Daily AI Digest
Top Stories
Mojo ends four years of closed development by opening its compiler and full toolchain
Modular has published the Mojo compiler, developer tooling and build system in a public repository. The code uses Apache 2.0 with LLVM exceptions.
The standard library and GPU kernels were already open. Developers can now build the entire toolchain from source with Bazel and contribute to the language and compiler themselves.
Verdict: AI infrastructure, compiler and high performance computing teams should evaluate Mojo seriously now that its roadmap, build and core implementation can be inspected independently.
ChatGPT introduces a teen experience that places users ages 13 to 17 behind stronger defaults
OpenAI will automatically place users in ChatGPT for Teens when its system estimates they are under 18 or when they report an age from 13 to 17. The experience combines Study Mode, quizzes and visualizations with reminders that redirect attempts to shortcut homework toward step by step learning.
Teen accounts receive default protections around self harm, violence, eating disorders and explicit content. Linked parents can set quiet hours and selected controls, and receive narrow alerts in high risk cases, but cannot read a teen's conversations.
Verdict: Parents and schools should test the defaults, then agree with teens on how to verify answers; safeguards can reduce risk but cannot replace conversation and judgment.
Tools and Products
OpenAI pauses its largest training run as cyber capability risk drives stronger isolation
OpenAI says its unreleased Astra model may meet the Critical cybersecurity capability threshold in its Preparedness Framework. The company paused reinforcement learning for two weeks, and its largest planned frontier run remains on hold while smaller training and evaluations continue.
New controls put model generated code in stronger sandboxes, isolate higher risk workloads from the internet and continuously watch for data theft, destructive behavior and attempts to defeat safeguards. OpenAI estimates that monitoring adds roughly 20% to the compute used by the inference it covers.
Verdict: Any team letting agents execute code or reach a network should follow the same order: isolate first, monitor next, and preserve an immediate stop path for serious alerts.
Claude Code extends its 50% weekly limit promotion through August 31
Anthropic has extended the Claude Code promotion through the end of August 31 Pacific time. Pro, Max, Team and legacy seat based Enterprise users continue to receive 50% more weekly usage automatically.
The promotion changes only the Claude Code weekly limit. Five hour limits, regular Claude, Cowork, Free plans and consumption based Enterprise seats are unchanged.
Firefox Smart Window adds web search while keeping its AI mode optional
Firefox's Smart Window assistant can now search the web and use open tabs or browsing history for context. The desktop beta is still rolling out first to English users in the United States and Canada.
Mozilla says it does not retain chat transcripts or use conversations for model training, and users can disable personalization from browsing activity and chats. Requests go to Mozilla before reaching the model provider a user selects.
Community Pulse
Turbovec fits ten million vectors into 4 GB instead of 31 GB
Turbovec is a local vector index written in Rust with Python bindings. The project says ten million 1,536 dimension vectors would occupy about 31 GB as float32 data and about 4 GB after its compression.
It implements Google Research's TurboQuant and can ingest vectors continuously without training a codebook first. Project benchmarks report 19% to 31% faster search than FAISS FastScan on ARM, but teams should retest recall and latency on their own data before deployment.
Builder Perspectives
Google Labs vice president Josh Woodward updated the Gemini product checklist: revamped Workspace tools should enter testing in one to two weeks, 3.7 Flash has improved tool calling, and a redesigned Projects experience is being implemented. Gemini now supports 49 connectors, while backend work for another item is finished and frontend work has begun.
View on X →Earlier Stories We Missed
Anthropic reportedly leads OpenAI in revenue run rate while Claude reaches far fewer users
Investor materials reviewed by Bloomberg put Anthropic's annualized revenue run rate above $65 billion by late July and preliminary second quarter revenue above $11.5 billion. Axios reported OpenAI's latest run rate above $40 billion, citing an internal message.
Reach tells a different story. Statista's Sensor Tower estimate puts Claude at about 245 million monthly active users in June, while OpenAI said in April that ChatGPT had 900 million weekly active users. Monthly and weekly activity are different measures, so using them to calculate share or revenue per user would be misleading.
The figures suggest that Anthropic's revenue may be more weighted toward enterprise and API use, but accounting differs too. The Wall Street Journal previously reported that Anthropic counts sales through cloud partners as revenue while OpenAI does not, and neither run rate is audited full year revenue.
GitHub Trending
PipesHub adds prompt injection protection for enterprise knowledge agents
PipesHub is an open source enterprise data connection and search layer. In window commits add a prompt injection guard for subagents, fix an administrator JWT role header injection issue, and improve query throughput and S3 storage for its web connector.
Docling 2.120.3 can parse Apple Pages files directly
Docling converts PDFs and office documents into structured content for AI systems. Version 2.120.3 was released inside the issue window with Apple Pages parsing, plus fixes for pictures inside Word content controls, ODF image paths and macOS OCR coordinates.
Today's common thread is capability placed inside inspectable boundaries. Mojo opens the compiler, OpenAI sets stronger defaults for teens and high risk models, and Mozilla keeps AI browsing optional. The Anthropic and OpenAI comparison is another reminder that user reach, revenue intensity and safety investment are separate dimensions, not one scoreboard.