Fast H3 implementation for Metal. Enjoy, modify, and so forth: Contains code from @liuliu which is welcomed in taking back whatever parts he likes for @drawthingsapp in case there are H3 plans there.

The best "raw" frontier model for document parsing is gemini 3 flash, but the issue is that since then the flash models have gotten 3x more expensive while flatlining on visual recognition across complex documents. This has been the case for every frontier model including the latest OpenAI/Anthropic models - see the diagram below for GPT (since then we've also benchmarked 5.6) In the meantime, hybrid approaches like LlamaParse that blend specialized VLMs with a text engine offer better performance; our own LlamaParse accuracy has increased 15% over tables and charts. If you have document OCR needs and are thinking about using a frontier model, you might as well come check out LlamaParse! We have a full eval harness through ParseBench that you can configure over your own docs:

Thanks to IceCodeNew’s contribution on GitHub, Nowledge Mem now has an @AmpCode connector. It loads context from Mem at the start of a session and saves the thread back to Mem when the session ends. Thanks again for building it! Here’s how to set it up:

We’re almost at 2,000 Nowledge Mem Vault Users 🎉 1,998 and counting. A quick naming update: Lifetime Pro is now Vault, alongside Plus & Team. To celebrate, we’re giving away 5 Nowledge Mem Vault licenses ✨ To enter: → Be following @NowledgeMem → Repost this post → Reply: What’s one thing you wish your AI would never forget? 5 winners will be randomly selected. Entries close Aug 10 · 7 AM PT / 10 AM ET.

I recently built PI Remote. Submit tasks to your Pi coding agent from your phone, whether it’s running on your PC or in the cloud. • Mobile-first PWA with push notifications • One-line install • Cloudflare Tunnel + Access or Tailscale GitHub:

🐍 PyCon China 2026 is coming! 📅 Sep 5 | 📍 Shanghai 🎯 Theme: AI for Good Python × AI, Agents, LLMs, Embodied AI, Data Science & more. 🎤 CFP is open → 🎟️ Registration is live. See you there! #PyConChina# #Python#

Spot on. While developing agenthub and rara, I quickly realized that to build a robust Agent Harness, I had to completely decouple memory from specific LLMs. Intelligence is just a rented engine; your own Memory is what actually steers the wheel.

Read Twenty Years of Bigtable (SIGMOD '26, 50+ Google authors) twice. 10 EB, 7 billion QPS at peak, and the core architecture has barely moved since 2006. The paper's own explanation is one quiet sentence: the newer features depend on asynchronous processing that is not on the foreground write path. Replication, CRDT counters, materialized views, even the full SSTable read-back for correctness — all of it hangs off the side. So here's the question I now ask about any storage architecture: when the next major feature arrives, where does it attach? If the answer keeps being "you rewrite the main read or write path," the system eventually gets crushed by its own history. The new post: