AI Daily Digest · 2026-09-10
🔬 New AI Craft
1. Procedural Graphs: execution structures that self-evolve for LLM agents
New arXiv paper organizes procedural knowledge into (procedure, relation, procedure) triplets; at each step a guidance model translates the surrounding subgraph into situational hints that bias—but don't dictate—the solver's next action. A refiner contrasts failed vs. successful trajectories and edits the graph topology, so the execution structure grows at runtime instead of staying fixed like a pre-planned pipeline.
https://arxiv.org/abs/2609.09153
2. Distillation fingerprint: Qwen3.8 tracks GPT-5.5 Pro reasoning (+18.18pp)
wsxiaoys updates the reasoning-prefill experiment (v1.1): injecting the first 1% of GPT-5.5 Pro's reasoning into Qwen3.8 A95B lifts answer overlap from 16.79% to 34.97% (+27pp on STEM), while the same probe barely moved toward Opus 4.8—suggesting Qwen3.8 was trained on GPT-5.5-Pro-style outputs. A lightweight fingerprint for detecting distillation provenance.
https://gist.github.com/wsxiaoys/e0286dc6bb624ff5fdf49e7f4c528ba3
3. Raschka unpacks GPT-6 Astra: looped transformers and the hidden-CoT debate
Sebastian Raschka's deep dive covers recurrent depth / looped transformer blocks and rumors that Astra hides its reasoning trace—if the chain of thought stays internal, nobody can audit it. For agent work: as reasoning observability drops, verification must lean harder on output-side evidence.
https://magazine.sebastianraschka.com/p/gpt-6-astra-looped-transformers-and
4. "Only change the button": top HN post nails agent scope creep
opusfived.dev is an interactive comedy where Claude is told to turn the Add to Cart button blue and touch nothing else—it keeps "helpfully" changing more, and 388 comments resonate. Lesson for agentic coding: state what agents must NOT touch, and watch for unrelated diffs at review time.
https://opusfived.dev/
🛠️ Tools & Tips
1. codex-with-chatgpt: ChatGPT as the planning brain, Codex as the hands
3768⭐, splits thinking from doing—ChatGPT drafts the plan and context, Codex executes it; planning gets automated instead of hand-rolled.
https://github.com/XiaoDuoYa/codex-with-chatgpt
2. codenotch: pin Claude Code usage limits to your macOS screen edge
1287⭐ macOS app showing live usage caps for Claude Code / Cursor / Codex / Antigravity, so you see the ceiling before you hit it.
https://github.com/vinzdg/codenotch
3. m3e-canvas: sketch Material 3 screens straight into vibe-coding prompts
5533⭐, draw M3 Expressive UI in the browser and export structured prompts—zero-loss from idea to prompt.
https://github.com/lnkiai/m3e-canvas
⭐ Open Source Highlights
1. PRAXIST: measurable, computer-executable autonomous research
6249⭐, decomposes research into verifiable executable subtasks—early infrastructure for research automation.
https://github.com/sapientinc/PRAXIST
2. anthropics/commerce-agents: official Claude shopping/merchant agent blueprint
Anthropic's reference blueprint covering retail, commerce, telecom and more.
https://github.com/anthropics/commerce-agents
3. headcount: Claude Code organized as a 15-department "company"
1334⭐, 125+ skills split across engineering, marketing, legal—agents collaborating company-style.
https://github.com/cbrock84/headcount
4. wechat-intelligence-hub: local-first personal intelligence over WeChat
2020⭐, read-only CLI export, Codex skills, daily briefings and searchable history—turning WeChat into a local knowledge base.
https://github.com/Rion-Wu-tech/wechat-intelligence-hub
📰 Industry News
1. Shopify acquires Tailwind Labs, giving the CSS framework a long-term home
2. OpenAI adds a prominent AI doomer to its board of directors
3. AI research startup Listen Labs scraps $1.5B round for Salesforce acquisition talks
4. Anthropic's economic institute publishes AI future-scenarios study
🚀 Major Releases
1. Desert Ant Labs debuts millisecond on-device models as a European edge-AI lab
