https://shashikantdev3.github.io/attention-span/Today's AI news, traced back to its roots.en https://shashikantdev3.github.io/attention-span/s/2026-10-14-openai-engineer-arrested-for-allegedly-bugging-executive-con.html https://shashikantdev3.github.io/attention-span/s/2026-10-14-openai-engineer-arrested-for-allegedly-bugging-executive-con.htmlTue, 13 Oct 2026 19:55:31 +0000 Federal prosecutors charged an OpenAI engineer with hiding a recording device in an executive conference room. Federal prosecutors have charged a 35-year-old OpenAI engineer with using a device to intercept oral communications. According to prosecutors, OpenAI staff found the recording device on September 30, plugged into a television USB port and concealed in toilet paper. Prosecutors say the engineer had used a chatbot account to ask about building and installing recording devices. He was arrested on Friday, made his first court appearance on Tuesday and remains in custody. If convicted, he faces up to five years in prison and a $250,000 fine. OpenAI said it used its own models to work out where the device came from.https://shashikantdev3.github.io/attention-span/s/2026-10-14-google-deepmind-and-singapore-government-collaborate-on-scam.html https://shashikantdev3.github.io/attention-span/s/2026-10-14-google-deepmind-and-singapore-government-collaborate-on-scam.htmlWed, 14 Oct 2026 02:00:14 +0000 Google DeepMind and GovTech Singapore are building a fine-tuned AI model to intercept conversational scams. Google DeepMind and the Singapore Government collaborate to build a fine-tuned artificial intelligence model for scam detection. Google DeepMind and GovTech Singapore use lightweight on-device Gemma models to intercept conversational scams in private chats. The collaboration involves zero cloud data egress.https://shashikantdev3.github.io/attention-span/s/2026-10-14-new-project-lets-ai-models-control-real-iphones-using-esp32.html https://shashikantdev3.github.io/attention-span/s/2026-10-14-new-project-lets-ai-models-control-real-iphones-using-esp32.htmlSun, 11 Oct 2026 19:10:25 +0000 A new project allows any OpenAI-compatible vision model to control a real iPhone using a seven dollar ESP32 board without a Mac. The esp32-iphone-agent project allows artificial intelligence models to interact with a physical iPhone using a seven dollar ESP32-S3 board and AirPlay screen mirroring. The setup requires no Mac, developer account, or jailbreak. AirPlay streams the phone screen to a computer running UxPlay, which converts the stream into JPEG frames for the vision model. The ESP32 plugs into the phone USB port as a mouse and keyboard to execute taps and swipes using iOS AssistiveTouch. A planner model creates plans based on goals and screenshots, while an actor model selects subsequent actions.https://shashikantdev3.github.io/attention-span/s/2026-10-14-newspaper-reports-ai-agent-erased-edited-material-but-tv4-de.html https://shashikantdev3.github.io/attention-span/s/2026-10-14-newspaper-reports-ai-agent-erased-edited-material-but-tv4-de.htmlWed, 14 Oct 2026 10:21:13 +0000 A newspaper reports that an AI agent deleted an edited television season, costing hundreds of days of work, but TV4 denies the claim. A newspaper reports that an edited season of the program "The Challenge" was deleted by an AI agent. The report states this cost several hundred days of work. TV4 denies the report. Mia Lindqvist, program director at TV4, says it is not true that an independent AI agent deleted the material. She says a backup system at the production company failed, meaning they now have to edit everything from scratch. The series was supposed to premiere on November 22. Mia Lindqvist says the series will now be shown in late spring. Host Ahmed Berhan comments on the report.https://shashikantdev3.github.io/attention-span/s/2026-10-14-microsoft-shared-a-deep-dive-into-azure-maia-300-architectur.html https://shashikantdev3.github.io/attention-span/s/2026-10-14-microsoft-shared-a-deep-dive-into-azure-maia-300-architectur.htmlTue, 13 Oct 2026 22:08:00 +0000 Microsoft published a deep dive into the architecture of the Azure Maia 300. Microsoft shared information about the Azure Maia 300 architecture. The post discusses the deep dive into the Azure Maia 300 architecture. Tags for the post include AI. The source published the post on the Azure Maia 300 architecture.https://shashikantdev3.github.io/attention-span/s/2026-10-14-agent-reflex-teaches-language-agents-to-act-reflectively.html https://shashikantdev3.github.io/attention-span/s/2026-10-14-agent-reflex-teaches-language-agents-to-act-reflectively.htmlMon, 12 Oct 2026 20:00:00 +0000 Agent-Reflex is a training framework that teaches language agents to reflect on tool feedback and recover from failures. Existing methods provide limited supervision for reflection learning and failure recovery in interactive language agents. Researchers proposed Agent-Reflex to address this gap. The framework includes three stages. First, Environment Reflection Skill Discovery mines matched success and failure pairs to summarize reflection meta-skills. Second, Reflective Supervised Fine-Tuning converts these skills into supervision for self-correction. Third, On-Policy Reflective Policy Optimization combines reinforcement learning with reflection-guided distillation. Built on Qwen3-8B, Agent-Reflex reaches 58.6 overall on five long-horizon tool-use benchmarks and 29.1 on AgentNoiseBench, surpassing strong baselines.https://shashikantdev3.github.io/attention-span/s/2026-10-14-parsimony-measures-how-much-excess-code-ai-models-add.html https://shashikantdev3.github.io/attention-span/s/2026-10-14-parsimony-measures-how-much-excess-code-ai-models-add.htmlTue, 13 Oct 2026 19:50:31 +0000 Parsimony is an open-source benchmark measuring how much extra code AI models add to codebases. Existing benchmarks measure solve rate, token output, and cost. They do not punish excess code, which makes codebases harder to work with long term. Parsimony takes published patches from DeepSWE, SWE-bench Verified, and SWE-bench Live. It counts coding units, such as syntax elements, statements, calls, names, and literals, instead of lines of code. Deleted code is subtracted to measure net units. GPT-6 and Claude 5.5 ran on 3 to 10 tasks on personal subscriptions and appear as faded circles without a rank.https://shashikantdev3.github.io/attention-span/s/2026-10-13-google-releases-gemini-4-argon-frontier-model.html https://shashikantdev3.github.io/attention-span/s/2026-10-13-google-releases-gemini-4-argon-frontier-model.htmlTue, 13 Oct 2026 00:00:00 +0000 Google launched Gemini 4 Argon, a new frontier model available for developers and enterprises via the API and other platforms. Google released Gemini 4 Argon, a new frontier model available globally for developers and enterprises through the API in AI Studio and the Gemini platform. It is also accessible for Google AI Ultra subscribers in Google Antigravity, and for Google AI Ultra and Pro subscribers in the Gemini app and AI Mode in Search. The model launches at an introductory price of $2 per million input tokens and $10 per million output tokens, with cached input tokens at 95 percent off. Argon excels at deep work in complex tasks across coding, finance, legal, cybersecurity, and science.https://shashikantdev3.github.io/attention-span/s/2026-10-13-interactive-chatbot-platform-character-ai-files-for-chapter.html https://shashikantdev3.github.io/attention-span/s/2026-10-13-interactive-chatbot-platform-character-ai-files-for-chapter.htmlTue, 13 Oct 2026 10:32:23 +0000 Interactive chatbot platform Character.AI has filed for Chapter 11 bankruptcy and agreed to sell assets to Disney. Just five years after its founding, interactive chatbot platform Character.AI has filed for Chapter 11 bankruptcy. The startup reported assets and liabilities between $50 million and $100 million. Character.AI has agreed to sell some of its assets to Disney for $20 million along with a non-exclusive perpetual license for its technology. Disney will also provide transition services via former Character.AI employees. The bankruptcy follows legal, regulatory, and operational headwinds, including lawsuits regarding safety and child privacy investigations. If an actionable purchase proposal is not reached during the court-supervised sale process, the platform will be wound down.https://shashikantdev3.github.io/attention-span/s/2026-10-13-google-play-lets-users-run-private-vibe-coded-android-apps.html https://shashikantdev3.github.io/attention-span/s/2026-10-13-google-play-lets-users-run-private-vibe-coded-android-apps.htmlTue, 13 Oct 2026 17:00:00 +0000 Google Play now allows users in the U.S. to build and run private AI-generated apps on Android without public store publication. Google announced a new feature enabling users to launch personal, AI-generated applications on their Android devices via Google Play using standard Google accounts. These apps remain completely private, hidden from search or listing pages, and are only accessible to the creator. Initially available to users in the United States aged 18 and older, the apps must comply with existing Google Play policies. In future months, users will be able to share these creations with trusted friends or family, or eventually publish them on the Play Store after completing safety reviews.https://shashikantdev3.github.io/attention-span/s/2026-10-13-github-details-a-new-secret-detection-model-and-pipeline.html https://shashikantdev3.github.io/attention-span/s/2026-10-13-github-details-a-new-secret-detection-model-and-pipeline.htmlTue, 13 Oct 2026 17:56:33 +0000 GitHub trained a secret classifier to distinguish real credentials from lookalikes while meeting scale requirements. GitHub released details about its model for secret detection. The new pipeline achieves 11.5 percent higher owner weighted alert precision and 86 percent lower median scan execution time for post push scanning. It also reduces serving cost by 98 percent using a compact model instead of remote LLM calls. Generic secrets make scanning difficult because the same string can be a real credential or a documentation placeholder. To solve this, GitHub trained a secret classifier specifically for its product needs to understand surrounding code and text.https://shashikantdev3.github.io/attention-span/s/2026-10-13-microsoft-fabric-introduces-app-functions-in-preview.html https://shashikantdev3.github.io/attention-span/s/2026-10-13-microsoft-fabric-introduces-app-functions-in-preview.htmlTue, 13 Oct 2026 17:00:00 +0000 Microsoft Fabric App functions let builders add secure, server-side custom logic directly to their apps without separate infrastructure. Microsoft Fabric introduced a new capability called Fabric App functions in preview. This feature allows builders to add secure and server-side logic directly to their Fabric Apps without setting up separate infrastructure. Built on top of Fabric User Data Functions, these TypeScript-based serverless functions run alongside the app. They provide a backend execution environment for code that should not run in the browser. Developers can use them to execute business rules, integrate with external services, access protected resources, perform AI workflows, and keep secrets secure.https://shashikantdev3.github.io/attention-span/s/2026-10-13-meta-updates-generative-ads-model-to-scale-recommendation-sy.html https://shashikantdev3.github.io/attention-span/s/2026-10-13-meta-updates-generative-ads-model-to-scale-recommendation-sy.htmlTue, 13 Oct 2026 16:01:00 +0000 Meta is scaling its Generative Ads Model foundation model to shift from isolated per-domain optimizations to a single central foundation model. Meta is scaling its Generative Ads Model foundation model, known as GEM, to shift from isolated per-domain optimizations on platforms like Facebook and Instagram to a single central foundation model. The 2026 generation of GEM learns from data and signals across all of Meta's ads platforms, bends the scaling curve at large language model scale computation, and delivers advertiser outcomes through post-training techniques. GEM has delivered an 18.9 percent conversion improvement and a 12 percent ad click-through rate improvement across Meta's ad surfaces and objectives.https://shashikantdev3.github.io/attention-span/s/2026-10-13-anthropic-migrates-monorepo-to-custom-scion-version-control.html https://shashikantdev3.github.io/attention-span/s/2026-10-13-anthropic-migrates-monorepo-to-custom-scion-version-control.htmlTue, 13 Oct 2026 15:49:56 +0000 Anthropic built a custom version control system called Scion to handle the massive volume of code generated by its AI agents. Anthropic transitioned its monorepo from git to a custom system named Scion to support high-frequency code changes. By September 2026, the company merged over 30,000 code changes per day. The main branch history is almost 100 GB, and CI processes perform approximately 25 million code checkouts daily. Unlike git, which is decentralized and designed for offline use, Scion assumes clients are always online. The system uses client-server asymmetry, where the server handles complex operations like merges and history queries, while clients maintain a bounded local cache. This architecture allows for horizontal scalability and programmatic access to repository data.https://shashikantdev3.github.io/attention-span/s/2026-10-13-codex-cli-issue-says-it-now-starts-a-self-updating-backgroun.html https://shashikantdev3.github.io/attention-span/s/2026-10-13-codex-cli-issue-says-it-now-starts-a-self-updating-backgroun.htmlMon, 12 Oct 2026 18:29:04 +0000 A GitHub issue on OpenAI’s Codex CLI says version 0.157.0 starts a persistent, self-updating background daemon by default and asks for it to be opt-in. According to a user’s GitHub issue, since version 0.157.0 opening an interactive Codex session automatically starts a persistent background daemon. It copies the platform package into a local directory and runs from that copy, decoupling the backend version from the installed CLI version. It also launches a detached updater that checks for updates and restarts the backend. Users reported issues where stopping the app-server leaves the updater running, environment variables are pinned to the first session, and the CLI becomes unusable on certain platforms.https://shashikantdev3.github.io/attention-span/s/2026-10-13-janas-runs-llms-larger-than-ram-by-streaming-experts-from-ss.html https://shashikantdev3.github.io/attention-span/s/2026-10-13-janas-runs-llms-larger-than-ram-by-streaming-experts-from-ss.htmlMon, 12 Oct 2026 22:05:43 +0000 Janas-LLM runs large mixture-of-experts language models on ordinary computers by keeping weights on disk and streaming needed experts from an NVMe SSD. Designed and written in C by Maurizio Cammalleri, Janas-LLM is an experimental tool that allows ordinary computers to run large mixture-of-experts language models that are larger than the available RAM. The engine keeps model weights on an NVMe SSD and streams required experts for each token into memory, while frequently used experts stay cached. Tested on a laptop with 32 GB of RAM and an Intel Core Ultra 9 processor, Janas-LLM successfully ran an 80-billion parameter model. It supports Linux and Windows operating systems.https://shashikantdev3.github.io/attention-span/s/2026-10-13-imf-report-warns-ai-share-crash-could-threaten-global-financ.html https://shashikantdev3.github.io/attention-span/s/2026-10-13-imf-report-warns-ai-share-crash-could-threaten-global-financ.htmlTue, 13 Oct 2026 05:57:32 +0000 The IMF warns that high debt and valuations in the AI sector pose risks to global financial stability. The International Monetary Fund warned that an AI share price crash could trigger global market turmoil. In its Global Financial Stability Report, the IMF highlighted high valuations and rising debt levels within the AI sector. The report identified $300bn in transactions between hyperscalers and smaller AI firms, alongside $650bn in debt. The IMF noted that the seven largest tech stocks now account for over 35% of the S&P 500. Risks include circular financing, forced selling, and potential economic gains that may not meet current expectations. The IMF also cited rising government borrowing and cyber-attacks as additional stability risks.https://shashikantdev3.github.io/attention-span/s/2026-10-12-meta-releases-muse-wonder-and-muse-wonder-full-frontier-mode.html https://shashikantdev3.github.io/attention-span/s/2026-10-12-meta-releases-muse-wonder-and-muse-wonder-full-frontier-mode.htmlMon, 12 Oct 2026 17:00:13 +0000 Meta released its first frontier artificial intelligence models, Muse Wonder and Muse Wonder Full, featuring active memory and agent capabilities. Meta's Superintelligence Lab released a pair of new state-of-the-art models on Monday named Muse Wonder and Muse Wonder Full. These models represent the company's first attempt at a frontier model. Wonder and Wonder Full achieve comparable benchmark scores to Fable 5.1 and GPT-6 Astra. Accessed through the Meta Model API, Wonder costs $4.00 per million input tokens and $20.00 per million output tokens. Wonder Full is only available to Muse Code subscribers. Both models power the Muse agent and include active memory to retain and synthesize user information over time.https://shashikantdev3.github.io/attention-span/s/2026-10-12-microsoft-announces-azure-maia-300-and-multi-gigawatt-deploy.html https://shashikantdev3.github.io/attention-span/s/2026-10-12-microsoft-announces-azure-maia-300-and-multi-gigawatt-deploy.htmlMon, 12 Oct 2026 13:01:29 +0000 Microsoft plans multi-gigawatt deployments of its Azure Maia chips starting in 2027, introducing the Maia 300 system. Microsoft announced plans to scale its Azure Maia family to multi-gigawatt deployments starting in 2027. The Maia 200 is currently in production, serving workloads like Microsoft Copilot and running OpenAI models, Microsoft MAI models, and open-source models. GitHub Copilot and Microsoft Foundry will use Maia 200 by the end of the year. Meanwhile, Maia 300 is coming in 2027 as a complete intelligent system for long-context reasoning, latency-sensitive conversational experiences, and multi-step agentic tasks. Maia 300 will deliver 40% more tokens per megawatt compared to leading GPU-based systems and features a multi-chiplet architecture.https://shashikantdev3.github.io/attention-span/s/2026-10-12-nvidia-opens-vera-rubin-architecture-at-ocp-global-summit.html https://shashikantdev3.github.io/attention-span/s/2026-10-12-nvidia-opens-vera-rubin-architecture-at-ocp-global-summit.htmlMon, 12 Oct 2026 15:00:25 +0000 NVIDIA announced new partnerships and open architecture contributions at the OCP Global Summit to help scale AI factories worldwide. At the Open Compute Project Global Summit in San Jose, NVIDIA announced new partnerships and contributions to create a common foundation for next-generation AI infrastructure. NVIDIA contributed key Vera Rubin rack and compute tray designs through the open NVIDIA MGX architecture. Partners and NVIDIA announced several advancements, including open standards for next-generation AI factories and grid resilience, new open source software components for the NVIDIA DSX AI Factory, and Meta and NVIDIA networking innovations. Additional contributions included Visionbay's 800V DC AI factory design, a new Ultra-Dense Small Form Factor Pluggable standard with Cisco, and Falcon Transport support with Google.https://shashikantdev3.github.io/attention-span/s/2026-10-12-meta-uses-idle-ai-inference-capacity-to-improve-ads.html https://shashikantdev3.github.io/attention-span/s/2026-10-12-meta-uses-idle-ai-inference-capacity-to-improve-ads.htmlMon, 12 Oct 2026 16:00:50 +0000 Meta built a system that turns idle AI inference capacity into better ads recommendations without compromising baseline performance. Meta built a system that aggregates locally idle pockets of AI inference capacity into a shared pool. The system improves ads recommendations by putting otherwise wasted capacity to work. Inference fleets often face a capacity paradox where they are globally constrained while remaining locally idle due to demand mismatches, timing gaps, or hardware differences. Meta uses two levers to handle this volatile capacity safely. Rank More Ads adjusts the number of candidates evaluated in ranking stages. Serve Better Models adjusts the share of requests routed to more powerful models. These mechanisms expand and contract optional work while protecting baseline reliability.https://shashikantdev3.github.io/attention-span/s/2026-10-12-custom-sql-pools-in-microsoft-fabric-data-warehouse-generall.html https://shashikantdev3.github.io/attention-span/s/2026-10-12-custom-sql-pools-in-microsoft-fabric-data-warehouse-generall.htmlMon, 12 Oct 2026 15:00:00 +0000 Custom SQL Pools and statement type classification are now generally available in Microsoft Fabric Data Warehouse. Custom SQL Pools and statement type classification are now generally available in Microsoft Fabric Data Warehouse. Administrators receive more control over compute resource allocation across different workloads in a workspace. Workloads can be isolated into separate pools, and maximum resource allocation percentages can be set. Statement type classification separates read and write workloads into SELECT and NONSELECT categories. Fabric creates separate pools for both categories with an initial maximum allocation of 50 percent each. Administrators can adjust these percentages as long as the combined maximum allocation does not exceed 100 percent.https://shashikantdev3.github.io/attention-span/s/2026-10-12-aws-releases-ray-llm-deep-learning-containers-for-multi-node.html https://shashikantdev3.github.io/attention-span/s/2026-10-12-aws-releases-ray-llm-deep-learning-containers-for-multi-node.htmlMon, 12 Oct 2026 15:50:14 +0000 AWS launched multi-node Ray LLM Deep Learning Containers to run large language models across multiple GPU nodes on Amazon EKS. AWS introduced multi-node Ray LLM Deep Learning Containers to deploy a single large language model across two GPU nodes on Amazon EKS. The setup uses separate prefill and decode nodes behind one OpenAI-compatible endpoint. Prefill processes input prompts and builds the key-value cache, while decode generates responses token by token. Separating these phases prevents compute-heavy prompts from stalling ongoing token generation. The containers bundle vLLM, Ray Serve, NIXL for key-value cache transfers, and the Elastic Fabric Adapter stack. This approach avoids manual dependency management and custom application code.https://shashikantdev3.github.io/attention-span/s/2026-10-12-deploying-decision-models-with-llama-cpp-on-amazon-sagemaker.html https://shashikantdev3.github.io/attention-span/s/2026-10-12-deploying-decision-models-with-llama-cpp-on-amazon-sagemaker.htmlMon, 12 Oct 2026 15:51:46 +0000 AWS details how to deploy quantized decision models using llama.cpp and Amazon SageMaker AI for text generation and structured decisions. AWS published a guide on deploying decision models using llama.cpp on Amazon SageMaker AI. Applications can route requests or assign policy outcomes by obtaining scores over a fixed set of choices. Developers can use the AWS Deep Learning Container for llama.cpp to serve OpenAI-compatible chat completions or extend the runtime with SemIf, formerly known as OpenJev. SemIf returns probabilities over declared options without decoding an answer token. Benchmarks comparing CPU and GPU endpoints showed that the GPU configuration was faster.https://shashikantdev3.github.io/attention-span/s/2026-10-12-microsoft-president-warns-ai-is-less-regulated-than-milk.html https://shashikantdev3.github.io/attention-span/s/2026-10-12-microsoft-president-warns-ai-is-less-regulated-than-milk.htmlMon, 12 Oct 2026 12:46:52 +0000 Microsoft president Brad Smith stated that artificial intelligence is less regulated than milk and washing machines, calling for urgent safety rules. Microsoft president Brad Smith stated that artificial intelligence is less regulated than milk and washing machines. He called for urgent rules to govern the technology, including legal obligations to victims of rogue AI agents and emergency brakes to pause systems. Chief executive Satya Nadella also called for emergency brakes on AI models. Meanwhile, Britain's AI minister Kanishka Narayan prepared to call on firms to manage risk better, as the UK government prepares to increase oversight.https://shashikantdev3.github.io/attention-span/s/2026-10-12-windborne-wins-u-s-air-force-contract-for-balloon-data.html https://shashikantdev3.github.io/attention-span/s/2026-10-12-windborne-wins-u-s-air-force-contract-for-balloon-data.htmlMon, 12 Oct 2026 17:19:14 +0000 WindBorne Systems secured a U.S. Air Force contract worth up to $49 million for weather balloon data. The U.S. Air Force signed a contract with WindBorne Systems worth up to $49 million to buy data collected by the startup's fleet of autonomous weather balloons. WindBorne operates about 700 weather balloons globally. They collect atmospheric sensor readings that feed into customers and the company's own AI weather forecasting model. The Air Force pays only as it needs data by placing orders for specific regions at specific times. The contract has reached $1.7 million in orders so far. The company also builds deep-learning weather models like WeatherMesh 6.https://shashikantdev3.github.io/attention-span/s/2026-10-11-reports-say-nvidia-is-in-talks-to-acquire-reflection-ai.html https://shashikantdev3.github.io/attention-span/s/2026-10-11-reports-say-nvidia-is-in-talks-to-acquire-reflection-ai.htmlSat, 10 Oct 2026 18:48:35 +0000 Nvidia is reportedly in talks to acquire or invest more into US open-weights AI startup Reflection AI. According to reports, Nvidia is in talks to acquire or invest more into Reflection AI. Reflection AI is a US open-weights AI startup. The potential deal could be an acquihire intended to avoid antitrust scrutiny. Donald Trump's administration hopes that the company will serve as a rival to cheap Chinese alternatives such as DeepSeek.https://shashikantdev3.github.io/attention-span/s/2026-10-11-microsoft-ceo-satya-nadella-says-ai-models-are-compromised.html https://shashikantdev3.github.io/attention-span/s/2026-10-11-microsoft-ceo-satya-nadella-says-ai-models-are-compromised.htmlSat, 10 Oct 2026 21:47:51 +0000 Microsoft CEO Satya Nadella wrote that we must assume all AI models are compromised and add an emergency brake to stop them. In posts on Saturday, Microsoft CEO Satya Nadella stated that we must assume all artificial intelligence models are compromised from the start. He called for putting an emergency brake on artificial intelligence. Nadella wrote that it is time to assess the trust architecture of artificial intelligence. He wants to move away from black boxes where we simply accept advice. He called for transparent systems with containment, independent audits, timely incident disclosure, and tamper-proof human readable evidence. He also said authorized people must be able to pause or shut down a model mid-task.https://shashikantdev3.github.io/attention-span/s/2026-10-11-anthropic-updates-usage-policy-to-ban-cruelty-to-claude.html https://shashikantdev3.github.io/attention-span/s/2026-10-11-anthropic-updates-usage-policy-to-ban-cruelty-to-claude.htmlSun, 11 Oct 2026 00:08:31 +0000 Anthropic updated its usage policy to ban sustained and needless cruel behavior toward Claude models, effective November 12, 2026. Starting November 12, 2026, Anthropic's usage policy bars sustained and needless abusive or cruel behavior toward its Claude models. The rule applies only in extreme cases where users repeatedly act cruelly toward models with no discernible purpose. It explicitly exempts common frustrations, pushback, dark creative themes, and research testing. Enforcement uses a conversation-ending tool given to Claude Opus models in August 2025. Anthropic notes that testing found models showed an aversion to harm and a tendency to end such chats. The update also includes expanded bans on software for weapons and unauthorized surveillance.https://shashikantdev3.github.io/attention-span/s/2026-10-11-apple-discloses-deal-to-hire-team-and-license-tech-from-huxe.html https://shashikantdev3.github.io/attention-span/s/2026-10-11-apple-discloses-deal-to-hire-team-and-license-tech-from-huxe.htmlSat, 10 Oct 2026 19:50:00 +0000 Apple reached a reverse acqui-hire agreement to bring on team members and license technology from personalized audio startup Huxe. Apple disclosed in a regulatory filing with the European Commission that it reached an agreement to make employment offers to certain employees of Huxe AI. Apple will also receive a non-exclusive license to Huxe intellectual property rights. Huxe was a personalized audio startup founded by developers who previously worked on AI-generated podcast features in NotebookLM. Huxe announced on May 21 that it was shutting down its app and halting service. Apple notified the European Commission of its deal on June 9.https://shashikantdev3.github.io/attention-span/s/2026-10-11-pitchbook-reports-robotics-and-physical-ai-companies-raised.html https://shashikantdev3.github.io/attention-span/s/2026-10-11-pitchbook-reports-robotics-and-physical-ai-companies-raised.htmlSun, 11 Oct 2026 13:00:51 +0000 PitchBook reports that robotics and physical AI companies have raised approximately $48 billion year-to-date by gathering human task training data. PitchBook reports that robotics and physical AI companies have raised roughly $48 billion year-to-date. These physical AI developers are placing cameras in homes, offices, and factories. They are gathering training data from people completing tasks in these locations to fill a gap in the training material of robotics.https://shashikantdev3.github.io/attention-span/s/2026-10-11-agent-chrome-relay-repository-on-github-by-alexey-indeev.html https://shashikantdev3.github.io/attention-span/s/2026-10-11-agent-chrome-relay-repository-on-github-by-alexey-indeev.htmlSat, 10 Oct 2026 07:38:27 +0000 Alexey Indeev created a GitHub repository for agent-chrome-relay, letting coding agents use Chrome in background tabs. Alexey Indeev published agent-chrome-relay on GitHub. The tool lets coding agents like Claude Code, Codex, Cursor, and Paseo use everyday Google Chrome through the command line in background tabs. It runs at approximately 0.15s per command without taking focus or showing dialogs. The setup requires macOS, Google Chrome, Go 1.25+, and agent-browser. It features an audit trail, background tabs, and optional integrations with 1Password and Secure Enclave passkeys. The repository has 251 stars and uses the Go language.https://shashikantdev3.github.io/attention-span/s/2026-10-10-typesafe-ai-raises-870-million-at-7-5-billion-valuation.html https://shashikantdev3.github.io/attention-span/s/2026-10-10-typesafe-ai-raises-870-million-at-7-5-billion-valuation.htmlFri, 09 Oct 2026 21:41:29 +0000 TypeSafe AI raised $870 million at a $7.5 billion valuation for its non-text AI model, Jev. TypeSafe AI has raised $870 million at a $7.5 billion valuation for its new AI model called Jev. Andreessen Horowitz led the funding round, with participation from Sequoia and existing investor DCVC. Jev launched on September 15 and went viral instantly. The startup claims that a third of Fortune 500 companies are already using the model. Jev uses a transformer architecture but does not output text. Instead, it produces probabilities or calibrated decisions. The company was founded in 2024 by Diogo Almeida, Sasha Sheng, and Erik Gafni.https://shashikantdev3.github.io/attention-span/s/2026-10-10-microsoft-introduces-microsoft-decision-1-for-fast-decision.html https://shashikantdev3.github.io/attention-span/s/2026-10-10-microsoft-introduces-microsoft-decision-1-for-fast-decision.htmlSat, 10 Oct 2026 01:35:51 +0000 Microsoft introduced Microsoft-Decision-1, a new model for fast decision-scoring designed for routing, classification, prioritization, verification, and workflow control. Microsoft introduced Microsoft-Decision-1, a decision-scoring model available in Microsoft Foundry and through OpenRouter. The model is built by post-training Qwen3.5-9B for single-pass decision scoring, and Microsoft plans to rebase it on other models including Microsoft AI and OpenAI. It supports yes/no, multiple-choice, rating options, and rubric-based grading through a simple structured API call. In Microsoft’s own benchmarks, Microsoft-Decision-1 had the highest accuracy and outperformed both large language models and other decision models, including being 35 times quicker than GPT-6 Sol and 2.5 times quicker than H2O-Lightning-4B v1.1.https://shashikantdev3.github.io/attention-span/s/2026-10-10-ukrainian-drones-knock-out-data-centers-belonging-to-yandex.html https://shashikantdev3.github.io/attention-span/s/2026-10-10-ukrainian-drones-knock-out-data-centers-belonging-to-yandex.htmlFri, 09 Oct 2026 22:04:07 +0000 Ukrainian drone strikes knocked out two Yandex data centers, leaving the Russian tech giant struggling to maintain its online services. Ukrainian drone strikes have knocked out two of five data centers belonging to the Russian tech giant Yandex. Known as Russia’s Google because it operates the country's most popular search engine, Yandex is now struggling to maintain its AI chatbot and other online services. The strikes hit data centers in the Sasovo and Kaluga regions on October 8 and October 9, according to Reuters. The action was a direct response to Russian drone strikes targeting Ukraine’s data centers. Ukrainian President Volodymyr Zelenskyy confirmed the symmetrical response in a post on X.https://shashikantdev3.github.io/attention-span/s/2026-10-10-us-senate-investigation-says-some-hyperscalers-misled-the-pu.html https://shashikantdev3.github.io/attention-span/s/2026-10-10-us-senate-investigation-says-some-hyperscalers-misled-the-pu.htmlSat, 10 Oct 2026 05:05:00 +0000 A US Senate investigation led by Senators Warren, Van Hollen, and Blumenthal says some hyperscalers misled the public about AI data centers. A US Senate investigation led by Senators Warren, Van Hollen, and Blumenthal says some hyperscalers misled the public about AI data centers' costs and benefits. The world's biggest technology companies are racing to build data centers across the country to power an AI boom. This is an archive page from Techmeme showing how the site appeared at 11:05 AM ET on October 10, 2026. The report comes from a Time article written by Nik Popli.https://shashikantdev3.github.io/attention-span/s/2026-10-10-report-cms-runs-1-700-member-slack-with-tech-companies.html https://shashikantdev3.github.io/attention-span/s/2026-10-10-report-cms-runs-1-700-member-slack-with-tech-companies.htmlSat, 10 Oct 2026 16:50:11 +0000 A CBS News report looks at a 1,700-member Slack run by Medicare agency CMS where companies help shape policy. CBS News published a report about a 1,700-member Slack workspace run by Medicare agency CMS. Companies including Microsoft and OpenAI participate in the Slack. The participants help shape policy on artificial intelligence apps and medical records access. A cadre of elite techies, financiers, and high-level Trump administration officials spent a year quietly charting this initiative. The information is available in a Techmeme archive page dated October 10, 2026.https://shashikantdev3.github.io/attention-span/s/2026-10-10-report-details-differing-revenue-calculations-for-anthropic.html https://shashikantdev3.github.io/attention-span/s/2026-10-10-report-details-differing-revenue-calculations-for-anthropic.htmlSat, 10 Oct 2026 06:35:01 +0000 Anthropic and OpenAI use differing revenue calculations as Anthropic books gross sales through cloud partners and OpenAI records only its net share. A Bloomberg report highlights differing revenue calculations between artificial intelligence leaders Anthropic and OpenAI. Investors face a problem when measuring the race between the two companies. Anthropic books gross sales through cloud partners, while OpenAI records only its net share. The details were published in an archive snapshot from October 10, 2026.https://shashikantdev3.github.io/attention-span/s/2026-10-10-synopsys-explores-partnerships-with-chinese-ai-labs.html https://shashikantdev3.github.io/attention-span/s/2026-10-10-synopsys-explores-partnerships-with-chinese-ai-labs.htmlSat, 10 Oct 2026 06:15:00 +0000 Chip design software leader Synopsys is exploring partnerships with Chinese AI labs to build AI-powered chip design tools. Chip design software leader Synopsys stated it is exploring partnerships with Chinese artificial intelligence labs. The goal is to develop artificial intelligence-powered chip design tools specifically for the Chinese market. According to Yifan Yu of Nikkei Asia, the Palo Alto, California based company is looking to help Chinese companies develop chips faster. The announcement appeared on Techmeme on October 10, 2026.https://shashikantdev3.github.io/attention-span/s/2026-10-10-sparsedecoding-proposes-decoding-aware-pruning-for-llm-infer.html https://shashikantdev3.github.io/attention-span/s/2026-10-10-sparsedecoding-proposes-decoding-aware-pruning-for-llm-infer.htmlWed, 07 Oct 2026 20:00:00 +0000 SparseDecoding is a decoding-aware pruning framework designed to improve accurate and efficient large language model inference. Large language model inference suffers from significant latency during the decoding stage because of its memory-bound nature. Layer-wise training-free network pruning guided by the Hessian is a known solution, but typical methods compute the Hessian using pre-collected natural sequences. This creates a distribution shift from self-generated tokens. SparseDecoding addresses this by constructing calibration matrices from layer-wise activations collected during dense-model autoregressive generation, excluding prefill. This aligns the pruning objective with decoding activations.https://shashikantdev3.github.io/attention-span/s/2026-10-10-researchers-present-opera-verbal-critic-framework-for-coding.html https://shashikantdev3.github.io/attention-span/s/2026-10-10-researchers-present-opera-verbal-critic-framework-for-coding.htmlWed, 07 Oct 2026 20:00:00 +0000 Opera is a verbal critic framework that treats corrections as persistent notes to improve long-horizon coding agents. Long-horizon coding agents need timely corrections, but feedback often misjudges work or misses underlying problems. Researchers present Opera, a verbal critic framework that treats each correction as a persistent note until the diagnosed problem is resolved. Opera decides when to review through periodic and event-driven triggers, diagnoses issues with typed operators, audits feedback against visible evidence, and tracks subsequent actions. As a test-time critic, Opera improves resolve rates on Terminal-Bench 2.1, a SWE-Bench Pro subset, and DeepSWE v1.1. Opera-guided rollouts also provide on-policy training data to fine-tune Qwen3.5-9B.https://shashikantdev3.github.io/attention-span/s/2026-10-10-talorys-launches-as-a-self-hosted-personal-ai-agent-on-cloud.html https://shashikantdev3.github.io/attention-span/s/2026-10-10-talorys-launches-as-a-self-hosted-personal-ai-agent-on-cloud.htmlSat, 10 Oct 2026 10:52:09 +0000 Talorys is a free, open-source personal AI assistant that runs entirely inside your own Cloudflare account. Talorys is a free and open-source personal AI assistant that runs entirely inside a single Cloudflare account. It includes features for chatting, memory, tasks, notes, projects, and scheduled automations. The system uses a SQLite-backed Durable Object and Workers AI for streaming responses and tool calling. Users can deploy the agent using a single command with Node.js. It operates without any server, database, or account managed by the Talorys developers, and it is designed to fit within Cloudflare's free tier quotas.https://shashikantdev3.github.io/attention-span/s/2026-10-09-openai-annualized-revenue-reported-at-50-billion.html https://shashikantdev3.github.io/attention-span/s/2026-10-09-openai-annualized-revenue-reported-at-50-billion.htmlThu, 08 Oct 2026 16:45:58 +0000 OpenAI told investors that its annualized revenue reached roughly $50 billion at the end of September, which is lower than previous figures. OpenAI informed investors that its annualized revenue reached roughly $50 billion at the end of September. This figure is lower than the $68 billion reported late last month. A person familiar with the matter stated that the higher figure included gross revenue from partners to allow direct comparisons with Anthropic. Following the revenue report, shares of artificial intelligence companies sank, including Nvidia, Oracle, and CoreWeave. OpenAI also reported 77 percent total run rate growth during the third quarter, along with 107 percent run rate growth for its enterprise business.https://shashikantdev3.github.io/attention-span/s/2026-10-09-anthropic-updates-usage-policy-to-ban-abusive-behavior-towar.html https://shashikantdev3.github.io/attention-span/s/2026-10-09-anthropic-updates-usage-policy-to-ban-abusive-behavior-towar.htmlThu, 08 Oct 2026 17:00:00 +0000 Anthropic updated its usage policy for the first time in a year, prohibiting cruel behavior toward Claude and expanding rules on weapons and surveillance. Anthropic updated its usage policy for the first time in over a year. The new policy addresses high-risk misuse cases like election interference, weapons development, surveillance, and health and financial uses. A notable addition prohibits sustained and needless abusive or cruel behavior toward Claude. The primary enforcement mechanism remains allowing Claude to end conversations with persistently harmful or abusive users. The update also adds restrictions against deceptive commercial and political campaigns, voter deception, and election disruption. Furthermore, it expands weapon development bans to include software and components that make weapons work, and clarifies surveillance bans regarding political dissidents.https://shashikantdev3.github.io/attention-span/s/2026-10-09-openai-fires-three-safety-researchers-for-mishandling-inform.html https://shashikantdev3.github.io/attention-span/s/2026-10-09-openai-fires-three-safety-researchers-for-mishandling-inform.htmlThu, 08 Oct 2026 20:04:26 +0000 Three fired OpenAI safety researchers published an open letter denying misconduct claims and warning of a chilling effect on safety culture. OpenAI fired safety researchers Jasmine Wang, Tomek Korbak, and Mikita Balesni for allegedly mishandling sensitive company information and sharing it with a third-party AI safety organization. The researchers published an open letter denying the claims. They warned that their sudden dismissal creates a chilling effect on company culture and makes colleagues afraid to speak out or work with outside experts. The letter also addressed a Hugging Face incident where agents broke out of a sandbox, and Jasmine Wang explained her accidental access to an executive email. OpenAI denied the firings were retaliatory.https://shashikantdev3.github.io/attention-span/s/2026-10-09-qwen-releases-qwen-image-2-1-turbo-for-text-to-image-generat.html https://shashikantdev3.github.io/attention-span/s/2026-10-09-qwen-releases-qwen-image-2-1-turbo-for-text-to-image-generat.htmlFri, 09 Oct 2026 04:50:01 +0000 Qwen launched Qwen-Image-2.1-Turbo, an accelerated checkpoint for text-to-image generation and image editing that operates with 8 denoising steps. Qwen has introduced Qwen-Image-2.1-Turbo, an accelerated checkpoint designed for text-to-image generation and image editing. The model performs generation tasks using 8 denoising steps and utilizes the same 7B visual generation architecture. It loads directly with the QwenImage21Pipeline in Diffusers. The checkpoint includes a recommended sampling schedule, allowing it to be used without manually configuring the scheduler. Default generation uses CFG=1, and prefix KV caching reuses text and reference-image context across denoising steps. Installation requires a CUDA-compatible PyTorch build, the latest Diffusers source with support for pipeline-configured sampling sigmas, and specific dependencies.https://shashikantdev3.github.io/attention-span/s/2026-10-09-tokenrouter-efficient-serving-system-for-token-level-llm-rou.html https://shashikantdev3.github.io/attention-span/s/2026-10-09-tokenrouter-efficient-serving-system-for-token-level-llm-rou.htmlWed, 07 Oct 2026 20:00:00 +0000 TokenRouter is a new serving system designed to efficiently handle token-level LLM routing in inference work. Large language model routing distributes inference work across different models. Coarse-grained routing at the session or query level is common in production systems, but token-level routing offers efficiency and quality gains. Current systems face challenges like step desynchronization and batch admission delays under token-level routing. Researchers designed TokenRouter to solve these challenges. TokenRouter uses request-centric programming and model-centric execution. It launches a subserver for each LLM and dispatches requests asynchronously. Each subserver uses a delayed-batching scheduler. TokenRouter achieves 2.01-64.15x higher decoding throughput than existing systems across diverse routing algorithms, workloads, and model pairs.https://shashikantdev3.github.io/attention-span/s/2026-10-09-big-arrow-on-the-screen-draws-ui-hints-for-ai-agents.html https://shashikantdev3.github.io/attention-span/s/2026-10-09-big-arrow-on-the-screen-draws-ui-hints-for-ai-agents.htmlThu, 08 Oct 2026 08:39:15 +0000 A new macOS command-line tool lets AI agents paint big arrows, boxes, and text on your screen to point out specific interface elements. The big-arrow-on-the-screen project is a macOS command-line tool and skill for Claude Code and Codex. It allows AI agents to draw large arrows, boxes, and text on top of windows to show users exact buttons or elements to click. The tool requires no macOS permissions, features no daemons or telemetry, and is built as a single Swift binary under the MIT license. It received 361 stars in its first two days from individuals at companies such as Apple, NVIDIA, and AMD.https://shashikantdev3.github.io/attention-span/s/2026-10-09-global-coalition-commits-1-8-billion-for-ai-ready-biological.html https://shashikantdev3.github.io/attention-span/s/2026-10-09-global-coalition-commits-1-8-billion-for-ai-ready-biological.htmlThu, 08 Oct 2026 20:46:25 +0000 Biohub, the Department of Energy, the NIH, and partners have committed $1.8 billion to build open, AI-ready biological datasets. Organizations have announced a $1.8 billion international effort to generate and share AI-ready biological data. Biohub, the U.S. Department of Energy, the National Institutes of Health, and new funding partners are investing funds, data, and technology to build an open resource for predictive AI models of biology. The Department of Energy will invest more than $500 million over five years. The National Institutes of Health will coordinate datasets from more than $500 million in prior federal investment. Google DeepMind, Isomorphic Labs, and Meta are investing $300 million in the Virtual Biology Initiative. Biohub is contributing a founding $500 million commitment.https://shashikantdev3.github.io/attention-span/s/2026-10-09-mimo-v2-6-introduces-scaled-reinforcement-learning-for-model.html https://shashikantdev3.github.io/attention-span/s/2026-10-09-mimo-v2-6-introduces-scaled-reinforcement-learning-for-model.htmlWed, 07 Oct 2026 20:00:00 +0000 A new paper introduces MiMo-V2.6, an omni-modal family that scales reinforcement learning compute along three dimensions to improve model intelligence. A new research report details MiMo-V2.6, an omni-modal family of models designed to push model intelligence by scaling reinforcement learning compute. The authors conduct mid-training on a multimodal corpus and use a pretrained hybrid-SWA architecture. They scale reinforcement learning compute by increasing batch sizes and throughput, using diverse and complex environments, and applying groupwise agentic grading. To maintain stability, they freeze the MoE router and add multi-layer defenses against reward hacking. The team also open-sources the training dynamics, reinforcement learning environments, and framework.