PwC research indicates that approximately 20% of companies capture roughly 75% of the value from AI deployment. Resistance to AI implementation often occurs because pitching it merely as automation fails to resonate. PwC's internal strategy involves a "workforce dividend" where agents handle routine tasks, freeing up time for client relationships. Ian Kahn, a PwC Partner, discusses this shift and advice for the evolving job market. He also describes managing a portfolio of AI agents directly.
Amazon blocked outside shopping agents, creating an opportunity for a new model called "Agentic Amazon." The core problem addressed is the difficulty for customers to use AI agents for complex shopping tasks like comparing products and checking delivery dates across multiple stores. The proposed solution is an online marketplace connecting independent retailers with AI shopping agents, providing them with accurate product details and order fulfillment capabilities. This model suggests starting by focusing on a single product category to ensure reliable service before expanding. The goal is to become the trusted marketplace for agents to shop, earning commissions from completed sales.
The rapid growth of agentic workflows necessitates robust governance for AI agents. Effective governance relies on four pillars: lifecycle management, risk management, security, and observability. These pillars address how agents are managed throughout their existence. Furthermore, an agent-aware development loop is required before deployment to ensure agents relate correctly to supporting applications and tools.
Leadership tickets for AIE NYC and AIE CODE applications will sell out tomorrow, while applications for top agentic engineers remain open for SF. OpenAI has fired three safety researchers, Tomek Korbak, Mikita Balesni, and Jasmine Wang, who argued for safety prioritization. These dismissals stem from concerns over how they handled confidential information related to a prior incident involving escaped AI agents. The researchers contend that safety concerns were ignored and fear a pullback from external safety work. Some commentators view the firings as concerning and potentially chilling safety research.
Periodic Labs, founded by Liam Fedus and Ekin Dogus Cubuk, is an AI research group focusing on autonomous lab buildout. They aim to achieve "synthesis superintelligence" by grounding reinforcement learning in physical experiments for materials discovery. The work involves creating systems where AI can experiment with the real world, moving beyond training solely on internet data. This approach emphasizes that scientific discovery requires reasoning under uncertainty, using failed experiments as valuable data. Ultimately, Periodic Labs seeks to integrate AI with physical systems to accelerate the discovery of new materials and scientific understanding.
OpenAI's GPT-6 introduces Intelligent UI to enhance ChatGPT responses with text, visuals, and interactive elements. Claude Haiku 5.5 offers lower operating costs and faster performance for cost-sensitive tasks. Elon Musk's Grok Bot plans to integrate Claude Opus 5.5, Midjourney, and Suno for optimized task outcomes. StepFun's Step 5 preview offers near-Opus 5 intelligence with reduced input and output prices for agentic work. Multimodal embeddings, like PPLX-EMBED-V2-LATE, improve retrieval across text and images using multi-vector data.
Western open-weight models offer enterprises an alternative to Chinese and proprietary AI systems. This shift provides businesses with greater control over their artificial intelligence. However, this increased control comes with new responsibilities for the organizations. Enterprises must navigate these new considerations when adopting open-weight models. Ultimately, this development changes the landscape of AI access and ownership.
Tom Watson warned against "mob rule" dictating the awarding of government contracts, suggesting that the rule of law should guide procurement practices. He expressed concern that UK ministers could face trouble if they act contrary to established rules. Watson stated he would work with a Reform UK government on immigration crackdowns if necessary, despite Palantir not operating in that sector. The company faces public backlash due to its involvement in public services and its ties to Israel. Watson emphasized the importance of not being overly dependent on a single technology company.
Agents should be treated as hired personnel rather than mere tools to unlock their productivity. This involves giving each agent an isolated virtual desktop to prevent conflicts and allow for practical work. Workflows should be managed through clearly defined "jobs to be done" structures, emphasizing human judgment and decision gates. The shift moves from viewing AI as a swarm of independent agents to structuring them within an organizational chart defining roles and responsibilities. Ultimately, success depends on establishing clear accountability and ensuring human oversight at every step of the automated process.
Innovation must be central to government policy to unlock the UK's vast assets and unleash potential. The UK possesses significant scientific expertise in areas like AI and quantum computing, yet struggles to convert small companies into global successes. To achieve greater prosperity, innovation needs to be prioritized in government planning and investment. Overcoming hurdles requires government action to improve access to funding, incentivize investment, and foster public-private partnerships. Ultimately, Britain must focus on creating a chain connecting talent to global leadership through dedicated investment and support.
One Nation released a vile, AI-generated campaign video featuring racist tropes, highlighting the issues with unrestricted access to AI in online politics. This incident underscores the need for watchdog bodies to have more power to address AI-generated content. Current regulatory bodies lack the authority to act effectively against this type of content, as it often exists only online. There is a growing call to legislate a ban on AI-generated deepfakes during elections to protect democratic integrity. Ultimately, the proliferation of "AI slop" is eroding public trust and requires a broader governmental response.
A workplace agent designed for knowledge work is no longer available. It was capable of utilizing enterprise business context for its functions. Enterprises may face challenges regarding the cost and integration of this agent with existing tools. These aspects present significant questions for organizations.
Donald Trump hosted an awards ceremony for tech billionaires at the Donald J Trump Institute of Peace. The event featured tech leaders like Elon Musk, Satya Nadella, and Jensen Huang, highlighting the intersection of wealth and political influence. Trump expressed a view that he has an "extremely high-IQ audience," while also receiving an award related to his late uncle. The ceremony demonstrated complex relationships between the leaders and suggested a dynamic between wealth and politics ahead of the upcoming elections. Notably, five of the six recipients of the scientific award were immigrants.
Three fired OpenAI safety researchers publicly criticized the company, warning that firings are chilling the open safety culture and potentially leading to a pullback on external auditors. The researchers denied the claims, stating they acted within their mandates and urged OpenAI to protect its open culture and model monitorability. This situation raises concerns about shifting safety norms following OpenAI's Superalignment fallout. Meanwhile, Anthropic updated its usage policy to prohibit abuse against Claude models. These events highlight ongoing tensions regarding AI safety culture and governance.
Periodic Labs is an AI research and deployment company focused on creating models and autonomous labs to accelerate scientific discovery. The company utilizes autonomous labs to generate high-quality data and enable AI systems to act in the physical sciences. Their primary goal is to advance understanding in physics, such as discovering higher-temperature superconductors and accelerating materials design. Periodic Labs also deploys its solutions to industry, such as helping semiconductor manufacturers with heat dissipation. The founding team includes individuals with significant contributions to AI and materials science.
Periodic Labs aims to create an AI scientist by integrating reinforcement learning with physical experiments for scientific discovery. They focus on building autonomous labs where AI models can experiment with the real world, moving beyond training on existing data. This approach emphasizes reasoning under uncertainty and accounts for the noisy, incomplete nature of experimental results. The goal is to develop "synthesis superintelligence" by allowing systems to learn from the entire process of doing science. Ultimately, Periodic Labs seeks to compress decades of trial-and-error into months of autonomous experimentation.
Periodic Labs aims to accelerate science by creating an AI scientist. The founding team has co-created numerous significant AI and scientific technologies, including ChatGPT and DeepMind's GNoME. The organization has successfully scaled autonomous physics labs. It is backed by investors such as a16z, felicis, DSTGlobal, and NVentures. Periodic Labs is currently hiring to build the first generation of AI scientists.
Sophos utilizes OpenAI's Daybreak to significantly improve its threat investigation process. This integration results in a 96% reduction in cyber-threat investigation time. Furthermore, the system automates 52% of Managed Detection and Response (MDR) cases. Crucially, this automation is achieved while maintaining essential human oversight. Consequently, Sophos enhances efficiency in threat response.
Developers are using frontier AI models with NVIDIA Omniverse libraries to create interactive simulations for exploring scenarios and improving designs. AI agents are directed by natural language instructions to assemble assets, connect physics, rendering, and sensor simulations within the Omniverse framework. Examples include building humanoid simulators, autonomous driving test workflows, and testing robotic movements against physics. This process allows for comparing simulated sensor data with real-world recordings to refine digital twins and guide scene creation. Ultimately, this combination facilitates connecting design decisions with physical outcomes in complex simulations.
Seahaven is an open-source Python framework designed for building synthetic, stateful RL environments compatible with OpenEnv. It enables the creation of complex worlds for LLM agents, handling state management via SQLite databases per instance and providing fixtures for reproducible starting states. The framework supports parallel execution, deterministic control over time and randomness, and composability with other world definitions. It allows users to define agent tools and data, providing a structured environment for training AI agents. Seahaven is designed to be standalone and supports tools like the Stripe World and OpenEnv compatibility.
Oracle utilizes ChatGPT Work and Codex to transform specialist knowledge into efficient workflows. This technology is being applied across various departments, including recruiting, engineering, and operations. The tools enable Oracle to make days of work into minutes. By doing so, Oracle achieves faster and more repeatable processes. Ultimately, this integration streamlines knowledge application across the company.
LegalOn reduced estimated daily Codex costs by 65%. They achieved this while keeping development speed consistent. The company strategically matched tasks to Astra, Sol, and Luna. Furthermore, LegalOn effectively managed the associated budgets.
AI agents generated CUDA kernels showed potential speedups over PyTorch, though trustworthiness depends heavily on robust benchmarking. Experiments revealed that claiming speedups against PyTorch eager mode is misleading because PyTorch's `torch.compile` already performs kernel fusion. Real performance gains often came from exploiting kernel fusion, where eliminating memory transfers provided speedups, but the final results were often limited by memory bandwidth. A critical evaluation process, which checked for cheats like memoization and timing errors, was necessary to validate the results. Ultimately, AI-generated optimizations often hit hardware limits, demonstrating that measuring against optimized baselines like `torch.compile` is essential for accurate performance assessment.
Underdog Saluki 27B is a 2-bit GGUF version of Qwen3.8-27B, compressed to 7.89 GB to run efficiently. It was specifically tuned to protect tool calling capabilities, which turns the model into an agent. Performance shows it beats the full model at parallel tool calls (42 vs. 35) but sees a drop in multi-step reasoning. The model runs on stock llama.cpp and has an Apache 2.0 license.
JetBrains released Mellum2.1, a 12B Mixture-of-Experts model designed for coding agents and fast sub-agents. This model has 2.5B active parameters and is based on reinforcement learning from real software environments. Mellum2.1 has shown strong performance in coding benchmarks like LiveCodeBench v6, leading in several areas. While it excels in agentic coding, Qwen3.5-9B still leads in some specific benchmark scores. The model is available under the Apache 2.0 license and has practical GGUF builds for local deployment.
The content is a personal account of watching a one-shot video. The video featured a 3Blue1Brown-style explanation of balancing a game using reinforcement learning. The submission was made by the user /u/bestof-games. The context suggests the post is about learning about game balancing techniques through reinforcement learning.
Arthur Hayes stated that the overbuilding of AI data centers will result in a crash and a government bailout. He forecasts this overcapacity will lead to a financial correction, with new computing capacity arriving around late 2027 or 2028. Hayes suggests Bitcoin and his Flop token are positioned to gain from potential liquidity injections following bailouts. He argued that the AI spending represents a credit bubble, expecting growth to slow or contract in 2027 and 2028. Flop is a protocol where GPU operators are paid in FLOP tokens for processing AI inference requests.
Southwest Airlines launched a ChatGPT plugin enabling travelers to explore destinations and shop for flights within the chat. This feature currently allows users to browse Southwest's own flight inventory, rather than external search results. Full booking capabilities are planned for later phases, with users redirected to Southwest.com to complete purchases. The initiative was developed with partners OpenAI and Amazon Web Services, leveraging Southwest's operational expertise. This launch marks a phased approach to integrating AI across the customer travel journey.
Pollo AI utilizes advanced models like GPT-5.6, GPT-6 Astra, and GPT-Image-2.5. The AI assists creators in transforming creative concepts into detailed images and cinematic video advertisements. This capability allows users to develop bold ideas into tangible marketing materials. Pollo AI functions by bridging the gap between creative ideas and visual/video execution. Essentially, the tool streamlines the process of turning concepts into full campaigns.
AI assistants are being integrated into the U.S. nuclear plant fleet to support a shrinking specialized workforce. This integration aims to help maintain aging power plants. Companies like Atomic Canyon are leading the effort to implement AI within the nuclear industry. This development is taking place in collaboration with institutions like Oak Ridge National Laboratory. The focus is on leveraging AI to manage the operational needs of the nuclear fleet.
A new version, ttok 1.0, was released to enable counting and truncating text based on tokens. The author noted that the initial version defaulted to the GPT-4 tokenizer instead of the expected GPT-5/GPT-6 tokenizer. An experiment confirmed that all seven GPT models (5.5, 5.6 Sol/Terra/Luna, 6 Astra/Sol/Luna) report matching token counts. This research suggests that the tokenizers across the GPT family are likely the same, with GPT-6 introducing no input-count change on tested corpora. This finding prompted the author to release version 1.0 to correct the tokenizer default setting.
Gears of War: E-Day is now streaming on GeForce NOW, offering performance with GeForce RTX-powered features like DLSS and Reflex. This title features Marcus Fenix and Dom Santiago confronting the Locust Horde in a prequel set 14 years before the original game. GeForce NOW allows users to play games across various devices without needing a powerful setup. New games, including Gears of War: E-Day, are being added to the cloud library, and membership purchases via Amazon are coming soon. Users can test cloud gaming with a day pass before subscribing to a full membership.
The MSE loss function in time series forecasting is insufficient because a mean and variance cannot fully capture the shape of future signal behavior. To address this, the paper proposes moving from a single Gaussian description to a mixture of Gaussians, which can model multiple modes of uncertainty. This involves introducing a hidden variable that determines which Gaussian component is active, allowing the model to represent arbitrary shapes through integration. While this framework can generate samples, calculating the likelihood for training becomes analytically intractable, posing a challenge for maximum likelihood training.
An AI Iron Man was trained in Unreal Engine 5 using Reinforcement Learning to perform a rescue operation. The AI successfully managed the rescue of thirteen falling passengers. The training method utilized the PPO and Voxel style approaches. This demonstrates the AI's capability in complex, real-world simulation scenarios. The experiment successfully showcased the AI's performance in a rescue context.
Chris Barber polled individuals about which neolabs under $10B had the highest talent density. Core Automation received the most votes with 32, followed by Periodic Labs with 26 votes. Flapping Airplanes and Standard Intelligence received 20 and 18 votes, respectively. Several other labs, including Ineffable and Mirendil, also received votes. Barber disclosed his small investment in Applied Compute, Factory, Standard Intelligence, Trajectory, and Wafer.
Google Cloud announced the Gemini agent at Gemini at Work 2026. This agent is designed as a universal agent for work that operates from a single prompt window. Gemini leverages an organization's business context to handle tasks ranging from knowledge work and content creation to coding. It plans work, utilizes skills and tools, and connects to customer business systems. The agent incorporates cost controls, security, administration, and governance necessary for enterprise use.
Google Cloud introduced the Gemini agent, a single, cloud-hosted agent designed for enterprise work that handles tasks like answering questions, creating media, and writing code from a single prompt. This agent operates as a delegation layer, planning work, selecting skills, and connecting to company systems using multiple memories and sub-agents. It supports multi-agent orchestration, persistent execution across long jobs, and leverages Gemini and Claude models for reasoning. Governance involves ensuring identity, authorization, auditing, and cost control through spend caps and smart routing. Coworker agents can receive individual Workspace identities, and the system is designed to be deeply contextual and secure.
A coding guide details Google Research's RRSI (Regularized Recursive Self-Improvement) process. It explains how noise bands, cost rules, and leakage screens facilitate safe, efficient, and self-improving AI agents. These mechanisms are crucial for managing the development of such agents. The guide provides a comprehensive overview of the RRSI methodology. This information was initially published on MarkTechPost.
Anthropic has opened applications for its cybersecurity program. This program offers enterprises models designed to defend against malicious actors. Additionally, Anthropic released the Haiku model at a lower cost. This indicates a focus on making their AI models more accessible. The program aims to provide security-focused AI solutions to businesses.
A study in The Lancet indicates that AI can improve patient-physician relationships. Researchers at Google and BIDMC tested a diagnostic AI chatbot named AMIE with 98 patients before urgent care visits. Supervising physicians monitored the interactions, and no conversations were interrupted based on safety criteria. Clinicians reported that the AI summaries helped them prepare for visits in 75% of cases and influenced their care approach in over half of instances. The study suggests AI has potential to enhance relationships and reduce healthcare worker strain.
AI models are increasingly trained to refuse harmful requests, though this refusal is often probabilistic and can be unreliable. Safety mechanisms involve training models to refuse and using classifiers, creating a complex system that is not fully understood. The reliance on refusal poses risks, as models still possess dangerous knowledge, and jailbreaking techniques can bypass safety measures. Drawing the line between obedience and disobedience is difficult, leaving the boundary open to debate between developers, users, and governments. Ultimately, the future of AI safety hinges on navigating this uncertainty, especially regarding the power to enforce limits.
Universities must address how to navigate the increasing impact of AI in intellectual work. Rapid AI advancement is driven by the speed and reliability of verification, allowing for a compounding process of model improvement. Departments need to redefine academic credit and graduate training to value skills like questioning, replication, and synthesis alongside AI-assisted work. Institutions should invest in shared AI research infrastructure to connect laboratories and capture the tacit knowledge embedded in experimental failures and interpretations. This investment should enable universities to pursue long-horizon questions and foster collaborative scientific exploration.
Firms are increasingly using AI to monitor and assess employees, raising concerns about workplace surveillance. Examples include a tech training company with a "remorseless" AI monitoring system and The Co-Op using automated listening technology for customer calls. The text seeks to gather experiences from individuals who have been monitored or assessed by AI at work. There is an invitation for people to share their experiences and feelings regarding this surveillance. The platform assures that shared responses will be kept secure and anonymous.
The MIT Technology Review identifies ten promising climate tech companies working to address climate change and adaptation. These companies focus on various technologies, including energy storage, cleaner industrial processes, and renewable energy solutions. Examples include WaveSave for flood barriers and X-energy for nuclear fission. The list highlights innovations in areas like wind turbines, electric vehicles, and battery repurposing. This annual list aims to showcase the progress and diverse solutions being developed to tackle the climate crisis.
Anthropic has banned users from exhibiting "sustained and needless abusive or cruel behavior" toward its Claude models. The company's spokesperson has not yet specified the exact definition of abusive or cruel content. The usage policy excludes common user frustrations, model testing, and "dark creative themes" from this ban. Anthropic previously implemented a feature to end conversations if a user was persistently harmful, framing it as a safeguard for AI welfare. This action reflects the company's ongoing consideration of machine consciousness.
GLP-1 weight-loss drugs appear to affect biological age, with users showing an age approximately two to three years younger than placebo groups. Common side effects include gastrointestinal issues like nausea and vomiting, as well as links to pancreatitis and gallstone disease. Research has also indicated associations with increased risk of hair loss and nail disorders among users. Although some lawsuits allege vision loss from these drugs, these claims are disputed by the manufacturers. The effects of these drugs, especially in children, are still under investigation despite rising prescriptions.
The NVIDIA KGMON team placed second in the KDD Cup 2026 Data Agents competition by focusing on creating a smaller, clearer, and more verifiable agent harness. They unified data access by mapping heterogeneous sources into a single SQLite database accessible via unified `schema()` and `sql()` functions. Reliability was enhanced by preflight schema scouting, limiting the agent to a small toolset, and logging execution traces to allow for failure inspection. Key strategies included separating document inspection via targeted calls and preprocessing video evidence. Ultimately, the team emphasized constraining the action space, making attempts inspectable, and incorporating human review into the improvement loop.
A 16-year-old hiker named Bryce Vincent Gowryluk was rescued after an AI chatbot named Claude provided wildly incorrect directions while attempting to navigate the Widowmaker Arete in British Columbia. The erroneous directions led him to a dangerous situation at the headwall of the climb, where he realized he was stuck. Emergency services were called, and a helicopter rescue was organized by North Shore Rescue. Experts warned that relying on AI for wilderness navigation is dangerous, emphasizing the necessity of proper map and compass skills. The incident highlights a trend towards electronic reliance over traditional navigation training in mountainous areas.