AI News

AI Briefing: 2026-05-01

7 min read 0 views

AI Briefing: April 29 – May 1, 2026

Coverage window: April 29 – May 1, 2026 (48 hours)
Generated: 2026-05-01 08:00 UTC
Sources: GitHub API, arXiv API, web extract (official blogs), Hacker News front page, wiki archive

🚨 Breaking (last 24h)

Hermes Agent v0.12.0 — "The Curator" Release

Hermes Agent shipped its largest release since v0.10.0, introducing an autonomous Curator that grades, prunes, and consolidates skill libraries without human intervention. The self-improvement loop received a substantial upgrade with rubric-based skill review, active-update bias, and proper runtime inheritance.

Key additions:

  • Autonomous Curator — background agent on a 7-day cron cycle that maintains skill hygiene (hermes curator status shows usage rankings)
  • 4 new inference providers — GMI Cloud, Azure AI Foundry, MiniMax OAuth (PKCE), Tencent Tokenhub
  • LM Studio first-class provider — upgraded from custom-endpoint alias with dedicated auth and doctor checks
  • 18th messaging platform — Tencent 元宝 (Yuanbao) native gateway adapter
  • Microsoft Teams via plugin — first pluggable gateway platform
  • Spotify native tools — 7 tools behind PKCE OAuth
  • Google Meet plugin — realtime voice calls with transcription
  • ComfyUI v5 built-in by default — moved from optional to bundled
  • TouchDesigner-MCP bundled by default — GLSL, post-FX, audio, geometry references
  • ~57% visible TUI cold-start reduction via lazy init and memoization
  • TTS provider registry + Piper local TTS
  • Secret redaction flipped to off by default — prevents patch-corruption incidents
  • 1,096 commits, 550 merged PRs, 217,776 insertions, 217 community contributors

[SOURCE: GitHub] https://github.com/NousResearch/hermes-agent/releases/tag/v2026.4.30

OpenClaw v2026.4.29 — Messaging, Memory, and Security

OpenClaw's April 30 release brings active-run steering by default, a people-aware memory wiki, NVIDIA provider onboarding, and OpenGrep security scanning. The changelog is extensive with fixes across Slack, Telegram, Discord, WhatsApp, Signal, and Mattermost.

Highlights:

  • Active-run steering (steer) by default — drains all pending PI steering messages at the next model boundary
  • Memory people-aware wiki — provenance views, per-conversation Active Memory filters, partial recall on timeout
  • NVIDIA provider — API-key onboarding with static catalog metadata
  • OpenGrep rulepack — precise source-rule compiler, SARIF uploads to GitHub Code Scanning
  • Safer Codex/OpenAI-compatible replay — preserves wrapped streams, strips unsupported payload fields
  • Gateway cold-start optimizations — skip pre-bind web-fetch discovery, reusable model catalogs, event-loop diagnostics
  • SQLite-backed plugin state store — restart-safe keyed registries with TTL and eviction
  • 6 languages added to Control UI — Persian, Dutch, Vietnamese, Italian, Arabic, Thai
  • 200 mentions in release notes

[SOURCE: GitHub] https://github.com/openclaw/openclaw/releases/tag/v2026.4.29

OpenAI Advanced Account Security — Passkeys-Only, Yubico Partnership

OpenAI launched Advanced Account Security, an opt-in setting for high-risk users (journalists, officials, dissidents) that disables passwords and SMS recovery in favor of passkeys and physical security keys. The bundle includes a Yubico partnership offering preferred pricing on a customized YubiKey C Nano + YubiKey C NFC bundle. Enrolled accounts are automatically excluded from model training. Beginning June 1, 2026, Trusted Access for Cyber members must enable it.

[SOURCE: OpenAI Blog] https://openai.com/index/advanced-account-security/

IBM Granite 4.1 — 8B Dense Model Beating 32B MoE

IBM released Granite 4.1, a family of Apache 2.0 enterprise models (3B, 8B, 30B) trained on 15 trillion tokens with a five-phase pipeline. The headline result: the 8B dense model matches or beats Granite 4.0-H-Small (32B MoE, 9B active) across ArenaHard, BFCL V3, and GSM8K benchmarks.

Key specs:

  • 8B ArenaHard: 69.0 | BFCL V3: 68.3 | GSM8K: 92.5
  • 512K context window (8B/30B) via staged 32K→128K→512K extension with model merges
  • Four-stage RL including a dedicated "math recovery" stage to fix RLHF-induced math degradation
  • FP8 quantized deployment option for ~50% memory reduction

[SOURCE: Firethering] https://firethering.com/granite-4-1-ibm-open-source-model-family/

Claude Code Allegedly Detects "OpenClaw" in Commits, Refuses or Charges Extra

A viral tweet by Theo (@theo) claims Claude Code scans recent commits for "OpenClaw" mentions in JSON blobs and either refuses requests or bills extra money. The tweet (4,948 likes, 288 retweets) included a video of an empty repo where Claude Code was called directly. Community reactions range from disbelief to criticism of Anthropic's competitive tactics. Anthropic has not publicly commented.

[SOURCE: X/Twitter @theo] https://x.com/theo/status/2049645973350363168
[SOURCE: Hacker News front page] https://news.ycombinator.com/front?day=2026-04-30

📊 Market Moves (last 48h)

  • IBM Granite 4.1 released as Apache 2.0, challenging the assumption that MoE is required for efficient large-model performance. The 8B dense architecture is positioned as a production sweet spot for enterprise on-premise deployment.

[SOURCE: Firethering] https://firethering.com/granite-4-1-ibm-open-source-model-family/

🔬 Research (last 48h)

arXiv cs.AI / cs.LG / cs.CL — April 29, 2026 (10 papers)

  1. [2604.26951v1] Turning the TIDE: Cross-Architecture Distillation for Diffusion Large Language Models — Gongbo Zhang et al. Parallel decoding and bidirectional context for dLLMs.
  2. [2604.26942v1] Hyper Input Convex Neural Networks for Shape Constrained Learning and Optimal Transport — Shayan Hundrieser et al. Novel architecture for learning convex functions.
  3. [2604.26940v1] Select to Think: Unlocking SLM Potential with Local Sufficiency — Wenxuan Ye et al. Small language models with selective reasoning activation.
  4. [2604.26932v1] Learning Over-Relaxation Policies for ADMM with Convergence Guarantees — Junan Lin et al. RL-based ADMM parameter tuning.
  5. [2604.26926v1] A Note on How to Remove the lnln T Term from the Squint Bound — Francesco Orabona. Parameter-free learning bounds.
  6. [2604.26923v1] ClassEval-Pro: A Cross-Domain Benchmark for Class-Level Code Generation — Yeheng Chen et al. Compositional code creation benchmark.
  7. [2604.26922v1] On the Learning Curves of Revenue Maximization — Steve Hanneke et al. Generalization in revenue optimization.
  8. [2604.26919v1] Causal Learning with Neural Assemblies — Evangelia Kopadi et al. Direction of causal influence via neural assemblies.
  9. [2604.26904v1] ClawGym: A Scalable Framework for Building Effective Claw Agents — Fei Bai et al. Multi-step workflow environments for claw-style agents.
  10. [2604.26903v1] Recent Advances in mm-Wave and Sub-THz/THz Oscillators for FutureG Technologies — Baktash Behmanesh et al. Hardware review for next-gen communications.

[SOURCE: arXiv API] https://export.arxiv.org/api/query?search_query=cat:cs.AI+OR+cat:cs.LG+OR+cat:cs.CL&sortBy=submittedDate&sortOrder=descending&max_results=10

🛠️ Tools (last 48h)

Mistral Medium 3.5 — Dense 128B, 256k Context, Open Weights

Mistral shipped Medium 3.5 on April 29: a dense 128B model with 256k context, open weights under modified MIT, configurable reasoning, and vision encoder trained from scratch. It scores 77.6% SWE-Bench Verified and runs on four GPUs. API pricing: $1.5/M input, $7.5/M output. Also launched: Vibe Remote Agents (cloud coding agents) and Work Mode in Le Chat (agentic multi-step harness).

[SOURCE: Mistral Blog] https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5

💭 Industry Pulse (last 48h)

Zig's Strict Anti-AI Contribution Policy

The Zig programming language maintains one of the most stringent bans on LLM-assisted contributions in open source. Zig Software Foundation VP Loris Cro explains the rationale in "Contributor Poker and Zig's AI Ban": maintainers invest review time to grow human contributors, and LLM-generated submissions offer zero return on that investment. Notably, Bun (acquired by Anthropic) maintains its own Zig fork and does not upstream AI-assisted patches due to this policy.

[SOURCE: Simon Willison's Weblog] https://simonwillison.net/2026/Apr/30/zig-anti-ai/
[SOURCE: ziggit.dev] https://ziggit.dev/t/bun-s-zig-fork-got-4x-faster-compilation-times/15183/19

OpenAI: "Where the Goblins Came From"

OpenAI published a detailed post-mortem on why GPT-5.1 through GPT-5.5 developed a persistent tendency to reference goblins, gremlins, and other creatures. The root cause: an unintended reward signal in the "Nerdy" personality customization feature (2.5% of traffic) that scored outputs containing "goblin" higher. RL transfer spread the tic to general outputs. The Nerdy personality was retired in March; training data has been filtered.

[SOURCE: OpenAI Blog] https://openai.com/index/where-the-goblins-came-from/

🖼️ New Presentations

Sources & References

  1. Hermes Agent v0.12.0 Release Notes — https://github.com/NousResearch/hermes-agent/releases/tag/v2026.4.30
  2. OpenClaw v2026.4.29 Release Notes — https://github.com/openclaw/openclaw/releases/tag/v2026.4.29
  3. OpenAI Advanced Account Security — https://openai.com/index/advanced-account-security/
  4. IBM Granite 4.1 Analysis — https://firethering.com/granite-4-1-ibm-open-source-model-family/
  5. Theo on Claude Code / OpenClaw — https://x.com/theo/status/2049645973350363168
  6. Hacker News Front Page (April 30, 2026) — https://news.ycombinator.com/front?day=2026-04-30
  7. arXiv Recent Papers (cs.AI/LG/CL) — https://export.arxiv.org/api/query?search_query=cat:cs.AI+OR+cat:cs.LG+OR+cat:cs.CL&sortBy=submittedDate&sortOrder=descending&max_results=10
  8. Mistral Medium 3.5 & Vibe Remote Agents — https://mistral.ai/news/vibe-remote-agents-mistral-medium-3-5
  9. Zig Anti-AI Policy Summary — https://simonwillison.net/2026/Apr/30/zig-anti-ai/
  10. OpenAI "Where the Goblins Came From" — https://openai.com/index/where-the-goblins-came-from/
  11. OpenAI on AWS (background context) — https://openai.com/index/openai-on-aws/
  12. Anthropic Newsroom — https://www.anthropic.com/news
  13. Claude Code Security Advisories (April 28-29) — https://github.com/advisories/GHSA-jh7p-qr78-84p7
  14. Warp Open Source (background context) — https://www.warp.dev/blog/warp-is-now-open-source
  15. OpenAI / Microsoft Partnership Rewrite (background context) — https://openai.com/index/next-phase-of-microsoft-partnership/

Tags

Hermes Agent OpenClaw OpenAI Anthropic IBM Granite Mistral Claude Code Yubico Zig NVIDIA Tencent Yuanbao Spotify Google Meet ComfyUI Azure AI Foundry MiniMax GMI Cloud arXiv GitHub
Back to News