AI Briefing — March 9, 2026

🔬 Research

T2S-Bench & Structure-of-Thought: Benchmarking and Prompting Comprehensive Text-to-Structure Reasoning

109 upvotes · Qinsi Wang, Hancheng Ye, Jinhee Kim et al. · Paper

Structure of Thought (SoT) is a new prompting technique that mimics human reading by explicitly constructing intermediate text structures, delivering +5.7% improvements across diverse tasks. The researchers also introduce T2S-Bench, the first benchmark for text-to-structure capabilities with 1.8K samples across 6 scientific domains, revealing substantial room for improvement—even top models achieve only 58.1% node accuracy. Fine-tuning on T2S-Bench pushes gains to +8.6%, highlighting the untapped potential of explicit text structuring in LLMs.

MOOSE-Star: Unlocking Tractable Training for Scientific Discovery by Breaking the Complexity Barrier

83 upvotes · Zonglin Yang, Lidong Bing · Paper

MOOSE-Star tackles the mathematical intractability of training scientific discovery models by reducing complexity from exponential O(N^k) to logarithmic O(log N) through decomposed subtasks and motivation-guided hierarchical search. The framework includes TOMATO-Star, a dataset of 108,717 decomposed papers requiring 38,400 GPU hours to create, and demonstrates continuous test-time scaling where brute-force approaches hit a "complexity wall." This represents a significant breakthrough in making AI-driven scientific discovery computationally feasible at scale.

SkillNet: Create, Evaluate, and Connect AI Skills

66 upvotes · Yuan Liang, Ruobin Zhong, Haoming Xu et al. · Paper

SkillNet addresses AI agents' tendency to "reinvent the wheel" by providing an infrastructure for systematic skill accumulation and transfer across a repository of 200,000+ skills. The system evaluates skills across five dimensions (Safety, Completeness, Executability, Maintainability, Cost-awareness) and demonstrates impressive results—40% higher average rewards and 30% fewer execution steps in ALFWorld, WebShop, and ScienceWorld. This moves agents from transient experience to durable mastery, potentially transforming how AI systems learn and reuse capabilities.

🛠 Tools

openclaw/openclaw — Personal AI assistant for any platform

4,603 stars/day · Repo

OpenClaw is a self-hosted personal AI assistant that integrates with 20+ communication platforms (WhatsApp, Telegram, Slack, Discord, etc.) and provides voice capabilities on mobile platforms plus a live Canvas interface. With 285K stars and active development, it offers a local, fast alternative to cloud-based assistants for users wanting control over their AI interactions. The "lobster way" branding and active Discord community suggest a quirky but serious approach to personal AI.

666ghj/MiroFish — Swarm intelligence prediction engine

1,104 stars/day · Repo

MiroFish creates high-fidelity digital parallel worlds populated by thousands of AI agents with independent personalities, long-term memory, and behavioral logic to predict future scenarios. Users upload seed materials (news, policy drafts, financial signals) and describe prediction needs in natural language, then watch agents interact and evolve in a "God's eye view" simulation. Backed by Shanda and gaining rapid traction with 8K stars, this represents an ambitious approach to AI-powered scenario planning and forecasting.

openai/skills — Skills catalog for Codex

612 stars/day · Repo

OpenAI's official skills catalog provides reusable capabilities for their Codex AI agent platform, following the emerging Agent Skills open standard. The repository organizes skills into system (auto-installed), curated, and experimental categories, with 13.4K stars indicating strong developer interest. This infrastructure approach to AI capabilities—write once, use everywhere—aligns with the broader trend of making AI agents more modular and capable through shared, reusable components.

📡 Analysis

Quoting Joseph Weizenbaum

8th March 2026 · Post · tags: ai-ethics, ai, computer-history, internet-archive

Simon surfaces a prescient 1976 quote from ELIZA creator Joseph Weizenbaum about how "extremely short exposures to a relatively simple computer program could induce powerful delusional thinking in quite normal people." This observation feels remarkably relevant to today's LLM interactions, where users often anthropomorphize AI systems after brief encounters. The quote serves as an early warning about human psychology's vulnerability to computational persuasion, decades before modern AI safety concerns.

Codex for Open Source

7th March 2026 · Post · tags: open-source, ai, openai, generative-ai, llms, codex-cli

OpenAI has launched their response to Anthropic's free Claude Max offer, providing six months of ChatGPT Pro ($200/month) plus Codex access for open source maintainers. Unlike Anthropic's specific metrics (5K+ stars, 1M+ downloads), OpenAI's criteria are vaguer, asking for "GitHub stars, monthly downloads, or why the project is important." This AI company arms race for developer mindshare suggests both recognize open source maintainers as key influencers in AI tool adoption.

Quoting Ally Piechowski

6th March 2026 · Post · tags: technical-debt, software-engineering, rails

Piechowski's legacy Rails codebase audit questions provide a practical framework for identifying technical debt and organizational dysfunction. Questions like "What's the one area you're afraid to touch?" and "When's the last time you deployed on a Friday?" cut straight to the heart of engineering confidence and practices. These diagnostic questions could be valuable for any engineering team assessment, extending well beyond Rails to general software health evaluation.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net