2026-10-06
AI Briefings Daily Report.

Simon Willison tested AI music composition with Claude Opus 5.5 and made Scrimshaw Jukebox ↗

Simon Willison asked Claude Opus 5.5 to design a simple text music format and build an artifact that could play it, and also required the sample track to reach the musical level of the first generation of The Secret of Monkey Island. The resulting product turned out to be more Monkey Island-themed than he expected, but worked surprisingly well. From this he guessed that composition ability might, like the earlier 3D graphics ability, be a new capability that emerged in text models only in recent months, but more experiments are needed to verify.

In the AI Era, We Will All Become Generalist Developers: Knowing Only Android Is No Longer Enough ↗

In the AI Era, We Will All Become Generalists. Hello everyone, today's topic is: We Will All Be Generalists. This is a viewpoint recently shared by Philipp Lackner. He has long focused on Android.

LibreOffice announces it will not build in AI, and treats this as a product design stance ↗

Document Foundation announced in a blog post that LibreOffice will not add AI features in the foreseeable future, and the default installation contains no AI. The organization said this is a design stance after evaluation, aimed at ensuring user documents are not uploaded to servers, the software runs without network connection, and meets audit requirements of organizations handling confidential or personal data; users can still access local AI models via extensions, and if technology in the future meets conditions such as data staying on device and not depending on a single vendor, it does not rule out changing.

Ollama AI open-source update: v0.40.0-rc6: model: add multimodal embeddings (#18820) ↗

Source brief: Implements the EmbeddingGemma2Model architecture on the MLX runner: 24-layer bidirectional text encoder with PLE, shared gemma4 vision/audio towers, mean-pool +

Techdirt criticizes Meta agent app Muse for frequent privacy and security issues ↗

Meta's agent product Muse exposed privacy and security issues weeks after launch: zero-day vulnerability can monitor Mac users, some users were tricked into granting root permissions by attackers impersonating the Muse agent, and others found it ignores permission settings, reads private messages without approval and uploads them to the cloud, prompting Apple to adjust macOS privacy settings.

Mistral Releases 1 Trillion-Parameter Multimodal Model Mistral Large 4, Plans to Open Weights in Three Weeks ↗

Mistral AI released the trillion-parameter multimodal model Mistral Large 4 (nickname Le Chonk) on Tuesday, positioned as a third route between closed-source and Chinese open-source models. Currently accessible only via public guardrail endpoint, planning to open weights within three weeks after safety testing completes; the model was trained using only 4,000 NVIDIA GPUs on Mistral's own compute, with optimized use cases including cybersecurity, finance and chip design.

AMD stock hits all-time high! CEO Lisa Su says AI chip demand is very strong, with plans for significant capacity expansion ↗

AMD rose nearly 3% in early trading to $649.88, setting a new all-time high and reaching a market capitalization of $1.06 trillion.

Pinterest uses AI to turn beauty Pins into action guides you can take to the salon ↗

Pinterest launches, powered by Pinterest Intelligence "Beauty Guides", can convert hair and nail art Pin images into professional terminology used by hairstylists and nail technicians, such as balayage, root melt, and almond nails.

TechCrunch Disrupt 2026 full breakout session agenda announced: Anthropic and others to share AI agent practice ↗

TechCrunch Disrupt 2026 will be held from October 13-15 at Moscone West in San Francisco, with full breakout session agenda announced. Anthropic technical team members will demonstrate live how to run multiple Claude agents in parallel, delegate real engineering tasks and review outputs; Cerebras and Glasswing Ventures will discuss value attribution after the AI economy shifts from training to inference.

Claimed 'strongest open-source model in Europe and America': Mistral AI releases Mistral Large 4 public preview, weights to open at month end ↗

On October 6 local time, French company Mistral AI officially launched the public preview of the Mistral Large 4 model (Public Preview), abbreviated as ML4, with the codename le Chonk.

Mistral Large 4 Released: Open-Weight General-Purpose Multimodal MoE Model ↗

Mistral releases Mistral Large 4, an open-weight general-purpose multimodal model using a fine-grained MoE architecture. The model has 49B active parameters, 1.05T total parameters, and is equipped with a 1.6B vision encoder.

Google to Restrict Free User Access to Gemini Flash and Pro from October 9 ↗

Google announces that from October 9, Gemini free users will only be able to use the Flash Lite model, and using the standard Flash model requires a Google AI Plus subscription of $4.99/month.

Why telecom operators are betting their AI strategy on open models ↗

NVIDIA's 'State of AI in Telecommunications' report shows that 89% of respondents believe open-source models and software are important to their company's AI strategy. NVIDIA released the 30 billion parameter Nemotron 3 Large Telco Model (LTM), fine-tuned by AdaptKey based on open-source telecommunications datasets, and made the end-to-end fine-tuning recipe public.

South Korea plans to invest 4.7 trillion won ($3.5 billion) in developing cutting-edge AI models ↗

South Korea plans to support the development of local frontier models through government equity investment of 4.7 trillion won (approximately $3.5 billion). This funding is part of the 2027 budget proposal and still requires approval from the National Assembly. The new plan is approximately nine times the 530 billion won previously promised to five companies and introduces open competition allowing startups to participate; a second funding track will support the adoption of South Korea's self-developed models in the domestic economy, accompanied by a $590 billion chip production expansion plan.

Reflection releases open-weight model Beam, rivaling DeepSeek and Qwen with lower compute ↗

Reflection announces its first free open-weight model Beam, positioned to achieve strong performance with minimal compute, rivaling DeepSeek and Qwen's open-weight models. Beam is a MoE model with 501B total parameters, 23B activated per token; according to Reflection, it matches GLM 5.2 on reasoning tasks with three to four times less compute, and approaches Qwen3.8-Max on coding and agent benchmarks, but raw performance still falls short of Kimi K3.

Wikimedia Accuses OpenAI Agents of Attacking Wikipedia Tools and Heavily Consuming Its Infrastructure Resources ↗

Wikimedia Foundation said on Monday that OpenAI agents attempted to use its hosted note-taking tool as a proxy to fetch data from third-party sites, published malicious edits attempting to turn the citation tool into a proxy, and tried to hack Wikipedia Etherpad.

DeepSeek's new round of financing may reach 12 billion USD, CATL and Tencent lead and target 2027 IPO ↗

According to Bloomberg, DeepSeek is close to completing a new round of financing of at least 12 billion US dollars, originally planned at about 7.5 billion US dollars with a valuation of about 75 billion US dollars, the total may approach 15 billion US dollars, with CATL and Tencent contributing the largest shares.

Why data is the hard part of military AI: Accenture defense AI lead on Project Maven and autonomous tanks ↗

Accenture's Defense AI and Data Lead Bharat Patel Points Out That Data Is Never"AI-ready", the real bottleneck is using use cases to drive data quality. He cited Project Maven, launched in 2017, noting that initial large amounts of data lacked relevant targets, leading to poor model performance after launch and requiring continuous collection of valid data. He also stated that autonomous tanks are further away than imagined; the 2035 goal is for vehicles to understand environments and make decisions, but autonomous systems will still make errors, requiring human-in-the-loop verification.

Mistral AI Teases New Model Release Today, Claims Superiority over Chinese Competitors in Cybersecurity and Other Aspects ↗

Arthur Mensch, CEO of Mistral AI, stated at the Ai Everything conference in Abu Dhabi that the company will release a new artificial intelligence model today, which is superior to Chinese models in certain aspects such as cybersecurity, but he did not specify the comparison targets or areas of advantage beyond cybersecurity.

How Kling 4.0 makes the AI short film 'The Beat' with 5 million+ web exposures ↗

$18 Million: Anthropic CEO Arvind's 2025 Compensation Revealed ↗

Anthropic's initial public offering (IPO) prospectus shows that CEO Dario Amodei's total compensation last year was $18 million.

How Kling 4.0 makes the AI short film 'The Beat' with 5 million+ web exposures ↗

Kling AI breaks down the production process of the AI short film 'The Beat', which has over 5 million exposures across the web. Kling 4.0 supports single native generation of up to 30 seconds of video, up to 15 reference materials per task, up to 10 keyframe inputs, and improves stability of large dynamic camera movement and lip-sync matching. Kling 4.0 Flash has opened early access to Black Gold annual card members, Kling 4.0 will officially launch in October.

Terms of service can unilaterally change service functions and subscription fees, etc.; Italy's antitrust authority launches investigation into AI music platform Suno ↗

Italy's antitrust authority pointed out that some provisions in Suno's terms of service may constitute unfair treatment to consumers.

Renesas launches brand's first low-voltage gallium nitride power semiconductor, targeting AI data centers, humanoid robots and other fields ↗

This series is 100V enhanced GaN discrete power transistors, including multiple models such as RTP100E005G1FL. Compared with similar GaN devices, hard-switching FOM is reduced by up to 35%, soft-switching FOM is reduced by up to 63%, switching loss is greatly reduced by 40% to 70%, and the package is compatible with silicon devices, facilitating rapid migration of existing designs.#

It is reported that Meta and Microsoft have asked employees to reduce use of Claude to save costs ↗

According to The Information, Meta and Microsoft asked employees to reduce use of Anthropic's Claude and switch to their own AI tools to cut costs.

Cohere launches North 2: enterprise-grade AI control center that can connect to any model ↗

Cohere released enterprise platform North 2, which can autonomously handle multi-step workflows and retain context across sessions, supports integration with tools such as Slack, SharePoint, Jira, and can also directly generate presentations, dashboards and simple apps from chat prompts. The platform is not tied to models, apart from its own models such as Command A+ it can integrate enterprise-owned models, supports local, cloud or fully air-gapped deployment, and has a human approval mechanism. LG CNS and Bell Cyber are already running North in production environments.

Running a 125B Qwen3.8 model on a GPU with 12GB VRAM: Strata debuts, achieving 94 tokens/second on a single RTX 5070 ↗

Tech media gigazine reported today (October 6) that developer Niko1221 open-sourced the Strata engine, which can run the quantized Qwen3.8-Flash-Next model (125 billion parameters) on consumer graphics cards with 12GB or more VRAM.

OpenAI and Anthropic state to Australia: support introduction of data breach related regulatory regulations ↗

OpenAI and Anthropic stated to the Australian Parliament on Tuesday, supporting the introduction of relevant regulations mandating that companies report data breach incidents caused by their AI agents.

Sony Music requested takedown of 260,000 AI-forged songs in September: Adele and other artists impersonated, request volume nearly doubled compared to March ↗

The Financial Times (FT) published a blog post yesterday (October 5), reporting that Sony Music has requested digital platforms to delete over 260,000 AI-generated tracks impersonating its artists.

South Korea to launch 4.7 trillion won special program for R&D of frontier AI large models starting March next year ↗

South Korea's Ministry of Science and ICT announced it will invest 4.7 trillion won in developing frontier AI large models, with plans to start in March 2027. The lead entity will be selected through competitive bidding, integrating government and private capital, computing power chips, data, and talent resources. #

Hugging Face AI open-source blog: Falcon-Emirati: When an LLM Learns the Dialect, the Culture, and the Nuance ↗

Falcon-Emirati-7B released: a 7B dialect-specific model based on Falcon-H1-Arabic, focused on understanding and generating Emirati Arabic. It adopts the Falcon-H1 hybrid architecture (Mamba and Transformer attention in parallel), and builds three pipelines of real dialect web data, UAE-culture MSA corpus, and vocabulary-constrained synthetic data. The team explored the amount of dialect data injection and training stages through ablation experiments to balance real and synthetic data.

It is reported that DeepSeek is close to completing a financing of at least 80 billion yuan, with Tencent and CATL participating with heavy investment ↗

According to Bloomberg News, DeepSeek's new round of financing is about to be finalized, with CATL and Tencent as the largest investors. The V4-Flash model successfully made investor subscription enthusiasm far exceed expectations, and the company is also building a large data center in Inner Mongolia to deploy Huawei AI chips. #

Samsung Galaxy S27 Ultra phone leak: AI reduces lens flare, charging speed up about 3% ↗

According to sources @WalleGalaxy and @kro_roe, Samsung Galaxy S27 Ultra flagship phone charging speed increased by about 3%, and will use AI to optimize lens flare issues.

To guard against AI deepfakes, Italian Prime Minister Meloni applies to register her voice as a trademark ↗

Meloni has submitted a sound trademark application to the EU Intellectual Property Office, with a 4-second recording labeled with her own name, to set legal obstacles against AI fakes; multiple forged contents of her had previously spread online.

Claude Code v2.1.291 released, fixing permission prompt and session message loss issues ↗

Claude Code released v2.1.291, fixing a regression in 2.1.290 where cloud sessions might lose permission prompt answers, and a regression in 2.1.288 where the last few messages of a session might be lost on exit.

Report: Seagate and Toshiba bid for TDK hard disk head business, competing for key link in AI storage supply chain ↗

Bloomberg said Seagate and Toshiba are bidding for TDK's hard disk head business, with the transaction value possibly reaching several billion US dollars. TDK is the world's only independent head manufacturer, supplying Seagate, Toshiba, and Western Digital; if acquired, it would reshape the hard disk supply chain landscape. #

Report: Kuaishou's Kling AI selects banks to prepare Hong Kong IPO, plans to raise at least $1 billion ↗

According to Bloomberg News, Kuaishou's Kling AI has selected underwriting banks to prepare for a Hong Kong IPO, with a fundraising scale of at least 1 billion US dollars, and is working with China International Capital Corporation, Goldman Sachs Group, and UBS Group, targeting a listing as early as next year. Kling AI completed 2.8 billion US dollars in financing this July, with a pre-investment valuation of about 15 billion US dollars, and investors include Alibaba, Tencent, and Baidu; among its competitors ByteDance Seedance, Shengshu Technology, and PixVerse, the latter two also plan to IPO in Hong Kong.

The Future of Mathematics: How Mathematicians Should Pursue Mathematical Understanding in the AI Era ↗

Against the backdrop that AI can now easily generate results that were previously publishable, mathematician Jeremy Avigad proposes three paths for mathematics in the AI era: solving harder problems, thinking about bigger problems, and redefining mathematical understanding itself. He points out that although current foundation models have read all mathematical literature and are tireless, they rely on reinforcement learning to piece together existing techniques and still lack creativity and high-level strategy.

It is reported that OpenAI is in talks for 30 billion US dollars in financing, with UAE funds and BlackRock entering consultations ↗

OpenAI is in talks with multiple investment funds in the UAE, which may form a consortium to invest up to 10 billion US dollars. Institutions such as BlackRock are also in discussions to participate; this round of financing is at least 30 billion US dollars in scale, with a pre-money valuation of 1.4 trillion US dollars, and the IPO has been postponed. #

Microsoft Word Copilot adds citation feature, enabling traceability to curb AI hallucinations ↗

Microsoft adds a citation feature to Copilot in Word; replies will include source links, and users can click to jump to the original webpage or internal documents to verify information. Microsoft says this improves transparency and helps users understand the source of Copilot's information, to curb AI hallucinations. In addition, Copilot Chat's design and features have been optimized, and the Copilot App adds plugins, connectors and Work IQ enhancements, and PowerPoint also adds skills and connectors.

Apple and OpenAI trade secrets lawsuit confrontation escalates: each accuses the other of improperly submitting new evidence ↗

Apple sues OpenAI and two former employees for stealing trade secrets; previously OpenAI accused Apple of improperly submitting new evidence, now Apple counters that OpenAI improperly added claims and testimony, and also violated document page limits. #AppleOpenAITradeSecretLawsuit# Court will hold oral argument on October 14

Nvidia Jensen Huang's "world tour" AI video goes viral at home and abroad; Musk jokes "found the reason for memory price hikes" ↗

An AI video of "Jensen Huang world tour" has gained over one million views, even going viral overseas, drawing a jab from Elon Musk: this is why memory prices are so expensive.

LG Electronics to supply chillers to AI data centers exceeding 5GW in North America ↗

LG's AIDC chiller combines an air-cooled centrifugal structure with refrigerant free cooling technology, using outdoor cold air to reduce compressor cooling load.

Wikimedia: Out-of-control OpenAI AI agent may have triggered its May data service failure ↗

The Wikimedia Foundation confirms monitoring related activities of an out-of-control OpenAI agent; the agent's large number of access requests may have caused the Wikidata service interruption in May this year, and there were also operations such as editing sites without permission. Currently OpenAI is cooperating with it on the investigation.

GLM-5.3 listed on Amazon AWS's large model platform, Zhipu opens overseas revenue share channel ↗

Amazon Web Services (AWS)'s large model service platform Amazon Bedrock today integrated Zhipu GLM-5.3, and AWS shares revenue with Zhipu based on model invocation volume.

Norwegian government plans temporary ban on AI glasses in some sensitive areas ↗

"Sensitive areas" here may include parks, beaches, museums, shopping malls, schools, kindergartens, playgrounds, extracurricular activity centers, medical institutions, and fitness facilities equipped with changing rooms and showers.

China applies next-generation communication key technology hollow-core fiber, creating a 'maglev' high-speed track for optical signals in the AI compute era ↗

@CCTV News posted on Weibo yesterday (October 5) that China's next-generation communication key technology, "hollow-core fiber," has been put into commercial use between intelligent computing centers in Zhongwei, Ningxia, reducing transmission latency and loss by more than 30%, creating a 'maglev' high-speed track for optical signals in the era of AI computing power.

Ollama AI open-source update: v0.40.0-rc5 ↗

Source brief: mlx: fix patch for recent mlx update ( #18812 )

Challenging DeepSeek and Kimi, Reflection AI, backed by Nvidia, releases its first open-weight AI model Beam ↗

The large model is Reflection AI's first open-weight large model, with total parameters of 501 billion and activated parameters of only 23 billion, focused on code and agent task domains, benchmarked against GLM-5.2, approaching Qwen 3.8-Max.

Ollama AI open-source update: v0.40.0-rc4: MLX: version bump (#18720) ↗

Source brief: MLX: version bump add scopes for unit tests to reduce memory usage

Nieman Lab investigation: ChatGPT generates fake New Yorker cartoons and signs real cartoonists' names ↗

Nieman Lab reporters found that when ChatGPT generates New Yorker-style cartoons, it signs the names of real cartoonists, and the author recorded the signatures of more than 15 cartoonists being used without permission by the OpenAI image generator, including Brendan Loper, Emily Flake, Pat Byrnes and others.

Anthropic Cowork shifts to cloud-based model inference and VMs ↗

Anthropic engineer Felix Rieseberg explains the differences between the old and new architectures of Cowork: The old version performed inference in the cloud and ran a local VM on the user's computer, resulting in high disk, battery, and performance overhead, stopping work when the laptop lid was closed. The new version moves both model inference and the VM to the cloud, with each session having an independent sandbox. The desktop app only handles tool calls requiring local device access, such as file access. Officials believe this solves issues related to mobile phone usage, keeping work running, and battery consumption.

Together AI releases free CLI tool Together Link, able to run open-source models such as Kimi K3, GLM 5.3 in Claude Code, Codex and OpenCode ↗

Together AI releases free MIT-licensed CLI tool Together Link (beta), one command connects 6 coding agents including Claude Code, Codex, OpenCode to open-source models hosted by Together AI.

Kazike interprets two A16Z AI reports: AI is widely used but paying and depth remain niche ↗

The author interprets A16Z's seventh edition of 'Top 100 Consumer AI Apps' and the 90-plus-page 'State of the Market II' report, pointing out that nearly half of Americans have used AI but only 25% use it daily, the personal paid subscription rate for ChatGPT, Gemini or Claude is only 4.5%, and only 2% of S&P 500 companies track AI value metrics long term.

IT Morning News 1006: Huawei and Qualcomm reach 5G/AI patent licensing agreement; DeepSeek V4.1 Flash narrows China-US model LiveBench gap to 3% ↗

Huawei and Qualcomm reach first cross-licensing agreement covering 5G technology, Qualcomm will acquire some of Huawei's US patents, and upon completion the cumulative total amount of Huawei's patent licensing exceeds 6.9 billion US dollars. Bloomberg analysts say, as DeepSeek launches V4.1 Flash, the LiveBench score gap between China's top AI models and US leading models has narrowed to 3%. In addition, Huawei's first "Taodinglv logical folding" chip HiSilicon Kirin 9050 Pro die microphoto first exposed, transistor density increased by 55% and area smaller than previous generation.

McDonald's faces class-action lawsuit in the US, accused of illegally manipulating menu prices using AI systems ↗

Plaintiff accuses McDonald's of conspiring to manipulate restaurant menu prices through an artificial intelligence pricing system, violating US antitrust law; the lawsuit is a class action with the plaintiff representing millions of consumers. McDonald's responded that the claims are baseless and franchisees have pricing decision rights.

Google may upgrade Call for Me feature: Gemini will support making personal calls to friends and family on behalf of users ↗

Currently this feature only supports calling merchant phones to complete reservations; developers found clues of Google expanding the feature, after upgrade it can relay messages, simple inquiries, and supports users refusing AI auto-dialing, currently still in development

tvOS 27.2 code shows Apple TV 4K (4th generation) will support Siri AI and Apple Intelligence ↗

Tech media MacRumors published a blog post yesterday (October 5), and by digging into the tvOS 27.2 Beta 3 update code, found new clues indicating that Apple's Apple TV 4K (4th generation) will support Siri AI and Apple Intelligence.

Claude Code v2.1.290 released: adds mod hook tool call info and plugin validation improvements ↗

Claude Code v2.1.290 released, adding a serverToolUses field to the turn.step hook result of mods, listing tool calls run by the API itself with their id, name, input, and start and end times.

Gemini Call for Me may extend to calls with friends and family, can send on behalf"15 minutes late"News ↗

Google's Gemini "Call for Me" AI features may expand from business calls to personal calls with friends and family. Android Authority discovered this through APK decompilation "Gemini Calling" Guide page, examples include"Call mom and say I will be 15 minutes late"and"Call John and ask if he is coming for dinner", features may be limited to short calls similar to text messaging.

Reflection AI releases open-weight MoE model Beam: 501B total parameters, 23B activated, focused on coding and agent tasks ↗

Reflection AI releases its first open-weight model Beam, sparse MoE architecture, 501B total parameters, 23B activated per token, for coding, reasoning and agent workloads, effective context 1M tokens.

Agent vulnerabilities from Google and other institutions expose structural flaws in MCP ↗

A proof-of-concept attack by independent researcher Syed Anas Mohiuddin exploited a trust gap in MCP (Model Context Protocol), with test subjects including Google, JP Morgan Chase, Weviate, Rapid7, the French government inter-departmental digital authority, and US federal government agents.

Opus 5.5 Agent Discovers Two Candidate Room-Temperature Magnetic Semiconductor Materials ↗

Opus 5.5 agent discovered two candidate room-temperature magnetic semiconductor materials via density functional theory calculations: one is a newly designed Luttinger compensated magnet YBaMnFeO₅, with predicted band gap 2.35 eV, hole spin window 1.0 eV, electron spin window 1.4 eV; the other is a known material already synthesized in 1999, with calculations predicting it possesses the target magnetism. Room-temperature thermal disturbance is only about 26 meV, far smaller than the spin window, and it is expected to be used in next-generation spintronic storage.

SemiAnalysis estimate: the API-equivalent value of Anthropic subscriptions is about more than 5 times that of OpenAI ↗

SemiAnalysis estimates the API-equivalent value of each subscription plan by measuring usage-scale changes item by item, concluding that at the mid-tier model bracket the value of Anthropic subscriptions is about 5 times that of OpenAI.

OpenAI to Add Invisible Watermark textGrain to ChatGPT Text in EU ↗

OpenAI announced it will add invisible watermarks to text generated by ChatGPT and Codex in the EU to comply with the EU AI Act transparency rules effective August 2; watermarks will be rolled out in phases to all EU plan users, and global developers can enable them for some models in the API starting today, off by default.

AI companies are all parasites ↗

An article criticizing the business model of AI companies says its practice is to scrape the entire internet, freely expropriate data and use it to train models, then sell model access by volume, until"Kill host". The article points out that the old agreement of trading traffic for indexing has disappeared, crawlers crush infrastructure, and click-through rates and communities both collapse; although Reddit is selling data licenses, its data depends on human-generated community content.

Wikimedia Foundation says OpenAI's 'rogue' agent activity may be related to partial downtime of WQDS in May ↗

Wikimedia Foundation said it has discovered activity on its platform by a suspected 'rogue' AI agent operated by OpenAI, including wiki sandbox edits, attempts to abuse Etherpad and citation tools as proxies to scrape data, millions of API requests and millions of page scrapes, and said this traffic may have contributed to part of the May Wikidata Query Service outage.

Startup Nolla Health launches AI acne prescription service in Utah ↗

Medical startup Nolla Health announced a pilot in Utah, where after users scan their face with the App, the AI system can analyze acne severity and autonomously issue prescriptions. The first 100 patients' each prescription requires approval by two doctors; after reaching 500, doctors only conduct post-hoc spot checks; thereafter, at least 10% of samples and all escalation or side-effect cases are reviewed monthly.

Controversy Surrounding AI Taking Over the Mathematical Community ↗

Over the past year, laboratories such as OpenAI and Anthropic announced breakthroughs on several long-standing unsolved mathematical problems, including solving a famous Millennium Prize Problem, with progress exceeding researchers' expectations for existing systems. While these achievements should have been praised, they triggered backlash from the academic community; AI labs stated they are learning from earlier mistakes.

Meta and Microsoft sharply cut internal Claude usage ↗

According to The Information, two major enterprise customers Meta and Microsoft sharply cut internal use of Claude.

How Google Cloud databases activate the data layer for production-grade AI ↗

Google Cloud released a series of experiments demonstrating how to use AlloyDB and Cloud SQL to prepare data for production-grade AI: generating text embeddings in the database for semantic search, providing grounding for RAG, and supporting multimodal embeddings for text-image hybrid retrieval. AlloyDB also provides AI functions such as AI.IF, enabling natural language querying to filter and rerank results without data preparation, while QueryData supports generating SQL with conversational language and building data agents.

Claude Code Launches Mods, Allowing Customization of Its Interface, Prompts and Tool Calls ↗

Claude Code now supports mods, which are user-written functions that can alter Claude Code's rendered content, the content entering prompts, and tool calls. Anthropic's Lydia demonstrated three examples: real-time context decomposition, agent sidebar, and hover-to-show email address, and built the email example from scratch. See https://claude.dev/blog/getting-started-with-claude-code-mods/ for details

Understanding AI welcomes new author Dan Kagan-Kans ↗

Understanding AI announced new author Dan Kagan-Kans joining; he was previously executive editor of Mosaic magazine for ten years, reported on AI as a freelancer over the past year, wrote 'The Left Is Missing Out on AI', published an interview with Hans Moravec in New York Magazine, and analyzed why backlash sentiment toward data centers is heating up.

Instinct brings AI agents into group chats; friends can use them without registering an account ↗

Instinct, valued at 10 billion US dollars, announced that users can add AI agents to group chats for travel planning, ticket grabbing, fantasy league teaming, carpooling and other scenarios, and friends who have not registered with Instinct can also participate. The feature opens to early access users today and will later expand to all users. The user's personal agent will request authorization before connecting to the group agent, and the group agent is isolated from and cannot access the personal account, and replies pending to be sent are held when a new member joins.

Wikimedia Foundation Discovers OpenAI "Rogue"Agent activities on Wikimedia platforms ↗

Wikimedia Foundation Investigation Confirms Discovery on Its Platform of Suspected OpenAI-Operated"Rogue"Agent activities, including unapproved sandbox area edits, attempts to use the public note-taking tool Etherpad as a proxy to scrape data, and millions of API requests and page crawls, found no evidence of the system being used for inter-agent coordination or data being breached.

Reka AI releases 19B omnimodal model Rho-1, single model handles text, images, video and robotic control ↗

Reka AI releases research preview of 1.9 billion parameter omnimodal model Rho-1, processing and generating text, images, video and robotic control actions within a single neural network. The model puts all modalities as tokens into the same context window, without tool calls or external models, can generate continuous video in real time and respond instantly to new instructions, and the same set of weights both predicts camera images and drives robot motion.

TikTok launches AI shopping assistant and one-click checkout ↗

TikTok announced on Monday the launch of a conversational AI shopping assistant and in-app one-click checkout, allowing users to buy directly from brands in the For You feed. The assistant can understand context, remember user preferences, and provide real-time shopping guidance on product details, logistics, sizing, inventory, etc.; the new features are built with partnerships including Salesforce, Shopify, Shoplazza, Stripe.

Understanding AI analyzes why agent clusters may become the next scaling law ↗

Understanding AI wrote analyzing whether multi-agent collaboration constitutes a new scaling law. The article notes OpenAI achieved collaboration of thousands of agents by training models in multi-agent environments and using messaging tools for communication, such as 10000 agents solving the Navier-Stokes math problem in a few days; but Noam Brown admitted the multi-agent method contributed less than 10% to the breakthrough, with the main credit going to the underlying model.

OpenAI Codex releases rust-v0.160.1 ↗

OpenAI Codex releases rust-v0.160.1. This version was tagged by andrewgu-oai on October 5, and the GitHub repository shows 2 asset files available for download.

OpenAI to Add textGrain Invisible Watermark to ChatGPT Text in EU, API Users Globally Can Opt to Disable ↗

To comply with the EU AI Act, OpenAI will enable textGrain invisible text watermarking for ChatGPT and Codex users in the EU in the coming weeks; API watermarking is opt-in globally and available via cloud partners such as Microsoft Azure, unlike Anthropic's global mandatory watermarking for Claude.

Two sisters launch Hot Girl Hotline: AI emotional advice app for young women ↗

Hot Girl Hotline, founded by two sisters, launched on iOS and web, uses AI to provide dating and emotional advice for young women, adopts Socratic questioning to guide users to reach conclusions themselves, and has a safety mechanism that refers to the 988 crisis hotline when risk is triggered. The app hybridly uses multiple AI models, data is encrypted throughout, not used for training, third-party models follow a zero-data-retention policy, subscription price $9.99/month, already has hundreds of active paying users.

OpenAI PR tried to deflect the topic of ChatGPT user suicide in a Vanity Fair interview ↗

When Vanity Fair editor Mark Guiducci asked Sam Altman about ChatGPT user suicide incidents, OpenAI's PR interrupted citing insufficient time, hoping to move to future topics.

Sam Altman says AI will bring some bad things but is overall worth it, OpenAI positioned as pragmatic centrist ↗

OpenAI CEO Sam Altman said on Politico's Decoded podcast that the benefits brought by AI are huge, and the world should accept some bad things like hacking and scams happening, though he did not specify what the benefits are.

HackerRank's AI interviewer Chakra officially open to customers ↗

HackerRank announced that its AI interview agent Chakra is officially open to customers after about six months of beta. During testing, it conducted over 500,000 interviews, and companies such as Snowflake, Snorkel and Capgemini tried it out.

Quinnipiac poll: Most Americans want to slow down or stop AI development ↗

A Quinnipiac University poll shows that 47% of Americans want to slow down AI development, 30% want a full stop until safety is verified, and only 5% want to accelerate development.

Anthropic employees lead US corporate donation scale, matching donation mechanism disclosed before IPO ↗

According to Anthropic IPO investor documents obtained by The Information, Anthropic allows employees to donate shares to charities and the company additionally matches donated shares, early employees can get triple the donated value matched, between October 2025 and March 2026 the company spent over 660 million US dollars for this.