Frontier labs, such as OpenAI and Anthropic, are facing challenges from lower-cost, highly-performant open-weight AI models, which could threaten their ability to charge sufficient prices for training, staff, and inference costs. These models, like Z.ai's GLM-5.2 model and Moonshot's Kimi K3 model, have improved performance at a smaller training cost, sparking concerns about the sustainability of frontier lab pricing methods. American frontier labs must contend with not only open-weight models but also competition from local inference models and first-party hyperscaler models with established distribution channels.
Thinking Machines has released Inkling-Small, a 276B-parameter mixture-of-experts model with 12B active parameters. The model retains Inkling's multimodal reasoning, variable thinking effort, and 1M-token context window while using substantially less compute. OpenAI reduced GPT-5.6 Luna pricing by 80% and Terra pricing by 20%, improving Sol's API speed and extending efficiency gains across API usage, Codex, and ChatGPT Work subscriptions.
The Federal Communications Commission (FCC) recently announced a ban on foreign-made robots and power inverters, which also extends to robot vacuum cleaners due to their software-controlled autonomous nature and wireless connectivity. The ban definition encompasses devices that can perceive its environment, weigh more than 4.4lbs, travel across the ground, and offer wireless connectivity. Existing US-sold robot vacuums are exempt from the ban but new models require a waiver application.
TLDR AI is a free daily newsletter covering AI news, research, and tools in a 5-minute read, reaching 1,100,000 subscribers weekly. It's geared towards engineers, researchers, and technical professionals who want to track AI without hype. TLDR AI differs from The Rundown AI, which has a more general focus on practical coverage. Both are free dailies targeting different audiences.
OpenAI has strengthened its approach to responsible AI in Europe by developing governance frameworks for safety, security, transparency, accountability, and provenance. The company contributed to the EU's General-Purpose AI Code of Practice and has implemented a Preparedness Framework to identify, evaluate, and manage risks from advanced AI systems. OpenAI is also expanding its approach to transparency and trust through its work on provenance and collaboration with third-party experts.
AI bots and chatbots now generate more web traffic than humans, forcing publishers to rethink how content gets built and monetized. Automated traffic grew almost eight times faster than human activity in 2025, with a BBC study finding 51% of AI-generated news summaries contained significant inaccuracies. This shift threatens the fundamental economics of running a website, with media executives expecting search traffic to drop by 43% over the next three years.
Researchers at OpenAI have made significant breakthroughs in mathematics and theoretical computer science, addressing long-standing open problems in geometry, cryptography, and complexity. These advances demonstrate progress in these fields, shedding new light on fundamental questions in mathematics and computer science.
A full-stack approach aims to make advanced artificial intelligence (AI) more capable by integrating multiple layers of technology. This approach involves combining hardware, software, and data processing capabilities to create a comprehensive system. By doing so, it can increase affordability and accessibility of AI solutions. The goal is to make AI more widely useful across various industries and applications.
Raph, a creative technologist at Google, built an app called Glanceboard with Gemini 3.6 Flash and Nano Banana to simplify morning routines. The app checks family's Google Calendar and local weather, generating a personalized illustration of children wearing forecast-friendly clothes. This image is displayed on an e-ink screen, providing a serene and organized way to stay on schedule.
Gemini is introducing a new feature that allows users to speak directly into any active window, creating, editing, and summarizing text with Dictate. Gemini Spark is now globally available 24/7 across all devices. The latest Flash models, Gemini 3.6 and Gemini 3.5 Flash-Lite, offer improved reasoning and speed. Users can add their avatar to images and link favorite apps to Gemini for seamless prompts.
Co-designing AI model attention for fast, interactive long-context inference is governed by group size (query heads per KV head), head dimension, and sequence length, each affecting prefill (compute-bound) and decode (memory-bound) phases differently. Choose G for decode efficiency and push it high to improve decode speed and GPU utilization. Use a head dimension of 128 or 256 to align with the GPU tiles and memory alignment. Sequence length affects prefill and decode differently; prefill scales quadratically, while decode scales linearly. Reduce effective KV state through KV-cache compression, sparse or sliding-window attention, or hybrid models. Match parallelism to KV head count (KH), and keep TP ≤ KH to avoid KV duplication. Get started co-designing AI model attention by using the four guidelines as a model design checklist: choose G for decode efficiency, use Hsz = 128 or 256, reduce effective KV state, and match parallelism to KH.
OpenAI has reduced prices of its models amid concerns from enterprises about AI spend, indicating a price war for AI models has begun.
The European Commission has pledged $11.5 billion for seven artificial intelligence (AI) gigafactories as part of an effort to compete with U.S. and Chinese companies. The funding will be used to support the development of AI technology in Europe. A tender has been launched to determine which companies will build these facilities, following a bid by European businesses to close the gap with tech giants from the United States and China.
DataFlow-Harness achieves a 93.3% observed end-to-end pass rate on a 12-task data-engineering benchmark, reducing API costs by up to 72.5% and response latency by 49.9%, compared to standard Claude Code. The framework guides an LLM agent to build structured, visual data-processing workflows step-by-step, making AI-generated pipelines easier to manage and integrate into existing architectures.
Supabase has open sourced Supabase Evals, a benchmark that scores AI agents like Claude Code, Codex, and OpenCode on real Supabase tasks. The framework runs coding agents against real tasks, scoring the results and providing an internal regression suite monitored daily. Agents pass most scenarios with no skills loaded, but three weaknesses surfaced: manual migration generation, authentication verification, and documentation usage variability.
MiniMax releases MiniMax H3, a general-purpose multimodal generation model that reads text, images, video, and audio as one unified context and returns video with native stereo sound. The model produces 2K output clips of 4–15 seconds with integer durations only. MiniMax H3 is deployable through the API, with the live model available on the platform's API and in the consumer Hailuo AI app.
Configure the input source, model behavior, inference mode, point-cloud generation, and export settings. Probe the available GPU and automatically adjust frame limits, camera iterations, scale frames, and KV-cache parameters according to the detected VRAM. Install the repository and its dependencies, download the pretrained checkpoint, preprocess image or video frames, and construct the GCTStream model with streaming attention and long-range trajectory memory.
A New York school district has paused plans to bring an android teacher's aide named Sally into the classroom due to public outcry. The $60,000 program aimed to have Sally assist actual instructors in high school classrooms, but concerns arose about teenagers' skepticism and lack of respect for a robot teacher. The superintendent believed Sally would provide opportunities for students to learn about robotics and technology maintenance, but critics argue that it's unlikely teens would take the android seriously or engage with it meaningfully.
The FTSE 100 is on track to finish the month up by around 3.5% at 10,840 points, having reached a new record earlier in the day. The index had previously hit a year low of 9,700 points due to the Iran attacks. The broader European Stoxx 600 index is up 1.2% for the month, following a 8% drop in March.
The Model Context Protocol (MCP) has been updated to a stateless version, which simplifies client and server implementations. The new specification reduces the complexity of implementing clients and servers, making it easier for smaller models to drive MCP tools. This change has inspired the creation of mcp-explorer, a stateless Python CLI tool that allows users to interactively probe an MCP server. Another project, datasette-mcp, is a Datasette plugin that adds a /-/mcp endpoint to any Datasette instance, allowing for SQL query execution against hosted Datasette instances using LLM tools like ChatGPT and Claude.
llm-mcp-client 0.1a0 is a release that enables access to tools from MCP servers as LLM (Large Language Model) tools, according to Simon Willison's Weblog, published on July 31, 2026.
Simon Willison was invited to join Bryan Cantrill and Adam Leventhal's podcast to discuss recent events, including Kimi K3's open weight models outperforming proprietary ones. The conversation also covered cybersecurity incidents, predictions made earlier in the year, and notable exceptions to those predictions. A new prediction added by Simon Willison is that the Pope will make a statement about open models by the end of the year.
Smevals is a small evaluation suite for assessing models, prompts, and harnesses. It allows users to create eval suites by defining directories with YAML files, which can then be run against different model configurations using the "run" command. The tool also provides grading operations, enabling users to evaluate results against defined checks, and offers options to serve local web servers or generate static HTML reports.
datasette-agent 0.4a0 has been released, featuring an LLM-powered agent for Datasette that enables the execution of custom JavaScript in the user's browser through a new mechanism called await context.browser_task(). This capability makes it easy for Datasette Agent plugins to provide tools that run code directly in the browser.
A seemingly harmless move to a multi-agent architecture quietly tripled LLM costs, due to the introduction of orchestration, which costs tokens. The issue was not fan-out or retry logic, but rather the supervisor agent's decision-making process, which required rebuilding upstream work. To fix it, three changes were made: 1) routing by task complexity, where smaller models are used for simpler steps; 2) context trimming between handoffs, reducing the amount of unnecessary input tokens; and 3) parallel execution, which reduced retry cascading. Token usage per model dropped by 40% and end-to-end latency dropped by 45-55%.
In 5-10 years, the next sharp manager in a company might not be human, will sleep less, and exist entirely in shared GPU memory. Most middle management is currently running an internal routing protocol, making sure teams communicate and turning strategy into actionable tasks. However, with the advancement of AI, this layer can collapse into mechanism design. Companies are already testing autonomous decision loops that can beat humans on latency, but they need to build a bare-metal execution fabric for these agents to think on. This shift will not come from better models or prompts, but from improving communication protocols and governance layers.
To debug AI coding agents when they change the wrong thing, record model tool requests, real function results, patches, checks, screenshots, and a saved run log. The agent should prove what it inspected, changed, ran, and verified to fix the issue. The main questions an agent must answer are: What files did it inspect? What selector or component did it identify as the target? What exact patch did it apply? Did the app still build? Did the right check prove the reported problem changed? To adapt this to your own agent, change tools to match the bug type (e.g., API, test failure), and modify the proof to match the original request. Keep the run log as a useful record of model tool requests, function arguments, function results, patch, command output, and final check. Tools such as Playwright's Chromium browser, MCP server, Weave, LangSmith, Phoenix, Langfuse, Dynatrace AI coding agent monitoring, Arize coding agent tracing, and OpenAI tools can be used to wrap the model-driven coding task so a developer can see the exact tool requests, file changes, checks, screenshots, and failures that led to the final fix.
Master problem: min 6x1 + 4x2 + θ, subject to θ ≥ 0, θ ≥ 16 - 10x1 - 10x2, x1, x2 ∈ {0,1}. Subproblem: min y1 + y2, subject to y1 ≥ 10 - 8x1 - 4x2, y2 ≥ 6 - 2x1 - 6x2, y1, y2 ≥ 0. Dual subproblem: max (10 - 8x1 - 4x2)u1 + (6 - 2x1 - 6x2)u2, subject to u1 ≤ 1, u2 ≤ 1, u1 ≥ 0, u2 ≥ 0. First Benders cut: θ ≥ 16 - 10x1 - 10x2. New master objective value after adding the first Benders cut: LB = 10.
Researchers and students of Multi-Agent Reinforcement Learning (MARL) are seeking resources that focus on coding rather than theoretical explanations, with a preference for Python/PyTorch implementations. Many struggle to understand the practical application of MARL concepts due to a lack of code examples and tutorials. The user is looking for video tutorials, GitHub repos, or guides that provide step-by-step instructions on building MARL projects.
AlphaFlowSeven's reinforcement learning process evaluates changes made to its agent configuration by using an optimizer with four council slots, where one slot runs the current best config as a control and the other three run variants. Variants are generated by an LLM that reads previous configs and scores, restricted to changing one or two things at a time. A variant must beat the control in at least two windows before replacing it, with ties going to the cheaper config.
The industry is mistakenly focusing on increasing context windows and memory for agents, when in reality the brain processes a tiny fraction of sensory information. The forgetting function is the key to intelligence, allowing an agent to discard irrelevant information and operate at the right level of reality. A mathematical rate distortion curve would be more effective than simply storing everything, with a specific compression ratio determining what history is sufficient for a task. When agents learn to modify their own forgetting functions, they can achieve self-sustaining growth without needing orchestration or supervisors.
Kris DeVault's son Brody, diagnosed with creatine transporter deficiency, a rare condition that affects brain and muscle development, is seeking experimental treatment through Montana's new "right to try" law, which allows access to unapproved drugs that have completed phase I clinical trials. The treatment, developed by French biotech company Ceres Brain Therapeutics, has shown promising results in animals but not yet tested in people with the condition or children. DeVault is considering accessing treatment through a private clinic in Honduras due to concerns about FDA regulations and the drug's lack of approval. Experts caution that phase I trials do not prove safety or efficacy, while DeVault argues he has the right to make decisions about his son's life.
NVIDIA Video Codec SDK 13.1 introduces zero-copy transcoding with shared CUarray buffers, reducing SM and memory bandwidth utilization, higher throughput, greater customizability, and simplified deployment using an official Docker-based development environment. The new SDK supports Hierarchical Reference Mode for AV1 with up to 31 B-frames, combined UHQ tuning info with iterative encoding, and significant bitrate savings in both CQ and VBR modes. Decode improvements include per-macroblock statistics for H.264/HEVC, frame-accurate seek via a GOP-aware architecture, and expanded MV-HEVC stereoscopic support.
A researcher built an OmniRay project, a CPU-only Active SLAM framework with AVX2 SIMD acceleration and Deep RL, on a commodity hardware setup. The project combines various components, including raycasting, particle filtering, and autonomous exploration, and is designed to be modular and reproducible. The goal was to experiment with Active SLAM and reinforcement learning while optimizing systems. Feedback from the robotics community is sought on aspects of architecture, benchmarking, documentation, and future improvements.