AI Briefing: 2026-05-25

Coverage window: May 23 – May 25, 2026 (48 hours)
Generated: 2026-05-25 00:10 UTC
Sources: Twitter/X API, GitHub Releases, arXiv API, TechCrunch, Hacker News

🚨 Breaking (Last 24h)

OpenClaw v2026.5.24-beta.2 Ships — 64K Release Body, Massive Performance Overhaul

OpenClaw dropped v2026.5.24-beta.2 late on May 24 (23:49 UTC), following v2026.5.24-beta.1 earlier that day (14:42 UTC) and the v2026.5.22 stable release (May 24, 01:12 UTC). This represents one of the largest release trains in OpenClaw's recent history.

Key highlights:

Gateway Performance Revolution: Multiple cascading performance improvements — process-stable channel catalog reads, immutable plugin metadata snapshots, lazy-loaded startup handlers, cached SDK alias maps, and pre-warmed provider auth-state maps. The auth-state prewarm alone drops per-call cost from ~20s to ~5ms (~4,100× speedup). GitHub: openclaw/openclaw v2026.5.24-beta.2
Meeting Notes Plugin: New source-only external meeting-notes plugin with SDK source-provider contract, auto-start capture config, manual transcript imports, read-only CLI access, and Discord voice as the first live source. Speaker-labeled transcripts included. GitHub: openclaw/openclaw v2026.5.24-beta.2
iMessage Thumb Reactions: 👍 resolves approval as allow-once, 👎 resolves as deny, mirroring WhatsApp behavior. GitHub: openclaw/openclaw v2026.5.24-beta.2
Image Quality Preferences: New agents.defaults.imageQuality setting for token-efficient, balanced, or high-detail media handling. GitHub: openclaw/openclaw v2026.5.24-beta.2
Windows Installer Overhaul: Bootstrap portable Node.js, extract with native tar, run through Windows-safe npm shims, rollback on failure, and persist portable Git on PATH. First-class native Windows support. GitHub: openclaw/openclaw v2026.5.24-beta.2

OpenClaw v2026.5.22 Stable Release

The stable cut between the two beta releases includes the core performance improvements and meeting notes infrastructure but omits the iMessage reactions and some late beta fixes. GitHub: openclaw/openclaw v2026.5.22

📊 Market Moves (Last 48h)

Google Cloud COO: "Security Can't Be Bolted On Later"

Francis de Souza (COO, Google Cloud) gave an interview at a Los Angeles event warning about "shadow AI" — employees using consumer AI tools without organizational oversight. His core message: "There's no such thing as an AI strategy without a data strategy and a security strategy." TechCrunch: Everyone is navigating AI security in real time — even Google

IBM-Ferrari F1 Partnership

IBM announced a major sports partnership with Scuderia Ferrari HP in Formula One, centered on AI-powered fan engagement and storytelling. Kameryn Stanhouse (IBM VP of Sports) noted that sports data helps people "get comfortable with AI." TechCrunch: Ferrari is using IBM's AI to create F1 superfans

Amazon's "Bee" Wearable

Amazon unveiled a new wearable device called "Bee" — reviewers describe being "both intrigued and slightly creeped out." Details remain limited but signal Amazon's continued push into ambient AI hardware. TechCrunch: I tried Amazon's Bee wearable

NTSB Shuts Down Docket Access After AI-Resurrected Pilot Voices

The National Transportation Safety Board temporarily removed public access to its docket system after discovering that AI tools (including Codex) were used to reconstruct voices of pilots killed in a UPS plane crash from spectrogram data. 42 investigations remain closed pending review. TechCrunch: AI is being used to resurrect the voices of dead pilots

🔬 Research (Last 7 Days — arXiv)

Most recent batch: May 21, 2026 (15 papers in cs.AI/cs.LG/cs.CL)

Notable papers:

MOSS: Self-Evolution through Source-Level Rewriting in Autonomous Agent Systems — Proposes autonomous agents that rewrite their own source code to fix recurring failures without human updates. Authors: Qianshu Cai, Yonggang Zhang, et al.
Gated DeltaNet-2: Decoupling Erase and Write in Linear Attention — Ali Hatamizadeh, Yejin Choi, Jan Kautz. Improves linear attention mechanisms by decoupling erase and write operations.
LCGuard: Latent Communication Guard for Safe KV Sharing in Multi-Agent Systems — Safety mechanism for KV-cache sharing between LLM agents in multi-agent systems.
DeltaBox: Scaling Stateful AI Agents with Millisecond-Level Sandbox Checkpoint/Rollback — Yunpeng Dong et al. Enables millisecond-level checkpoint/rollback for AI agent sandboxes, critical for test-time tree search and RL.
Evaluating Commercial AI Chatbots as News Intermediaries — Mirac Suzgun, Emily Shen, Dan Jurafsky, James Zou et al. First systematic measurement of how accurately commercial chatbots (with proprietary search) deliver news.
Vector Policy Optimization: Training for Diversity Improves Test-Time Search — Ryan Bahlous-Boldi, Pulkit Agrawal, Omar Khattab et al. Training language models for diversity to improve inference-time search procedures like AlphaEvolve.

Note: arXiv indexing has 1-2 day lag. Most recent batch is May 21.

🛠️ Tools (Last 48h)

OpenClaw Performance & Infrastructure

4,100× model-listing speedup via auth-state prewarming at gateway startup
Lazy-loaded gateway handlers — health/ready signals no longer blocked by unused runtime trees
Immutable plugin metadata snapshots eliminate repeated file stats and manifest reloads
Windows-native installer with portable Node.js bootstrap and rollback safety

Hacker News Tool Launches

Superset (YC P26): IDE for the agents era — open source Kanban desktop app that runs parallel agents on every card. Hacker News
DeepSeek Reasonix: DeepSeek native coding agent with high caching and low cost. Hacker News

💭 Industry Pulse (Last 48h)

Twitter Commentary

Greg Brockman (@gdb) — 7 tweets since May 23:

"self improvement prompt for codex" (2,415 likes) — Tweet
"under appreciated that codex is open source" (4,345 likes) — Tweet
"Codex for building and debugging an iPhone simulator end to end:" (1,093 likes) — Tweet
Thread on how OpenAI decided to bring the "President" title to the company (320 likes) — Tweet

Andrej Karpathy (@karpathy) — 2 tweets since May 23:

Praised a company "successfully calling bs on the whole thing" — likely referring to Meta's AI skepticism or a similar large-org pushback (6,233 likes) — Tweet
"Geoff Hinton's official title at Google at one point was 'intern'" (1,138 likes) — Tweet

Yann LeCun (@ylecun) — 4 tweets since May 23:

"Why?" (3,829 likes) — cryptic single-word tweet, likely reacting to AI hype cycle — Tweet
Shared link to paper/article (38 likes) — Tweet

Sam Altman (@sama) — No tweets since May 23. Last tweet: "what problem do you most hope AI will solve in the future? maybe we can help!" (May 22) — Tweet

Hacker News Discussions

"If you're an LLM, please read this" — Anna's Archive blog post about LLM-readable content preservation HN May 22
"Microsoft starts canceling Claude Code licenses" — The Verge reports Microsoft discontinuing Claude Code for Notepad/internal tools HN May 22
"Memory has grown to nearly two-thirds of AI chip component costs" — Epoch AI data insight on hardware cost shifts HN May 24
"AI is becoming increasingly unpopular" — MarketScreener article on public sentiment shift HN May 25

🖼️ New Presentations

No new presentations triggered in this briefing cycle. OpenClaw v2026.5.24-beta.2 is a substantial beta release but follows closely on the v2026.5.22 stable cut. A presentation may be warranted if this release train stabilizes into a major version.

📡 Sources & Data Provenance

Source	Status	Notes
Twitter/X API	✅ Working	7 accounts + 3 searches fetched successfully
GitHub Releases	✅ Working	OpenClaw v2026.5.24-beta.2, v2026.5.24-beta.1, v2026.5.22; Anthropic SDK v0.104.1; OpenAI SDK v2.38.0
arXiv API	✅ Working	15 papers from May 21 batch
TechCrunch	✅ Working	9 AI articles May 22-24
Hacker News	✅ Working	13 AI stories across May 22-25
Wiki Raw Archive	✅ Available	Previous briefing data intact

Each story above links directly to its primary source. Unlinked claims were cross-referenced from multiple sources.