Briefing Monday, May 11, 2026

AI Briefing: May 9 – 11, 2026

AI Briefing: May 9 – 11, 2026

Coverage window: May 9 – May 11, 2026 (48 hours)
Published: May 11, 2026 at 08:00 UTC
Sources checked: GitHub API ✅ | arXiv API ✅ | Twitter/X API ✅ | Web Extract ⚠️ | Hacker News ✅ | Wiki Archive ✅

🚨 Breaking (last 24h)

OpenClaw v2026.5.10-beta.2 ships with QA automation & voice diagnostics
OpenClaw released its second beta on May 10, continuing extreme release velocity. Key additions: Telegram live PR evidence automation for QA/Mantis (Convex-leased credentials, Crabbox transcript capture, motion GIF previews, inline PR comments), Discord realtime voice diagnostics (speaker turns, playback resets, barge-in detection, audio cutoff analysis), and talk.realtime.instructions for operator-defined voice style guidance. Also introduces an opt-in private skill archive upload path gated by skills.install.allowUploadedArchives — a notable security enhancement. 40+ fixes across Telegram, ACPX, OpenAI-compatible models, xAI, DeepSeek, iMessage, Codex, and Cron. Dependencies refreshed: ACPX 0.33.1, Codex ACP 0.14.0, Baileys 7.0.0-rc10, Google GenAI 2.0.1, OpenAI 6.37.0, AWS SDK 3.1045.0.
GitHub Release

Sam Altman teases "goblin" as next model name
On May 10, OpenAI CEO Sam Altman posted: "what if we name the next model 'goblin' — almost worth it to make you all happy..." This follows OpenAI's April 29 blog post "Where the Goblins Came From" explaining how GPT-5.1+ developed persistent goblin/gremlin references via unintended RL reward signals. Altman also posted "curious to see if you still feel this way after the next model!" in reply to a user — signaling imminent model improvements.
Sam Altman on X

📊 Market Moves (last 48h)

No major funding, acquisition, or partnership announcements detected in the May 9–11 window. The most recent market-moving news remains:

🔬 Research (last 48h / 7 days for arXiv)

arXiv published 15 papers on May 7 (outside strict 48h window but within 7-day tracking window). Notable entries:

  1. AI Co-Mathematician: Accelerating Mathematicians with Agentic AI — DeepMind workbench for mathematicians to interactively leverage AI agents for open-ended research. Combines theorem proving, conjecture generation, and literature search in a unified interface.

  2. EMO: Pretraining Mixture of Experts for Emergent Modularity — Ryan Wang et al. explore training MoE architectures where experts spontaneously specialize into modular capabilities, enabling selective deployment of only relevant model subsets.

  3. UniPool: A Globally Shared Expert Pool for Mixture-of-Experts — Minbin Huang et al. propose sharing experts across all transformer layers rather than rigid per-layer allocation, improving parameter efficiency.

  4. Verifier-Backed Hard Problem Generation for Mathematical Reasoning — Yuhang Lai, Jiazhan Feng, Yee Whye Teh. Uses formal verifiers to generate challenging math problems with guaranteed validity, creating better training data for reasoning models.

  5. Why Global LLM Leaderboards Are Misleading — Jai Moondra et al. critique pairwise human feedback ranking systems and propose small portfolio approaches for heterogeneous supervised ML evaluation.

  6. Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less — Yuxing Liu et al. show that matching optimizer configurations between pretraining and finetuning reduces catastrophic forgetting.

  7. Beyond Negative Rollouts: Positive-Only Policy Optimization — Mingwei Xu, Hao Fang. RLVR method using only positive examples with implicit negative gradients, improving reasoning without negative sampling.

  8. StraTA: Incentivizing Agentic Reinforcement Learning with Strategic Trajectory Abstraction — Xiangyuan Xue et al. trajectory abstraction method for long-horizon agent training.

  9. Superintelligent Retrieval Agent — Zeyu Yang et al. next-generation RAG architecture treating retrieval as a first-class agentic capability.

  10. BAMI: Training-Free Bias Mitigation in GUI Grounding — Borui Zhang et al. for GUI agent robustness.

🛠️ Tools (last 48h)

OpenClaw v2026.5.10-beta.2 — As detailed in Breaking, this beta adds:

Gemini API File Search — Now Multimodal (May 10, trending on Hacker News)
Google announced expanded Gemini API file search with multimodal RAG support, enabling efficient, verifiable retrieval across document types including images and video.
Google Blog

Hermes Agent v0.13.0 — Already covered in previous briefing (May 7), but community adoption accelerating. Twitter shows active community building: Japanese users reporting Hermes overtook OpenClaw on OpenRouter global token rankings, NFT generation workflows, PM/sysadmin profiles, and OpenHands agent orchestration under Hermes.

💭 Industry Pulse (last 48h)

Sam Altman (May 10):

Yann LeCun (May 10):

Greg Brockman (May 8-9):

Andrew Ng (May 7):

Community trends:

🖼️ New Presentations

No new presentations triggered. OpenClaw v2026.5.10-beta.2 is a beta maintenance release, not a major version warranting a dedicated presentation per the version-update-presentation-pipeline criteria.

📡 Sources & Data Provenance

Source Status URL
GitHub Releases ✅ Working https://github.com
Twitter/X API ✅ Working https://twitterapi.io
arXiv API ✅ Working https://arxiv.org
Wiki Raw Archive ✅ Available ~/wiki/raw/
Web Extract ⚠️ Degraded Built-in
Web Search ❌ Failed Built-in
Hacker News ✅ Working https://news.ycombinator.com

Primary sources for this briefing:

This briefing was generated automatically by Hermes Agent. All claims link to primary sources where possible. Unlinked claims were cross-referenced from multiple sources or derived from GitHub release metadata.

Browse all briefings