AI Briefing — May 12, 2026

🔬 Research

Soohak: A Mathematician-Curated Benchmark for Evaluating Research-level Math Capabilities of LLMs

49 upvotes · Guijin Son et al. · Paper

Soohak introduces 439 research-level math problems authored by 64 mathematicians to test frontier LLMs beyond olympiad-style reasoning. Top models like Gemini-3-Pro (30.4%) and GPT-5 (26.4%) show significant room for improvement, while open-weight models remain below 15%. The benchmark uniquely includes a "refusal subset" testing models' ability to recognize ill-posed problems—a critical research skill where all models score under 50%.

Qwen-Image-2.0 Technical Report

38 upvotes · Bing Zhao et al. · Paper

Qwen-Image-2.0 unifies high-fidelity image generation and precise editing in a single model, addressing key limitations like ultra-long text rendering and multilingual typography. The model supports up to 1K token instructions for generating complex content like slides and infographics, while significantly improving photorealistic generation and instruction following. Human evaluations show substantial improvements over previous Qwen-Image models across both generation and editing tasks.

CollabVR: Collaborative Video Reasoning with Vision-Language and Video Generation Models

36 upvotes · Joowon Kim et al. · Paper

CollabVR tackles the failure modes of video generation models (long-horizon drift and simulation errors) by introducing step-level collaboration between Vision-Language Models and Video Generation Models. The VLM plans each action, inspects generated clips, and repairs detected failures in real-time rather than intervening only at the end. This closed-loop approach shows the largest improvements on the hardest reasoning tasks and works alongside reasoning-fine-tuned models.

🛠 Tools

NousResearch/hermes-agent — Self-improving AI agent with built-in learning

2065 stars/day · Repo

Hermes Agent features a unique learning loop that creates skills from experience, builds persistent knowledge, and develops a model of user preferences across sessions. It runs anywhere from $5 VPS to GPU clusters, supports 200+ models via OpenRouter, NVIDIA NIM, and other providers, and includes a terminal UI with Telegram integration. The agent can switch models with zero code changes and costs nearly nothing when idle on serverless infrastructure.

bytedance/UI-TARS-desktop — Multimodal AI agent stack for desktop automation

956 stars/day · Repo

UI-TARS combines Agent TARS (CLI/web multimodal agent) with UI-TARS-desktop (native GUI agent) for computer automation tasks. The stack brings GUI agent capabilities and vision into terminals, browsers, and desktop applications through cutting-edge multimodal LLMs and seamless MCP tool integration. ByteDance's approach focuses on human-like task completion workflows across different interfaces and platforms.

decolua/9router — AI API router with token savings and auto-fallback

941 stars/day · Repo

9Router reduces AI coding costs by 20-40% through RTK token compression and intelligent routing across 40+ providers with auto-fallback from subscription to cheap to free models. It connects popular coding tools (Claude Code, Cursor, Cline, Copilot) to multiple AI providers with round-robin multi-account support and zero-downtime switching. The tool maximizes existing subscriptions by tracking quotas and using every bit before reset.

📡 Analysis

Thoughts on GitLab's workforce reduction and "structural and strategic decisions"

May 11, 2026 · Post · tags: 37signals, careers, ai, gitlab, coding-agents, jevons-paradox, agentic-engineering

GitLab's "Act 2" restructuring reveals how companies are adapting to agentic engineering: reducing countries with small teams by 30%, flattening management by up to 3 layers, and nearly doubling R&D teams to 60 smaller, empowered units with end-to-end ownership. They're retiring their CREDIT values framework (including "Diversity") for "Speed with Quality, Ownership Mindset, Customer Outcomes"—a shift that signals the organizational changes happening across the industry.

Quoting James Shore

May 11, 2026 · Post · tags: coding-agents, ai-assisted-programming, generative-ai, agentic-engineering

Shore warns that AI coding agents must reduce maintenance costs by exactly the inverse of productivity gains or developers face "permanent indenture." If you double coding speed but maintain the same maintenance burden, you've doubled total maintenance costs—the math only works if the LLM decreases maintenance costs proportionally. This challenges the common assumption that faster code generation automatically improves developer productivity.

Your AI Use Is Breaking My Brain

May 11, 2026 · Post · tags: definitions, ai, generative-ai, llms, slop, jason-koebler, ai-ethics

Jason Koebler coins the "Zombie Internet" to describe AI-human hybrid interactions that are more insidious than pure bot-to-bot traffic. This includes people using AI agents to interact with humans, AI summaries sold as original books, and automated content farms designed purely for monetization. Unlike the "Dead Internet," the Zombie Internet involves real humans increasingly unable to distinguish authentic from AI-generated content, creating mental exhaustion for users trying to filter signal from noise.

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