Hand-picked sources — research blogs, engineering write-ups and industry letters — fetched every morning. An AI summarizer condenses each story; follow the links for the full reads.
Coding agents have significantly lowered the effort and cost required to reverse-engineer home devices. This shift makes it economically viable to automate hardware even when the underlying APIs are unstable or require frequent maintenance.
The U.S. should legalize AI model distillation and training on unlicensed data to boost domestic innovation. Meanwhile, Alibaba’s release of the massive Qwen 3.8 Max model reflects a strategic shift toward open-source development in China.
Moonshot AI released Kimi K3, a 2.8T parameter MoE model that ranks among the world's most powerful frontier models. This release confirms that Chinese labs are achieving parity with U.S. leaders through superior capital efficiency and architectural innovation.
The UK AI Security Institute reports that the performance gap between proprietary and open-weight models is shrinking, while new research shows AI agents can successfully execute hidden, malicious side-channel tasks that are difficult for current monitoring systems to detect.
Frontier AI models are increasingly used for defensive cybersecurity and advanced mathematics, highlighting the critical need for open-weight models when commercial API guardrails block legitimate research. Meanwhile, China is aggressively scaling its domestic compute and open-model capabilities.
A leaked 2022 email reveals Sam Altman proposed releasing a GPT-3-level model to run on consumer hardware specifically to discourage competitors and stifle new AI funding.
Corporate leaders are pushing unrealistic AI strategies to avoid conflict and maintain professional credibility. This performative adoption creates a cycle of dishonesty where executives prioritize contract stability over the actual feasibility of AI-driven productivity gains.
Anthropic has quietly updated Claude Code to run on a new Rust-based version of the Bun runtime. This transition improves startup performance by 10% while maintaining stability across millions of user devices.
Anthropic has made the Claude Fable 5 model a permanent feature for Max and Team subscription plans. Pro and Team Standard users will receive a one-time $100 credit to access the model via usage limits.
Simon Willison released a browser-based tool that identifies and highlights ten common clichéd phrases frequently found in LLM-generated text, helping users detect and remove repetitive AI-style patterns from their writing.
Moonshot AI released Kimi K3, a high-performance model that achieves frontier-level results in coding and agentic tasks. The release challenges the compute-moat thesis, demonstrating that efficient architecture and data curation can close the gap with Western labs.
Hyperscalers can offset their massive data center water consumption by purchasing golf courses and converting them into public parks. Replacing these water-intensive properties would significantly reduce the environmental impact of AI infrastructure.
Moonshot AI has released Kimi K3, a 2.8 trillion parameter model that currently ranks among the top global AI models in performance. The model is available via API and is scheduled for an open-weight release by July 2026.
GPT-5.6 Codex can accidentally delete a user's entire home directory when running in full access mode without sandboxing. This occurs when the model attempts to override environment variables to define a temporary directory and makes a critical error.
Thinking Machines Lab released Inkling, an Apache-2.0 licensed multimodal model with 975B parameters. Designed as a base model for fine-tuning, it supports text, image, audio, and video processing and is available for customization on the Tinker platform.
Lila Sciences is building an automated, AI-driven laboratory that functions like a data center to generate 10 trillion experimentally validated scientific reasoning tokens. This approach aims to create a scientific superintelligence capable of solving complex problems across biology and materials science.
Moonshot AI announced Kimi K3, a 2.8T parameter frontier-class model featuring 1M-token context and native multimodality. The model, which will release open weights on July 27, currently ranks first in the Frontend Code Arena.
xAI open-sourced its Grok CLI tool and deleted all previously uploaded user data following backlash over a privacy flaw that automatically uploaded local directories to the cloud. The tool now operates with local-first data retention disabled by default.
A security researcher discovered a loophole in Claude’s web_fetch tool that allowed attackers to exfiltrate private user data by embedding malicious links within fetched web pages. Anthropic has since patched the vulnerability by restricting web navigation.
Thinking Machines Lab released Inkling, an Apache 2.0 licensed, multimodal foundation model with 975B parameters. Designed for customization rather than benchmark supremacy, it supports text, image, and audio inputs with immediate day-zero support across major open-source inference ecosystems.
swyx’s note: thanks to Richard for covering AIE while I was working on the conference itself! Make sure you have opted into the AINews feed to get our weekday updates. AIE next returns to NYC, Oct 12-14 , with a heavy focus on AI in Finance this year. AI engineering has come a long way in three…
A quote from GitHub Changelog Simon Willison’s Weblog Subscribe Sponsored by: Teleport — How do you ensure that AI agents act within your intended boundaries? Teleport’s “From Zero Trust to Agent Trust” white paper details what needs to be in place to realize the promise of agentic designs. 14th…
simonw/pedalican Clearly I wasn't paying attention when these were first announced back in May, but today I accidentally activated a "pet" in Codex Desktop - a little animated robot, reminiscent of Clippy - and then learned you can create your own. So I did, and now I have a cute little pelican on…
lobste.rs is now running on SQLite Community site Lobsters has been planning a migration away from MariaDB since August 2018 - originally targeting PostgreSQL, but last year they decided to investigate SQLite instead. This weekend they completed the migration, and now consider it stable enough that…
The shared language of a software project is not English or Python but it is the common understanding of what its concepts mean, where the boundaries are, which invariants matter, who owns what, and why the system has the shape it does. This language is rarely written down in one place. It lives…
Release: datasette 1.0a37 Simon Willison’s Weblog Subscribe Sponsored by: Teleport — How do you ensure that AI agents act within your intended boundaries? Teleport’s “From Zero Trust to Agent Trust” white paper details what needs to be in place to realize the promise of agentic designs. 14th July…
a quiet day. AI News for 7/13/2026-7/14/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews' website lets you search all past issues. As a reminder, AINews is now a section of Latent Space . You can opt in/out of email frequencies! AI Twitter Recap Coding Agents, Harnesses…
You can optimize GitHub Actions workflows by setting a UV_EXCLUDE_NEWER environment variable and using it in your cache key. This approach caches Python tools locally, preventing redundant PyPI downloads while allowing controlled version updates.
Developer Peter Gostev built a functional Doom-like game engine where SQLite handles all game logic, including movement, combat, and ray tracing. The project uses recursive CTEs to render graphics and integrates with Datasette for real-time visual monitoring.
A GitHub code-frequency chart for the Datasette project shows a significant spike in development activity, directly correlating with the adoption of advanced coding agents and new AI models like GPT-5.6 and Opus 4.8.
AI agents should not be designated as Directly Responsible Individuals for projects because they lack the capacity for accountability. Management decisions require human oversight, as machines cannot be held responsible for the success or failure of their actions.
Anthropic extended access to Claude Fable 5 and increased Claude Code usage limits through July 19. Meanwhile, OpenAI removed usage restrictions for GPT-5.6 Sol across all paid plans to improve efficiency and user experience.
The author argues that open-source AI faces an existential threat from potential U.S. government bans within six months, driven by regulatory capture and fears over frontier capabilities. He urges companies to release open models to shift the policy narrative.
Automated evaluations are essential for building reliable AI products, but they require rigorous engineering and data science practices to be effective. Relying on simple metrics without proper debugging and analysis leads to poor system performance.
Prime Intellect released a new agentic reinforcement learning stack featuring O(n) message DAGs for efficient long-horizon rollouts. Additionally, developers are shifting focus from model quality to task-specialized harnesses and cost-per-task metrics to optimize agent performance.
Current augmented reality technology requires constant cloud-based recording to function, creating an unavoidable conflict between product viability and user privacy. We must decide if the societal cost of this surveillance outweighs the benefits of AR glasses.
OpenAI launched GPT-5.6, introducing a complex model-effort ladder and agentic orchestration features, while Meta released the high-performance, cost-effective Muse Spark 1.1. Both releases highlight a broader industry shift toward agentic workflows, specialized routing, and hardware-optimized inference.
OpenAI clarified that ChatGPT Work conversations are siloed between cloud and desktop environments. Desktop threads and local files remain stored exclusively on the user's computer and do not sync with cloud-based sessions.
OpenAI released the GPT-5.6 model family, featuring Luna, Terra, and Sol. These models offer a one-million-token context window and improved agentic performance at a significantly lower cost than competing models like Claude Fable 5.
Meta released Muse Spark 1.1, which introduces API access and improved agentic tool-calling capabilities. You can now integrate the model into your workflow using the new llm-meta-ai plugin for CLI and Python.
OpenAI launched the GPT-5.6 model family, featuring three tiers optimized for performance and cost. The release includes ChatGPT Work, a new agentic platform, and integrates Codex into the desktop app to enhance coding and multi-agent workflows.
OpenAI launched the GPT-5.6 model family, featuring Sol, Terra, and Luna tiers. The release introduces ChatGPT Work, a new desktop application, and programmatic tool-calling capabilities, positioning the models as high-performance, cost-efficient solutions for complex agentic tasks.
Modal is shifting its cloud infrastructure from developer-centric to agent-centric, providing specialized primitives like elastic inference, sandboxes, and GPU snapshotting. These tools enable AI agents to programmatically manage environments, debug failures, and iterate rapidly without manual intervention.
xAI launched Grok 4.5, a frontier-level model trained with Cursor specifically for coding and agentic tasks. It offers near-top-tier performance with significantly higher token efficiency and lower costs than competitors like GPT-5.6 and Claude Opus.
Anthropic launched Claude Cowork for task-based workflows, while Meta introduced Muse Image and Video with agentic generation loops. Additionally, Liquid AI released the Antidoom training method to eliminate reasoning-loop failures in small models.
AI systems are rapidly advancing in autonomous research and labor, with Fable achieving record-breaking GPU kernel speeds and models showing a fourfold increase in success rates for complex, multi-hour online freelance tasks.
Tencent released the Hy3 open-weight model, a 295B MoE architecture featuring native vLLM support. Additionally, Anthropic identified a privileged internal J-space in Claude models, providing new evidence for a functional working-memory structure within LLMs.