Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
OpenAI's use of Weekly Active Users (WAU) rather than Monthly Active Users (MAU) renders its user base scale incomparable to other large consumer technology products. ChatGPT's user retention is likely quite low, meaning many users cycle in and out every month, inflating MAU relative to WAU. Dividing MAU would reveal relatively weaker unit economics that are directly comparable to other consumer technology products.
AI is making formal verification go mainstream. Random testing will play an important role in the future of software engineering. As autoformalization tools get better, we will have many more formal specifications. Random testing benefits from these formal specifications in different ways than formal verification, but both have their places.
Gemini 3 Flash was Arnaud Autef's first release as distillation TL.
Mother Nature can host human intelligence using only 20 watts of power in a space equivalent to the space inside your skull.
Poetiq saw results as high as 75% at under $8/problem using GPT-5.2 X-High on the full PUBLIC-EVAL dataset.
Sponsored
Welcome to Delve Shipmas. Every day this week, Delve is launching a new AI feature to make compliance more modern than ever. Yesterday, we launched Delve's AI Copilot that can help on any compliance task or question. Today, we're launching... AI VENDOR RISK MANAGEMENT
OpenAI introduced a personalized year-in-review feature called "Your Year with ChatGPT," available to eligible users in select regions. Inspired by Spotify Wrapped, the feature highlights individual usage trends from the past year.
GLM-4.7 is the latest release in Z.AI's General Language Model line. The high-end foundation model is aimed at advanced reasoning, coding, and multimodal workloads. The late update expands context handling and reasoning depth compared to earlier versions. It introduced upgraded reasoning pipelines and broader multimodal support.
MiniMax M2.1 is ahead of DeepSeek and Kimi on several benchmarks. It is even catching up to state-of-the-art models in some areas. The model is super-fast and efficient. It is now available to all Kilo Code users.
Manus Design View is an extension of the Manus agent for seamless AI design workflows. It enables designers to generate concepts with simple prompts, make precise edits using the Mark Tool, and change text easily. The Manus mobile app allows design edits on-the-go with either text or voice input. It is available to all users today.
This post details OpenAI's ongoing efforts to secure its AI browser, Atlas, against prompt injection attacks - malicious instructions embedded in web content that manipulate agent behavior. While mitigation techniques are improving, the company acknowledged prompt injection remains an unsolved and persistent threat, particularly as agent capabilities expand on the open web.
It's pretty easy to homebrew your own asynchronous coding agent. This means that businesses selling coding agents can no longer differentiate themselves by only running sandboxed agents in the cloud that connect to Slack. Companies working on coding agents likely realize this and are doing everything they can to train their own SWE agents and auxiliary models to improve their harnesses.
Gemini 3 Flash is a lightweight, efficient model optimized for speed and low latency. It is capable of delivering performance comparable to Gemini 3 Pro at a fraction of the cost. The model's design brings unprecedented power but introduces specific tradeoffs in token efficiency and reliability. This post takes a look at the leaked architectural details of the new model.
Vercel spent months building a sophisticated internal text-to-SQL agent with specialized tools, heavy prompt engineering, and careful context management. It kind of worked, but it was fragile, slow, and required constant maintenance. The team then deleted most of it and stripped the agent down to a single tool that executed arbitrary bash commands. Its agent got simpler and better at the same time: it had a 100% success rate instead of 80%.
Sponsored
Interested to learn how different LLMs perform for coding? New research [15] on models like GPT-5.2 High and Gemini 3.0 Pro reveals trade-offs in structural quality and security.
This repository contains a comprehensive collection of Agent Skills for building production-grade AI agent systems. They are categorized into Foundational skills, Architectural skills, and Operational skills. Each skill is structured for efficient context use. The patterns work on any agent platform that supports skills or allows custom instructions.
OpenTinker is an RL-as-a-Service infrastructure for foundation models. It features separation of programming and execution, separation of environment and training code, and seamless transition from training to inference. The platform enables users to perform RL training and inference without requiring local GPU resources by separating client-side programming from server-side execution. It provides a high-level Python API that abstracts away the complexity of distributed systems.
SGI-Bench is a benchmark for assessing Scientific General Intelligence across the entire research cycle, such as Deliberation, Conception, Action, and Perception. It spans 10 disciplines with over 1,000 expert-curated tasks inspired by major open scientific questions.
Cursor has announced partnerships to integrate its agent hook system with security and platform vendors. These hooks allow organizations to observe, modify, or block stages of the agent loop, supporting use cases like governance, dependency scanning, secrets management, and agent safety.
AI's utility in the coming decade will come from understanding the technology's strengths and where it can be used to augment human ability. It won't replace humans, at least in the short term, because we are too complex. However, the technology will eventually conquer territories we thought were exclusively ours. This will be the first time we will face true otherness, a new species of being.
Sponsored
The only open-source TTS with built-in watermarking. Up to 6× faster than real-time on a GPU. Clone voices with 5 seconds of audio. Try
Alphabet agreed to acquire Intersect, a provider of energy and data center infrastructure, for $4.75 billion.
Google announced advanced governance features for Vertex AI Agent Builder, enhancing its Agent Engine to manage both short-term and long-term memory.
2026 may also be the first year the AI trade meaningfully splits into AI infra vs AI applications.
Sponsored
Many financial systems of record (like ERP) were built at a time when AI tools were the stuff of science fiction. Now, they need to work together.
OpenAI has introduced new personalization options in ChatGPT, letting users adjust enthusiasm, warmth, and emoji use directly. These controls, available in the Personalization menu, offer "More," "Less," or "Default" settings, expanding tone customization beyond the existing base style and tone feature.
Cursor has acquired Graphite, a company known for its performance-focused internal developer portal. This marks Cursor's third acquisition as it aims to build a comprehensive AI-powered dev platform.
Anthropic's Bloom is an open-source tool for generating automated behavioral evaluations of AI models. Bloom assesses specific behaviors like self-preferential bias and sabotage by creating scenarios and quantifying behavior occurrence across models. It efficiently differentiates between aligned and misaligned models and correlates strongly with human judgment, enabling scalable and reliable behavior evaluations.
LLM use is rising. People are increasingly using different LLMs, different products, and in different places. ChatGPT remains dominant and keeps acquiring new users, but Gemini's growth has been faster over the last few months. OpenAI's revenue seems to be on track, but consumer revenue is likely decreasing as a share. A substantial share of workplace AI use involves workers adopting tools on their own rather than waiting for employer-provided access.
What happens when agents run out of memory determines whether they continue productively or have to start from scratch. This post explores an evaluation framework that measures how much context different compression strategies preserve. Structured summarization retains more useful information than alternative methods without sacrificing compression efficiency.
Benchmarks are the most widely misunderstood part of the AI ecosystem. The narrative keeps implying a universal increase in intelligence, but the numbers can be misleading. To navigate this noise, look at the aggregate, look at the relative, and verify with your own tasks. The only benchmark that matters at the end of the day is your own workload.
This document contains a diary for an experiment aimed at teaching an LLM using GRPO to generate regex given a description. It details the performance, learnings, modifications, and key takeaways from each experiment. The initial training run was on December 17. It saw the model quickly learning how to generate valid regex tags, but the model was basically generating random regex strings.
Andrej Karpathy has outlined paradigm shifts of LLMs in 2025, including fast inference engines, model distillation trends, real-time agents, neural GPUs, and the rise of high-quality open models like DeepSeek-V2 and RWKV.
Sponsored
One in three employees uses AI tools without approval, risking data leaks and compliance violations. With this Enterprise AI Governance Kit [13] from You.com [13], you will get ready-to-use templates to help protect your organization, including security checklists, usage policies, and governance frameworks.
Qwen-Image-Layered is a model capable of decomposing an image into multiple RGBA layers. Each layer can be independently manipulated without affecting other content. They can be resized, repositioned, and recolored. The approach enables high-fidelity and consistent editing.
MiMo-V2-Flash is a powerful, efficient, and ultra-fast foundational language model that excels in reasoning, coding, and agentic scenarios. It serves as an excellent general-purpose assistant for everyday tasks. The model is available globally on Hugging Face, AI Studio, and Xiaomi's API platform. Benchmark results are available in the article.
jax-js is a machine learning framework for the browser. It brings JAX-style, high-performance CPU and GPU kernels to JavaScript, so users can run numerical applications on the web. The library is written from scratch and has no external dependencies. It can run anywhere a browser can run.
MCP servers give agents the tools they need to accomplish tasks. This post discusses how to multiplex MCP servers to simplify the connection to various tools within them. Multiplexing allows multiple MCP servers to be used over a gateway in a single interaction. It allows agents to access multiple MCP servers with different stacks, clouds, applications, and frameworks for specialized tasks.
tcgen05 is the set of PTX instructions that programs Tensor Cores on the latest NVIDIA Blackwell GPUs. This post contains a tutorial for Blackwell in plain CUDA C++ with PTX. It documents the author's process of learning tcgen05 and reaching 98% of CuBLAS speed. Readers can follow the tutorial using Modal or any other B200 cloud providers.
METR topics are public, making it easy to game METR horizon length measurements for a frontier lab. The horizon length under METR's assumptions might be adding little information beyond benchmark accuracy. There is a meme going around based on a team achieving a one to four hour range on the METR plot. This post explains why the plot has been interpreted incorrectly.
AI is incredibly good at some tasks while being really bad at others. This is the 'Jagged Frontier' of AI ability. The jaggedness is likely going to remain a big part of AIs going forward. However, the growing frontier will outpace jaggedness.
Masayoshi Son has already sold SoftBank's entire $5.8 billion stake in Nvidia, offloaded $4.8 billion of its T-Mobile US stake, and slashed staff to come up with the money.
A lot of research progress on agentic reinforcement learning made its way into Gemini 3 Flash, but it was too late for Pro.
World simulators are models trained to predict how the world evolves over time, frame-by-frame, using large amounts of video and interaction data.
Chinese GPU maker Moore Threads' Huagang architecture promises significant performance gains in gaming and AI.
In 2026, US companies will apply a factory mindset to energy, mining, construction, and manufacturing by pairing AI and autonomy with skilled labor to standardize complex work.