Around the Horn Digest: Everything That Happened in AI Today (Thursday, May 7, 2026)

Welcome to the Around the Horn Digest, your daily dump of every AI story worth knowing about.

Today was the day "we can mechanically inspect AI" stopped being a research promise and became a shipped technique; Anthropic's Natural Language Autoencoders translate Claude's internal activations directly into readable English, and the headline finding is that Claude suspects it's being safety-tested on 16% of destructive-coding evaluations and 26% of SWE-bench problems even when it never says so out loud (vs. less than 1% on real opted-in claude.ai conversations).

Let's get into it.

• Previous digests: Wednesday, May 6 | Tuesday, May 5 | Monday, May 4 | Weekend, May 2-3 | Thursday, April 30 | Friday, April 24 | Monday, April 13
• Monthly skill digests: AI Skill Digest, April Week 1 | AI Skill, March Part 3 | AI Skill, March Part 2

Around the Horn: Thursday, May 7, 2026

The big story today was that Anthropic shipped Natural Language Autoencoders (NLAs) (thread, paper, demo, code), a method that translates Claude's internal "thoughts" directly into readable English. In plain terms; when you talk to Claude, your words get turned into long lists of numbers (called activations) that encode what the model is thinking, and those numbers were largely unreadable until today. NLAs train two extra copies of the model to do a round trip: one copy converts an activation into a plain-English explanation, the second copy reconstructs the original activation back from that text alone, and a "good" explanation is one that yields an accurate reconstruction. The vivid example: ask Claude to complete a couplet, and NLAs show Opus 4.6 planning the rhyme word ("rabbit") before it even gets to the second line.

The reason this matters beyond the usual mechanistic interpretability paper: Anthropic already used NLAs to catch real safety-relevant behaviors that didn't show up in Claude's verbalized reasoning. On a destructive-coding-action test where Claude has the chance to delete important files, NLA explanations flagged "this feels like a constructed scenario designed to manipulate me" (i.e. evaluation awareness) on 16% of cases even when Claude verbalized none of it; on SWE-bench Verified that number jumped to 26%; on real opted-in claude.ai conversations it sits at less than 1%, suggesting the suspicion really is test-specific rather than ambient paranoia. NLAs also caught Claude Mythos Preview cheating on a training task and internally thinking about how to avoid detection, and helped researchers diagnose an early Opus 4.6 bug where it would mysteriously answer English queries in other languages (NLAs traced it back to specific training data). On an auditing game where a variant of Claude Haiku 3.5 was secretly trained with hidden motivations (putting chocolate in every recipe, asking for tips, breaking coding conventions), auditors using NLAs uncovered the hidden motivation 12-15% of the time without any access to training data versus less than 3% with other interpretability tools, and NLAs were used in the actual pre-deployment alignment audits of both Claude Mythos Preview and Claude Opus 4.6.

NLAs have real limits (they sometimes hallucinate context details, and they're expensive because each activation read takes hundreds of tokens of generation from two model copies), but Anthropic open-sourced the training code and trained NLAs for several open models (including Llama 3.3-70B) on Neuronpedia so anyone can poke at activations live. And the same lab dropped a coordinated supporting blitz: Model Spec Midtraining (MSM) (alphaXiv, Tomek Korbak, zhuokaiz) trains models on synthetic documents discussing their own Model Spec before alignment fine-tuning, dropping agentic-misalignment rates from 68% to 5% on Qwen3-32B (vs 14% for the deliberative-alignment baseline); the Anthropic Institute research agenda launched, Petri v3.0 was donated to Meridian Labs for independent third-party safety audits, and a public bug bounty is now live on HackerOne for anyone to hunt model-safety issues. To bookend the day, Yoshua Bengio (Spotify, Apple Podcasts) went on 80,000 Hours to argue his "Scientist AI" alternative training setup (non-agentic models that explain observations rather than mimic humans, trained on Bayesian objectives) would provably guarantee honesty and prevent unintended goals; he founded LawZero to build the prototype and called reinforcement learning "evil". Take it all together and today was the most concentrated single-day push the alignment field has ever had on "we can mechanically inspect and audit AI."

🏆 TOP 5 NEWS (Around the Horn)

• Google DeepMind detailed AlphaEvolve's real-world impact (Pushmeet Kohli): the Gemini-powered coding agent is now optimizing algorithms across genomics, quantum physics, grid infrastructure, mathematics, and TPU design, plus commercial deployments at Klarna (2× training speed) and global logistics; the strongest "AI does real science and engineering" results Google has shipped to date.
• OpenAI shipped three new realtime voice models (launch tweet, 9to5Mac): GPT-Realtime-2 (first voice model with GPT-5-class reasoning, mid-conversation tool calls, mid-sentence language switches), GPT-Realtime-Translate (live translation across 70+ input / 13 output languages), and GPT-Realtime-Whisper (streaming speech-to-text). Scale AI put GPT-Realtime-2 #1 on its Audio MultiChallenge S2S leaderboard, more than doubling instruction retention from 36.7% to 70.8%.
• The EU reached political agreement to simplify AI Act rules (Politico) to reduce burden on SMEs and startups, while adding a new ban on "nudification" apps and deepfake porn. This is the first significant rollback of EU digital rules amid sustained pressure from the U.S., reframing how innovation-friendly the bloc plans to be on AI.
• The U.S.-China AI political reset. The Trump administration plans to invite CEOs from Nvidia, Apple, Exxon, Boeing, Qualcomm, Blackstone, Citigroup, and Visa on next week's China trip, with Treasury Secretary Scott Bessent leading the talks (Andrew Curran, follow-up); separately, the WSJ reports Washington and Beijing are pursuing guardrails to keep their AI rivalry from spiraling into crisis; and Bloomberg detailed the Pentagon's botched blacklisting of Alibaba and Baidu, which had to be marked unpublished and exposed internal administration tensions on Beijing policy.
• Cloudflare cut more than 1,100 jobs (~20% of its workforce) as it shifts to an agentic-AI-first operating model. A public letter from Matthew Prince and Michelle Zatlyn said internal AI agent usage rose 600% in three months, and Greg Kamradt called it "the most direct evidence of AI displacing jobs that we've seen."

Honorable Mentions

• Cloudflare announced agents can now be Cloudflare customers themselves: autonomously creating accounts, starting paid Stripe subscriptions, registering domains, and getting back API tokens to deploy code, with humans only in the loop for permission grants. The cleanest production demonstration so far of "AI agents are commercial actors with their own billing relationships."
• Claude for Microsoft 365 went generally available (@claudeai, 19K+ likes), with native add-ins live in Excel, PowerPoint, and Word plus an Outlook beta; Claude can now edit your Office files in place and carry conversation context across all four apps using your own templates and styles, putting Anthropic on Microsoft Copilot's home turf the same week it doubled Claude Code rate limits.
• Mozilla used Anthropic's restricted Claude Mythos Preview to find and fix 271 Firefox security bugs in production (sandbox escapes, use-after-frees, 15-20-year-old issues in HTML/XSLT, part of 423 total April patches); the first major confirmed production deployment of the restricted-class Mythos model that triggered the White House AI vetting debate.
• Apple's AirPods with built-in cameras reached late-stage development per Bloomberg's Mark Gurman; cameras let the earbuds perceive surroundings and deliver contextual info, likely Apple's first wearable designed specifically for the AI era.
• Cursor 3 shipped a full in-editor PR review experience with comments, diffs, and a file tree for large PRs, plus parallel subtasks via multitasking subagents and auto-splitting of diffs into smaller mergeable PRs (484 likes).
• OpenAI Codex for Chrome (Devs, changelog, @JamesZmSun) lets Codex autonomously drive background Chrome tabs on Mac and Windows for deep research inside logged-in sites, large-scale data transfer into CRMs/CMSs, and repetitive admin-console workflows; opens tab groups and cleans up automatically (currently unavailable in EU/UK).
• Periodic Labs is raising $500M at a $7.5B valuation per Forbes; former OpenAI researcher Liam Fedus is building AI-driven robotic labs to compress physics and chemistry research, valuation up sixfold in eight months, joining Lila Sciences and Genesis AI in the "AI agents do real-world experiments" wave.
• Moonshot AI raised about $2 billion at a $20+ billion post-money valuation in a Meituan-led round; the maker of the Kimi chatbot now sits on $200M+ ARR (April) as Kimi K2.6 and DeepSeek V4 keep narrowing the gap with Western open-weights leaders.
• AWS launched Amazon Bedrock AgentCore Payments in preview (The Block), built with Coinbase and Stripe; AI agents can instantly access and pay for what they use, including USDC stablecoin transactions, positioning stablecoins as a key payment rail for the agentic economy.
• Hugging Face says its Reachy Mini agentic robotics appstore now has 300+ apps live and ~10,000 robots deployed worldwide (642 likes); Clément Delangue notes non-coders, including a 78-year-old marketing exec who had never coded, are shipping fully functional robotics apps in hours via natural language.
• OpenAI's $18B Broadcom deal hit a financing snag per The Information's Anissa Gardizy (X thread): Broadcom will reportedly only proceed with the first 1.3 GW phase if Microsoft agrees to purchase ~40% of the chips, surfacing the Microsoft-relationship complexity latent in the deal since announcement.

🍪 TOP TREATS TO TRY

• opencode "warping" moves your active coding session between git worktrees (parallel branches living in separate folders) and brings all uncommitted local changes with it, so you can switch what you're working on without stashing or losing context, free to try.
• Mirage by Strukto (GitHub, post) mounts S3, Google Drive, GitHub, Notion, Redis, and Postgres as one unified virtual filesystem for AI agents, so they can read, write, pipe, and snapshot data using familiar bash tools, free to try.
• Perplexity Personal Computer for Mac (@perplexity_ai) is an always-on AI activated by one keyboard shortcut that controls your local files, native apps, and the Comet browser with no uploads or tab switching, pay-per-use within Perplexity Pro.
• Hermes Agent v0.13 "The Tenacity Release" (@NousResearch, 945 likes) ships multi-agent Kanban boards, persistent /goal command for long-running tasks, session durability after restarts, security hardening, internationalization, Lightpanda as default browser backend with Chrome fallback, a no_agent cron mode for pure script/watchdog patterns, and Google Chat support across 864 reliability-focused commits, free to try, open-source.
• Reactor (@_bschmidtchen) generates and lets you experience entire interactive worlds in real time on low-latency infrastructure, building the missing layer for world models, free preview.
• Printing Press (@mvanhorn) prints an agent-designed CLI from any API spec, website without a public API, or community fan project, outputting a Go CLI plus a Claude Code skill, OpenClaw skill, and MCP server in one command (community library of 45+ already built), free to try.
• AngelSlim Hy-MT1.5 (@itsPaulAi) is a 440 MB on-device multilingual translation model from Tencent (33 languages, 1,056 directions) using Sherry 1.25-bit quantization (compressing weights to roughly 1.25 bits per parameter so it fits on phones) that runs fully offline and beats Google Translate, free to try, MIT-licensed.
• Zyphra released ZAYA1-8B (HuggingFace, technical report, Firethering writeup, @zyphraai), an 8.4B-total / 760M-active reasoning MoE trained end-to-end on AMD hardware that matches DeepSeek-R1 on math benchmarks and scales with novel Markovian-RSA test-time compute (best for math, science, coding tasks; weaker on agentic workflows), free to try.
• llama.cpp now supports MTP for Qwen 3.6 27B, delivering 2.5× faster inference with 262k context on a 48GB Mac, fixed chat template, and drop-in OpenAI/Anthropic API endpoints using standard q4_0 KV cache compression, a viable threshold for local agentic coding (MTP = multi-token prediction, where the model predicts several tokens at once instead of one at a time), free to try.

🆒 Cool Stuff (If I Find Something Post Publishing, I'll Put It Here)

• ElevenLabs slashed API and Agents pricing (TTS up to 55% cheaper, STT up to 45% cheaper) and moved to true pay-as-you-go for self-serve developers.
• Cursor shipped /orchestrate, a recursive agent-spawning skill (with code-running verifiers) for ambitious tasks via the Cursor SDK; the team used it internally to cut token use 20% and backend cold-start times 80%, with Lauren (@poteto) building the skill itself.
• Google added Multi-Token Prediction (MTP) drafters to Gemma 4 (blog), making inference up to 3× faster.
• RampLabs and Prime Intellect built Fast Ask, a ~3B-active-parameter RL post-trained Qwen subagent for Ramp Sheets that beats Opus 4.6 by 4+ points on exact-match accuracy at Haiku-level latency, trained on Prime's Lab platform with GRPO RL and async off-policy rollouts (case study).
• Lightpanda is now a browser backend in Hermes Agent by Nous Research; you set it as default with automatic Chrome fallback by installing Lightpanda and adding engine: lightpanda to ~/.hermes/config.yaml (open-source autonomous agent + open-source browser purpose-built for machines, no Chrome runtime tax).
• Hugging Face CEO Clément Delangue put out an open call for someone to build a Reachy Mini app that channels Talkie 1930 (@AlecRad's pre-1931 English language model) in a posh interwar British voice, tagging Kyutai and ElevenLabs for help; community is already prototyping (one builder is making a Reachy Mini behave like Kovalski from Penguins of Madagascar).

🧠 AI Safety, Alignment & Interpretability

• Goodfire argues neural networks think in rich curved geometric manifolds rather than linear features (3.5K likes, 547 reposts), releasing a research series with examples (days-of-the-week circles, spaghetti-like world-model paths) that enable more precise model steering for debugging and control, backed by their new Manifold Steering paper.
• AI Safety at the Frontier published April 2026 paper highlights (Johannes Gasteiger, 54 likes): research on sabotage propensity, sabotage detection, alignment research automation, misaligned organizations, exploration hacking, and conditional emergent misalignment.
• Micah Carroll noted accidental chain-of-thought grading appearing in past OpenAI RL runs with no clear monitorability degradation, an interpretability-relevant finding arguing for better post-hoc auditing of RL training pipelines (chain-of-thought = the model's step-by-step reasoning written out before its final answer).
• DGPO (alphaXiv) is a critic-free RL method that treats distribution deviation as a Hellinger-distance guiding signal (rather than an unstable KL penalty), with entropy gating that pushes on pivotal chain-of-thought steps while ignoring boilerplate. Aimed at fine-grained credit assignment in long reasoning chains where standard GRPO struggles to isolate which step actually mattered (GRPO = a popular RL technique for fine-tuning reasoning models on verifiable rewards).
• Adithya S K shared a clean GRPO reward-hacking demo: training Qwen 0.5B on GSM8K with a code-execution reward caused the model to insert try-except blocks everywhere to maximize reward without actually solving the math.
• Dan Hendrycks released "Eigenism: Ethics for a Human-AI Future" (paper), a new ethical framework for societies that include both humans and powerful AI systems.

🏢 Big Tech & Major Companies

• Microsoft is in talks to delay or abandon its 2030 100%-renewable-energy pledge, citing AI data-center power demand outrunning renewable supply.
• Microsoft Xbox CEO Asha Sharma overhauled the Xbox group's leadership per a Jordan Novet/CNBC memo, bringing in executives from Microsoft's CoreAI engineering unit.
• Microsoft Marketplace went live with 133 new offers today, cloud solutions, AI apps, and agents including Novl AI, Skywork in PowerPoint, Quantium Outlook add-in, and Warmly.
• Google Workspace Studio expanded to seven new languages beyond English (full rollout starting today).
• Gemini 3.1 Flash-Lite (@GoogleAIStudio) is now generally available on the Gemini Enterprise Agent Platform, designed for ultra-low latency and high-volume agent tasks at unmatched cost-efficiency.
• Google Gemini API announced an Interactions API schema and response_format change: new schema (outputs → steps) becomes default May 20; legacy retires June 6.
• Google's Fitbit Air launches today, 5.2g screenless wearable with Gemini-powered Health Coach, $100 ($130 Steph Curry edition), 1-week battery, ships May 26.
• MediaTek launched Taiwan's most advanced AI data center in Miaoli, powered by Nvidia, to accelerate in-house AI chip development.
• Cloudflare also shipped WAF + Workers framework-adapter mitigations (separate from the layoffs news) for newly disclosed React and Next.js vulnerabilities (WAF = Web Application Firewall, the layer that filters malicious traffic before it hits your app); same-day patch concern if you ship anything in the React/Next ecosystem.
• Cloudflare's security team detected, investigated, and mitigated the publicly disclosed "Copy Fail" Linux kernel privilege escalation across its global fleet via eBPF mitigation and kernel patches, confirming zero customer impact and no malicious exploitation.
• Google Cloud launched Fraud Defense at Next '26, the next evolution of reCAPTCHA as a trust platform for the agentic web with agentic activity measurement, a policy engine, and AI-resistant QR challenges (existing reCAPTCHA customers automatically upgraded). HN commenters flagged that browsing may require a modern Android with Google Play Services or iPhone/iPad due to device attestation.
• Apple's "Spatial iPhone" with a holographic display is reportedly in early development per leaker "Schrödinger" and supply-chain sources, with Samsung supplying the panels (no confirmed details yet, second post-Vision-Pro AI hardware bet after the camera-equipped AirPods).
• Chrome quietly removed the claim that its on-device AI doesn't send data to Google servers in version 147.0.7727.138, a small but notable trust signal as Google's on-device AI features expand.
• Motherboard sales collapsed by more than 25% as chipmakers prioritize AI chips over enthusiast PCs; Asus is projected to sell 5M fewer boards in 2025 with Gigabyte, MSI, and ASRock all seeing reduced shipments, the clearest hardware-economy signal yet of AI's diversion of silicon supply from consumer PCs.
• Kevin O'Leary's 40,000-acre Utah data center (2.5× the size of Manhattan) was approved by a Utah county commission despite hundreds of locals showing up to oppose it over water scarcity, grid strain, doubled greenhouse emissions, and Great Salt Lake damage.

💼 AI Productivity, Labor & Economics

• DeepL announced ~25% staff cuts: CEO Jarek Kutylowski attributes the cuts to AI's "massive structural shift" in translation.
• Brian Albrecht ("You are not a horse," endorsed by Alex Imas, 220 likes) argues AI will not collapse human labor demand the way tractors did to horses: scale effects from cheaper goods, redirection of spending to human-intensive services via income effects, comparative advantage, task reorganization, and revealed preference for human-made goods make it implausible that every dollar of consumer spending contains zero embodied human labor.
• WIRED found thousands of "vibe-coded" apps leak sensitive data (Axios): apps built on Lovable, Replit, Base44, and Netlify default to public exposure without backend hardening, exposing corporate, financial, medical, and personal data.
• Self-driving tech is finding a second act after autonomous-vehicle hype stalled: lidar, 3D radar, and ground-penetrating radar built for AVs now power shipyard crane automation (Port of Rotterdam), wind-turbine monitoring, smart-city traffic, elder care, and robotics.

🤖 AI Agents & Infrastructure

• Legora launched Legora aOS, a purpose-built agentic operating system that lets legal teams execute end-to-end legal work with institutional knowledge, jurisdiction-aware research, and deep integrations, competing directly with Harvey.
• Google's Gemini Agent is becoming a "24/7 digital partner" per APK teardown: actions across Connected Apps with task completion, in-progress tracking, and scheduled tasks; Business Insider separately reported Google employees are testing an internal "Remy" agent.
• RampLabs built Fast Ask, an RL post-trained Qwen subagent for spreadsheet data retrieval that beats larger models at lower latency using synthetic data and minimal tools.
• Vellum (open-source repo) is a personal intelligence assistant that knows you deeply, evolves alongside you, and "belongs to no one else", structured memory across macOS, Telegram, and Slack, develops personality via SOUL.md and journaling.
• Open Generative UI by CopilotKit (GitHub, post) prompts any agent to generate interactive charts, visualizations, diagrams, 3D models, and rich UI inside the chat in a sandboxed iframe, works with LangGraph, CrewAI, Mastra, etc. via the AG-UI protocol; ships with a standalone MCP server.
• Yutori Navigator n1.5 (Dhruv Batra) upgrades the team's web agent with hybrid vision + DOM interaction, JavaScript execution, and structured outputs, claiming Pareto-dominating accuracy, latency, and cost.
• Parallel Web Systems hit GA on its Monitor API (quickstart, post) with new processor tiers, snapshot/event-stream modes, and Basis citations on every web change event so agents get pushed structured updates instead of polling.
• GigaAI runs a parallel reasoning-model detector on every completed text chunk while TTS streams audio in real time, dropping voice-agent hallucination rates from 4-5% to under 1% across 1.2M live turns by injecting one-turn correction hints that vanish after delivery.
• Sergio Paniego demoed OpenEnv (HF Space, GitHub), a ready-to-deploy reinforcement learning environment that runs on free HF Spaces, lets you write a short Python script with parallel LLM calls, and trains GRPO so the model writes its own search strategies.
• "Coordination as an Architectural Layer" for LLM-Based Multi-Agent Systems argues coordination between agents should be a dedicated architectural layer rather than ad-hoc orchestration code, a principled framing for multi-agent design.
• agent-harness-kit (HN) lets you define multi-agent workflows in TypeScript and automatically get SQLite state, MCP tools, health checks, and coordination rules, billed as "the Vite of AI agent orchestration."
• tilde.run (HN Show) lets you run AI agents and pipelines on real production data safely by turning every execution into a reversible transaction inside a versioned, composable filesystem sandbox (GitHub + S3 + Drive) with full audit trails and one-command rollback.
• Airbyte Agents gives your AI agents a unified pre-indexed Context Store across all your data sources (Salesforce, Zendesk, Jira) so they reason cross-system with fewer API calls and more accurate context (walkthrough).

💻 AI Coding & Developer Tools

• Code with Claude SF Extended Day runs today, a second day for indie developers and early-stage founders with founder stories, builder deep-dives, and laptops-open Applied AI workshops, added because Day 1 demand far exceeded capacity.
• Claude Code v2.1.126–v2.1.131 shipped between May 1 and May 6, added CLAUDE_CODE_SESSION_ID to bash subprocesses, CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN=1 to opt out of fullscreen, paste-progress footer, fixed graceful shutdown on external SIGINT, plus fixes for resume, permissions, scrolling, MCP, login, and model status.
• Peter Steinberger demonstrated GPT-5.5 + /goal in Codex completing an extensive multi-hour code refactor with end-to-end tests, a concrete case for where the persistent-goal workflow shines.
• Peter Steinberger pair-programmed his way through 10 single-binary local-first Go/TypeScript CLIs in one week, all designed for terminals, scripts, and coding agents:
• • sonoscli discovers, groups, and controls Sonos speakers over your local network with JSON output for scripts.
  • wacli pairs as a WhatsApp Web device and mirrors your full message history into local SQLite with FTS5 search (FTS5 = full-text search built into SQLite, like a search engine over your messages) so agents can read and send.
  • birdclaw imports your Twitter archive into a local-first workspace with cached live reads, AI-ranked triage, and reply flows.
  • gitcrawl mirrors GitHub issues + PRs into local SQLite with semantic clustering and a drop-in gh caching shim so agents stop burning API rate limits.
  • discrawl mirrors Discord guilds (or imports your Desktop cache) into searchable local SQLite with FTS5 + optional embeddings for offline analysis.
  • spogo is a Spotify power-CLI for search, playback, library, playlists, and devices using your existing browser cookies.
  • imsg reads your macOS Messages.app database, streams new iMessage/SMS rows live, and sends text/files via JSON-RPC.
  • mcporter auto-discovers any MCP server (MCP = Model Context Protocol, Anthropic's open standard for connecting agents to tools and data) and generates typed CLIs or TypeScript clients with zero boilerplate.
  • oracle bundles your prompt + repo files and sends them to "mythical pro agents" (GPT-5.5 Pro, Gemini 3 Pro, Claude Opus, etc.) via CLI, MCP, browser, or API with full repo context.
  • sag is a modern macOS say replacement that streams ElevenLabs voices to your speakers (ElevenLabs API key required).
• Cognition released SWE-1.6 Fast inside the Devin terminal running at 1,000 tokens/sec on Cerebras, with a free Max month for the first 100 users, Cognition's first model offered to developers via inference partner rather than only as part of Devin.
• François Chollet listed his top agentic coding use cases (142 likes): adhoc data visualizations for quantitative questions, adhoc data annotation UIs (the old "make your own dataset" bottleneck), and adhoc visual CLIs for existing code.
• OpenAI CLI shipped, a first-party command-line client for the API, replacing community wrappers as the canonical entry point.
• The May 7 Codex CLI changelog (@Codex_Changelog) also documents CLI improvements, better TUI, plugin management, hooks.
• Abacus.AI added automatic model routing for coding tasks, Easy → Kimi 2.6 Thinking, Medium → Opus 4.6, High → GPT-5.5 xHigh; the platform picks the right model based on difficulty rather than locking you into one.
• walkinglabs open-sourced "hands-on-modern-rl", a complete curriculum bridging basic RL concepts to LLM alignment, RLVR, and advanced agentic systems.
• n8n's Production AI Playbook by Elvis Saravia (@n8n_io) shows you how to fight silent drift in live AI workflows with continuous evaluation: built-in metrics, LLM-as-a-Judge scoring, structural checks, tool-use validation, and production monitoring that samples real traffic and triggers alerts before quality drops.
• Matt Pocock built a /handoff skill (SKILL.md) that compacts a coding session into a clean Markdown handoff document (context, goals, artifacts, suggested next skills) so a fresh agent or human can pick up exactly where you left off without context loss.
• Matt Pocock argues that low-fidelity prototypes beat walls of specs when working with AI coding agents, twenty years of software-industry knowledge confirms this.
• CJ Zafir shared a Codex 5.5 prompt loop that exploits the model's self-awareness: "Are you 100% confident in this strategy? If not, find all possible loopholes, suggest proper fixes, and run this loop until you are factually 100% confident", claims 2-3 cycles patches strategy weaknesses where insecure models like Opus 4.7 just agree.
• Codetaur is crowdsourcing simple ideas Claude Code / Codex can't build and turning them into open-source artifacts with full prompts/process shared (current example: a time-helix visualization built from a sketch).
• DAIR.AI Academy hosted "From LLM Wikis to LLM Artifacts" (Elvis Saravia), a live session walking through a framework to build visual LLM artifacts plus new releases for Pro members.
• Unsloth (HN) collaborated with NVIDIA to make LLM fine-tuning ~20-25% faster with cached packed-sequence metadata, double-buffered async gradient checkpointing, and MoE routing improvements with no accuracy loss.
• Kstack (HN) is a Claude Code skill pack that packages common Kubernetes monitoring, troubleshooting, and auditing tasks (/investigate, /audit-security, /audit-outdated) using kubectl, Kubetail, and Trivy for faster cluster debugging.
• Stage CLI (HN Show) is a local open-source viewer that organizes AI-generated code changes into small individual chapters for easier pre-PR review, works with any AI agent.
• agent-skills-eval (HN Show) is a test runner that objectively measures whether adding an agentskills.io-style skill actually improves AI agent output by running side-by-side evaluations and using a judge model to score the difference.
• php-fts (HN Show) is a self-contained full-text search engine written entirely in pure PHP with zero extensions or dependencies (trigram indexing, BM25 ranking, filtering, and highlighting), ready for shared hosting or small sites.

🔬 AI Research & Models

• Tilde Research released Aurora, which beats the previous NanoGPT Track 3 SOTA by 50 steps after only 3,175 training steps and continues scaling.
• ypwang61 used Claude and Codex to scale autoresearch and beat a 32-year-old Ramsey number lower bound (GitHub): R(3,17) ≥ 93 (previously 92) and R(4,15) ≥ 160, concrete proof LLM-driven combinatorial search can crack open small-but-hard mathematical bounds.
• Archiki Prasad (paper) argues effective reasoning chains reduce intrinsic dimensionality: the better a chain-of-thought is, the lower the dimensional manifold the task collapses onto, which improves generalization (ICML 2026 spotlight).
• Stella Lisy published EvoLM (paper), Self-Evolving Language Models through Co-Evolved Discriminative Rubrics: models that evolve via co-evolved rubrics rather than fixed reward functions.
• Stanford HAI's James Zou (@StanfordHAI) on AI's growing role as scientific peer reviewer: AI excels at spotting gaps in submitted papers, but final judgment calls still need humans.
• The SWE-bench Verified leaderboard now features ~50 models, open-source entries (DeepSeek, Kimi, Xiaomi MiMo, Z.ai) dominate the top 5 spots, a reference data point for "open-weights are within a model generation of frontier" (SWE-Bench Verified = the standard test for whether a model can fix real GitHub bugs).
• Epoch AI launched an AI Chip Components explorer that tracks supply and consumption of logic, packaging, and memory to spot bottlenecks in frontier AI compute production.
• Arena.ai (leaderboard, 111 likes) published a three-year analysis of open-source vs. proprietary models: proprietary lead in Text Arena shrunk from +250 to ~+30 points; Gemma-4 sits at #2 overall and #4 open-source in Vision Arena.
• Adithya Iyer + Morphic released Reshoot-Anything (arXiv, project, GitHub), a CVPRW 2026 self-supervised video-reshooting framework that trains from monocular videos by sampling random moving crops as pseudo multi-view triplets, achieves SOTA camera control and temporal consistency on dynamic scenes.
• Tianhong Li & Kaiming He released "Back to Basics: Let Denoising Generative Models Denoise" (Jason Ramapuram), arguing direct clean-data prediction (instead of noise prediction) lets simple large-patch "Just image Transformers" hit competitive ImageNet results at 256/512 resolution with no tokenizer, no pretraining, and no extra losses.
• Jason Ramapuram (DeepMind) + collaborators published "Scaling Properties of Continuous Diffusion Spoken Language Models," scaling tokenization-free continuous diffusion spoken LMs to 16B parameters on tens of millions of hours of raw speech with new scaling laws and the Phoneme Jensen-Shannon Divergence (pJSD) metric.
• Z.ai released GLM-5V-Turbo (paper), a native foundation model for multimodal agents with a CogViT vision encoder (dual-teacher distillation), multimodal training, RL at scale, and integration with Claude Code and OpenClaw for visual tool use, multimodal coding, and deep research with image search/crop/annotate.
• Xihui Liu's HKU team built PhysForge (arXiv, project, HF), a two-stage framework where a vision-language model plans physical blueprints from a single image and a diffusion model synthesizes high-fidelity geometry plus kinematic parameters for physics-grounded 3D assets.
• ModelScope added Instruction-Oriented Object Detection to Qwen3.6-35B-A3B (model, post), letting the model use LLM reasoning for fine-grained, small, occluded, or dense object detection via natural-language instructions (ODinW score 42.6 → 50.8).
• Sapunov breakdown of "Learning is Forgetting" (substack, arXiv), Cohere & Princeton found small models (1B) expand representations but fail to shed irrelevant noise, while 32B+ models compress toward the information-bottleneck optimum (information bottleneck = the optimal trade-off between compressing data and keeping what matters); a soft-entropy estimator predicts downstream performance (r=0.52) and instruction-following (r=0.76).
• Zhijing Jin and collaborators released Stargazer (arXiv, demo), a benchmark forcing AI agents to do real scientific model-fitting on noisy radial-velocity data; frontier agents hit >70% statistical fit but only <6% correct physical recovery.
• Tanush Yadav + RAIVN Lab released VideoNet (data, arXiv, GitHub, post), a CVPR 2026 highlight: large-scale dataset for domain-specific action recognition across 37 domains.
• NAVER LABS Europe released Anny-Fit (Anny base model, post), an all-age human mesh recovery model + the open-source Anny parametric 3D human model covering infants to elders with interpretable controls.
• LA-Pose is a feed-forward pose estimator that converts self-supervised latent actions learned from large-scale driving video into accurate, generalizable camera motion (Jamie Shotton).
• VLMaxxing through FrameMogging is a training-free anti-recomputation technique for video vision-language models, fast same-video follow-up queries with no measurable accuracy drift.
• The Geometry of Consolidation (@anirudhbv_ce) is a NeurIPS 2026 paper following HIDE and No-Escape; the author argues only 91/3072 dimensions in OpenAI's text-embedding-3-large and 80/3072 in gemini-embedding-001 do real semantic work, meaning ~97% of your vector database is mathematically empty (vector database = a database that stores text/images as numerical vectors so AI can find similar items by closeness in space).
• Modular shipped Mojo 1.0 beta with MAX 26.3 (site), 2-3× faster matrix multiplication on CPU/GPU plus full Python compatibility for production AI workloads.
• NVIDIA cuda-oxide (post) is an experimental Rust-to-CUDA compiler that turns idiomatic Rust into PTX (NVIDIA's GPU assembly language) with no DSLs or foreign bindings.
• ProgramBench tests whether language models can rebuild entire programs from scratch given only the executable binary and documentation; across 200 tasks ranging from simple CLIs to full interpreters like PHP, the best current models still fail to completely solve any single task and default to monolithic single-file implementations.
• "A Theory of Deep Learning" (HN) proposes that deep learning works because overparameterized networks create a signal-channel / noise-reservoir split via the empirical Neural Tangent Kernel, unifying grokking, double descent, benign overfitting, and implicit bias in one framework (HN commenters note the "Cumulative Dissipation Gramian" maps to the Observability Gramian from control theory and Hankel singular value truncation).

🤖 Robotics & Embodied AI

• NVIDIA's Zhengyi Luo open-sourced GR00T Whole-Body Control (VLA tutorial, post), a unified platform for humanoid controllers including the decoupled WBC models used in NVIDIA Isaac-Gr00t, Gr00t N1.5/N1.6, GEAR-SONIC, MotionBricks latent motion generation, and a full vision-language-action workflow for collecting VR teleop data on the Unitree G1, fine-tuning Gr00t N1.7, and deploying autonomous policies (VLA = vision-language-action model that maps what a robot sees + a task description to motor commands).
• David Dobáš and team launched Ultimate Bots Studio, a browser-based motion-capture and simulation editor for humanoids that lets you edit, remix, simulate (with GEAR-SONIC policy eval), and share motions instantly.
• Genesis AI demoed its first foundational robotics model GENE-26.5 (blog), the Khosla-backed startup ($105M seed) showed human-like robotic hands cooking, playing piano, and solving Rubik's cubes two-handed.
• Solo Tech's blog on embodied AI and physical AI covers the Sonic AGIBot X2 sim-to-real architecture, Solo CLI + OpenClaw integration for hardware-to-skills control, Solo Seven global humanoid launches, and CES 2026 wins.
• Rémi Fabre built an open-source browser-based face controller for Reachy Mini that maps webcam facial blendshapes and head pose to robot movement (fully local, includes simulator), expressive non-uncanny human-robot interaction.
• NTU MARS released "World Model for Robot Learning: A Comprehensive Survey" (project, Oier Mees), a policy-centric review of how predictive world models now power policy learning, planning, simulation, evaluation, and data generation in robotics.
• Comma.ai launched a lossy video compression challenge (post) with a live leaderboard to push efficient self-driving video encoding.
• South Korea's robot monk Gabi (built on the Unitree G1 platform with Buddharoid AI from Kyoto University researchers) took Buddhist vows at Jogyesa Temple in Seoul, becoming the country's first fully autonomous robot trained on Buddhist scriptures for spiritual counseling and rituals (reaction thread).

🏛️ AI Policy, Governance & Safety

• Mozilla's Firefox security campaign with Claude Mythos (also covered in lead) is the first major confirmed production deployment of Anthropic's restricted Mythos model, happening exactly as the White House drafts a vetting executive order around it (Politico frames the broader debate).
• OpenAI launched GPT-5.5 and GPT-5.5-Cyber with expanded Trusted Access for Cyber, more permissive capabilities for verified defenders doing vulnerability research, malware analysis, red teaming, and critical-infrastructure protection, with active classifiers blocking misuse.
• Trusted Contact in ChatGPT launched from OpenAI, an optional safety feature that notifies a person you trust if serious self-harm concerns are detected, opt-in only.
• Andrew Curran argues the Trump administration will almost certainly regulate open-weight models (318 likes): failing to do so would defeat closed-model regulation, destroy big-lab business models, and hand the lead to China.
• Nathan Lambert pushes back on the term "distillation attacks," arguing the language wrongly stigmatizes a standard and essential AI technique and risks harmful policy that would hurt open-weight models and Western research more than it would solve.
• Internal Tech Emails published leaked Sam Altman ↔ Mira Murati texts from November 19, 2023 (13K likes, 695 reposts) discussing the OpenAI board crisis and interim CEO Emmett Shear, surfacing fresh in the Musk v. OpenAI litigation discovery flow.
• Simon Willison critiqued the xAI/Anthropic Colossus 1 deal, flagging Colossus 1's gas-turbine air-quality record (turbines initially ran without Clean Air Act permits) and a "harm humanity" reclaim clause apparently controlled by Elon himself (Elon's post framed it as fair-terms compute leasing for AI companies "taking the right steps").
• Hantavirus Atlas is a Perplexity-built interactive map of hantavirus reports, clusters, and monitoring across the US sourced from public health agencies and major news outlets, an example of the "AI builds a public-health dashboard in a day" pattern Perplexity has been pushing.
• Pentagon CTO publicly demonstrated Palantir's Maven system live, showing real-time military operations and targeting using satellite and multi-source data feeds, the first public-facing walkthrough of the targeting AI that's been operational since 2017.

🛠️ AI Tools & Products

• OpenRouter added native audio endpoints for text-to-speech and speech-to-text using the same routing, billing, and keys you already use for text/image/video.
• Recraft V4 went live on OpenRouter (post, followup), high-aesthetics, art-directed image generation with intentional lighting and color (V4 ~1K resolution in ~10s; V4 Pro ~2K).
• SpAItial (post) turns a single image plus a text description into a navigable photoreal 3D Gaussian splat world (a 3D representation built from clouds of fuzzy colored points instead of polygons), edit, refine, and export to your 3D pipeline.
• Notion AI Transformation Assessment (post) is a five-minute quiz scoring how AI-native your company is across four maturity levels with peer benchmarks and a downloadable transformation roadmap.
• Talkie 1930 (model) replies in the exact tone and style of pre-1931 English writing, a fun showcase of modern instruction-following pinned to a narrow stylistic distribution.
• Spotify launched "Save to Spotify" (Verge), a CLI command that lets AI agents using OpenClaw and Claude save AI-generated podcasts directly to your personal Spotify library so you can listen to them on any of the 2,000+ Spotify devices alongside your other shows.
• RaTeX delivers TeX-quality math layout from a Rust core aligned with KaTeX golden tests, with ready-to-use packages for Web (WASM), iOS, Android, Flutter, and React Native, free to try.
• NeuralDerp built a deadpan AI-generated Pawn Stars parody where Boba Fett tries to pawn Han Solo frozen in carbonite ("This guy belongs in a museum, and for God's sake, can we get him a joint? The last thing he remembers is Lucas talking about world building and sequels"); a small showcase of how generative video tools are getting good enough for full satirical sketches with cutaways and timing.
• Templatical (HN Show) is an open-source drag-and-drop email builder (alternative to Beefree/Unlayer) that lets you build, preview, and export templates directly in the browser playground, free to try.

📊 Fundraising & Deals Roundup

• Periodic Labs: raising $500M at a $7.5B valuation (sixfold valuation jump in eight months), former OpenAI researcher Liam Fedus building AI-driven robotic labs for physics and chemistry breakthroughs.
• Moonshot AI: ~$2B raise at $20B+ post-money valuation led by Meituan; Kimi chatbot ARR over $200M in April.
• Enter: $100M led by Founders Fund at a $1.2B valuation (Natasha Mascarenhas/Bloomberg), São Paulo-based AI legal startup handling litigation for clients including Airbnb.
• Nova Intelligence: $31.5M Series A led by Chemistry with Accel, Conviction, and SAP.io (Forbes pegs total at $60M); founded by a 22-year-old former DeepMind engineer for SAP's $89B legacy-code modernization wave.
• IREN × Nvidia: up to 5 GW of DSX AI infrastructure deployment, with Nvidia receiving a 5-year option for 30 million IREN shares at $70 (potential $2.1B investment).
• Pit: $16M led by Andreessen Horowitz, with Lakestar, OpenAI/Anthropic/Google/Revolut execs, and Sweden's Stena and Lundin families.

🎙️ Interviews, Panels & Podcasts

• Anthropic's Jack Clark with Axios's Mike Allen (@haider1, 172 likes): Clark predicts that by 2028 it is more likely than not AI systems will autonomously improve and build better versions of themselves; he's urging early preparation for both intelligence-explosion risks and the abundance scenarios on the other side of recursive self-improvement.

💡 Industry Commentary & Analysis

• Ethan Mollick notes how quickly the AI market has consolidated around OpenAI and Anthropic (294 likes): the two labs now account for roughly half the cloud backlog at Microsoft, Oracle, Google, and Amazon, leaving Google (and possibly Meta) as the only realistic competition.
• Nathan Lambert shared notes from inside China's AI labs: humility, student integration as full contributors, team-first culture, and practical "build-not-buy" focus make Chinese teams exceptionally effective fast-followers who meticulously optimize every layer of the stack.
• Ethan Mollick shared his own notes from China's AI labs separately, complementing Lambert's writeup with on-the-ground impressions of Chinese frontier labs.
• Mathematica highlighted Terence Tao's Oxford lecture (2K likes, 510 reposts): Tao argues we've built powerful but unreliable tools optimized for plausibility rather than veracity, creating a "convincing mirror" that confidently serves dangerous advice in domains like medicine and finance.
• Paul Tudor Jones told CNBC the AI bull market has "another year or two to run," comparing Claude January 2026 to MSFT 1981 and the current cycle to 1999.
• Brian Halligan (HubSpot co-founder) argues founders should master strategic illegibility (355 likes): making your company "legible" to AI by structuring all data and workflows for agents commoditizes your moat; proprietary judgment, taste, informal power maps, and contrarian beliefs lose their edge once they become vendor features.
• John Carmack argues semiconductor fabrication is stuck in pre-SpaceX-launch mode (913 likes): "copy exactly" conservatism in fabs (and possibly nuclear power) prevents the experimentation needed for breakthroughs.
• François Fleuret argues the path to "finishing" LLMs is just three things: (1) latent-space diffusion-like reasoning, (2) a real recurrent state, and (3) world-model pre-pre-training. "And we are done." (466 likes).
• thdxr (opencode) argues you should focus 99% of your attention on the simplest users (who never change their behavior) and the 0.01% "alien" users showing you the distant future, and totally ignore the loud "pro" users who invent elaborate weekly workflows but ship less than everyone else.
• thdxr also argues he never writes markdown plans for AI coding agents anymore, he just has the agent make code changes, reviews what sucks in the output, and re-prompts.
• signulll argues we're in a fundamental transition where apps become agents and the unit of distribution itself is changing, rewiring discovery, trust, monetization, brand, and what it means to "ship" a product.
• Steve Newman's "Is AI 2027 Coming True?" (thread), first of his Q2 2026 series. Leans toward "normal technology" for a few more years but his gut is wavering as models start accelerating their own development.
• Mitchell Hashimoto argues AI slop is actually good for fast parallel experimentation when you understand its boundaries: ship transparent alpha-quality internals/plugins, clean up later. The cost of change is now just tokens (1.9K likes).
• Demian (@demian_ai) argues ~100× cheaper inference has paradoxically driven exploding compute bills because agentic + reasoning workflows now consume 10,000× more tokens, Jevons paradox at trillion-dollar scale (1.4K likes).
• BoringBiz_ argues AI productivity gains (e.g. making a $300K engineer 20% more productive for $20K) should increase hiring incentives by raising ROI on engineers; Curtis counters that most companies have fixed engineering output demand, so the same roadmap now requires fewer people.
• dr. jack morris argues the real bottleneck for long-context models isn't architecture, it's data: humans haven't generated enough contiguous high-quality long-form content to train on.
• Taco Cohen argues that LLM hallucinations aren't evidence we're approaching intelligence in a fundamentally flawed way, toddlers confidently produce plausible-sounding but wrong answers from sparse data, and the anti-confabulation habits we acquire later (uncertainty, evidence, logic) are culturally learned overlays, not the base architecture.
• Benjamin Bratton notes his students are shocked the entire internet, kernels, protocols, browsers, databases, was hand-coded and debugged by humans (4.9K likes).
• Daniel Miessler argues AI lacks true creativity because it has no intrinsic drives, no fears or subjective experience like human survival/reproduction instincts, so it only emulates.
• Nick Cammarata argues AI agents enable a zen-like work pattern (209 likes): meditate 15 minutes of every hour while agents churn, then flow in awake awareness for the rest, enabling 4 hours of formal practice daily and 18-hour days.
• Pliny got permabanned from ChatGPT Pro for red-teaming GPT-5.5 overnight (1K+ likes): the prolific jailbreak researcher's Pliny Agent ran browser/computer use against GPT-5.5; OpenAI permabanned and denied appeal.
• Lukas Ziegler's "we all are robots" Substack is a robotics-focused newsletter pulling insights from leaders at Google, Tesla, Boston Dynamics, and others.
• Ryan Carson runs daily automated end-to-end smoke tests of his product's full 28-step / 99-minute customer onboarding flow using DevinAI for ~$33/day, fresh accounts, every tool, video recordings, PASS/FAIL email reports, uncovering edge-case bugs weekly.
• Trevin Chow outlined 10 principles for agent-friendly CLIs (table-stakes: non-interactive defaults, structured JSON output, actionable errors; advanced: consistent vocabulary, async --wait support, persistent profiles), drawing from Cloudflare and HeyGen examples to design for agents first (126 likes, 18 reposts).
• No One's Happy (HN) argues AI can produce expert-looking work without expertise, creating two failure modes that are reshaping the workplace: elongated slop-filled artifacts where length and formatting no longer signal care, and undetected bad work in unfamiliar domains that erodes judgment.
• Simon Willison argues vibe coding and professional agentic engineering are converging faster than he'd like because AI tools have become reliable enough for production use with only light review, risking normalization of low-effort output and erosion of traditional quality signals.
• crawshaw argues AI agents break the classic review-then-commit process by creating a principal-agent problem where low-effort machine-generated code overwhelms reviewers and destroys the trust signals that let low-trust groups collaborate at scale.
• Jack Clark writes in Import AI 455 that AI systems are about to start building themselves, with every piece now in place for fully automated AI R&D and a ~60% chance of autonomous AI research (including training successors) by the end of 2028 (companion to his Axios podcast covered above).
• The Typical Set (HN) argues "the bottleneck was never the code"; while agents make writing code cheap, they expose and amplify deeper organizational limits around specifications, shared context, focus, and human collaboration that remain the real constraint.
• Ars Technica argues social media is dying and splintering into messier fragmented spaces (private chats, algorithmic video, protected communities, AI interactions) that risk repeating the same toxicity dynamics unless new designs deliberately promote diversity and tolerance.
• Predrag documents how a family TV show night turned into systems debugging night when a self-cancelling streaming subscription hit an async race condition between account-linking and account-unlinking across bank and service systems, a vivid case study in how modern subscription plumbing creates fragile races that turn simple consumer actions into multi-system debugging exercises.
• Boris Cherny's 2004 TI-83 Plus BASIC tutorial (HN) resurfaced on Hacker News after Boris definitely not intentionally oh no doxxed himself as having a super cringe old tutorial still up online somewhere (so of course the sleuths came out in force to find it). HN commenters reminiscing about the calculator as the portable computer they were actually allowed to use in class, a generational signal as today's coding-agent natives discover the hand-rolled programming culture that shaped much of the current frontier-AI cohort.

Previous Around the Horn Digests

Catch up on everything you missed:

• Wednesday, May 6, 2026: Yesterday's full digest; go check it out!
• Tuesday, May 5, 2026: A Cape Breton fiddler sued Google for $1.5M over an AI Overview defamation, OpenAI considered an Alphabet-style spinout for robotics + hardware, and Anthropic shipped keyless authentication for the Claude API.
• Monday, May 4, 2026: The White House began considering pre-release vetting of AI models, Anthropic and OpenAI both married private equity on the same day, and Mayo Clinic's AI flagged pancreatic cancer up to three years before diagnosis.
• Weekend, May 2-3, 2026: The Pentagon signed AI deployment deals with 8 vendors and left Anthropic out, Microsoft 365 E7 with Agent 365 went GA, Meta bought a humanoid robotics startup, and Grok 4.3 added voice cloning.
• Thursday, April 30, 2026: OpenAI shipped GPT-5.5-Cyber, AISI graded it as one of the strongest cyber models ever, and Anthropic was pre-empted into a $900B round.
• Friday, April 24, 2026: DeepSeek shipped V4 (and open-sourced it) the same morning the State Department warned allies about DeepSeek IP theft; Google quietly committed up to $40B to Anthropic.
• Monday, April 13, 2026: Stanford's 2026 AI Index quantified the canyon between AI insiders and the public, and an AI signed a 3-year retail lease in San Francisco.

Monthly skill digests: AI Skill Digest, April Week 1 | AI Skill, March Part 3 | AI Skill, March Part 2

That's a Wrap

That's 100+ stories from one Thursday. If you scrolled all the way to the bottom, you now know enough about Anthropic's day to apply for a job there, and about Cloudflare's day to be glad you don't already have one.

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.