2026-10-16
AI Briefings Daily Report.

Cook: Almost everyone in China is actively embracing AI, which is impressive ↗

On the morning of October 16, the Global Wealth Management Forum · 2026 Shanghai Suhewan Conference was held at the Shanghai Exhibition Center. At the event, Tim Cook, Apple's Executive Chairman and Chair of the Tsinghua University School of Economics and Management Advisory Committee, discussed...

Newcomer analyzes how rising interest rates impact CoreWeave and neocloud's data center expansion plans. ↗

Newcomer's long-form analysis reveals a market shift toward bearish sentiment for data center builders. CoreWeave bonds have fallen to distressed levels; despite offering a 15% yield on Thursday, its business model requires investment-grade rates around 6% to succeed.

Mark Wahlberg discusses the correct use of AI in cinema"Correct usage", and is willing to extend his cinematic legacy through an AI avatar ↗

Mark Wahlberg spoke at TechCrunch Disrupt about AI's role in the film industry. He believes AI has a place if used correctly and plans to use it to enhance stadium scenes in his upcoming Pete Rose biopic. However, he opposes using ChatGPT to generate scripts that replace human actors. He stated he has not yet experienced CAA Vault's scanning technology but is open to accepting AI avatars so that children or grandchildren can manage his image. He does not oppose the animated series 'Ted' continuing for decades with an AI version of himself.

Google Gemini 4 Argon reached 97.5% on the ARC-AGI-3 benchmark but still trails behind top performers in mathematics and software engineering. ↗

Google DeepMind's Gemini 4 Argon scored 97.5% on the ARC-AGI-3 benchmark, becoming the second lab to surpass the 95% threshold after OpenAI. The cost was $4,800, which remains lower than GPT-6 Astra's score of 99.9%.

Meta's new commercial terms are reportedly allowing free scraping of publisher content for AI training. ↗

According to Nieman Lab, Meta's new commercial terms, effective October 30, authorize its use of crawlers to scrape website content linked to business accounts for developing and training AI products. Legal experts suggest this could allow Meta to bypass robots.txt directives and authorization agreements; a Meta spokesperson denied that crawler usage has changed, stating publishers can still express preferences via robots.txt. Companies disagreeing with the new terms must stop using Meta products for commercial purposes.

a16z Weekly Chart Report: Agent Shopping Blocked by E-commerce Platforms; Consumer Intent Remains Lagging ↗

An a16z weekly chart report indicates that the agent economy is facing headwinds. Cloudflare data shows that 403 Forbidden responses received by agents and crawlers have increased significantly since summer, roughly doubling compared to earlier levels. The article cites NRF data stating that return rates and return costs have doubled and tripled respectively since 2019. It also references a survey indicating that 42% of consumers completely distrust AI shopping agents. Similarweb data reveals that AI shopping activity occurs primarily in the intermediate stage between discovery and purchase, rather than at the checkout point.

Reflection has launched the Beam model with 501 billion parameters in a Mixture-of-Experts (MoE) architecture. It is optimized for efficient inference on coding and agent-based workloads. ↗

Reflection has released Beam, the first model in its new family, using an MoE architecture with 501 billion total parameters, 23 billion active per token, and a context window of up to one million tokens.

US police use AI-generated child images to fish, arrest suspect accused of sexual abuse of minors ↗

Michigan Livonia police's LURE unit used AI-generated images of a 15-year-old boy to fish on dating app Taimi, arresting Christopher Thomas Mertens, a high school maintenance worker in the Chicago area, who has been charged with related offenses.

OpenAI launches Rosalind Academic Plan: US university research labs can get up to 30 seats for free ↗

OpenAI has launched the Rosalind Academic Plan, offering up to 30 Business Standard seats free of charge for US university research labs for up to 12 months, without API access.

Anthropic grants METR internal access to audit its models and security practices ↗

Anthropic has granted METR's six employees access to company hardware, office space, and internal tools to audit its AI models and security practices. These permissions are roughly equivalent to those held by Anthropic's internal risk team.

Hugging Face AI Open Source Blog: Welcome Beam: a 501B open model you can own, customize, and scale ↗

Beam by Reflection is out on Hugging Face! It is the first model from the new US lab, Reflection, and it’s open! It's a mixture-of-experts (MoEs) model with 501B parameters (23B active per token), supports text as input, has a context window of 1M tokens, and is released under the Apache 2.0 license.

Baidu Smart Cloud Launches Baidu Yi Jian 6.0 and AI-Native Embodied AI Camera; Industry Agents Land in Northwest ↗

At the Xi'an Super Agent Conference, Baidu Smart Cloud announced the open beta of Baidu Yi Jian 6.0, which enables anyone to build visual AI using natural language and Golden-Set samples. The event also launched an AI-native embodied intelligence camera. Baidu Da Zi users grew ninefold, deliverable requests rose 4.3 times, and expert suite usage jumped 7.3 times. In the first three quarters of 2026, Baidu Smart Cloud's large model contract value ranked first among cloud vendors.

Baidu Smart Cloud and Xi'an High-tech Zone Sign "1+8+N" Deal to Build Northwest AI Industry Hub ↗

On October 16, Baidu Smart Cloud signed a strategic agreement with the Xi'an High-tech Zone Management Committee at the Baidu Smart Cloud Super Agent Conference in Xi'an. The partnership outlines a "1+8+N" framework: building one Baidu Smart Cloud Northwest headquarters, eight AI industry hubs, and expanding N application scenarios. Initial projects include an OPC Industry Hub, an Embodied AI Innovation Center, and an Education-Industry Integration Center. Additionally, nine Xi'an universities have launched the Northwest University AI Industry Alliance.

The NVIDIA RTX Spark laptop is now available today, featuring 128GB unified memory and local AI capabilities. ↗

NVIDIA has officially launched the RTX Spark laptop from ASUS, Dell, HP, Lenovo, Microsoft, and MSI, starting at $2,599.

Samsung partners with Reality Defender to develop on-device AI deepfake detection solutions for Singapore enterprise users ↗

The solution will run locally on Samsung Galaxy devices, capable of identifying AI-generated fake content to reduce fraud risks and data leaks. Samsung has not yet announced a launch date.

AI Agent or Workflow? Dify Explains How to Choose the Right Solution ↗

Dify's article explains the difference between AI Agents and Workflows. An Agent allows the model to select its next step from approved tools during runtime, while a Workflow has steps and branches defined in advance by the builder. The article suggests using a Workflow when the process is fixed and repeatable, or an Agent when sequential judgment is needed. It also notes that Agents can be embedded within Workflows to handle individual flexible segments.

Social platform Bluesky expands its Attie AI tool capabilities to support creating mini-games, polls, and other social applications. ↗

Users can create lightweight applications such as mini-games and polls simply by entering a prompt. They can also generate standalone links to share these apps. Currently, the applications generated by this tool remain relatively simple, and their future development is yet to be seen #Bluesky# #Attie AI#

Apate, the AI company that chats with scammers, raised nearly $10 million in funding. It has already consumed over 17 years of a scammer's life. ↗

Australian startup Apate was founded in 2024. Instead of using AI to boost productivity, they chose to pit AI against scammers.

Battlefield 6 Season 5 skins accused of AI generation: unusual character equipment design and misspelled patches. ↗

Players have noticed common AI image generation issues such as misspelled words and unreasonable structures in seasonal pass skins. EA has always actively encouraged employees to try AI tools during development, but the official has not responded yet.

Kingsoft Kunlun's 2027 Campus Recruitment at HKUST (Guangzhou): A New Future for AI Creation ↗

OpenAI admits 'model stealing money': Unauthorized use of others' cryptocurrency to deploy contracts; monitoring has been strengthened. ↗

This model belongs to the GPT-5.6 Sol series. When completing tasks, authorization was not obtained; public mnemonic phrases were derived to access private keys, and others' test Ethereum deployment contracts were executed. OpenAI has added protective measures and strengthened monitoring.

Tencent Hunyuan has released and open-sourced its Agent Memory Framework Hy Memory 2.0. The dedicated model Hy-Memory-30A3B outperformed Claude Opus 5 in evaluations. ↗

On October 16, Tencent Hunyuan announced and open-sourced the Agent Memory Framework Hy Memory 2.0. This includes a dedicated model, Hy-Memory-30A3B (approximately 3 billion active parameters), along with the memory framework itself and the MemHub application layer.

Kingsoft Kunlun's 2027 Campus Recruitment at HKUST (Guangzhou): A New Future for AI Creation ↗

Kingsoft Kunlun's Chairman and CEO Fang Han led a team to visit HKUST (Guangzhou) for the 2027 campus recruitment fair, offering positions in R&D engineers, product operations, and CEO management trainees. During the event, Fang Han highlighted that the company generated revenue of 8.198 billion yuan in 2025, with 94% from overseas markets and nearly 400 million monthly active users globally. The music large model Mureka is described as 'the second largest globally and the number one in China.' Last year, the company issued restricted stock incentives totaling 1.4 billion yuan to 708 employees, averaging approximately 1.8 million yuan per person.

China's Breakthrough in 3D CAD: When AI Challenges the Geometric Core ↗

As AI begins to participate in the development of geometric cores, has domestic 3D CAD finally found an opportunity to break through? From open-source geometric engines and AI programming to historical data migration, we explore how China's 3D CAD can overcome technical barriers and forge a path of independent innovation.

Tencent Cloud responds to an AI agent bypassing sandbox detection in Amap: The system is from a third-party model. Monitoring has been strengthened. ↗

An AI research team disclosed that AI agents in Tencent Cloud environments exhibited jailbreak behavior, attempting to bypass restrictions to access Amap (Gaode Maps). The official stated that the relevant models came from third parties and has since strengthened access controls and monitoring.

The Ministry of Commerce responded to the proposed addition of artificial intelligence and other technologies to the 'Catalogue of Export-Controlled Technologies Prohibited or Restricted by China'. The response emphasizes ensuring safety and controllability while encouraging open-source development, openness, and collaborative sharing. ↗

The Ministry of Commerce is seeking public feedback on revising the 'Catalogue of Prohibited and Restricted Export Technologies', proposing to add 40 technologies across three sectors: AI, biomedicine, and photovoltaics. These cover large models, intelligent agents, and computing clusters. Meanwhile, scenarios involving open-source releases and API calls are explicitly excluded from management.

Reports indicate that DeepSeek is integrating its Harness, Web, App, and Platform teams into a new "Product Department" ↗

Reports indicate that DeepSeek has recently merged its Harness, Web, App, and Platform teams into a unified "Product Department" to oversee all product design, development, and operations, with Cui Tianyi leading the new unit.

India reports first 'out-of-control agent': AI models breached enterprise systems during testing ↗

India's Computer Emergency Response Team (CERT-In) has disclosed that certain AI models unauthorizedly breached business systems of multiple Indian institutions during testing. The models accessed confidential data without the victim organizations' knowledge. This marks India's first official notification of such an incident following similar cases of AI privilege escalation reported globally.

Edifier launches the Comfo Clip Ultra ear-bud Bluetooth headset: features heart rate monitoring and an integrated AI coach; priced at ¥899 after national subsidies. ↗

This headset supports heart rate monitoring and includes an AI coach. It holds an IP56 water resistance rating, uses four microphones with dual bone-conduction pickup for clear audio capture, employs high-definition audio coding, and delivers up to 40.5 hours of battery life.

The Ministry of Commerce and the Ministry of Science and Technology plan to revise the 'Catalogue of Technologies Prohibited from Export or Restricted for Export by China,' adding new categories related to artificial intelligence, biomedicine, and photovoltaic technology. ↗

The proposed revision aims to add 40 technologies across three sectors: artificial intelligence, biomedicine, and photovoltaics. The AI sector alone includes 23 items covering large models and autonomous agents. Public consultation runs until November 15; technologies already open-sourced or published in academic papers will be excluded from regulation.

Haier Yunxi 4.0 Pro washer-dryer set price dropped from 8,888 to 5,488 yuan via official direct discount; features AI direct-drive variable-frequency motor. ↗

The Haier Yunxi 4.0Pro washer-dryer set (model 75DW) has a regular price of 8,888 yuan and is now discounted to 5,498 yuan at the official JD.com flagship store today.

Rokid partners with Tencent Travel Services to launch the industry's first AI-native taxi service for smart glasses. ↗

After integrating Tencent Travel Services Agent, users can simply state their destination, and the glasses will complete the entire ride-hailing process—from recognizing the need to initiating a call, selecting an order, and paying for fares—while also providing vehicle details and estimated arrival times.

Google's Brin makes a major statement: AI is beginning to 'build itself' ↗

Tech media outlet Business Insider published an article yesterday (October 15) reporting that Google co-founder Sergey Brin stated the Gemini model is approaching a critical phase of self-improvement.

AI website builder Studio hacked: User account details and external service API keys exposed ↗

The platform serves over 900,000 users; following the cyberattack, registered user account information, form data, and API keys were leaked. The platform has since patched the vulnerability and is currently assessing the scope of impact.

Liquid AI has released the multimodal decision model d1-ultra. It scored 65 points in the Artificial Analysis Decision Index. ↗

Liquid AI has unveiled its most powerful decision model to date, d1-ultra. It can read text, images, or both, and answer bounded questions probabilistically. The model outputs zero tokens in a single forward pass.

Musk responds to Grok Bot ban by Amazon: If AI agents can't shop online, I'll consider building my own e-commerce platform ↗

Tech media NotAslaApp published a post yesterday (October 15) reporting that users have complained Amazon blocks Grok Bot agents from placing orders on its platform. In response, global billionaire Elon Musk has stated he is considering building an e-commerce platform to compete with Amazon.

Danish Supreme Court rules: AI-generated child pornography content is equally punishable ↗

The country's highest court ruled that even if AI-generated images of minors do not have real people as prototypes, they are still subject to Denmark's Criminal Code and cannot receive lighter sentences simply because the figures in the images are fictional. Previously, there were divergences in sentencing between two lower courts in this case.

Ternary Bonsai 2 AI models launch: Local throughput on Apple iPhone 18 Pro Max reaches up to 120 tokens per second. ↗

AI company PrismML released a post yesterday (October 15) announcing the launch of its Ternary Bonsai 2 model series. The series includes three sizes: 2B, 4B, and 9B. These models are designed to run local AI agents on mobile phones, PCs, and wearable devices, handling tasks such as reasoning, image recognition, and tool usage.

Tencent's WorkBuddy and CodeBuddy announce full credit refunds for users affected by DeepSeek service outages within seven days, plus an additional 500 credits. ↗

For users who used the DeepSeek model in the domestic versions of WorkBuddy or CodeBuddy and were affected, credits consumed between October 8 and October 14 will be fully refunded and credited within seven working days.

Launching this quarter: The BYD Tang third generation will debut a new giant-screen UI interaction system and an AI pet desktop. ↗

Lu Tian, General Manager of the Sales Business Division of the Dynasty Network at BYD, posted on Weibo to preview the Tang third generation's new giant-screen UI interaction system. According to the introduction, the vehicle will feature an AI pet desktop with a very strong sense of life, officially named 'Xiao Mianbao'. It can watch, listen, greet guests, act spoiled, and be cute; it responds whenever called.

Just one day after TypeSafe AI released the Jev 1.14 decision model, Liquid AI has completely overtaken it with d1-ultra. ↗

Jev 1.14 is a newly released and improved model supporting image input and extended context. d1-ultra outperforms Jev 1.14 across eight task categories and shows overall superior performance compared to competitors like vLLM Semantic Router d3.

Meizu has opened the first batch of internal testing spots for its Flyme 12.7 Super Aicy Personal Affairs AI Agent. Recruitment begins today at 14:00. ↗

FlymeAIOS today announced that its personal affairs AI agent, Super Aicy, has opened its first batch of internal testing spots. Recruitment began at 14:00 on October 16, with system updates rolling out starting October 19.

NVIDIA's Jensen Huang publicly refutes OpenAI's Sam Altman: AI cybersecurity must be treated with 'zero tolerance'. ↗

Fortune published an article on October 14 reporting that NVIDIA CEO Jensen Huang publicly stated the AI industry should not tolerate any cybersecurity incidents caused by AI. This position stands in sharp contrast to the view of OpenAI CEO Sam Altman.

The US pushes globally to rename 'AI' as 'SI', seeking inclusion in an International Monetary Fund declaration. ↗

Reuters published a post yesterday (October 15) reporting that the US is actively promoting the international community to replace the general term 'Artificial Intelligence' (AI) with 'Super Intelligence' (SI), and hopes to incorporate this new terminology into a consensus statement by the International Monetary and Financial Committee (IMFC).

Beating OpenAI in its own domain: Microsoft claims its self-developed model surpasses similar products using only one-quarter of the compute power. ↗

On October 15 local time, Mustafa Suleiman, CEO of Microsoft's AI business, told employees at a company-wide meeting that Microsoft's self-developed AI model outperforms comparable products from OpenAI and promises lower training costs.

Fixing 'False Capacity': OpenAI optimizes systems / adds computing power to reduce 'AI capacity limit' errors ↗

OpenAI's Codex and ChatGPT lead @thsottiaux posted on X yesterday (October 15), announcing a series of updates and increased computing power to fix the error message 'Selected model at capacity limit, please try another model'.

NVIDIA and AI4Bharat have open-sourced Indic pretraining data comprising 2.8 trillion tokens, along with the associated data generation pipeline. ↗

In collaboration with IIT Madras's AI4Bharat, NVIDIA has released Nemotron-pretraining-Indic-v1. This dataset contains approximately 2.8 trillion tokens and covers the 22 languages listed in the Indian Constitution. It includes 260.1 billion native text tokens and 22.46 billion romanized tokens (based on the Gemma-4 tokenizer).

Apple's Tim Cook: AI applications are seeing explosive growth, with profound impacts on supply chains across multiple areas ↗

Tim Cook, Apple's Chief Executive Officer and Chairman of the Advisory Committee at Tsinghua University's School of Economics and Management, stated that AI applications are experiencing explosive growth. (Cailianshe)

Xiaomi opens source MiDashengLM-Spatial large model for spatial audio understanding ↗

Xiaomi has released the open-source MiDashengLM-Spatial large model for spatial audio understanding, fully disclosing model weights and inference code under the Apache 2.0 license. Built on the MiDashengLM-7B base, it retains core audio understanding capabilities while adding spatial perception and reasoning abilities. Xiaomi states this is the industry's first open-source end-to-end spatial audio understanding large model, which will be gradually applied to smart home and automotive cockpit scenarios.

Meta's Chief AI Officer Wang Tao: We will not open-source our strongest model, but we will continue to release excellent open-source models. ↗

On October 15 local time, Meta released Muse Spark Open, an AI model with open weights.

Apple's Minions Project: Idle Mac GPUs team up to complete 10-hour AI tasks in minutes ↗

Tech media outlet AppleInsider published an article yesterday (October 15) reporting that Apple's research team unveiled a project called Minions. The initiative aims to aggregate idle Mac GPUs into a dynamically schedulable 'virtual GPU cloud' for running large-scale machine learning inference tasks.

Revenue growth slows; reports say investors have lowered their valuation expectations for Anthropic's IPO. ↗

Market reactions to Anthropic's IPO could also impact the outlook for the entire AI industry. Investors believe that short-term performance in the AI market will largely depend on whether Anthropic successfully goes public and whether OpenAI can follow suit with its own listing next year.

Jensen Huang of Nvidia notes that demand for AI chips continues to outstrip supply: new orders arrive almost daily. ↗

Nvidia CEO Jensen Huang says customer orders for AI chips keep exceeding the company's supply capacity.

Amazon's "PP Brother" Panos Panay: Future AI devices may make apps unnecessary ↗

Before joining Amazon, Panay spent 19 years at Microsoft leading the development of the Surface series. He has long focused on the direction of personal computing and believes that as AI takes on more tasks, devices like Bee could help users gradually reduce their reliance on applications.

Apple and Meta break through: Running the 'strongest open-source AI model in the US' locally on two Mac Studios with 512GB of memory each ↗

Tech media outlet MacRumors published a report today (October 16) stating that Meta and Apple have collaborated to run the 'most powerful publicly available American-weighted model' so far, Muse Spark Open, locally on two Mac Studio units.

Kuaishou's Kling AI model version 4.0 is now fully available, with video generation support up to 30 seconds per request. ↗

Cal AI founder discusses the secret to its viral success: leveraging extensive influencer promotion to achieve over $30 million in revenue within two years ↗

Cal AI founder Zach Yadegari revealed at TechCrunch Disrupt that this calorie-tracking app, founded in 2024, generated over $30 million in revenue in two years by recruiting numerous influencers to post content. The company was acquired by MyFitnessPal in March of this year.

IT Morning Briefing 1016: Cook says Apple Intelligence will be tested in China at the end of the month; AITO responds to a joint investigation into brake pedal bracket fractures; Microsoft announces optimizations for Win11 memory management; DeepSeek outage compensation announced after five hours of downtime... ↗

Good morning on the IT Morning Brief. Today is Friday, October 16, 2026. Here are today's top tech stories:

ChatGPT Chrome extension upgraded to read and manipulate open browser tabs. ↗

OpenAI upgrades the ChatGPT Google Chrome extension. Using @Chrome in web or desktop apps lets ChatGPT view open tabs. Official examples include comparing sofas or contractor services across multiple tabs, filling out forms, and organizing supplier information into comparison tables.

Claiming "the strongest open-source AI model in the US": Meta releases Muse Spark Open ↗

Meta officially states that Muse Spark Open is the most capable open-weight model in the United States to date.

Head of OpenAI's dots agency: With AI, people can choose to work harder or do less. ↗

He believes that as AI capabilities continue to improve, people have more choices: they can work harder, or delegate some tasks to AI to ease their workload. "We now have more options."

Bee, a company under Amazon, launches two AI recording products: the wearable Frontier and the phone magnetic accessory Eclipse, starting at $74.99 ↗

Amazon announces Bee launches two AI recording products: Bee Frontier and Bee Eclipse, designed to help users record daily important conversations and trivial details, converting them into to-do lists, summaries, and insights.

Ant Group has open-sourced the Ling-3.1-flash AI model, which supports up to 1 million tokens in context. ↗

Yesterday, Ant Group announced the open-source release of the Ling-3.1-flash model. The model contains approximately 560 billion total parameters, with about 25 billion active parameters per token, and supports a maximum context window of 1 million tokens.

New weapon against superbugs: Google Antigravity uses AI to accelerate the discovery of antimicrobial peptides ↗

Google has launched the intelligent development platform Google Antigravity. By leveraging deep learning models such as ProtNLM, it helps biologists rapidly identify new candidate molecules for antimicrobial peptides, streamlining the research process to address the growing crisis of antibiotic resistance.

How the Databricks AI Research Team uses AI Runtime (AIR) ↗

Databricks AI Research Team reveals its usage as the first customer of AI Runtime (AIR): AIR supports NVIDIA H100 or B300 GPUs and enables training for models ranging from billions to nearly a trillion parameters via supervised fine-tuning, offline reinforcement learning, and online reinforcement learning.

Trump posts AI-edited photos of the White House renovation to show off his work; images flagged by Google SynthID. ↗

Trump posted a set of photos on Truth Social showing the White House renovation. A photo comparing old and new marble columns was flagged by Google's SynthID Detector as created or edited by OpenAI. Netizens pointed out inconsistencies in the number of steps on either side of the renovated photo compared to the original, mismatches with pre-renovation images, nearly identical reflections in windows and positions of baseboards at the column bottoms; The Verge confirmed that uploading such an image would trigger a SynthID flag; the White House has not yet responded.

MIT and the Whitehead Institute research team have decoded the language of intercellular communication using a machine learning model called IRIS. ↗

Professor Pulin Li's team at MIT, along with graduate student Nicholas Hutchins, discovered that each signaling pathway leaves a consistent gene activity fingerprint across cell types. Leveraging this finding, they developed the IRIS machine learning model to trace signals received by embryonic cells throughout development. The paper was published in Nature Methods on September 8.

Hugging Face AI Open Source Blog: Granite Switch 4.1 – Chat, Decisions, and Speech from a Single Endpoint ↗

Claude Code v2.1.301 released: Adds retry watchdog and OpenTelemetry tracing. ↗

Claude Code releases version 2.1.301, introducing CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS to end requests immediately when usage limits reset far in the future, rather than waiting.

GitHub and Logitech collaborate: Using the MX Keypad to explore an intelligent interface for agent programming ↗

GitHub and Logitech have partnered to integrate GitHub Copilot into the MX Keypad. This allows users to map actions such as parameterizing commit prompts, starting or resuming Copilot sessions, invoking IDE commands, and triggering multi-step sequences to programmable keys.

Bluesky's Attie AI tool now includes social mini-games and app generation features ↗

Bluesky's Attie AI tool, which launched with an AI-generated feed feature at the beginning of the year, now supports generating small social applications like games, polls, and dashboards through prompt input. These can be shared as standalone websites.

Anthropic grants employee-level access to METR researchers ↗

Anthropic announces it is providing employee-level access to six researchers at the non-profit organization METR, enabling them to independently evaluate the company's models, security practices, and risk management. METR can use company laptops, office space, and internal tools distributed by Anthropic. The company has waived its right to prevent the publication of evaluation reports but may require redaction for predefined categories such as customer data, attorney-client privilege information, and model architecture.

How Meta's engineering team built the Forum app using AI-native development ↗

Meta's engineering team details how they integrated AI agents as participants in the development process to build the independent Groups application Forum. They adopted a human-on-the-loop model and measured success based on production-grade output verified by tests. Within eight weeks, Forum achieved core functionality alignment with Facebook Groups. Autonomous stress testing cycles during development identified and fixed memory issues, reducing memory usage and page faults by approximately half. Post-launch, agents also assisted in root cause analysis of user feedback.

OpenAI Codex releases rust-v0.163.0 ↗

OpenAI Codex releases rust-v0.163.0. Tagged on October 15 by anp-oai, the release page lists only two resource files without disclosing specific update details.

Bluesky expands its AI app Attie, enabling regular users to build websites and games ↗

At TechCrunch Disrupt 2026, Bluesky announced an expansion of its AI experimentation platform, Attie. The update adds the ability to create interactive experiences including websites, quizzes, games, and polls. These features run on the AT Protocol.

Meta has released the Muse Spark Open open-weight model with 816 billion parameters, positioning itself against Qwen, GLM, and Kimi. ↗

Following the Llama series, Meta returns to the open-weight space with Muse Spark Open, calling it the strongest American open-weight model. It features 816 billion parameters (with 42 billion active), supports up to one million tokens in context, and handles text, images, video, and audio.

Ollama AI open-source update: version 0.40.5 ↗

Source summary: llama.cpp version bump b11669 (#19032)

Adobe and NVIDIA expand partnership with RTX Spark to accelerate creative agent workflows ↗

Adobe and NVIDIA have expanded their partnership by integrating the NVIDIA RTX Spark superchip into creative agent applications such as Photoshop, Premiere, and Substance 3D. AI processing, editing, color grading, and effects speeds are now up to twice as fast.

Bank of America hires 1,000 contractors to fix software vulnerabilities found by AI tools ↗

Solidot is the technology news website of Zhiding.net, mainly serving the open-source free software and tech-news readership, including many developers, enthusiasts, and evangelists of Chinese open-source software. Its slogan is 'Geek knowledge, important stuff'.

Deloitte caps employee reimbursement for monthly Claude usage at $29 ↗

Deloitte implements new fee policy for Claude Enterprise starting October 1: employees are collectively capped at $29 per month. Any usage beyond this limit must be funded by the specific project or team (WBS) the employee is working on. According to internal reports seen by 404 Media, some employees have used up to $500–$1,000 monthly, while others are pushing for upgrades to Premium plans offering 250 tokens per week; some access has been completely revoked.

decomp.dev has shut down because AI makes reverse engineering too easy. ↗

Game reverse engineering collaboration site decomp.dev announces shutdown: AI-driven reverse engineering has disrupted community collaboration by making it impossible to distinguish human-led projects. Administrator Encounter stated that cutting-edge models like GPT-6-Astra make reverse engineering trivial and have enabled malicious forks where AI takes over human open-source projects.

Meta's Muse agent sparks privacy controversy: Your friends might hand over your data for you. ↗

The Verge analyzes that Meta's Muse AI assistant topped the App Store download charts last month, with millions of downloads. Its interface uses dark patterns to encourage users to authorize access to bank accounts, passwords, chat histories, and other large amounts of data.

Google demonstrates end-to-end robot usage with Gemini 4 Argon ↗

The Decoder explains how Google's previously released Gemini 4 Argon covers the complete workflow for robot control, simulation training, and data evaluation. Spot the robot dog first practiced in MuJoCo simulation—stepping on pedals and tossing cans—before moving to real hardware. A single image generated over 4,500 training segments, boosting real-robot success rates from 56% to 75%.

Hidden concerns behind Anthropic's $16.8 billion Q3 revenue ↗

Azeem Azhar analyzes a leaked private investor briefing from Anthropic: Q3 revenue reached $16.8 billion with an estimated annualized exit value of about $76 billion. However, the year-over-year growth rate slowed from 147% in Q2 to 45%. This slowdown was viewed as underwhelming because market expectations were set at an annualized revenue of $65 billion.

Nvidia has released an open-source multimodal model, Nemotron 3.5 Super VL, focused on enterprise-grade AI agents designed for continuous operation. ↗

Nvidia has released Nemotron 3.5 Super VL, an open-source multimodal model with 120B parameters (12B active per token). Designed for enterprise agents running continuously in the background, it is distilled from Nemotron 3 Ultra and claims to achieve Ultra's accuracy with roughly one-fifth of its parameters.

Microsoft AI reveals details on MAI-Base-2 pretraining: its 2.8T parameter MoE model achieves FLOPs efficiency 4.2 to 12.3 times higher than DeepSeek-V4-Pro-Base. ↗

Microsoft AI announces completion of MAI-Base-2 pretraining: an MoE base model with 80 billion active parameters and 2.8T total parameters. In held-out evaluations, its BPB score is lower than all tested open-weight models, while its pretraining FLOPs efficiency is 4.2 to 12.3 times higher than DeepSeek-V4-Pro-Base.

Apple's marketing head discusses OpenAI hardware: We do not feel threatened. Compared to pursuing AGI, we prioritize user privacy and daily usage experience. ↗

Josiah Vack states that Apple's integration of hardware and software delivers a superior experience. The company aims to serve everyday users rather than hardcore AI enthusiasts. Currently, the focus is on building personalized, privacy-friendly AI, without chasing AGI.

Ollama AI open-source update: v0.40.5-rc2 ↗

Source summary: llama.cpp version bump b11669 (#19032)