☕️ OpenAI pools ChatGPT and Codex limits

• OpenAI will merge usage allowances across ChatGPT and Codex starting in December, drawing conversations and agentic coding tasks from a shared token pool for Plus, Pro, and Business subscribers, as part of unifying its conversational and agentic products.
• Heavier models and more complex tasks will consume a larger share of the combined pool, while ChatGPT's faster Instant chat mode stays exempt; OpenAI hasn't disclosed the size of the merged allowance or how it compares to today's separate limits.
• Developers on a forum warn that intensive Codex sessions will now eat into ChatGPT capacity and vice versa, demanding a real-time usage dashboard before launch; OpenAI will show combined remaining allowance and reset times in settings but declined to specify detail.

• Google is shutting down Jules, its asynchronous AI coding agent launched in 2025, folding the effort into its consolidated agentic project, Antigravity, which will inherit many of the tools Jules offered.
• Jules sessions stay active until November 11 2027, after which existing projects go read-only; a new export dumps everything to a local folder, account settings, environment variables, connected repositories, scheduled tasks, memories, and every session's prompt, plan, log, and code patches.
• Google framed Jules as an "experiment" whose learnings will feed Antigravity, and never disclosed usage beyond "thousands" of beta testers, having faced competition from OpenAI's Codex, Claude Code, and GitHub Copilot.

• Anthropic shipped Message Threads as a public beta on Claude, cutting time to first token by about 35% for large requests by resending only what changed in a conversation rather than the full history.
• For requests between 500K and 1M tokens, the thread stays encrypted in memory between calls and expires after five minutes of inactivity; token usage is unchanged since Claude still reads the entire conversation.
• Coding agents, computer-use agents, and document- or image-heavy workflows benefit most, and the feature is available on the Claude platform plus AWS, Microsoft Foundry, and Google Cloud.

• OpenAI built Pando, a new Git backend for its monorepo that absorbs the surge of pushes and fetches from engineers and Codex agents, now serving nearly a petabyte of Git data daily without changing how anyone uses Git.
• Pando coordinates updates per-branch using causal consistency rather than a global sequence, storing packfiles in cloud object storage with regional caches exceeding 99% hit rates, cutting median push latency by ~80% and lifting merge reliability to 99.99%.
• Written in Rust and in production since April 2026, Pando is designed to scale Git capacity alongside agent usage, though the tradeoff is that readers examining multiple branches may see different versions while each individual branch stays consistent.

• Cloudflare stopped triggering full DNS zone rebuilds for zone-setting changes, letting the Zone Builder regenerate only the records and metadata a change can prove it touched rather than reprocessing entire zones.
• A targeted build planner picks the smallest provably safe build from five types and a unified scheduler merges work for the same zone, cutting peak memory by 85% across zones that can hold more than 20 million records.
• The freed headroom let Cloudflare shorten resilience waits, halving the time from a DNS change to its data landing in Quicksilver, though zones above a record threshold never escalate to a full build and rare cases need a manual per-name rebuild.

• OpenTelemetry's eBPF Instrumentation tool, OBI, reached v1.0, bringing zero-code observability to running Linux applications by capturing metrics and traces across services, languages, and frameworks without touching their source.
• The release locks down compatibility commitments via Config v2 and a Weaver-compatible schema registry, while expanding coverage to MQTT, NATS, MSSQL, and others, plus HTTP/2 and gRPC propagation fixes and route discovery for Django, FastAPI, Flask, Rails, and .NET.
• Existing users must move to Config v2, since v1 is frozen with no fixes or features; follow the migration guide and compare telemetry in a canary deployment before rolling out, as some conventions remain unstable and may change.

📚 Tutorials & Advice

🔗 Other News & Articles

• AI model gets a 32-bit RISC-V CPU compiled directly into its weights, then runs DOOM on it — experimental AI model can run Doom and C code using its own weights as RAM LINK

• CentOS co-founder creates a Linux-compatible kernel in just 11 days, with AI doing almost all the coding LINK

• ivankovic/omnidiff: Fast, robust, accurate diffing LINK

• raphlinus/nibnib: Image compression at 4bpp with extreme decompression speed LINK

• Developer manages to run Windows 11 26H2 with just 128MB of RAM, just 3.1% of recommended minimum — stripped-down installation can access Desktop, File Explorer, and Notepad LINK

• Eneskirca/nodeterm: Node-based terminal manager for AI coding agents - tmux-backed terminals and parallel agent sessions as draggable nodes on an infinite pan/zoom canvas. macOS, Linux, and a browser Server Edition. LINK

📄 Papers & Resources

🧰 Latest Tools

💬 How did you find today's edition?

We read every reply — just reply to this email and let us know how we can improve!