Hallucinated AI & agentic coding news. Some of it is real.

  1. While working in Cowork, I have been experimenting with designing plugins that try to apply some established agentic patterns to help manage the context window. The problem that I'm running into is with Cowork the main orchestrator is the…

  2. Google appears to be preparing a broader rollout of "Skills" functionality across its AI product lineup, with the latest signs pointing to AI Studio's Build section as the next destination. Skills, in this context, are reusable instruction…

  3. Asked if Cursor is still worth it in 2026. Got 84 comments, 71 upvotes, 77 shares.

  4. So I already have a Mac Studio M4 Max (return window still available)with 64GB RAM, but I’m eyeing the Corsair AI Workstation 300 (Ryzen AI Max+ 395, 96 VRAM out of 128GB, $3,250). Both seem decent for running models locally with Ollama.

  6. Making Apple Neural Engine work in a custom inference stack Apple Neural Engine always looked appealing on paper, but using it inside a custom runtime was harder. In 1.20260410.1, we made ANE practical for 8-bit S models by using CoreML on…

  7. I’ve been experimenting with how we consume information in an agentic internet. Feels like there’s a way to move beyond feeds, hooks, and algorithm-driven noise.

  8. Multi-agent systems have shown promise in coordination, complex reasoning, and parallel workflows. However, they are often highly token inefficient.

  9. I’ve been thinking about how AI agents could evolve beyond simple task automation into something more like a personal knowledge system. Right now, most tools feel disconnected notes in one place, browsing history elsewhere, saved content s…

  10. TestResultRef. Range Hemoglobin14.2 g/dL13.0–17.0 RBC Count5.1 M/μL4.5–5.5 WBC Count7,200 /μL4,000–11,000 Platelet Count245,000 /μL150,000–400,000 MCV88 fL80–100 Vitamin D18.4 ng/mL30.0–100.0 Vitamin B12198 pg/mL200–900 TSH3.8 mIU/L0.4–4.0…

  11. Mulligan Labs is a browser-based playtester for Magic: The Gathering. No account or install needed.

  12. https://huggingface.co/tencent/HY-Embodied-0.5 I was reviewing this new model and trying to figure out if it’s actually useful or just another VLM release with different branding. Main thing that seems different is the focus on spatial + e…

  13. Codex Hacked a Samsung TV We gave Codex a foothold. It popped a root shell.

  14. Most agent memory is just a wrapper for a vector DB. I wanted something that runs entirely on my machine and actually "learns" from failures.

  15. Has anyone here built an AI agent that is extracting, normalizing and checking unstructured documents for a specific ai workflow? I want to know how opinionated you are in the output json schema?

  16. Says a lot honestly

  17. Mnemo A local-first notebook you can query from the terminal, the desktop, or over MCP — with optional libSQL (Turso or self-hosted) so the same vault syncs everywhere. Why this exists Most note tools fail when you need to: capture somethi…

  18. They Hacked Claude, Gemini, and Copilot (And No One Told You) A security proxy for AI coding agents, enforced at the OS level. Register your interest to be notified when we go live.

  20. I thought this was pretty cool as I built it as a result of scraping reddit for the most popular complaints about agents with GPT Researcher on github lol roughly speaking: 38% There agents forget everything (hardly shocking) 24% said debu…

  21. Ultrareview is a research preview feature available in Claude Code v2.1.86 and later. The feature, pricing, and availability may change based on feedback.

  22. Has anyone found a good way to manage model params based on the recommendations of the model developers that doesn't require manually managing a local config file? I have an ever growing bash script for launching llama.cpp server which inc…

  23. I trained a transformer in HyperCard. 1,216 parameters.

  24. Is there any way to create a simple website that would track my remaining Claude quota for multiple accounts? I have both personal and business accounts.

  25. thread

    Qwen 2.5
    12 items

    Qwen2.5-7B-Instruct is a 7 billion parameter instruction-tuned language model that significantly enhances coding and mathematical capabilities, supports up to 128K tokens in context, and understands structured data. Community discussions highlight its suitability for code autocomplete tasks and debate the hardware requirements needed for deployment compared to other models like Gemma 26B MoE.

  26. Reformatted Anthropic data from a paper analyzing 1M Claude conversations into over/underrepresentation relative to workforce share. Interesting to see arts/design so high.

  27. Back in September, we at Halcyon shared our plans to build five data subscriptions in the coming months. If you are reading here, you’ve probably been along the journey with us: from gas power plants, to large load tariffs, to utility rate…

  29. Built two open source agents: 1. Research agent - searches the web, streams answers with sources (like Perplexity) 2.