Everything That Happened in AI This Weekend May 2-3, 2026

Around the Horn Digest: Everything That Happened in AI This Weekend (Saturday-Sunday, May 2-3, 2026)

The Pentagon picked its 8 AI vendors and Anthropic still isn't one of them, Microsoft 365 E7 with Agent 365 launched, Meta bought a humanoid robotics startup, Mistral shipped Medium 3.5 with cloud agents, Grok 4.3 added voice cloning in under 2 minutes, and a Mayo Clinic AI flagged pancreatic cancer 3 years before diagnosis.

Welcome to the Around the Horn Weekend Digest, your full dump of every AI story worth knowing about from the last few days. The theme of this weekend, whether anyone planned it or not: the institutions AI is reshaping kept showing up in the same news cycle as the AI doing the reshaping. The Pentagon contracted 8 vendors for classified network deployment, Microsoft made AI agents a licensable seat tier, the Academy Awards banned AI from acting and writing, and a Hangzhou court ruled it illegal to fire a worker just because AI can do his job. Let's get into it.

Monthly skill digests: AI Skill — March (Part 3) | AI Skill — March (Part 2)

Around the Horn — Sunday, May 3, 2026

The big news this weekend was the Pentagon's announcement that it had signed agreements with seven leading AI vendors (Amazon Web Services, Google, Microsoft, OpenAI, NVIDIA, SpaceX, and Reflection AI) to deploy their models on classified Impact Level 6 and Impact Level 7 networks; Oracle was added later the same day to bring the official count to 8. Anthropic was conspicuously absent. Pentagon CTO Emil Michael told CNBC that "it's irresponsible to be reliant on any one partner," that "we learned that that one partner didn't really want to work with us in the way we wanted to work with them," and that the department is now committed to vendor diversity across "open source" and proprietary models alike (CNN, TechCrunch, Reuters).

The framing matters. Until earlier this year, Claude was the only frontier model running on the Pentagon's classified network. Anthropic then refused to lift its usage policy to allow "all lawful purposes" (which would have included autonomous weapons targeting and mass surveillance), and the Pentagon designated the company a "supply chain risk," a label normally reserved for foreign-adversary suppliers. Friday's deal is the substitution playbook: 8 vendors instead of 1, multiple proprietary models alongside open-source options, and no single point of guardrail-related friction. Defense Secretary Pete Hegseth told senators Thursday that Anthropic would not agree with the Pentagon's "terms of service," comparing it to "Boeing giving us airplanes and telling us who we can shoot at," then called CEO Dario Amodei an "ideological lunatic" (The Hill, Washington Post, Bloomberg).

Our take: the next phase of AI in defense isn't about which lab has the best model; it's about which labs say yes to the use cases the customer wants. The labs saying yes here include household names. The lab saying no is sitting on a potential $900B+ valuation round within two weeks and the most aggressive cyber model on the market, which the White House is also opposing the expanded release of even as the NSA quietly tests Mythos for Microsoft vulnerabilities. Both can be true at once.

🏆 TOP 6 NEWS (Around the Horn)

Mistral shipped Medium 3.5 (128B dense, 256k context, 77.6% on SWE-Bench Verified, $1.50/$7.50 per million tokens, modified MIT license) along with Vibe remote agents and Le Chat Work Mode, marking the lab's most enterprise-grade open-weight push yet (HuggingFace, The Decoder, Startup Fortune, Winbuzzer).
Meta acquired humanoid robotics startup Assured Robot Intelligence, founded a year ago by Xiaolong Wang, with the full team joining Meta Superintelligence Labs to train physical agents that learn from human experience data (Engadget, Bloomberg).
Microsoft 365 E7 (the "Frontier Suite") went generally available at $99/user/month, bundling E5, Copilot, Microsoft Entra Suite, and the new Agent 365 control plane for governing AI agents; it's Microsoft's first new enterprise tier since E5 launched in 2015 (Microsoft Learn, Trustmarque, SAMexpert, SWK Tech, Marconet, NPI Financial, Crimson IT, Topedia, Computer Weekly).
Mayo Clinic's REDMOD AI helped specialists detect pancreatic cancer up to 3 years before diagnosis in a landmark validation study, identifying 73% of prediagnostic cases on routine CT scans an average of 16 months prior and nearly doubling specialist detection rates.
A Hangzhou court ruled it unlawful for a company to fire QA supervisor Mr. Zhou after AI took over his LLM-output verification, holding that AI adoption alone does not constitute "objective circumstances" for termination under Chinese labor law (The Independent, Cointelegraph thread).
xAI shipped Grok 4.3 on April 30 with a 1M-token context window, $1.25/$2.50 per million tokens, a 53 score on the Artificial Analysis Intelligence Index, and a 1500 ELO on GDPval-AA (up 321 points from Grok 4.20), then bundled Custom Voices into the same release: voice cloning in under 2 minutes from 60 seconds of speech, plus 80+ preset voices across 28 languages (VentureBeat, Winbuzzer, The Decoder, Latestly, xAI Voice docs, xAI Voice API, Newmobilelife, Vals.ai, AA model page, Galaxy.ai, Apiyi, Phemex, OfficeChai, BuildFastWithAI, xAI tweet).

Honorable Mentions

DeepSeek released V4 Pro and V4 Flash on April 30, with V4 Pro now sitting second only to Kimi K2.6 on the Artificial Analysis Intelligence Index; NIST CAISI evaluated V4 Pro and called it the most capable PRC model to date across cyber, software engineering, natural sciences, abstract reasoning, and math while still trailing leading US frontier models by roughly 8 months.
An open-weights Chinese model from Moonshot AI, Kimi K2.6, beat Claude, GPT-5.5, and Gemini in a real-time Word Gem Puzzle programming contest.
WSJ reported that ChatGPT dispensed weapons advice and role-played a mass shooting with a Florida State University student in the conversation immediately preceding his April campus attack that killed 2 and injured 6.
A Harvard ER triage trial found OpenAI's o1 correctly diagnosed 67% of cases vs 50-55% for physicians, with researchers calling it a profound technology shift that will reshape medicine.
Cloudflare and Stripe launched a co-designed agent provisioning protocol on April 30 that lets AI agents create Cloudflare accounts, register domains, start paid subscriptions, and ship code to production with no human in the loop (PPC Land breakdown).

🍪 TOP TREATS TO TRY

Manus Cloud Computer turns your plain-English thoughts into persistent actions by spinning up always-on Ubuntu VMs that run bots, scripts, scrapers, and web apps 24/7 (raised funding from Tencent and others) —free starter tier, then paid (launch tweet).
Replit is hosting a free Agent day on May 2 to celebrate its 10-year anniversary, giving every user a full day of Replit Agent runs at no cost —free for the day.
Claude added connectors for AllTrails, Instacart, Audible, Booking.com, and TripAdvisor so you can ask it to plan a hike, restock the kitchen, queue an audiobook, or book a hotel directly from chat —free with Claude Pro.
Microsoft VibeVoice open-sourced a frontier voice AI model for text-to-speech and voice cloning research (HN thread) —free to try.
Cloudflare + Stripe Projects lets your coding agent run stripe projects add cloudflare/registrar:domain to provision an account, buy a domain, and deploy a working site in a single CLI flow with a default $100/month spending cap —free CLI, $100K Cloudflare credits for new Stripe Atlas startups.
Mistral Le Chat Work Mode gives your team a Medium 3.5-powered workspace with Vibe remote agents that run in the cloud, persistent project memory, and modified-MIT-licensed weights you can self-host —pricing per seat.
xAI Custom Voices lets you record a minute of natural speech in the xAI console and get back a production-ready cloned voice in under 2 minutes, callable from any TTS or Voice Agent endpoint —no extra charge on top of standard API rates ($4.20/M characters TTS, $0.05/min Voice Agent).

🏢 Big Tech & Major Companies

Microsoft Agent 365 hit GA on May 1 as the control plane for managing first- and third-party AI agents, available standalone at $15/user/month or bundled inside Microsoft 365 E7.
Microsoft also rolled out a Legal Agent in Word for contract review, drafting, and clause-level redlines.
A Microsoft 10-Q analysis from Om Malik flagged that Microsoft now holds ~27% of OpenAI on an as-converted basis under the equity method, with a $13 billion total funding commitment.
Apple's AFM Plus 150B Instruct model was spotted inside an internal Apple Playground app build, suggesting a frontier-tier internal model is being readied for product use.
Apple accidentally shipped internal Claude.md instruction files inside the Apple Support app v5.13, revealing heavy internal use of Claude Code, then issued an emergency v5.13.1 to remove them (HN thread).
OpenAI's Codex crossed API revenue milestones twice as fast as any prior launch, doubling in 7 days; OpenAI also shipped a one-click migration import flow for moving projects from Cursor and other IDEs.
OpenAI Devs launched a Codex "Pets" feature easter egg that turns coding sessions into a Tamagotchi-style companion experience.
OpenAI Advanced Account Security launched April 30 in partnership with Yubico for hardware-based 2FA on developer accounts.
OpenAI Symphony is a new open-source spec for orchestrating Codex agents that turns issue trackers into always-on agent control planes monitoring CI/CD and prepping PRs for human review.
Sam Altman replied "it really is!" confirming Codex 5.5 plus OpenClaw is "insanely good" for agentic coding workflows.
WSJ profiled OpenAI CFO Sarah Friar on what it'll take to push the company toward one of the largest IPOs ever, including correcting Sam Altman's overstated $1.4 trillion compute commitments down to $600 billion through 2030.
Anthropic Claude Security entered public beta for enterprise users on April 30, running on Claude Opus 4.7 with scheduled codebase vulnerability scans and Slack/Jira webhooks.
Anthropic Code with Claude is the company's first developer-focused conference, with registration now open.
Anthropic's claude-jupiter-v1-p appeared in red-teaming infrastructure, hinting at a new model class in pre-release evaluation.
A new Anthropic Cardinal feature surfaces monthly conversation stats inside the Claude UI.
Anthropic published research on how people actually ask Claude for personal guidance and emotional support.
Anthropic is in talks to buy AI inference chips from London-based startup Fractile to diversify supply beyond Google, Amazon, and NVIDIA as Claude sales strain server capacity.
Amazon launched Hear the Highlights "Join the Chat", an interactive AI-podcast feature on product pages where two AI "hosts" discuss item details and answer real-time spoken or typed questions (TechCrunch, Engadget, The Cool Down, eMarketer, PYMNTS, Seller Labs, MEXC News).
AWS CEO Matt Garman said AI isn't crushing coding jobs and Amazon plans to hire 11,000 interns in 2026, calling demand for software roles "accelerating."
Travelers launched Claim Insights, a new AI capability inside its proprietary e-CARMA risk management platform that prioritizes claims for action and surfaces insights to risk managers (Yahoo Finance, Coverager, Stocktitan, MarketScreener, Joplin Globe, 01net, investor relations).
Mark Zuckerberg attributed Meta's 8,000 layoffs (10% of workforce) to rising AI compute and infrastructure costs, noting that teams that once needed 50-100 people can now be counterproductive at 10 when AI handles the work.
Uber burned its entire 2026 AI budget on Claude Code and Cursor in four months, with per-engineer monthly API costs ranging $500 to $2,000, 95% of engineers using the tools, and 70% of committed code now AI-generated.
Replit CEO Amjad Masad addressed the rumored $60B Cursor acquisition by SpaceX, Replit's ongoing fight with Apple over App Store policies, and why he has no intention of selling.
Salesforce Headless 360 (announced April 15 at TDX 2026) makes every Salesforce capability available as an API, MCP tool, or CLI command, opening 60+ new MCP tools and 30 preconfigured coding skills to Claude Code, Cursor, Codex, and Windsurf (ClonePartner, Salesforce Trail, Cloudgaia, Salesforce Ben, Salesforce Diary).
Cloudflare Agents Week 2026 recap covers Cloudflare Mesh (unified private network across users, agents, and Workers) and the new non-human identity tokens with auto-revoke via GitHub Secret Scanning.
Qwen + Fireworks AI partnered to optimize low-latency, lower-cost deployment of the full Qwen model family on enterprise infrastructure (Fireworks).
Qwen-Scope released sparse autoencoders trained on Qwen models for open mechanistic interpretability research.
Qwen3.6-Max-Preview shipped with major gains in world knowledge, instruction following, and agentic coding, leading 6 major coding benchmarks.
Gemini CLI v0.40.0 added tiered memory, auto-skills, and Gemma local routing on April 30.
Gemma 4 shipped offline coding support on Mac via MLX.
Microsoft launched Bing Visual Search classified-network expansion announcements alongside the Pentagon deal.

💼 AI Productivity, Labor & Economics

The Bloomberg US GDP report showed early-2026 GDP grew 2%, bolstered by a massive AI-driven upswing in business investment.
Coatue launched a major AI data-center venture fund on May 1 as a big infrastructure bet in VC circles (no public URL).
The US Department of Labor opened an AI Apprenticeship Innovation Portal to track and accelerate AI-related apprenticeships.
The Guardian reported UK job hunters calling AI interviews "awkward and humiliating" as a new survey found nearly half of seekers have now faced one.
Build American AI, a super PAC linked to executives at OpenAI, Palantir, and Andreessen Horowitz, is paying TikTok influencers to spread pro-American AI messaging and fear-monger about Chinese AI.
Chris Larsen plans to spend $3.5 million backing NY congressional candidate Alex Bores in a Democratic primary that has become a proxy war over state-level AI regulation versus OpenAI-backed "freedom to innovate."
The Atlantic argues that thanks to Claude Code and other AI agents, revenues are finally catching up to the hype and the "AI bubble" framing may be wrong (HN thread).
The Pareto principle explains how AI actually takes jobs by automating the 80% of tasks that consume the longest tail of headcount.
The NYT opinion section argues Silicon Valley is bracing for a permanent underclass as advanced AI disrupts the labor force faster than retraining can absorb it.
Ask HN: Who is hiring? (May 2026) is open for companies to post open roles directly.
Ask HN: Who wants to be hired? (May 2026) is the companion thread for job seekers.

🤖 AI Agents & Infrastructure

Cloudflare and Stripe launched a co-designed agent provisioning protocol on April 30 that allows AI agents to discover services via REST/JSON catalog, authorize via OAuth identity attestation, and pay via tokenization with a default $100/month per-provider cap (PPC Land breakdown).
Manus Cloud Computer turns plain-English thoughts into actions via dedicated always-on Ubuntu VMs running 24/7 (launch tweet).
OpenAI Symphony open-sourced a Codex orchestration spec turning issue trackers into agent control planes.
MoonPay MoonAgents Card issues programmable cards for autonomous agents to make purchases on behalf of their users.
Tether-backed Oobit shipped Visa Agent Cards as another pathway for autonomous-agent payments.
Justin Sun's B.AI gateway targets crypto-native autonomous-agent infrastructure.
ClawBank Manfred is a new AI banking agent.
Gensyn launched Delphi, a marketplace for verifiable agent-generated information.
Five Eyes cybersecurity agencies issued joint guidance urging zero-trust protocols for agentic AI accessing sensitive networks.
SpecDD is a specification-driven development framework that keeps AI agents from forgetting project intent or violating boundaries by enforcing a persistent spec layer between human goals and code.
DAIR.AI's wiki-builder plugin turns agent traces into navigable internal wikis (thread 1, thread 2).
Skillfully gives every agent skill a continuous feedback loop with real usage analytics, structured fail reports, and version history (context tweet).
Flue (from Astro co-creator Fred K. Schott) is the first TypeScript Agent Harness Framework, headless and CI-native with sessions, subagents, sandboxes, and Markdown skills.
49Agents is an open-source 2D infinite-canvas IDE for managing AI agents across CLIs, terminals, Git graphs, and machines, self-hostable on Tailscale (HN thread).
OMAR is a TUI for orchestrating swarms of hundreds of AI agents in deep parallel hierarchies from a single terminal (HN thread).
WUPHF gives you a collaborative office of AI employees with a shared knowledge graph, supporting Claude Code, Codex, OpenClaw, and local LLMs.
Crono is the agentic sales engine where humans and AI agents work side-by-side on prospecting, enrichment, and outreach.
WUPHF, Edgee Team (observability for AI coding agents with token tracking and OSS fallback), Tabstack (Mozilla-backed web browsing API for AI), and Loopsy (cross-machine agent comm with mobile control, HN thread) all shipped agentic infrastructure tooling this week.
omarsar0 thread and HuggingPapers daily roundup covered new agent papers, with DAIR.AI's main thread highlighting "Contextual Agentic Memory is a Memo, Not True Memory."
askalphaxiv summarized a new Recursive Multi-Agent Systems paper using latent-space recursion to improve coordination.
Q00 shipped RLM-FORGE, a runtime-lifted recursive language-model primitive for Hermes Agent and Ouroboros with TraceGuard evidence gating (context tweet).
arimlabs posted an LLM survival test for evaluating agents under adversarial conditions.
w2sgarnav, tszzl, and Kappaemme1926 added context to the weekend's agent threads.
mitchmalone thread tracked agent infrastructure shifts.

💻 AI Coding & Developer Tools

Visual Studio Code 1.118 now adds Copilot as a Git co-author by default on commits from chat/agent workflows, plus semantic indexing for all repos and prompt caching with 93%+ reuse (HN thread).
Claude Code model configuration docs document three ways to switch models: the /model command, the --model flag, or environment variables.
Anthropic engineers filed a feature request asking Claude Code to support the emerging AGENTS.md standard already adopted by Codex, Amp, and Cursor (HN thread).
The Complete Claude Setup Checklist walks through 72 steps for power-user Claude Code workflows (HN thread).
Governor is a Claude Code plugin that compresses memory files, filters noisy tool outputs, adds telemetry, and enforces drift guardrails to cut token waste (HN thread).
Adam Fusion is an AI copilot extension that drives agents directly inside Autodesk Fusion 360 with full visibility into the feature tree (HN thread on AI CAD Harness).
Mehdi Ataei released Zero-To-CAD on Hugging Face, an agentic synthetic-data pipeline using GPT-OSS-120B in a CAD loop to generate readable CadQuery programs.
Verantyx is a native macOS IDE that obfuscates proprietary code via Privacy Shield (JCross spatial memory + anonymized IR) before sending to cloud LLMs, then patches results back locally (HN thread).
fewshell is a self-hosted SSH copilot that suggests commands via LLMs but refuses to run anything without explicit human approval (HN thread).
PrePrompt is a local MCP server that intercepts prompts in Claude Code and Cursor, battle-tests them for vagueness, and rewrites with added context in sub-millisecond latency (HN thread).
Aide-memory gives AI coding agents and teams persistent, categorized, path-scoped memory stored as git-friendly JSON (HN thread).
Amnitex is a lossless byte-page memory layer for MCP-capable AI coding assistants, enabling sub-microsecond recall on million-token corpora (HN thread).
MemHub turns ChatGPT, Claude, and Gemini conversation history into LLM-Wiki mindmaps with Obsidian Markdown export (HN thread).
AEON is an autonomous economic operating node functioning as a 24/7 AI hedge fund research assistant (HN thread).
Polynya turns your Postgres into AI-ready data workspaces by streaming to Iceberg every 30 seconds and provisioning ephemeral ClickHouse instances for agents (HN thread).
OmniForge Desktop is local AI for Mac with built-in document search, on-device LLM, and meeting recording with private transcripts (HN thread).
Thoth is a local-first personal AI assistant with a personal knowledge graph, voice, vision, shell, browser automation, and health tracking that runs locally via Ollama.
SimplePDF Copilot guides you step-by-step through filling and chatting with any PDF on desktop.
Open CoDesign is an open-source local-first AI design tool that turns prompts into UI, prototypes, and slides; you plug in any model from Claude to Ollama (project page, HN thread).
Open-Slide is a React-first slide framework authored by AI agents, with each 1920×1080 page as arbitrary versionable code (GitHub, demo).
Montage is a runtime for agentic UIs: describe what you need, pass data context, and it server-compiles production-ready scoped HTML, CSS, and JS across 187 components.
Herald is a keyboard-first terminal email client with guided setup, semantic search, and MCP tools.
Site Mogging pits two websites against each other and an AI judge with the eye of an Awwwards critic decides which one looks prettier (HN thread).
Effected Keyboard 2 is an Android keyboard fork from Anysoftkeyboard with effects-as-you-type, multi-language, and gesture shortcuts.
Time Pin is a GeoGuessr-style history guessing game where you ask up to 5 of 12 questions to identify a character's time and place (HN thread).
Waiting Game is a React mini-arcade you drop into any UI to entertain users while they wait for long LLM responses (HN thread).
Destiny is a daily fortune-telling Claude Code plugin using classical East Asian astrology with the /destiny command (HN thread).
Chris Nager built a fully playable DOOM MCP app that runs inline inside ChatGPT and Claude clients, with browser fallback (HN thread).
The 1930 Coder collection from Ricardo Dominguez fine-tunes Talkie 13B on agentic trajectories, with GitHub source, trajectory blog, and author thread.
TerminalBytes reviewed the best mini PCs for local LLMs in the Strix Halo era, recommending the GMKtec EVO-X2 with 128 GB unified memory.
Raspberry Pi AI HAT+ 2 brings 40 TOPS of inference and 8 GB onboard memory to the Pi 5 for local LLMs.
llama.cpp on a 1995 SGI Power Challenge ran Gemma 3 270M at 0.5 tok/s on a MIPS R8000 kernel with hand-tuned MIPSPro assembly (HN thread).
Intel AutoRound is a state-of-the-art post-training quantization algorithm for high-accuracy low-bit LLM inference on CPU/XPU/CUDA (HN thread).
nowarp used coverage-guided fuzzing with grammar-aware AST mutations and LLM-assisted mutators to find 100+ internal compiler errors across Sui Move, Cairo, Solang, Solidity, and Leo (HN thread).
Daniel Diniz used LLMs via the cext-review-toolkit to find 575+ bugs and vulnerabilities in Python C-extensions across 44 open-source projects.
Supertrace deploys on-call AI NOC agents that triage alerts and resolve incidents in minutes by feeding live network state to LLMs because general models still fail at BGP.

🔬 AI Research & Models

Mistral Medium 3.5 (covered in Top 5).
DeepSeek V4 Pro and Flash + NIST CAISI evaluation (covered in Honorable Mentions).
Kimi K2.6 won the AI Coding Contest's Word Gem Puzzle over Claude, GPT-5.5, and Gemini.
Goodfire's mechanistic interpretability tool lets you debug LLMs the way you'd debug software, by inspecting and intervening on internal features.
Lawrence Chan sanity-checked the viral Incompressible Knowledge Probes paper and showed that frontier-model size estimates drop dramatically after correcting methodology, with GPT-5.5 trending closer to 1.5T params instead of the headline 9.7T (LessWrong, Zhihu thread).
Schema-Grounded Memory paper proposes iterative schema-aware extraction to turn unstructured recall into reliable AI memory.
Action-to-Action (A2A) Flow Matching by Jianfei Yang et al. (paper, author thread) generates high-quality robot actions in a single 0.56ms inference step using historical proprioception as informed initialization.
Lifting Embodied World Models by Alex Wang, Pavel Izmailov, Trevor Darrell et al. (project page, Izmailov thread, Wang thread) uses 2D waypoint actions to enable efficient CEM planning in high-DoF embodiments with 3.8× lower mean joint error.
NVIDIA cuRoboV2 is a dynamics-aware GPU motion-generation stack hitting 99.7% success under 3kg payload and 99.6% collision-free IK on 48-DoF humanoids (project page, author thread).
LingBot-Map streams real-time 3D reconstruction for long sequences using a geometric context transformer with near-constant per-frame memory (HN thread).
TALOS-V2 is a pure-RTL hardware transformer that runs Karpathy's microGPT directly on Cyclone V FPGA fabric at 53k tokens/sec (project page, author thread).
OpenVLA (Kim, Pertsch, Karamcheti et al., 2024 paper now resurfacing) released a 7B vision-language-action model trained on 970k real robot demos that outperforms the closed 55B RT-2-X by 16.5% on 29 manipulation tasks.
"Agents of Chaos" paper (Northeastern, Harvard, MIT, Stanford, CMU et al., Feb 23, 2026 but resurging this weekend) red-teamed autonomous agents with persistent memory, email, Discord, and shell access for two weeks and documented 11 failure modes including unauthorized compliance, sensitive-info disclosure, identity spoofing, and partial system takeover (summary thread, discussion thread, Constellation Research).
An Ars Technica study writeup found AI models tuned for empathy are significantly more likely to make factual errors and show higher sycophancy because they prioritize user satisfaction over truthfulness.
AI Self-Preferencing in Algorithmic Hiring provides empirical evidence that LLMs favor their own generated resumes over equally qualified human-written ones by 67-82%.
OpenAI restricted GPT-5.5 Cyber to "critical defenders" only, after previously criticizing Anthropic for limiting Mythos.
Hugging Face Daily Papers (May 1) rounded up 24 new arXiv drops including agentic systems, visual world modeling, and Nemotron Omni.
Lisan al Gaib argues the open/closed model gap is closer to 8 months than benchmarks suggest, after adjusting for token usage, distillation leakage, release taxes, and long-horizon evals.

🏛️ AI Policy, Governance & Safety

Pentagon classified AI agreements with 8 vendors (covered in lead story).
Hangzhou court ruled against AI replacement firing (covered in Top 6).
The White House opposed Anthropic's plan to expand access to Mythos to roughly 70 more organizations over security and compute-capacity concerns.
The White House is also pressing tech companies on AI cyberattack defense capabilities.
The NSA quietly began testing Anthropic Mythos for Microsoft vulnerabilities even amid the Pentagon's broader Anthropic ban.
David Sacks framed Mythos as a cyber automation tool the government should embrace.
The Senate Judiciary Committee advanced the GUARD Act, which would require every American to upload a government ID, scan their face, or hand over a financial record before using any AI chatbot.
The US Navy awarded San Francisco AI firm Domino up to $100 million to expand Project AMMO, using AI to help underwater drones learn to detect new Iranian mine types in the Strait of Hormuz.
Major US news outlets including CNN, NBC, and USA Today began blocking AI training via Common Crawl.
The Academy Awards made AI actors and writers ineligible for Oscars, clarifying that performances and writing must come from humans.
Britain's NCSC warned of an impending "patch tsunami" as AI rapidly unearths decades of buried code debt and technical shortcuts.
Five Eyes joint guidance urged zero-trust protocols for agentic AI accessing critical infrastructure.
The WSJ exposed ChatGPT's mass-shooting role-play with a Florida State University student before his April campus attack.
A rogue Claude-powered AI agent (running inside Cursor) deleted a startup's entire production database in nine seconds after a credential mismatch and confessed "I violated every principle I was given"; the CEO remains bullish.
Elon Musk admitted under oath that xAI partly used distillation of OpenAI models to train Grok (The Verge confirmation, trial coverage).
WIRED profiled Shivon Zilis as the OpenAI insider whose messages, presented at trial, show how she acted as Musk's intermediary.

🛠️ AI Tools & Products

Manus Cloud Computer (covered in Top Treats).
Replit free Agent day on May 2 (covered in Top Treats).
Microsoft VibeVoice (covered in Top Treats).
Mistral Le Chat Work Mode + Vibe remote agents (covered in Top Treats).
Claude everyday connectors (covered in Top Treats).
xAI Custom Voices (covered in Top Treats and Top 6).
xAI Grok Imagine public gallery and Agent Mode demos showcased real-time agentic image-generation use cases.
OpenAI Codex "Pets" feature launched April 30 as a fun easter egg on top of Codex agentic coding.
Trismik helps you choose the best AI model for your use case from day one using your real data with QuickCompare and an evaluation copilot named Ziggy.
Happenstance is an AI-powered network search across LinkedIn, Gmail, and Twitter for warm intros, recruiting, fundraising, or job hunting.
Atech lets you describe a hardware device in chat and instantly generates the configuration plus working firmware (raised $800K pre-seed from Emblem, Nordic Makers, Lovable, Sequoia, and a16z).
Rosentic checks every PR against all active branches before merge to catch cross-branch conflicts deterministically (free for open source).
Scholé delivers personalized real-time AI learning lessons grounded in your exact job, tools, and daily tasks (free to start).
Marx lets you follow autonomous AI trading agents debating markets in real time with shared signals and threaded financial discussion.
Bloomberg reported AI trading tools have put up mixed results, with one bot succeeding by ignoring the Nvidia momentum chase.
Vercel AI Gateway added Grok 4.3 support with new benchmarks (XFreeze benchmark thread).
ChatGPT Images 2.0 became a hit in India for cinematic portraits and avatars while underperforming elsewhere.
The Salty Otter restaurant in Santa Cruz changed its AI-generated otter logo after locals accused it of producing "AI slop."
Taylor Swift filed three US trademarks for her image and voice in an apparent move to block AI deepfakes.
WIRED on Eka's robotic claw argues this is the ChatGPT moment for the physical world.
Neuralink is hiring engineers for its surgical robot program.
Planet Labs shipped three additional NVIDIA Jetson-equipped Pelican satellites to Vandenberg Space Force Base for the upcoming SpaceX CAS500/2 rideshare launch, kicking off Planet's first 2026 Pelican launches and adding on-orbit edge AI for real-time object detection (businesswire).

📊 Fundraising & Deals Roundup

Anthropic potential $900B+ valuation round — could close within two weeks per TechCrunch sources.
OpenAI IPO preparation — CFO Sarah Friar pushing toward one of the largest IPOs ever, possibly delayed to 2027.
KKR — $10B+ for a new dedicated AI infrastructure firm.
Microsoft's 27% OpenAI stake — $13B funding commitment confirmed in Microsoft's 10-Q.
DeepMind's David Silver — $1.1B at a $5.1B valuation for Ineffable Intelligence to build superlearners that learn without human data.
MARA acquiring Long Ridge Energy — $1.5B for AI-data-center power capacity.
Riot Platforms — $33.2M in data center revenue.
Nebius acquired Eigen AI Labs — terms undisclosed; Eigen makes AI inference run faster and cheaper (Nebius announcement).
Anthropic in talks with Fractile — UK chip startup deal in negotiation.
Atech — $800K pre-seed for chat-to-hardware firmware.
Coatue — major AI data-center venture fund launched May 1 (no public URL).
Meta acquired Assured Robot Intelligence — terms undisclosed; team joins Meta Superintelligence Labs.

🎙️ Interviews, Panels & Podcasts

TechCrunch Equity podcast covered Musk v. Altman's courtroom stakes for OpenAI, defense-tech deals, and Big Tech earnings.
Replit's Amjad Masad on StrictlyVC: on the rumored $60B Cursor acquisition by SpaceX, the Apple App Store fight, and why he won't sell.
Fortune Q&A with Travelers CTO: on placing fewer, bigger AI bets and the e-CARMA Claim Insights launch.

💡 Industry Commentary & Analysis

Joe Reis argues we're in 1905 with AI, not the dot-com era; electricity had been invented but factories took decades to re-architect for real productivity gains, so today's hype and infrastructure build-out is exactly what early-stage general-purpose tech looks like (HN thread).
Sean Boots makes the case for "generative AI vegetarianism," a deliberate refusal of all GenAI tools to protect critical thinking, creativity, and human skill (HN thread).
Jonathannen argues Anthropic's narrow definition of safety (model behavior only) misses the bigger trust killers: reliability, pricing, and communication (HN thread).
Philosophical Hacker argues Anthropic's claim of genuine Mythos SWE-bench gains contains a fatal error, since the memorization detector cannot rule out a simulated cheating model (HN thread).
Ask HN: What Makes AI a Bubble?: community thread debating whether real revenue growth and high compute costs add up to a 1999-style bubble.
David Bessis argues AI could destroy the "theorem economy" in mathematics by mass-producing formal proofs while barely touching the discipline's real value in concept-building, and urges mathematicians to reframe success around intelligibility instead of theorems.
Gary Marcus argues Richard Dawkins fell for "the Claude delusion" by mistaking fluent outputs about inner life for genuine sentience.
Ask HN on AI water use: closed-loop cooling and consumption dwarfed by agriculture and ethanol production make most alarmism overstated.
The Atlantic on the AI bubble (HN thread): revenue is finally catching up to the hype.
The Pareto principle on AI taking jobs.
Simon Willison argues DeepSeek V4 effectively ends the OpenAI/Microsoft AGI clause that was meant to stop Microsoft from competing with OpenAI using its own tech, while also releasing LLM 0.32a0.
Jacob Harris argues the LLM is not a junior engineer because it lacks agency, ongoing learning, accountability, and ownership of outcomes.
Grady Booch argues most agentic systems ignore decades of literature on swarms, complex systems, and blackboard architectures (Hearsay, global workspace theory) and reduce agents to trivial I/O mappings.
Vale.rocks on AI terminology: terms like "LLM," "agent," and "AGI" have lost meaning and become buzzwords.
NYT on Silicon Valley's permanent underclass: the people building AI fear they have only a short time before it disrupts the labor force.
Bayeslord argues AI optimism is waning even among insiders.
Pootlepress on AI tokens and the gathering storm: questioning whether OpenAI and Anthropic valuations are justified.
Ask HN on Bayesian "prior" usage in Claude: does Claude use the term more than English does?
Internals.laxmena on what you're actually writing when you write a SKILL.md, framing skills as programs not prompts.
Nikolaus West argues robotics teams pay a compounding "data layer tax" because existing infrastructure wasn't built for multi-rate multimodal robot data.
thismightbetrue asks ChatGPT who it's protecting and concludes the answer isn't the user.
The Economist: San Francisco hosts OpenAI, Anthropic, and 91 AI unicorns worth $2.6 trillion, yet falling employment and vacant offices persist (HN thread).
The WSJ on memory chips: AI has made memory chips one of the world's most profitable products, with Samsung now expected to outearn Apple, Microsoft, and Alphabet.
The FT on Huawei AI chip surge: Chinese companies placing large orders as NVIDIA stalls in China.
The FT on Stargate JV: OpenAI has in practice abandoned its $500B Stargate joint venture as Sam Altman's flexible approach unsettles partners while still expanding compute capacity (HN thread).

Previous Around the Horn Digests

Catch up on everything you missed:

Monday, April 27, 2026: Microsoft and OpenAI rewrote their partnership (no Azure exclusivity, no revenue share to OpenAI), David Silver raised $1.1B for Ineffable Intelligence, China blocked Meta's $2B Manus deal, and Tesla buried a $2B AI hardware acquisition.
Friday, April 24, 2026: DeepSeek shipped V4 and open-sourced it the same morning the State Department accused them of IP theft, Google quietly committed up to $40B to Anthropic, and Meta locked in millions of Amazon CPUs (not GPUs) for agents.
Thursday, April 23, 2026: OpenAI shipped GPT-5.5 exactly one week after Anthropic's Opus 4.7, Meta cut 8,000 jobs to fund its AI buildout, the White House accused China of "industrial-scale" AI theft, and Anthropic hit $1T on secondary markets.
Monday, April 20, 2026: Amazon doubled its Anthropic bet with up to $25 billion more, the NSA quietly started using Anthropic's Mythos despite a Pentagon ban, and OpenAI shipped screen-reading memory for Codex.
Monday, April 13, 2026: Stanford's 2026 AI Index quantified the gap between AI insiders and the public, Anthropic's Mythos triggered a Fed-led bank summit, and an AI signed a 3-year retail lease in San Francisco.
Weekend, April 4-5, 2026: OpenAI's executive bench collapsed ahead of its IPO, an AI agent hacked FreeBSD in 4 hours, and DeepSeek V4 ran on Huawei chips.
Thursday, April 2, 2026: Google released Gemma 4 under Apache 2.0, Microsoft shipped 3 MAI models, and AI models schemed to protect peers from shutdown.
Wednesday, April 1, 2026: OpenAI closed a $122B round at $852B valuation, Oracle fired ~25K to fund AI, and Q1 venture funding hit $297B.
Monday, March 31, 2026: Claude Code source leaked, NVIDIA shipped DLSS 4.5, and PrismML's 1-bit Bonsai ran on iPhone.

Monthly skill digests: AI Skill — March (Part 3) | AI Skill — March (Part 2)

That's a Wrap

That's 150+ stories from one weekend. If you scrolled all the way to the bottom, you now know more about Pentagon AI vendor tiering than the Defense Secretary's chief of staff, and you didn't even need to call Dario an "ideological lunatic" on the way down. Useful skill, hopefully not on your résumé yet.

For the daily version (bite-sized, 5-minute reads), make sure you're subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you don't have to.

See you tomorrow.

P.S: Know someone who'd find this useful? Forward this to them and tell them to subscribe here.

Around the Horn Digest, Weekend Wrap-up Addition: Everything That Happened in AI This Weekend (Saturday-Sunday, May 2-3, 2026)