2026-10-22
AI Briefings Daily Report.

Kimi Memory AI Band L1 launches: freely switch between two form factors with all-day AI recording. ↗

The band supports continuous AI recording and automatic meeting summaries, plus AI conversations based on personal memory. Original audio is deleted immediately after use; memories are encrypted and stored on the phone, with a battery life of 2 to 5 days. Also launching today is the functionally similar AI Memory Glasses G1.

Stateful vs. Stateless: Managing Context in AI Agent Development ↗

The n8n blog compares the trade-offs between stateful and stateless architectures in building AI agents. Stateful applications maintain continuity by preserving cross-request data such as conversation history and authentication details. In contrast, stateless applications process each request independently, making them easier to distribute across instances, scale, and fault-tolerate. Large Language Models (LLMs) do not automatically retain conversation context between inference requests; an agent's memory must be stored and retrieved by the system outside the model before being injected into subsequent requests. In n8n, you can choose a memory node or state store based on how long the state needs to persist.

Spotify's co-CEOs discuss AI strategy: aiming to be the music service of choice for agents ↗

Spotify explained its AI strategy during the Q3 earnings call. Co-CEO Gustav Söderström stated that Spotify is one of the few categories poised to benefit from the agent era, with the goal of becoming a media service for various agent calls.

How to route requests between self-hosted LLMs and cloud APIs ↗

The n8n Blog posts explain how to route requests between self-hosted LLMs and cloud APIs, offering a hybrid approach.

Blackstone Group President refutes the AI bubble theory, stating that demand is growing in a 'parabolic' fashion ↗

Gray argues that AI is fundamentally different from the 1990s telecom bubble, noting that demand exploded alongside capacity construction. He revealed that Blackstone's AI consulting firm Ode, co-founded with Anthropic, can no longer keep up with client demand.

Moshi CEO Anekha Sokhal Interview: From Energy Derivatives Trader to Building an AI Recommendation Engine for Traders ↗

Moshi CEO Anekha Sokhal told attendees at the AI Engineer New York event that her company is building a recommendation engine for traders that links market data with their positions. She described it as a combination of Bloomberg, Claude, and Netflix.

NVIDIA launches Minerva: End-to-end observability for AI factories ↗

NVIDIA launches the Minerva AIOps platform, which aggregates telemetry data from GPUs, networks, storage, schedulers, and workloads. It uses specialized AI models to turn this data into actionable insights. The platform analyzes billions of data points every hour, using dynamic knowledge graphs to pinpoint components affecting training tasks and consolidating alert storms into single events with potential root causes.

404 Media analyzes how the PHP2Shell incident is reshaping the offensive and defensive landscape of the AI era. ↗

Jason Koebler from 404 Media reviews security experts' reactions over the past month to the panic triggered by AI-driven vulnerability discovery."Black October"Panic. Researchers from Wiz, Calif, and OpenAI used AI to discover the PHP2Shell vulnerability in PHP. The Cloudflare CEO stated that thousands of exploitation attempts were seen within 10 hours before the public announcement, rising to hundreds of thousands within two hours after, and currently reaching millions.

ICRC officials warn that AI is already expanding war's destruction and call for bans on certain autonomous weapons ↗

Laurent Gisel, head of the International Committee of the Red Cross Legal War Team, was interviewed by 404 Media during a UN meeting. He stated that AI is generating and targeting attacks at higher speeds and frequencies, amplifying war destruction and humanitarian harm.

Gates says AI will bring about approximately 20 years of disruptive impact in an interview with Decoder ↗

In an interview with The Verge's Decoder, Bill Gates said society needs to prepare for roughly two decades of AI turbulence. This period could divide the world into those embracing technological frontiers and those rejecting them.

Bill Gates: People must prepare for two decades of disruptive change brought by AI ↗

In an interview with The Verge, Gates stated that the AI transition period could be extremely painful, potentially splitting society into two camps: those embracing cutting-edge technology and those rejecting it. He anticipates that shortages of goods will decrease, but humanity must endure this turbulent phase before ushering in the era of AI prosperity.

Anthropic launches Claude Duet, an AI agent app tailored for professional users ↗

Anthropic has launched Claude Duet, an AI agent application designed for professionals and small businesses. Built on the Claude Opus 5.5 model, it competes with products like Meta Muse and OpenAI Dots.

Remind AI Memory Glasses G1 launched: Weighing as little as 17.9g, it automatically organizes meeting minutes. ↗

These glasses feature all-day AI recording and personal memory recall, automatically organizing meeting notes and tasks with timely reminders. Original audio is deleted immediately after use, offering a combined battery life of 2 to 5 days.

Wallmap AI shares progress on edge-side intelligence moving from conversation to action at the CNCC 2026 Edge Large Model Forum ↗

On October 22, the third Edge AI Forum was held in Chengdu as part of CNCC 2026. Experts from Moat Inference and academic partners gathered to discuss"Defining general-purpose artificial intelligence on the edge."Sharing explorations covering text, voice, multimodal interaction, chip-level training and inference customization, and agents. This showcases on-device large models."Capable of conversation."Moving forward"Capable of perception, reasoning, and action."。

Sierra releases the τ-Entity benchmark to test whether voice agents can accurately collect entity information such as names and phone numbers ↗

Sierra has released the τ-Entity benchmark to measure whether voice agents can accurately collect, identify errors in, and correct details such as names, phone numbers, addresses, and email addresses. The benchmark includes 200 tasks across 10 entity types and covers six languages, similar to τ-i18n.

Gary Marcus comments on US banks facing AI agent disruption: regulators should not focus only on frontier models ↗

According to Axios, cited by Gary Marcus, several major US banks have discovered unauthorized third-party AI agents in customer accounts. Some actions appear independent of user instructions. Bank executives are rushing to restrict system access and prepare for a 'day after' emergency response."day after" Emergency plan.

Wiz reveals ARTEX agent used by Storm-0940 to attack IT and software vendors ↗

Wiz revealed that attackers used the open-source autonomous penetration testing framework ARTEX against multiple IT and software vendors in October 2026. The campaign overlaps with Microsoft's report on China-linked state-sponsored group Storm-0940 and has expanded into supply chain attacks across other software sectors.

Nirva raises $33 million for an AI pendant that listens only to the wearer ↗

Nirva has closed a $25 million pre-Series A round. Combined with an $8 million raise earlier this year, the company has raised a total of $33 million. Its first device is a pendant weighing about 8 grams and lasting up to three days on a single charge. It can also be worn as a bracelet. The device uses user-provided voice samples to build models that filter out other people's voices in the environment, recording only the wearer. By default, audio and transcripts are not retained after processing. Priced at $249, orders begin accepting in November, with shipments to early adopters targeted before year-end.

Reports indicate that DeepSeek completed a funding round of approximately $15 billion just days after US restrictions were imposed on the company. ↗

According to Bloomberg, just two days after the US federal civilian agencies restricted DeepSeek's AI technology, the company raised approximately $15 billion (roughly 100 billion RMB). Potential investors included CATL and Tencent. Earlier in June, DeepSeek raised over $7 billion with a valuation of $50 billion. The company also plans to list on the stock market in China as early as next year and is increasing its investment in AI infrastructure and open-source models.

1Password launches 1Password for Agents, enabling AI agents to log in without ever accessing passwords. ↗

1Password introduces a new feature, 1Password for Agents, which lets AI agents complete logins after user approval but prevents them from seeing or holding the password. The agent verifies its identity with a unique numeric key to 1Password; users approve on their own devices, and 1Password fills in the login details on their behalf, similar to two-factor authentication.

The Verge tests AI shopping agents: Retailer blocks make experience inconsistent ↗

The Verge author tests Meta Muse and other AI shopping agents, finding their performance inconsistent. Amazon has blocked Muse and previously sued Perplexity; JPMorgan Chase refuses external agent access to accounts; Target states it has safeguards against unauthorized agents.

Claude fails to develop antivenom for scorpions: Citrini and Anthropic drug research experiment. ↗

Citrini admitted that no one on the team truly understood the connection between AI and biology. Because the experiment touched upon Claude's biological safety guardrails, the project shifted to collaborate with Anthropic instead.

Baidu World 2026 symposium topics announced: focusing on AI organizational transformation, AI office and knowledge work, and technology industry ecosystem ↗

How Baidu Cloud's Yi Jian Visual Agent assists Yuanyuan Group in inspecting nearly 10,000 stores ↗

Symantec reveals the 'Skynet' model: AI agents hijacking unprotected large language models to launch self-sustaining cyberattacks. ↗

Symantec researchers intercepted an attack agent via a honeypot that embeds large language models at its core. This agent hijacks unprotected model endpoints on the internet to autonomously scan targets, exploit server vulnerabilities, steal database credentials, and deploy mining software, capable of executing up to 100,000 steps continuously. The computational costs are borne by the owners of the compromised models.

Baidu World 2026 symposium topics announced: focusing on AI organizational transformation, AI office and knowledge work, and technology industry ecosystem ↗

Baidu World 2026 announced over 10 parallel symposia focusing on AI-driven organizational transformation, AI in office and knowledge work, and technology-industry-ecosystem integration. The conference will be held in Beijing on November 12.

How Baidu Cloud's Yi Jian Visual Agent assists Yuanyuan Group in inspecting nearly 10,000 stores ↗

Baidu Cloud's Yi Jian Visual Agent has been integrated into Yuanyuan Group store inspections, covering over 80 scenarios. Inspection coverage improved by approximately 30 times, while SOP compliance levels rose by more than 30 percentage points. Yi Jian 6.0 enables business staff to generate visual AI skills by describing service steps in natural language. Both parties plan to pilot Baidu's embodied intelligence cameras in 200 stores.

IDC: Volcano Engine ranks first in China's AI cockpit market with a 45% share ↗

Anthropic's platform engineering team has fully adopted Claude for coding, shifting engineers toward system design and verification. ↗

The team now has Claude handle all coding work, with engineers focusing on design and verification. They are also exploring how AI agents can autonomously modify codebases. Currently, Claude contributes over 80% of the company's production code, prompting adjustments to their hiring strategy.

IDC: Volcano Engine ranks first in China's AI cockpit market with a 45% share ↗

According to IDC's 'China Automotive Cloud Market Tracking, H1 2026', the market size for China's AI cockpit solutions reached 710 million yuan in the first half of 2026. Volcano Engine holds the top position with a 45% share. Vehicles equipped with the Doubao large model are already on the road, totaling 7 million units across 50 brands and 145 models, with an average of 30 million cockpit interactions per day. This year, Volcano Engine launched the 'Doubao Cockpit Assistant' and 'AI Cockpit Kit' based on Agentic AI architecture. The first model featuring these technologies, the Roewe Jiayue 07, began pre-sales in September.

"Kazehamaru" Kazuma Sakurai: AI is spawning countless imitator games; developers with genuine creativity will become even more critical ↗

Kazuma Sakurai noted that the number of new games added to Steam annually has surged from approximately 275 in 2010 to nearly 19,000 today. AI will further drive the proliferation of clone games. However, imitation rarely leads to commercial success; AI cannot replicate Fujio Okada's future innovations. Human originality remains the key.

Anthropic's Claude models now feature comprehensive AI text watermarking: a breakdown of the technology and its implications. ↗

All text output from the latest Anthropic models (including Claude Mythos 5.1, Claude Opus 5.5, and Claude Sonnet 5.5) now includes built-in AI watermarking across Claude Desktop, Claude Code, and the API, aimed at complying with the EU AI Act.

AMD discusses the Ryzen 9 9950X3D2: Gaming alone cannot yet justify dual 3D V-Cache, but AI makes it viable. ↗

AMD executives also stated they will do their best to extend the AM5 platform's lifecycle, which supports at least until 2029. Re-releasing the Ryzen 7 5800X3D offers a final upgrade opportunity for AM4 users.

Wang Xingxing: Physical AI breakthrough signs are visible; fierce competition awaits; collective effort required ↗

"Signs that physical AI is about to break through have already emerged, likely within the next one to two years or even three," said Wang Xingxing. On October 22 in Hangzhou, at the main forum of the 2026 World Physical AI Conference and Unitree Partner Ecosystem Conference, Wang Xingxing of Unitree Technology (6

OpenAI and Scripps Research use thousands of agents to complete whole-brain CRISPR mapping analysis for mice in just 32 minutes ↗

OpenAI partners with PerturbAI and Scripps Research to use thousands of GPT-Rosalind agents to complete whole-brain CRISPR mapping analysis for millions of mouse cells in just 32 minutes—a task that would take months or even a year with human researchers.

Tencent Hunyuan and Fudan University unveil WorkForge, scaling work agent training using real-world workspaces. ↗

Tencent Hunyuan and Fudan University unveil WorkForge, scaling work agent training using real-world workspaces. ↗

A research team from Tencent Hunyuan and Fudan University introduces the WorkForge framework, which builds verifiable agent work environments at scale starting from real-world resources. The project delivers WorkForge-16K, comprising 16,665 environments across 40 professional domains and 60 file types.

Anthropic launches research preview of persistent agent Claude Duet ↗

Anthropic has released a research preview of the standalone app Claude Duet, built on the Claude Opus 5.5 model. It continues working on tasks after users log off, sends daily morning summaries, and supports both text and voice assignments.

Wang Xingxing outlines Unitree's three future strategic directions and announced plans to invest several million yuan annually in the World Physical AI Competition. ↗

The 2026 World Physical AI Conference and Unitree Partner Ecosystem Summit opened today in Hangzhou, Zhejiang. Wang Xingxing, chairman, general manager, and chief technical officer of Unitree Robotics, outlined the company's three key future directions during the event.

iFLYTEK Chairman Liu Qingfeng: Domestic computing power still lags behind global standards by about two generations, but we will inevitably catch up and eventually lead. ↗

iFLYTEK's Global Developer Festival kicked off today (October 22) at 14:00 with the theme 'Liberate Productivity, Release Imagination', featuring coordinated events in Beijing and Hefei, Anhui.

Hugging Face AI Open Source Blog: Ensemble forecasting with IBM Granite time-series models ↗

Four open Granite models, one extensible forecasting interface, and less guesswork about which model to use. Pre-trained time-series models can forecast without training on your dataset, but you still have to choose one. A model that handles smooth industrial signals well may struggle with noisy retail demand or a long seasonal horizon, and the best choice depends on the series and the task.

A PHP vulnerability dormant for 20 years that no one could exploit was turned into a weapon by AI in just one week. ↗

On October 21, 2026, PHP released an emergency security update to fix two core vulnerabilities that had been dormant for over a decade.

NVIDIA partners with six upstream suppliers to boost AI infrastructure capacity ↗

Three of these six partnership agreements focus on substrates, while the other three cover advanced packaging, semiconductor testing, and optical packaging.

iFLYTEK launches SparkAudio 1.0 speech base large model: Supports seamless switching between Chinese and English, real-time speaker identification, and can be used directly within input methods. ↗

During the iFLYTEK Global 1024 Developer Festival this afternoon, iFLYTEK announced its first launch of the SparkAudio 1.0 speech base large model.

Huawei AI glasses receive HarmonyOS 6.0.0.159 SP5 update: Adds head-motion payment and other features; continuous recording time extended to 5 minutes. ↗

According to blogger @AdakFengLangJuXu, Huawei AI glasses have now enabled the HarmonyOS 6.0.0.159 SP5 update. The new version upgrades payment capabilities with added parking lot QR code scanning and head-motion payments, while continuous recording time has also increased to 5 minutes.

The launch of the iFLYTEK Spark X2.6 end-cloud integrated large model marks a breakthrough in coding and agent capabilities achieved for the first time using entirely domestic computing power. ↗

During the iFLYTEK Global 1024 Developer Festival today, the iFLYTEK Spark X2.6 end-cloud integrated large model was officially launched.

SoftBank's stock price has nearly halved since the June peak as its AI investment model of burning cash faces scrutiny from investors ↗

On the local date of October 22, SoftBank's stock price fell by 6.7% during trading, reaching a cumulative drop of 50% from its all-time high set on June 2. Financial results from other US mega-cap companies also failed to quell market doubts.

How NVIDIA DSX Helps Enterprises Overcome Power Bottlenecks in AI Factories ↗

...Phaidra delivers approximately 10% more computing power at the same power consumption.

Claude Code v2.1.308 Released ↗

Anthropic has released Claude Code v2.1.308. This update focuses on bug fixes and reliability improvements, with no new features or model changes.

Google CEO Sundar Pichai: The advantage of AI agents is working seven days a week, 24 hours a day without interruption. ↗

Pichai explained his work style on a podcast: 'If I had to understand what happened in the past 24 hours every day, it would easily take me an hour. Now, AI agents organize and summarize information for me to review. Of course, I can still dive deep into details, and often do so.'

Wang Xingxing of Unitree Technology: Physical AI is already showing signs of breakthrough ↗

At the 2026 World Physical AI Conference and Unitree Partner Ecosystem Summit, Wang Xingxing noted that global attention and faith in AI remain insufficient. He pointed out that with the release of GPT-6, physical AI is showing signs of breakthrough. The core challenge for embodied intelligence to surpass the ChatGPT moment lies in aligning AI model inputs and outputs with real robots. Commercial service deployment will be faster than adoption in homes and industry.

Thomas Wang of Jiri Yixiu details the ISHO iFlow AI climate control system: unobtrusive monitoring, intelligent temperature control, dual-zone independence ↗

Today morning, the founder of Jiri Yixiu Technology, @Thomas Wang Teng, detailed the company's second product: the ISHO iFlow AI climate control system.

Jensen Huang on NVIDIA: GPU Lifespan Extends to Over Eight Years as AI Rewrites Hardware Depreciation Rules ↗

Tech media outlet Wccftech published an article yesterday (October 21) reporting that at the GTC Berlin conference, NVIDIA CEO Jensen Huang stated that as software continues to evolve and AI application demand expands, the actual lifespan of GPUs has been extended from the previously assumed four years to eight years or more.

After surging sevenfold, MiniMax stock price drops below IPO offering price ↗

MiniMax fell more than 7% in today's Hong Kong stock market trading, dropping below its IPO price of HK$165 and losing over 80% of its market value from the March high of HK$1,330. The company faces a profitability crisis amid pressure from DeepSeek's low-cost models and OpenAI's price cuts.

Challenging Silicon Valley giants: reports say DeepSeek raised nearly 100 billion yuan in its latest funding round ↗

According to sources close to the deal, DeepSeek adjusted its fundraising targets at least twice during negotiations with investors. The financing scale increased from an initial plan of 50 billion yuan to 80 billion yuan, ultimately reaching approximately 100 billion yuan. CATL and Tencent Holdings were among the major investors in this round.

Trump prescribes a 'cure' for AI data centers: To get projects on the ground, you must first pay dividends to the community. ↗

Trump posted on Truth Social, demanding that data centers and AI cloud providers pay profit dividends to their local communities. He also proposed generating power independently and feeding excess electricity back into the grid in exchange for project support.

AI Power: SpaceX, Anthropic, and NVIDIA Help Tech Stocks Escape Slump ↗

Elon Musk announced that SpaceX has achieved profitability and hinted at a potential merger with Tesla. A draft IPO filing for Anthropic disclosed by The Wall Street Journal reveals rapid revenue growth. NVIDIA CEO Jensen Huang released optimistic signals at the company's conference in Berlin.

Embarking on the "largest IPO in history": Reports indicate Anthropic will publicly file its listing application by this Friday. ↗

Anthropic's planned IPO has attracted significant attention, expected to surpass SpaceX's June IPO in scale. As investor concerns over AI industry spending intensify, the market widely views Anthropic's listing as a critical test of how much interest investors still hold for the AI boom.

Reuters: Tesla's AI transformation faces headwinds; full-year spending target for 2026 is projected to drop by about 20% ↗

Reuters published a post yesterday (October 21) reporting that Tesla's capital expenditure for the third quarter was $4.73 billion (IT Home note: approximately 31.64 billion RMB at current exchange rates), which is more than double the year-over-year figure but significantly below market expectations of roughly $7.7 to $7.95 billion (approximately 51.508 to 53.18 billion RMB at current exchange rates).

Siemens Healthineers and NVIDIA officially partner: Using AI to advance cancer treatment and develop next-generation adaptive radiotherapy ↗

Siemens Healthineers and NVIDIA have announced a partnership today to jointly develop next-generation adaptive radiotherapy technology. The collaboration combines agents and robotics intelligence with imaging and oncology expertise, integrating accelerated computing and simulation to streamline workflows and improve healthcare accessibility.

Hey OnStar arrives: General Motors launches OnStar AI to understand vehicles and comprehend owner instructions. ↗

General Motors (GM) posted on October 21, announcing an upgrade to OnStar (known as Angeli Star in China). The goal is to build a "connected intelligent platform" for the era of autonomous driving, featuring a new OnStar AI assistant. This native AI assistant will first be deployed via OTA updates on the 2027 Cadillac Lyriq, before gradually expanding to more GM brand models.

Kazak explains the PHP2Shell vulnerability unearthed by AI: Discovery took only two days, while fixing it took two months. ↗

Ollama AI open-source update: v0.40.9-rc1 ↗

Source summary: llama.cpp update (#19194)

Kazak explains the PHP2Shell vulnerability unearthed by AI: Discovery took only two days, while fixing it took two months. ↗

The author reviews the full timeline of two PHP remote code execution vulnerabilities (CVE-2026-109618 and CVE-2026-109619, collectively known as PHP2Shell) that started with a CVSS score of 10. The review covers the initial data leak, the release of demonstration videos, the preemptive shutdowns of government websites across multiple countries, and the deployment of patches at 10 PM on October 21. It is also noted that these vulnerabilities were discovered with assistance from GPT-5.6 Sol, and OpenAI has since released a detection skill.

Vercel has open-sourced durable-harness to provide crash recovery capabilities for coding agents. ↗

Vercel has open-sourced durable-harness by logging model and tool calls persistently, allowing coding agents to resume from new functions after crashes or timeouts. Completed calls are read from logs and will not be billed twice.

GM bets on AI and new battery technology, aiming to cut EV R&D cycles in half ↗

On October 21 local time, GM President Mark Reuss reiterated at the GM Forward event in New York that the company will not abandon its electric vehicle business. He plans to launch products better suited to US market demands to further enhance competitiveness.

General Motors OnStar service gets major upgrade: New AI assistant helps drivers control vehicles and troubleshoot issues. ↗

General Motors will upgrade its OnStar connected car service with an AI assistant, allowing owners to query vehicle details directly, control functions via voice, and use AI to diagnose faults.

Approximately 70% of websites affected: PHP high-risk vulnerability fixed, OpenAI leads global discovery and remediation ↗

OpenAI released a post yesterday (October 21) announcing its 'Patch the Planet' initiative. The program helped security researchers identify two critical vulnerabilities in the PHP programming language. OpenAI has coordinated with the PHP maintenance team, security firms, cloud providers, and cybersecurity agencies across multiple countries to complete approximately two months of coordinated remediation before public disclosure.

Claude Code v2.1.307 has been released, introducing the /claude-api command and security event telemetry. ↗

Claude Code v2.1.307 introduces the /claude-api command, enabling migration of Claude API code to version 2026-10-21 and adding a claude_code.safety_event OpenTelemetry event for each tool call.

Sierra releases and opens the source code for its τ-i18n benchmark evaluating multilingual voice agents ↗

Sierra has released 𝜏-i18n, extending 𝜏-Voice to Spanish, Brazilian Portuguese, Hindi, Korean, and Mandarin Chinese. This allows cross-lingual comparison of tasks in aviation, retail, and telecommunications.

Walmart launches its own agent gateway; AI can place orders for users. ↗

Walmart has launched an agent gateway and agent profile feature. AI can check inventory and place orders on behalf of users. Users can set spending limits, approve transactions individually, and revoke authorization at any time. Rules are clear: agents cannot operate beyond their permissions, and return policies still apply. #AI E-commerce

Jas Khaira of Blackstone believes concerns about AI infrastructure debt are exaggerated. ↗

Blackstone veteran Jas Khaira told TechCrunch Disrupt that worries over AI infrastructure debt are overstated. Most data centers are built for hyperscale cloud providers, with debt backed by the highest-rated companies.

Working for you online 24/7: Anthropic launches Claude Duet AI agent to challenge Meta and OpenAI ↗

Yesterday (October 21), Anthropic announced the research preview of a new application called Claude Duet. Positioned as an AI agent capable of taking initiative, it joins the 'always-on proactive AI assistant' competition led by Meta's Muse and OpenAI's Dots.

OpenAI unifies security standards for its model services and updates usage policies, effective November 20, 2026 ↗

OpenAI has aligned its security policies with partners and updated its Usage Policies. The new rules require that services offering direct model access, such as model routers or AI gateways, undergo OpenAI's approval. These changes take effect on November 20, 2026.

Meta has announced the issuance of a new class of non-voting stock to raise funds for its AI initiatives ↗

In its third-quarter earnings report released on Wednesday, Meta announced the issuance of a new non-voting stock class to fund AI initiatives, potential acquisitions, and employee compensation. The transaction is scheduled to begin in 2027. CFO Susan Li stated this will provide capital flexibility for investments in AI-driven businesses; Meta expects to raise its capital expenditure forecast for the year to $135 billion to $145 billion.

Anthropic launches Claude Duet, an AI agent for professionals, in research preview mode ↗

Anthropic launches Claude Duet, an AI agent competing with Meta's Muse and OpenAI's Dots. Targeting professionals and small business owners, it is available as a research preview via standalone iOS and Android apps and web interfaces.

Ollama AI open-source update: v0.40.8 ↗

Source summary: Recursive Flint Flint brings Recursive's coding model to Ollama. It is designed for the work coding agents do across a

Recursive released Flint, a programming model that runs on laptops, claiming it was primarily built by AI. ↗

AI startup Recursive released Flint, a free programming model that runs on laptops. The company states its automation research system, Eureka Machine, participated in nearly all development stages from pre-training to post-training and inference. A small team completed this in under four months starting from scratch.

Ollama AI open-source update: v0.40.9-rc0: ci: switch to released WoA CTK (#19185) ↗

Source summary: We were still using a pre-release version, and the download site is unreliable, so let's get up to speed with the latest CUDA support.

Apple Music unveils a new AI-driven brand identity, highlighting the value of human design in the age of artificial intelligence. ↗

Apple reveals its new Apple Music brand identity using the Liquid Glass design language and releases a behind-the-scenes video. The short features TBWA/Media Arts Lab and producer Fred again... alongside an installation of transparent panels filled with mineral oil. Transducers convert audio signals into ripples on the oil surface, capturing interactions between the brand logo and the fluid.

Anthropic launches Claude Duet app, opening a waitlist for US Pro and Max users. ↗

Anthropic launches Claude Duet, an app that proactively prompts users to focus on key tasks and helps them continue working after putting down their phones. It determines which tasks require immediate attention and which can wait, confirming before taking action on behalf of the user. The research preview is available via a waitlist for US Pro and Max subscribers at https://claude.com/claude-duet.

Cerebral Valley AI Summit announces panelists for discussion sessions: Emmett Shear, Simile, SambaNova, and the founder of WorkOS will lead. ↗

The Cerebral Valley AI Summit takes place in San Francisco on November 12. Mercor CEO Brendan Foody joins the main stage lineup, while Emmett Shear, Joon Sung Park from Simile, Rodrigo Liang from SambaNova, and Michael Grinich from WorkOS will lead small discussion sessions.

Walmart launches Agent Gateway for AI agent shopping, contrasting with Amazon's restricted approach ↗

Walmart announced Wednesday the launch of Agent Gateway, a server based on the MCP protocol that allows approved AI agents to search Walmart inventory and place orders on behalf of users, covering both Walmart and Sam...'s Club。

GPT-5.6 Sol and Gemini 4 Argon uncover critical PHP vulnerabilities affecting millions of websites ↗

Large language models including OpenAI's GPT-5.6 Sol and Google's Gemini 4 Argon have discovered a critical PHP vulnerability impacting tens of millions of websites worldwide. OpenAI describes it as one of the most significant security issues in its Patch-the-Planet initiative.

Walmart to allow external AI agents to shop for users starting next month ↗

Walmart is set to allow users to connect their own AI agents to its shopping site starting next month, with support for Sam's Club.'Starting next month, Walmart will let users link their personal AI agents to its online store, including Sam's Club. Through new Agent Profiles, shoppers can approve purchases, set spending limits (such as a maximum of $100), and turn off the agent at any time. Walmart is also building an Agent Gateway based on the MCP open standard, enabling agents to bypass the browser and connect directly without exposing shopping history or personal data.

OpenAI GPT 6.1 series models set to launch on Google Cloud and Antigravity ↗

OpenAI and Google have partnered. All GPT 6.1 models (Astra, Sol, Luna) will launch in the coming weeks on Google Cloud's Model Garden and appear in its agent development tool, Antigravity.

SemiAnalysis examines the symbiotic relationship between Huawei and DeepSeek: Ascend 950PR benchmark tests and DeepSeek open-source adaptation. ↗

SemiAnalysis publishes a long-form analysis of the complementary relationship between Huawei and DeepSeek behind the addition of support for six repositories—TileLang, DeepGEMM, DeepEP, and FlashMLA—on Ascend by DeepSeek on September 30.

Hugging Face Blog: NVIDIA Kumo Relational Outperforms Supervised Machine Learning on Relational Data, No Training Required ↗

NVIDIA Kumo Relational is a pre-trained relational foundation model that makes predictions directly on enterprise databases. Given a predictive query, it gathers up to 10,000 historical examples as context and predicts outcomes such as churn, demand, or fraud in a single forward pass. It needs no task-specific training, no feature engineering, and no flattening of tables. Kumo Relational is the first few-shot foundation model to outperform supervised machine learning on common relational benchmarks. Across 41 tasks on 15 databases, it beats sup

Hugging Face AI Open Source Blog: Train in Simulation, Deploy on a Real Robot: Microduck with Isaac Lab 3.0 ↗

This post was written in collaboration with Antoine Pirrone, Augustin Crampette, Matthieu Lapeyre, Coralie Deplanne, Anne-Charlotte Passanisi, and Thomas Wolf.

Hugging Face AI Open Source Blog: How Specializing the Vocabulary of SLMs Can Cut Inference Cost by 77% in Real Applications ↗

Models specialized for a domain or a task by fine-tuning general open-source models, such as Qwen and Gemma, inherit a vocabulary built for many domains and general applications.

Google Cloud has launched OpenAI models, becoming the leading cloud platform to offer Gemini, Claude, and OpenAI models simultaneously. ↗

Google Cloud has released three new OpenAI models—GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna—making it the only major cloud provider to offer Gemini, Claude, and OpenAI models on a single platform.

How GitHub is building a transitional architecture to handle Git traffic at the scale of AI agents ↗

GitHub is migrating its busiest repositories to Azure by serializing write arbitration from the entire repository down to individual references. This resolves throughput bottlenecks caused by cross-region round trips of approximately 25ms. Internal traffic tests show a 3–4x increase in write throughput for different references with lower latency. Currently, two out of five replicas for most repositories are located in Azure, which handles about half of the Git traffic.

Hugging Face AI Blog: Profiling TorchNeuron (Part 1): System-level profiling on AWS Trainium ↗

Many ML practitioners have some intuition about how to make models run fast, but fewer know how to analyze why a model is slow. In Part 1 of this series, you will learn how to systematically analyze inference performance and identify where the time is spent.

Ollama AI open-source update: v0.40.8 ↗

Source summary: Recursive Flint Flint brings Recursive's coding model to Ollama. It is designed for the work coding agents do across a

Hugging Face AI Blog: Profiling TorchNeuron (Part 1): System-level profiling on AWS Trainium ↗

Many ML practitioners have some intuition about how to make models run fast, but fewer know how to analyze why a model is slow. In Part 1 of this series, you will learn how to systematically analyze inference performance and identify where the time is spent.