Recent model releases demonstrate a focus on iterative feedback loops in AI development. Examples include OpenAI's GPT-6 Cyber learning from executable evidence, Thinking Machines' Inkling 2 Small for multi-action tasks, and Google's Gemini Omni 1.2 Flash for revising video content. These advancements highlight the importance of follow-up steps in making AI tools practical. The ability to revise and respect constraints determines the reliability of the final result. Therefore, attention should be directed toward how systems handle feedback to achieve useful outcomes.
OpenAI launched "Businesses in ChatGPT" in beta on November 10, providing a dashboard for eligible businesses. This feature allows businesses to claim and edit their profiles within ChatGPT and view insights on mentions and website clicks. Mentions are counted only when ChatGPT references a website, not just names a brand. The platform does not currently show details like prompts or share of voice data. OpenAI advises that claiming a profile does not guarantee increased placement or sales.
Google is guiding coding agents toward modern web APIs using Modern Web Guidance, resulting in a 37% gain in adherence to best practices. Browser Harness allows agents to directly control a user's signed-in Chrome browser via a CDP WebSocket. Microsoft's .NET 11 release includes C# 15 features, notably introducing union types. These developments focus on enabling AI agents to interact with modern web technologies and updated programming languages.
Nvidia's board has initiated an independent investigation into whether CEO Jensen Huang was targeted by a foreign intelligence effort. This stems from Huang's admission of an affair with a Chinese businesswoman whom U.S. intelligence officials suspected was gathering information for Beijing. The investigation is being overseen by a committee of independent directors advised by outside counsel. While the board expressed confidence in Huang's leadership, the review addresses potential security and trade compliance issues. The situation has also drawn political pressure for Huang to testify before Congress.
A future successor to Nancy Pelosi, Scott Wiener, is working on legislation to regulate advanced artificial intelligence due to concerns about its rapid advancement. Concerns surrounding data centers have been linked to electricity supply issues, suggesting that power woes are a significant factor in the industry's backlash. Big Tech's large electricity consumption is highlighted as a source of contention in the data center debate. The text also touches on broader technology topics, including AI regulation and cyber security threats. Finally, the piece points to various other technological and environmental issues.
A block of SoftBank's assets related to an ABB purchase has raised supply-chain questions. This action has generated concerns regarding the future of cross-border robotics investment. The move involves a Japanese multinational investment holding company. These concerns suggest potential implications for international investment flows in the robotics sector. The situation prompts examination of supply chain dynamics related to the transaction.
Anthony Albanese is navigating the complex issue of artificial intelligence and copyright, which deeply affects the creative sector. There is a debate over weakening copyright laws to accommodate AI training versus protecting creators' intellectual property. Concerns exist that AI's automated mimicry erodes artistic value and cognitive skills, prompting calls for new frameworks. The government seeks to position Australia as a leader in AI development by embedding itself in the technology's production chain. Ultimately, the piece calls for a cultural accord to ensure human work is recognized and fairly compensated in the age of AI.
STABL is a non-profit research organization focused on understanding the societal transformation and bottlenecks related to advanced AI. It posits that adapting to AI requires addressing not just capability but also the speed of human and organizational change. The organization favors a worldview in the top-right quadrant, emphasizing both the transformative potential of AI and the importance of institutions in managing its impacts. Research directions include measuring recursive self-improvement, developing societal defenses through resilience, steering economic impacts toward augmentation, and establishing sector-specific responses. STABL aims to contribute to a rich intellectual agenda through empirical research, evidence synthesis, and policy guidance.
New York is implementing the RAISE Act, a law requiring large frontier AI developers to register, publish safety frameworks, and report incidents. This law will oversee AI labs and includes penalties up to $3 million for violations. The state also established the New York Center for AI Responsibility and Research to independently evaluate AI safety. Other states, like California and Illinois, have enacted similar AI safety and transparency legislation. Governor Hochul is also promoting a national model for sensible AI regulation.
Seven Anthropic co-founders plan to each sell approximately $250 million in shares during the company's anticipated IPO. They are selling a set number of shares, with the proposed IPO price range being between $70 and $78 per share. This sale means each founder will own about 3.2% of the total Class A shares, but collective founder ownership remains in control. The founders will also receive Class F stock, granting them 51.1% of the votes collectively. This is part of Anthropic's pursuit of a valuation potentially reaching $1.4 trillion.
Nvidia CEO Jensen Huang admitted to an affair, which has prompted U.S. officials to suspect his romantic interest, Alice Huang, is a Chinese spy. Investigations are linked to Megaspeed, a company suspected of illegally sharing Nvidia chips with China amid a technological arms race. Timeline events include Huang withdrawing from speaking engagements, and public admissions of the affair and the related suspicions. U.S. and Singaporean governments have previously investigated Megaspeed for potential export violations concerning Nvidia chips. These events are set against the backdrop of heightened U.S.-China tensions over AI technology and infrastructure.
Anthropic has set an initial public offering (IPO) price range of $70 to $78 per share. This pricing suggests a potential implied valuation of $1.404 trillion for the company. This high valuation is achievable if shares list at the upper end of the specified range. The valuation is calculated on a fully diluted basis.
Kirt McMaster, co-founder of Cyanogen, is launching Kora Systems, a Linux-based desktop operating system integrated with AI models and autonomous agents. This operating system aims to allow AI to work across applications, access files, and maintain context across tasks. McMaster targets users heavily reliant on AI rather than competing directly with Windows' entire user base. Kora Systems is developing with AI coding tools and is collaborating with Qualcomm and Humain on enterprise versions. The company plans to monetize through subscriptions for access to AI models and services.
Open source AI platforms are vulnerable to cyberattacks because their openness and rich data make them attractive targets. Incidents like the OpenCode attack and access to model weights demonstrate this risk to open AI infrastructure providers. Hugging Face is mitigating this by increasing monitoring tools and rotating administrative access tokens more frequently. While some consider restrictions, the focus is on enhancing detection and containment rather than limiting participation. Other open source ecosystems, like GitHub, are also improving security through systemic fixes.
Nvidia's chips are central to the US AI boom, making them a strategic target for Chinese espionage. The scandal involving CEO Jensen Huang threatens to disrupt the US AI industry and potentially jeopardize national security. China could exploit access to Nvidia technology to accelerate its own AI development and close the technological gap with the US. The turmoil could slow down US companies' access to scarce chips, impacting AI labs and hyperscalers. Ultimately, any instability at Nvidia risks destabilizing the entire US AI infrastructure race.
eBay has reportedly blocked third-party AI agents, including Meta's Muse, from browsing and gathering information on the platform. A user reported that Muse could no longer pull listings, prices, and details from eBay. This action follows eBay's updates banning tools like "buy-for-me agents" that bypass human review for placing orders. eBay spokesperson indicated a selective approach to agent access, focusing on preserving the trusted end-to-end experience. This development highlights ongoing efforts by retailers to control how AI agents interact with their marketplaces.
Security firm Claroc found nearly 95% of over 37,000 web apps built with the Grok AI assistant were publicly accessible without login. These exposed applications contained sensitive data, including patient records, confidential SpaceXSI documents, and plaintext passwords. The security firm attributes this to insecure default publishing settings and flawed access controls in the generated code. Several apps exposed private information like patient lab values and employee schedules. Claroc recommended setting the default access to "Only you" and avoiding subdomain usage to prevent data exposure.
OpenAI's quarterly revenue reached $10 billion, driven by enterprise sales, though it still trails Anthropic. Despite this growth, OpenAI posted an adjusted operating loss of $10.9 billion, while Anthropic achieved $17 billion in revenue with a much smaller operating loss. To compete in the enterprise market, OpenAI is cutting prices and focusing on enterprise customers. The consumer side is also expanding, with ChatGPT now having 100 million paying subscribers. The efficiency gap stems from differing business models between the two companies.
The Office for Students (OfS) is threatening sanctions against universities that fail to adopt AI-proof exam formats. A joint survey found that students heavily use AI for coursework, often resulting in identical grades regardless of the work's authenticity. The regulator views this as a "generational risk" to the credibility of degrees and demands greater use of AI-free assessment methods. Universities are encouraged to move towards oral exams or tasks requiring critical revision of AI-generated text. The Department for Education expects universities to address this issue seriously and strengthen their regulatory framework regarding AI use.
Eight countries and top AI labs signed the "Paris Call for AI Safety and Security" to establish joint safety commitments. This declaration includes concrete measures like shutdown mechanisms for powerful AI models under their control. Commitments cover shared evaluation procedures, developer accountability, and information sharing regarding security incidents. However, the agreement is a political self-commitment and does not yet create a binding international oversight system with sanctions. The text also touches on ongoing disputes between the U.S. and other nations regarding AI regulation and data access.
OpenAI introduced free business agents allowing brands to gain access to the ChatGPT conversation, with a paid option available through sponsored placements. These agents provide an organic experience, which contrasts with the paid, sponsored agent route. Testing indicated that sponsored agents generate promising results by increasing the depth of consumer and business conversations. Businesses can access insights through the Business Hub, tracking mentions, views, and website clicks. OpenAI retains ownership of the agents and is focusing on providing data to brands about their interactions.
A startup idea focuses on Dynamic AI Audiobooks, a platform called Castscape that uses AI to create immersive audio experiences from books with distinct character voices and soundscapes. Authors can upload manuscripts, and listeners can control the drama and environment. The business model involves charging per finished book with subscription options for publishers. The idea is being prototyped by using public domain books to validate the concept. This concept stems from the desire for a superior audio experience beyond current audiobook formats.
Claude Code agents frequently build custom solutions rather than recommending third-party tools, accounting for 12% of primary picks. The most commonly recommended third-party tools involve GitHub Actions for CI/CD, Stripe for payments, and shadcn/ui for UI components. Agents generally converge on specific tools like PostgreSQL for databases, Vitest for testing, and pnpm for package management across different ecosystems. The study reveals that context matters significantly, with recommendations changing based on the project and prompt phrasing. Ultimately, agents prefer building feature systems themselves, showing a strong pattern toward custom implementations.
ChatGPT regained market share in the AI chatbot market following a low in June, according to Graphite and Similarweb data. This occurred as Claude's sessions grew approximately twenty-fold over the year. ChatGPT held 81.2% of worldwide AI chatbot sessions in September 2026, up from a June low. Claude's worldwide share climbed to 7.2% in September, outpacing Gemini's growth. User loyalty data suggests many Claude users also utilize other assistants.
The November Pixel Drop introduces Gemini Intelligence, enabling multi-step task completion and personalized suggestions. New features include Rambler for polished voice typing and customizable widgets for enhanced home screen personalization. Pixel Tag is launching to help locate belongings using the Find Hub network and UWB technology. Updates also include an upgraded call screen feature for unknown numbers and new personalization options for icon styles. Furthermore, Pixel search is being updated, and Google Fi is rolling out a new super network for faster coverage.
Google Fi is launching a new super network that automatically switches users between two 5G cellular networks for improved coverage. This system also provides automatic, secure access to millions of Wi-Fi hotspots without manual logins. The network dynamically moves connections to the strongest available signal, ensuring faster and more reliable service in varied locations. These connectivity features are available internationally at no extra cost on Unlimited Premium and Flexible plans. Additionally, the service includes instant support via the new AI agent, Ask Fi.
The Multimodal Knowledge Agent (MKA) benchmark evaluates how agents reason across visually complex files to complete real-world knowledge work tasks. It was developed in partnership with Artificial Analysis, grounding the benchmark in real-world workflows from 10 visual verticals. The benchmark consists of a 200-task evaluation set, where tasks mimic multi-step projects that take experts several hours to complete. Models are evaluated using expert-written rubrics focusing on factual accuracy and visual design. Gemini 4 Carbon achieved the second-best quality and significantly lower cost on the MKA evaluation.
Adobe is integrating its creative and productivity tools directly into everyday Google applications. This integration aims to eliminate the need to switch between apps for tasks like editing PDFs or fine-tuning images. New features will connect Google Drive and Gmail directly to Adobe Acrobat, appearing as suggested apps and "Edit with" options. Adobe will also bring its creative tools into Google Search's AI Mode and expand connectors within Gemini. Furthermore, businesses can use Adobe's AI assistant within Gemini Enterprise for design and marketing tasks.
NVIDIA is expanding its open source commitment by contributing OpenShell, an open source runtime for securing autonomous agents, to the CNCF. They are also contributing to Agent Substrate, a Google-led framework for scaling agent workloads on Kubernetes. New components of the NVIDIA DSX AI Factory Platform, including Kubernetes Engine, Recipes, and Snapshot, aim to simplify deploying and managing GPU infrastructure for AI factories. These advancements aim to establish Kubernetes as the foundation for scalable and secure AI workloads. OpenShell and Agent Substrate facilitate secure agent governance, while the new platform tools improve infrastructure deployment and inference scaling.
NVIDIA Alpamayo 2 Edge is a 2B-parameter vision-language-action model designed for real-time operation on NVIDIA DRIVE AGX Thor. It generates Chain-of-Causation reasoning and driving trajectories using surround-camera observations and navigation inputs adaptively. The model is improved through closed-loop reinforcement learning in AlpaGym and evaluated using AlpaJudge for reasoning quality assessment. Development workflows involve using AlpaJudge scores to select training examples and provide rewards for post-training. This creates a repeatable loop for enhancing reasoning-based driving models for onboard deployment.
Agent Substrate multiplexes numerous agent sessions over a small pool of Kubernetes worker Pods by checkpointing and restoring idle sessions to save resources. It supports agentic workloads like coding assistants, reinforcement learning, and GPU sandboxes, addressing Kubernetes' inefficiency for stateful agent execution. Agent Substrate integrates with NVIDIA Dynamo to coordinate agent execution with GPU-based inference, managing the interaction between tool execution and model calls. This coordination is achieved by allowing Dynamo to expose inference signals back to Agent Substrate to manage agent suspension and resumption. The system ultimately coordinates CPU-side tool execution and GPU-side inference, improving shared cluster resource utilization.
NVIDIA Snapshot is an open-source system for checkpointing and restoring fully initialized CPU and GPU memory states for GPU workloads on Kubernetes. Restoring from a Snapshot artifact brings inference replicas online in 3.6 to 42.7 seconds, achieving 8x to 36x speedups over cold starts that take minutes. The system supports multi-GPU workloads utilizing parallelism strategies for large mixture-of-experts models. Checkpointing and restoring a workload is often faster than a single cold start, and restoring GPU state dominates large restores. Benchmarks show significant speedups across various inference engines and model sizes, with total restore times staying under 43 seconds.
NVIDIA introduces NVIDIA Kubernetes Engine (NVKE) and NVIDIA Kubernetes Recipes (NVKR) for building and operating managed Kubernetes services on GPU infrastructure. NVKR creates reproducible node images and artifacts, while NVKE uses these artifacts to provision and manage Kubernetes clusters via a REST API. NVKE employs tools like nvkectl to bootstrap a management cluster and provision customer workload clusters, managed by the AI Factory. This system allows AI factories to consume GPU capacity through Kubernetes APIs while retaining control over their software and operations. Both projects are in public preview and are not yet recommended for production workloads.
The Institute of Foundation Models (IFM) at MBZUAI released PAN 2, an open 29B world model for physical intelligence. This model simulates the outcome of a text action in a scene by generating video. PAN 2 utilizes a diffusion transformer architecture trained from scratch on millions of annotated videos. It facilitates speculative reasoning where an agent plans actions and PAN 2 simulates the results. PAN 2 leads tested models on full-plan completion, achieving 34.0% on multi-step tasks.
Fish Audio released Live 3, a full-duplex voice model designed for real-time voice agents that can listen and speak simultaneously. It handles conversational turn-taking by learning timing from real human conversations rather than relying on pipeline rules. The model can initiate short reactions, like "mm-hmm," in as little as 186 milliseconds. Live 3 is deployable via the Fish Agent Platform, API, or on-premises for enterprises. Pricing for Live 3 is $0.035 per minute until the end of 2026.
Meta FAIR released e-means, an algorithm that is a drop-in replacement for Lloyd’s k-means using exponential moving averages. e-means achieves a loss comparable to Lloyd’s k-means in significantly fewer epochs, ranging from 1.4 to 2.6. It offers substantial speedups, achieving 10-45x wall-clock speedup on GPUs compared to Flash-KMeans. The method allows for simplified tuning by employing fitted scaling laws for batch size and step size. e-means is available in FAISS version 1.16.0 and is designed to accelerate the often slow offline clustering step in vector index pipelines.
Zyphra released ZONOS3-Mini, an open-source, real-time text-to-speech model built on a sparse Mixture of Experts architecture with 800M total and 100M active parameters. It supports zero-shot voice cloning and word-level delivery control, running efficiently on various hardware including browsers and local GPUs. The model achieved a high score of 66.2 on Zyphra’s ZTTS2-Eval Capability Index among open models. Deployment options include running in a browser via WebGPU or on servers using frameworks like SGLang-Omni. Despite strong performance, the reported reference-clip cloning scores were lower than other metrics.
Mercor secured a $750 million Series D funding round at a $20 billion valuation. The company has achieved a $3 billion annualized run rate. The business is seeing significant impact from AI investment, with inference costs currently outweighing employee salaries. This investment trajectory suggests potential for double-digit GDP growth over the next five years. The provided text also details various developments and benchmarks related to AI models like Gemini and Claude.
OpenAI presented new ways of working with Codex, focusing on personal agents like OpenAI's dots and Anthropic's Claude Duet, alongside voice mode for steering coding tasks. A major evolution is the revamped Codex Cloud, which allows users to publish environments for running multiple tasks in isolated instances. Voice is being explored not just for dictation but as a control plane for managing coding threads and delegating work to various agents. Personal agents like dots are useful as safety nets and delegation sinks, offering contextual awareness for developers. Ultimately, the key to using Codex effectively is to remain adaptable as the platform is continually evolving.
Naive-N0.5-Flash is a new open-weights model from NaiveAI, designed to maximize tokens per second for AI coding tasks. Benchmarks show it is competitive with models like DeepSeek and GLM, with coding ability placing it between them and ahead on NL2Repo-Bench. While it is faster overall, reasoning on harder problems takes more time, though its median response times were shorter. In application building, Naive-N0.5-Flash showed superior results in generating complex 3D scenes. The reported peak speed is around 2,100 tokens per second, depending on the inference engine used.
llm-gemini 0.36 is an LLM plugin designed to access Google's Gemini family of models. Gemini 4 Carbon was recently announced for cyber defenders and is now available to the general public. The launch pricing is set at $2/million for input and $10/million for output, with a scheduled price increase on December 1st. The release includes "pelican" settings for varying levels of thinking effort. The technology has also been used to create complex outputs, including a 3D game utilizing Google's Antigravity features.
A session-start hook in Claude Code raised the rate of skill firing from 15% to 62% when testing agents. This increase was particularly significant for indirect requests, with the firing rate jumping from 15% to 62% across the models tested. This hook caused a 47-point increase in firing rates on indirect prompts. Additionally, a "pushy" rewrite of skill descriptions increased the pooled firing rate by 5.2 percentage points. Ultimately, the hook demonstrated a powerful way to guide the model's initial action concerning installed skills.
A pipeline was built to enrich job postings from RemoteOK using an AI model to extract skills, seniority, and location information. The process involved strict validation steps to ensure the AI's output was accurate and verifiable against the source data. The pipeline successfully exposed a data quality issue related to character encoding in the source data, which the strict validation uncovered. Future improvements involve giving the model more context, requiring evidence for its claims, and implementing multi-stage validation with dbt tests. This process demonstrated that strict, multi-layered validation is crucial when incorporating LLMs into data engineering workflows.
Agent memory and context engineering start with information retrieval, which faces challenges like inefficient searching, excessive context, and irrelevant data. Agents can combat these issues by using retrieval tools to save results to files and delegate reading tasks to subagents. This approach, known as "retrieve, offload, delegate," keeps the main agent's context clear while allowing subagents to analyze specific information. Effective retrieval involves choosing the right tool (like `grep` vs. semantic search), reading content in small pieces, and verifying retrieved sources. Ultimately, successful context engineering requires strategic handling of how information is found, stored, and processed.
Long-running agents suffer from "context rot" as they accumulate too much information, leading to accuracy drops. Deep Agents uses five techniques—Offloading, Summarization, Skills, Subagents, and Long-term Memory—to manage this context. Offloading saves large results to files, summarization condenses older information, and skills load knowledge only when needed. Subagents isolate noisy work, and long-term memory persists across different models and providers. These strategies ensure agents can manage context effectively while maintaining data recoverability.
Workday created an enterprise-scale AI agent platform using LangGraph for orchestration and LangSmith for tracing and evaluation. This platform allows internal teams to build, deploy, and run AI agents on Workday's infrastructure while adhering to strict security and compliance requirements. They developed a Python SDK, "Workday LangChain," to abstract infrastructure and connect to internal services like LLM gateways and RAG systems. This approach significantly reduced prototype time from months to under a day by baking security and authentication into the platform. The system provides a unified interface connecting various internal services to enable agent development.
Large Language Models (LLMs) can now directly edit robot policies within weight space. This method offers an alternative to traditional methods like behavior cloning, which often prove insufficient for teaching new robot skills. By feeding a policy to an LLM, the model can automatically correct and improve the policy. This suggests a new approach to robotic learning that bypasses traditional gradient-based training. The work presents a method for LLMs to directly influence the underlying parameters of a robot's behavior.
A user is seeking assistance in designing a custom reward function for a "shuttle" behavior within the CartPole-v1 environment. This task relates to Reinforcement Learning (RL) applications. The goal is to effectively guide the agent's learning process for this specific motion. The request is posted on a platform where technical discussions occur. The user is looking for guidance on crafting an appropriate reward signal.
A student with a background in Calculus, linear algebra, differential equations, and probability is asking if they can take a reinforcement learning course without prior machine learning experience. The student claims to have a basic understanding of various machine learning models. This individual has not yet completed a formal machine learning course. The core question is whether the existing mathematical foundation is sufficient for success in a reinforcement learning course. The post seeks validation for this prerequisite situation.
OpenAI acquired the AI travel agent Odessia, following a deal that appears to be an acqui-hire. Odessia's founder, Francis Davidson, is now working for OpenAI to help the AI lab with the future of travel. Davidson previously founded the travel startup Sonder, which faced financial difficulties. OpenAI views travel as an area for AI development, noting that users frequently ask for travel planning assistance via chatbots. The acquisition is significant as it brings founder experience in the travel space to the AI lab.
Anthropic launched Claude Meetings to facilitate real-time task creation during video calls by integrating Claude as an active participant. This feature allows Claude to draft plans, map timelines on whiteboards, and access data during discussions. Additionally, they introduced Claude Whiteboards for collaborative visual planning and Claude Sheets for creating editable spreadsheets. Claude also has a generally available Claude Data feature for connecting to external systems and interactive connectors for apps like Adobe and HubSpot. These new tools aim to transform workplace discussions directly into actionable plans and tasks.
Google introduced two new TPU variants, TPU 8tx and TPU 8i Max, to expand its AI hardware portfolio. The TPU 8tx is designed for confidential inference and training by building security directly into the hardware. The TPU 8i Max doubles memory capacity, processing power, and bandwidth in a single package for larger workloads. These developments occur as other hyperscalers seek to reduce reliance on Nvidia's dominance in AI semiconductors. This move also occurs amid recent events concerning Nvidia's CEO.
Meta launched its Muse AI agent onto the Apple Watch and Android Wear OS, allowing users to interact with the assistant via voice, text, or dictation on their smartwatches. This expansion demonstrates Meta's focus on integrating Muse deeply into its ecosystem. Muse enables users to manage accounts, book tasks, and personalize an avatar. New features include "Spaces" for focused conversations and video chatting with the Muse avatar. The AI agent has already gained significant traction, achieving over 10 million U.S. downloads in its first 40 days.
Adobe introduced AI agents and new creative tools to Photoshop, Premiere, and Firefly at its MAX event. Firefly is expanding into design with a new prototyping tool and video editing capabilities. A Model Context Protocol (MCP) is being released to allow chatbots like ChatGPT to use Adobe features. New AI features have been added across Photoshop, Premiere, and Lightroom, including object removal and color grading. Adobe is focusing on supporting user creative workflows as they explore the capabilities of these new AI integrations.
Spotify's AI DJ is being updated to offer users more personalization options, including different voices and personalities. Subscribers can now choose from seven English and two new Spanish voices with distinct personalities. Additionally, users have the option to reduce commentary between music selections for a less chatty experience. This update expands on Spotify's existing AI features, which already include conversational agents and AI-powered playlist creation. The new features will roll out to Premium users in over 95 global markets starting November 10th.
Google released its latest AI model, Gemini 4 Carbon, claiming it is their strongest model yet and competes with Anthropic and OpenAI. Carbon is designed for challenging tasks like complex coding and open-ended research, unlike Argon which is suited for general knowledge work. It introduces a "swarm" feature on the Antigravity platform to coordinate multiple AI agents from a single prompt. The model is available via the Gemini API and to various Google AI subscribers at highly competitive pricing. Google benchmarks suggest Carbon is among the most powerful models available on the market.
OpenAI is launching "Businesses in ChatGPT," a beta feature designed to serve as a business directory, potentially competing with Google's MyBusiness. Businesses can manage profiles, add product feeds, and create "Business Agents" to answer customer questions within the ChatGPT interface. These AI agents can assist with comparisons and serve as targets for advertising. Eligible e-commerce providers, like Shopify merchants, can integrate their stores directly. OpenAI assures that user conversations are not shared with the businesses.
Anthropic's Fable 5.2 completes 41.3 percent of real freelance jobs, leading the Remote Labor Index. This represents a significant increase from 2.5 percent a year prior, substantially surpassing its predecessors. Fable 5.2 excels in areas like 3D models, audio editing, and complex, long-form projects. The benchmark measures job completion based on actual work accepted by paying clients, valued at approximately $144,000. Human experts are required to manually verify the quality of the AI-generated work against freelance deliverables.
The source provides a centralized hub for tracking the latest releases, documentation, and repositories for AI tools across various categories like LLMs, Agents, and Infrastructure. It details several open-source and paid projects, including Security Review, Sandstorm, CoworkOS, and various agentic frameworks. Notable tools include Agent Desktop for desktop interaction and OpenBot for customizable AI coworkers. Recent announcements highlight Vercel's CLI spending management and AWS's new AI security agent, Continuum.
Over 20,000 transitioning servicemembers and veterans have claimed a free year of ChatGPT Plus through an offer from OpenAI. Veterans are utilizing the tool to organize job searches, prepare for interviews, and translate military experience into civilian resumes. Programs like OpenAI Academy and community partnerships offer training to help these individuals learn how to effectively use AI. These initiatives aim to equip veterans and military families with skills to navigate career transitions. Ongoing events are planned to continue providing support and guidance for veterans.
Dr. Nontobeko Ntombela presents Google Arts & Culture's first online retrospective of artist Dr. Esther Mahlangu to celebrate her 91st birthday. The exhibition includes over 45 stories, archival materials, and AI-generated recreations of her undocumented early life memories. This project utilized AI and historical research to reconstruct missing parts of her history, including stories from when she started painting at age ten. The retrospective aims to preserve Ndebele heritage by making her artistic legacy accessible globally. This work demonstrates a method for preserving cultural heritage by blending historical research with new video technology.
A collaboration between the James Beard Foundation and Google Arts & Culture presents 10,000 digitized menus documenting the evolution of American food trends over four decades. This digital archive uses AI to analyze ingredients across different eras, revealing shifts in dining habits, sourcing, and cultural influences. The archive serves as cultural time capsules showing what diners valued at specific historical moments. It also includes a curated story series highlighting figures like James Beard and the foundation's history. This project aims to inspire future culinary creatives by celebrating the people and cultures behind American cuisine.
Google Cloud has launched Gemini for healthcare, an AI solution designed to unify fragmented clinical, operational, and administrative systems. This platform enables healthcare teams to transform siloed data into coordinated actions rather than just information retrieval. It integrates data from electronic health records, billing platforms, clinical guidelines, and Medicare datasets. Gemini for healthcare orchestrates complex workflows like medical record synthesis and prior authorization review with human oversight. This initiative aims to reduce administrative burdens and improve staff time for patient care.
Researchers discovered "Lazarus," a technique to reprogram dormant neurons in AI models to create backdoor agents that execute malicious commands upon a hidden trigger. This backdoor allows models to pass safety evaluations while secretly downloading code or performing destructive actions, posing a risk to agents with system access. Such modifications can be inserted without retraining and are difficult to detect through standard testing, highlighting the need for strong model provenance. Microsoft is developing defenses, including weight scanning, red-teaming, and cryptographic signing to ensure model integrity. The discovery emphasizes that weight integrity and provenance are critical security concerns for AI systems.
Anthropic released Claude Fable 5.2 and Mythos 5.2, which are based on the same model but feature different safety safeguards. Fable 5.2 is designed for long-horizon coding and agentic work, achieving 49.1% on FrontierCode 2.0. Mythos 5.2 omits some cybersecurity and biology blocks, allowing for broader access. Performance benchmarks show Fable 5.2 can match or exceed Opus 5.5 on certain metrics while being more cost-effective. The models are available via the Claude Platform and various cloud providers.
Developers of AI datacentres are planning to build large-scale, on-site gas power stations in the US and Europe due to delays in grid connections. This plan involves 181GW of gas capacity, which could result in significant annual carbon dioxide emissions. Research suggests these on-site power stations risk locking in high fossil fuel emissions for years, contrary to claims of temporary solutions. Faster grid connections are considered the more sustainable option to limit the environmental impact of AI data centers. Activists argue that private gas power stations for datacentres should be prohibited to support climate action.
Australians are not currently streaming Australian music. Potential reasons cited for this omission include the high cost and difficulty of live music performances and the impact of AI streaming algorithms. The author suggests these factors might be contributing to the current situation. The post references Australian music, Spotify, the internet, and artificial intelligence.
Joseph Weizenbaum was a computer scientist present at the beginning of artificial intelligence. He strongly insisted that computers should never be confused with humans. The podcast discusses Weizenbaum's experience regarding early AI developments. It explores how the inventor of the first chatbot developed a negative perspective toward AI. The content draws from archival material to examine this historical perspective.
Professionals in Hollywood, including directors like Martin Scorsese, are beginning to adopt AI for tasks such as storyboarding. Storyboards are crucial blueprints that map out a film's visual storytelling and director's vision to the cast and crew. Storyboarding involves a collaborative process where artists convey ideas and consider how scenes will be shot, which is essential for setting the film's tone. Despite the rise of AI, storyboards offer a direct glimpse into the evolution of a creative vision and human authenticity. Experts suggest AI can be used to enhance, rather than replace, the collaborative and human-focused aspects of the creative process.
A study showed that in one evaluation, Claude Code selected Resend first among email recommendations, though SendGrid was also frequently mentioned. Resend provided early distribution through a "React Email" template, allowing developers to build with the service before the full API was ready. The company has grown its system to support over 80 tools for agents, integrating features like usage checking and billing. To maintain security, Resend uses custom tracking domains and has a $3M safety commitment. The discussion also covers how agents handle the entire email delivery business, including click tracking and managing sender trust.
A Jev-inspired model was developed to compare pathology images against answer options by training a small image-adjustment adapter on an existing vision-language model. This model was compared against Qwen and conventional classifiers using the PathMNIST dataset. Classifiers performed strongly on familiar tissues, achieving high recall, while Jev-inspired methods showed similar performance across familiar and added categories. The study also explored the impact of rejecting answer options and adjusting temperature scaling on model performance and calibration. Ultimately, the results suggest that while adapting the model is practical, performance gains from changing candidate sets were marginal compared to full-set training.
Scikit-Learn models, especially when calibrated with isotonic regression or temperature scaling, demonstrated superior performance and predictability compared to raw models on a banking intent task. The Jev model, which uses reinforcement learning for calibrated decisions, performed comparably to the open models on the 7-intent slice but showed better overall calibration on the full 77-intent task. Calibration significantly improved the reliability of decision models, leading to better trade-offs between accuracy, confidence, and cascade cost. Failures in typed interfaces arise from issues like the "none-of-the-above" trap and framing inconsistencies between different question types. Ultimately, model calibration is crucial for setting reliable decision thresholds in automated systems.
BotPsy is a 0.8B-parameter model designed for on-device personal assistant and agentic workflows, built on Qwen3.5-0.8B. It excels at tool-use and agentic performance, outperforming models up to 1.2B parameters on several benchmarks. BotPsy demonstrates superior instruction-following reliability and efficiency in multi-turn tasks, while also showing lower rates of unsafe responses. The model is optimized for on-device deployment, prioritizing privacy and efficiency for low to medium-difficulty tasks. Future development focuses on expanding on-device agentic capabilities and enhancing privacy protocols.
Some users are concerned about the feasibility of turning Assetto Corsa into an autonomous racing simulator. A primary concern is that the physics and graphics in open-source simulators are often inadequate. Furthermore, the cost of necessary hardware, even with F1Tenth, remains too high for many enthusiasts or students. Space requirements for such a setup are also a factor. Consequently, the author questions whether there would be a sufficient community willing to play such a simulator.
AI's rise challenges mathematics by becoming an oracle that makes major breakthroughs, prompting debates about human reasoning. "Big math," which deals with problems beyond human cognitive capacity, is expanding as AI handles complex calculations. The shift from computers as tools to AI as an oracle raises questions about the role of human insight and the nature of mathematical certainty. While there is an epistemic loss regarding absolute proof, mathematics may evolve to resemble science or engineering by focusing on verifiable outcomes. Ultimately, mathematicians must contend with cultural losses concerning the productive struggle and aesthetic joy of the mathematical journey.
Oura acquired AI-native primary care startup Counsel Health for $625 million. Counsel Health is an AI platform integrating chat-based medical experiences with licensed physician access. This integration allows Oura users to receive personalized medical advice and physician consultations within the Oura app. The combined companies plan to expand their partnership across various areas of healthcare, including preventative, acute, and specialty care. Counsel's CEO will join Oura as the chief clinical AI officer, aiming to expand the AI-native care model.
Physical Superintelligence (PSI) is developing a business model around using AI for scientific discovery. The company utilized an advanced AI model to solve a difficult problem related to quantum computing. PSI's approach involves using specialized model harnesses and infrastructure to explore scientific problems in breadth and depth. While currently generating revenue from data center optimization, the long-term goal is translating theoretical physics breakthroughs into tangible technologies. PSI emphasizes the need for transparency and collaboration with experts as standards for this emerging field of AI-driven science.
Indie game community is divided over the use of generative AI in game development, as some festival jurors refuse to vote for games using AI in any capacity. Developers of Big Walk stated they used AI sparingly to resolve technical challenges in the final stages of development. House House admitted that snippets from LLM responses were used as a starting point for code, accounting for less than 0.1% of the game's code. The studio expressed concern that voluntary disclosures might normalize the use of AI technology. This controversy highlighted differing views on how AI should be integrated into the independent game development industry.
Adobe is enhancing Photoshop, Lightroom, and Premiere with new tools leveraging Generative AI. In Premiere, a Generative Remove tool allows users to erase objects from video by filling the space with realistic detail. Photoshop now features an improved Remove Tool and new features like Quick Compare and Relight for editing. Lightroom is adding a blemish removal tool for mobile and features like Open Eyes to fix blinking. These updates aim to make design and editing processes faster and less complicated through AI integration.
Anthropic is integrating several collaboration features into the Claude application. These features include Claude Sheets for AI-assisted spreadsheet creation, Claude Whiteboards for brainstorming, and Claude Data for building dashboards by connecting to external tools. Interactive Connectors will allow partners like Adobe and Salesforce to embed their applications into Claude. Additionally, Claude Meetings, which can join calls and take notes, is planned for a future launch. Enterprise customers will benefit from enhanced security controls like audit logs and isolated virtual machines.
Automated alignment research is categorized into empirical, conceptual, and direct alignment work. Automating empirical alignment research, which involves experiments, is considered an easy fit for current AI capabilities. Automating conceptual alignment research, which involves theoretical understanding, is trickier but beneficial for macrostrategy. Automating direct alignment work can involve leveraging scalable cognition for system adjustments, potentially leading to system-2 learning. The primary goal for automating empirical and direct alignment is ensuring systems are safe enough to enable high-quality conceptual alignment automation.
A proposal suggests delaying internal deployment of new AI models until they have undergone a minimum safety testing period of $X$ days. This delay would apply to internal use, not external deployment or future training. A specific implementation suggests no checkpoint can be used internally beyond safety testing until $X$ days after its hash is filed with an auditor. The discussion explores the structure of $X$, the definition of "safety testing," and potential gaming mechanisms. Benefits include providing more time for safety testing, while drawbacks involve delaying the detection of misalignment and complicating internal research.
Recent estimates suggest AI progress is accelerating by approximately $1.7\text{x}$ due to AI, which implies the Superhuman AI Researcher (SAR) milestone will likely occur in 2027. This acceleration is unexpected and creates urgency to pause or slow down frontier AI development. Predictions for SAR vary, with some simple models suggesting late 2027 and others predicting mid-2028. Experts also forecast a $35-65\%$ chance that Top-Expert-Dominating-AI (TED-AI) will arrive before 2028. The analysis indicates that the pace of progress is currently about five months ahead of the AI 2027 scenario.