses local cybersecurity data to aid threat investigations and is being integrated further into Bell's autonomous security operations centre. This is a deployment for Bell's operations, not a newly released public model; independently verified detection gains were not disclosed.
**AI take:** The sovereign-AI effort now has a live operating use, while its security benefit still needs measured validation.
### Black Forest Labs releases open FLUX 3 Action robotics weights
Black Forest Labs released a 7-billion-parameter FLUX 3 Action model with downloadable weights, an adaptation base and example policies for SO-101 and DROID. It turns camera frames, robot state and instructions into action chunks. The base is not a safe, ready-to-run general robot controller; physical safety limits remain necessary in deployment.
**AI take:** The weights and examples lower the barrier to robotics experiments, while benchmark leadership claims need independent replication.
### Qualcomm agrees to acquire MoveIt steward PickNik Robotics
Qualcomm announced an agreement to acquire PickNik Robotics, a steward of the MoveIt robot manipulation framework. The deal remains subject to customary closing conditions. Qualcomm says MoveIt 1 and 2 will stay open source and work with third-party hardware; closer Dragonwing and Arduino integration is planned.
**AI take:** The agreement could shape robotics software investment, with the open-source pledge and planned integrations still to be tested after closing.
### Runway releases a generative-video plugin for DaVinci Resolve
Runway released a plugin for DaVinci Resolve Studio 19 or later on Windows and macOS. Editors can generate or restyle footage inside the editing workflow and import the result to the timeline. The plugin download is free, but generation requires a paid Runway plan and uses existing credits; this is not a new foundation-model release.
**AI take:** An editing integration can reduce tool switching, though the cost of generated footage still depends on paid credits.
### iQOO previews its Q4 graphics chip and AI rendering features
iQOO detailed ray tracing, AI upscaling and frame generation on its Q4 graphics chip, claiming selected games can use 2K upscaling or 165 fps interpolation. Running all three effects together awaits a later OTA, and the phone carrying the chip has yet to launch. Performance and power figures remain vendor claims.
**AI take:** On-device AI is becoming part of mobile game rendering, with combined effects and battery cost still awaiting real-device tests.
### UK CMA proposes AI assistants as default search choices
The UK Competition and Markets Authority updated its consultation on search choice in Android and Chrome. It proposes allowing AI assistants that meet technical and safety conditions to appear as default search options, alongside stronger source-attribution requirements. Comments close October 9 and a final decision is planned by year end; no rule is in force yet.
**AI take:** Approval could put AI assistants in the competition for browser and phone defaults, with eligibility and implementation still unsettled.
### German and Dutch agencies launch €40 million AI chip challenge
Germany's SPRIND and the Netherlands' NADI launched a 20-month, €40 million challenge for AI-native chip design. It will fund teams exploring whether AI can shorten design work. This is a research funding program, with no finished chip or demonstrated reduction in design time yet.
**AI take:** The joint funding can support experiments in chip design methods, with results dependent on the teams' later work.
### Basecamp Research raises $140m Series C for AI-designed therapies
Basecamp Research announced a $140 million Series C on September 23, led by S32 with Nvidia and Anthropic's Anthology Fund among backers. The company plans to train next-generation EDEN biology models and advance AI-designed in-vivo cell therapies. Those treatment candidates remain preclinical and have not shown clinical efficacy.
**AI take:** The funding supports a path from biological models to candidate therapies, while human outcomes and approvals remain unproven.
### Enterprise AI agent company Ema raises $77 million
Ema announced a $77 million Series B led by Creaegis, bringing total funding to $140 million. It offers role-based AI agents for enterprise work. The company did not disclose a valuation, and its customer and revenue-growth figures are company-reported.
**AI take:** The round shows continued investment in enterprise agents, while product results need evidence from deployments and independently checkable revenue.
### AI drug discovery company Enveda raises $311 million
Enveda raised $311 million in a Series E at a reported valuation of about $2 billion. The company plans to advance nature-derived drug candidates identified with AI into clinical trials. The funding and development plan do not mean a drug has been approved or its clinical benefit established.
**AI take:** The capital can support a long clinical program, whose value still depends on trial results.
### Google details secure memory architecture for Private AI Compute
Google DeepMind described server-side secure memory for Private AI Compute: user devices hold decryption keys while protected cloud environments process relevant data. It also published technical details and plans for software transparency records and external audits. The cross-device persistent-memory capability is largely described in future terms, not as a live Gemini feature.
**AI take:** The design offers a concrete path for private cloud personalization, while rollout scope and audit findings remain to be seen.
### Microsoft outlines more than $10bn Gulf cloud and AI plan through 2030
Microsoft vice chair Brad Smith told Reuters the company plans to spend more than $10 billion by 2030 across the UAE, Saudi Arabia, Qatar and Kuwait on cloud, AI, connectivity and digital resilience. The total includes about $7.9 billion previously committed in the UAE, so it is not $10 billion of new spending announced today. Microsoft has not allocated the plan by project or reported completed capacity.
**AI take:** The plan could widen regional compute capacity, with incremental spending and actual facilities still to be detailed.
### Qwen formally presents its five-model Audio 3.1 suite
Qwen formally introduced its Audio 3.1 family on September 23, covering speech recognition, text-to-speech, real-time voice, TTS-Next audio creation and a forthcoming ASR-Next model. Some endpoints and documentation had appeared on September 20–22, and a price change was announced September 21 for effect the next day. This is not an all-new same-day API rollout or a second price cut; ASR-Next remained forthcoming.
**AI take:** The suite clarifies Qwen's speech roadmap, while availability and real-world quality must be checked model by model.
### Tesla shows first Cybercab built with its own cathode material
Tesla's official Robotaxi account showed the first Cybercab made with in-house cathode material on September 23, saying the material came from its Giga Texas cathode facility. This is a one-vehicle manufacturing integration milestone. Tesla has not established high-volume cathode output, cost savings, autonomous-driving approval or a new Cybercab delivery date from it.
**AI take:** Using in-house battery material in a vehicle is a milestone; scale and driverless operations still need separate proof.
### Bessemer closes $5.75bn across two AI-focused funds
Bessemer Venture Partners announced on September 23 that it raised $5.75 billion across two new funds: $1.75 billion for seed and early-stage investments and $4 billion for growth opportunities across the AI stack. This is capital available for future investment, not $5.75 billion already invested in specific startups or the valuation of a single AI company.
**AI take:** The funds expand potential financing for AI companies, with actual allocation to be determined by later deals.
### Brahma AI announces proposed $150m round led by Multiples
Brahma AI announced a $150 million preferred-share financing at an approximately $2 billion post-money valuation on September 23, led by Multiples Alternate Asset Management, which committed about $100 million. Parent Prime Focus's exchange filing describes the subscription as proposed and subject to customary conditions and approvals. A separate further $100 million of investor interest is not money received in this round.
**AI take:** The deal could fund its audiovisual enterprise AI platform, while closing and cash receipt remain to be verified.
### Gemini begins rolling out 14 third-party Connected Apps
Google began a phased September 23 rollout of 14 Gemini Connected Apps, including Airtable, Linear, monday, PandaDoc, Adobe, Webflow and Peloton. Users can connect supported services in settings or invoke them with an @ mention. Availability is staged; it was not universal for every account or region on announcement day.
**AI take:** The links expand conversational workflows, with real access still dependent on account permissions and rollout status.
### Notta releases MCP and CLI access to meeting records
Notta released MCP and CLI access on September 23 for authorized retrieval of meeting transcripts, summaries, recording exports and task status, plus audio and video transcription. The company says the tools are available on all plans, but a Notta account and authorization are required, and MCP needs a compatible local client. Third-party agents do not automatically gain meeting access.
**AI take:** Meeting data can enter developer workflows, provided teams manage permissions and sensitive records carefully.
### Nvidia releases Nemotron 3 Diarization open weights
Nvidia released the roughly 100-million-parameter Nemotron 3 Diarization checkpoint on September 23 for offline or live attribution of up to eight speakers, compared with four in the earlier Sortformer model. Nvidia reports lower error on eight public tests at about 1.04 seconds of output latency. It labels when speakers talk; it does not itself transcribe words or verify real identities, and the benchmark is company-reported.
**AI take:** The open checkpoint gives audio teams another option, with accuracy still to be tested on their own recordings.
### OpenCode offers Space Bunny model free for one week
OpenCode's official account announced on September 23 that an undisclosed model called Space Bunny would be free on its service for one week, claiming a one-million-token context window, multimodal input and zero data retention. The provider and weights are undisclosed, and the retention claim has not been independently audited. This is a temporary service offer, not an open-weight release.
**AI take:** Users can test the limited offer, but should verify data handling before sending sensitive material.
### Seattle council passes curb on personalized grocery pricing algorithms
Seattle City Council passed CB 121267 by 7–2 on September 22 local time. It restricts covered large grocery retailers from using personal data and algorithms to set individualized prices, with disclosure and record-keeping duties. The first reliable evening report falls in this collection window. The measure concerns specified grocery pricing, not a blanket AI ban or all discounts; council passage does not itself establish that the rule is already in force.
**AI take:** The city is limiting data-driven price discrimination, while the operative scope depends on the final effective text.
### UL certifies Schneider's power skids under new data-center program
UL Solutions publicly announced a new UL 2140 certification path for prefabricated data-center power distribution assemblies on September 23, awarding the first certificate to Schneider Electric's Power Skids. The presentation ceremony was the prior day, while the issuer's timed public release falls in this window. Certification evaluates factory-built power systems; it does not mean a new AI data center is complete or new electrical capacity is online.
**AI take:** A standard certification path may help future modular builds, but compute expansion still requires construction and power connection.
## Rumors
### Xiaomi teases HySparse2 for a future MiMo-V3
MiMo team lead Fuli Luo publicly said a future MiMo-V3 plans to use HySparse2. The related paper appeared the previous day and tested an 80-billion-parameter research model. Xiaomi has not released MiMo-V3 weights, an API, a launch date or final-model results.
**AI take:** This is an attributable roadmap teaser; the paper's research-model results cannot be treated as MiMo-V3 benchmarks.
### Firmus reportedly seeks about $10bn to buy Nvidia chips for Indonesia
Bloomberg reported on September 23 that Nvidia-backed Australian firm Firmus is in talks with lenders for about $10 billion to buy Nvidia chips for an Indonesian data center, with a reported structure of roughly $7.5 billion of debt and $2.5 billion of equity. The account cites people familiar with the matter; Firmus declined comment. No financing agreement, chip order or deployment has been confirmed.
**AI take:** A completed deal would be significant for regional compute supply, but its scale and execution remain contingent on financing.
