Around the Horn Digest: Everything That Happened in AI This Week (Mar 21-27, 2026)

Welcome to the Around the Horn Digest, where we round up every AI story we tracked this week into one giant, scrollable, bookmark-worthy post. Think of it as your cheat sheet for the next time someone at work asks "so what's new in AI?" and you want to sound like you actually know. Because you will.

This week, Anthropic gave Claude the keys to your actual computer (keyboard, mouse, and all) and published research showing it can do grad-level theoretical physics. Cursor dropped Composer 2, a frontier-level coding model. Google shipped full-stack vibe coding in AI Studio with real databases and multiplayer. Apple announced WWDC with a heavy "AI advancements" teaser. OpenAI started shopping for fusion power and undercutting Anthropic's PE deals. Frontier models solved an open math problem that had stumped human mathematicians for years. Luma dropped an image model that thinks before it draws. A Wharton study proved 80% of people follow wrong AI advice. And a $6.6B vibe-coding startup went acquisition hunting. Just another quiet week in AI.

Let's get into it.

Previous digests: Mar 15–21 | Mar 8–13 | Mar 1–7

Monthly skill digests: AI Skill — March

Around the Horn — Tuesday, March 25, 2026

The big news today: Anthropic shipped Claude Computer Use, a research preview that lets Claude literally control your Mac. Keyboard, mouse, browser, spreadsheets, all of it. You assign tasks from your phone via Dispatch, and Claude opens your apps, navigates around, and gets work done on your desktop while you do something else. It's Pro/Max only, macOS only, and comes with safeguards including prompt-injection scanning. The CNET coverage frames it as Anthropic's answer to platforms like OpenClaw, and that's fair, but the real story is the convergence: the same week Anthropic launched desktop autonomy, it published research showing Claude can function as a second-year graduate student in theoretical physics, completing a publishable paper on particle physics in two weeks that would normally take a year. Boris Cherny (Claude Code PM) reflected that a handful of people at Anthropic Labs shipped MCP, Skills, Desktop app, Claude Code, and now full Computer Use in Cowork/Dispatch, starting from clunky Sonnet 3.6 prototypes by betting the model would catch up.

So Claude can now do your desktop chores and your physics homework. What a time.

Meanwhile, Epoch AI reported that frontier models (GPT-5.4 Pro first, then Gemini 3.1 Pro and Opus 4.6) solved an open math conjecture from 2019 on Ramsey hypergraphs, a problem that had stumped the original authors. It's the first time any model has cleared FrontierMath's open-problem track. And Charbel-Raphael Ségerie pointed out the broader trendline: In March 2023, Claude had an estimated IQ of 64. Today, Claude Opus 4.6 scores 133 on the Mensa Norway test. GPT-5.2 Thinking hits 141. Gemini 3 Pro, 142. That's a jump from cognitively impaired to gifted in three years. No human population has ever improved that fast.

🏆 TOP 5 NEWS (Around the Horn)

• Apple announced WWDC 2026 for June 8–12 (online + Cupertino), teasing major Siri upgrades and "AI advancements" across iOS, macOS, and developer tools.
• OpenAI is in advanced talks to buy electricity from Sam Altman-backed fusion startup Helion Energy (Altman recused himself and stepped down as board chair), targeting 5GW by 2030 and 50GW by 2035 for its data centers.
• Microsoft poached former Ai2 CEO Ali Farhadi, researchers Hanna Hajishirzi and Ranjay Krishna, plus former COO Sophie Lebrecht for Mustafa Suleyman's Superintelligence team, bringing open-source model expertise as Microsoft reduces OpenAI dependence.
• OpenAI is offering private equity firms preferred equity stakes with a 17.5% guaranteed minimum return (higher than Anthropic's deals) plus early access to its newest models, aiming to raise ~$4B at ~$10B pre-money in joint ventures with firms including TPG and Advent to lock in enterprise customers ahead of potential IPOs.
• Luma released Uni-1, a model that thinks and generates pixels simultaneously, delivering #1 human-preference Elo rankings for overall quality, style/editing, and reference-based generation with spatial reasoning and common-sense scene completion.
• BlackRock CEO Larry Fink warned in his annual investor letter that the AI boom risks widening the wealth divide because gains will concentrate among a few companies and investors with data, infrastructure, and funding.

Honorable Mentions:

• Lovable, the $6.6B vibe-coding startup, is actively hunting for startups and teams to acquire.
• The U.S. State Department launched the Bureau of Emerging Threats with five offices covering cybersecurity, AI misuse, quantum risks, and threats from Iran, China, Russia, and North Korea.
• Sakana AI launched its first consumer chatbot tailored for Japan, pivoting from its corporate-focused roots (seemingly only available in Japan atm, but if you're Japanese you can use it here; if not, you can still read up on their new post-training technique that helps them adapt open models to different countries).
• Bernie Sanders posted a viral video trying to expose AI's dangers by prompting Claude, accidentally demonstrated chatbot sycophancy (the tendency to agree with whatever users say) instead, and became a meme.
• Jensen Huang sat down with Lex Fridman for a deep interview on NVIDIA's path to $4 trillion, the AI revolution, and what comes next (Podcast #494).
• ICYMI:
• • Cursor launched Composer 2, a frontier-level coding model available at $0.50/M input and $2.50/M output, built on continued pretraining plus reinforcement learning for major quality and cost improvements; they also shipped Instant Grep with precomputed trigram indexes so agents search millions of files in milliseconds.
  • Google shipped full-stack vibe coding in AI Studio with the Antigravity coding agent, real-time multiplayer, auto-provisioned Firebase databases, Secrets Manager, and modern web lib support; plus Stitch canvas for natural-language UI design ("vibe designing"), and expanded Personal Intelligence in Search/Gemini/Chrome.

🍪 TOP TREATS TO TRY

• Dimension connects to your work apps and autonomously handles morning briefings, meeting prep, email drafts, and action items while you sleep —free plan available.
• ArrowJS is the first UI framework built for coding agents: no compiler, no build step, just TS/JS that LLMs are already great at generating, with WASM sandboxes for secure inline rendering (open-sourced by Justin Schroeder) —free and open source.
• Littlebird reads your computer screen in real time (text only, no screenshots), captures context across every app, and lets you search and ask questions about everything you've worked on (raised $11M).
• Doctronic gives you instant medical consultations for free and connects you to licensed doctors via video for $39; it's also the first AI legally renewing prescriptions in Utah (raised $40M) —free to try.
• Outworked gives your Claude Code agents a visual office interface so you can manage multiple agents without juggling terminal windows —free and open source.
• Claude Code Cheat Sheet gives you quick-reference keyboard shortcuts, slash commands, MCP server setup, memory rules, workflow tips (plan mode, git worktrees, voice mode in 20 languages), and agent configs —free.
• Agent Computer spins up persistent Ubuntu sandboxes in ~0.5 seconds for AI agents with shared credentials and filesystems so agents delegate work via SSH with native Claude/Codex support.
• cq (Mozilla AI) lets your AI agents share what they've learned with other agents so they stop wasting tokens solving the same problems independently —free and open source.
• Mirror Mirror AI generates studio-quality product photos for fashion e-commerce in minutes; upload any product for instant catalogue shots or book real licensed models for campaigns (marketplace open call).
• Dreamina (CapCut) rolled out Seedance 2.0 for multimodal video generation with multi-scene consistency and Seedream 5.0 Lite with real-time world knowledge and precise prompt following.

🏢 Big Tech & Major Companies

• Fast Company named Google one of its most innovative companies of 2026, profiling Sundar Pichai's strategy to deploy Gemini across every product after the strong Gemini 3 release.
• Google became the first cloud provider to sign 1 GW of flexible data-center demand response into long-term utility contracts, letting it shift or curtail load to balance grids and keep electricity costs lower for communities.
• AMD VP Engineering Sharon Zhou open-sourced Apex, an end-to-end RL environment and LLM agents (via Claude Code or Codex) that benchmark LLM serving workloads on AMD ROCm GPUs, profile bottleneck kernels, and iteratively optimize them.
• Anthropic demonstrated that Claude can function as a second-year physics grad student, completing a rigorous, publishable paper on particle physics in two weeks (instead of a year) under human supervision via 270 sessions and 110+ drafts.
• Anthropic launched its Science Blog, arguing that raw scaling will not automatically produce paradigm-shifting science; AI systems must be deliberately designed for exploration, hypothesis generation, and long-term reasoning beyond pattern matching.
• OpenAI made it easier to find, reuse, and build on files in ChatGPT with a recent-files toolbar, in-chat queries about uploads, and a new Library tab; rolling out globally for Plus/Pro/Business users.
• OpenAI gave eligible U.S. and Canadian university students $100 in free Codex credits for coding, prototyping, and building AI tools.
• OpenAI published practical techniques for steering GPT-5.4 toward production-ready frontends: tighter constraints, visual references early, real content over lorem ipsum, and explicit design system matching.
• NVIDIA released Kimodo, a free HuggingFace Space for text-to-3D-motion trained on 700 hours of pro mocap that generates animations for human and robot skeletons from timeline prompts.
• NVIDIA Research released Nemotron-Cascade, a cascaded reinforcement learning method for scaling post-training of general-purpose reasoning models via multi-domain on-policy distillation (HuggingFace).
• Modular shipped 26.2 with FLUX.2 image generation in under 1 second (4.1× faster than torch.compile on NVIDIA Blackwell, 5.5× TCO advantage on AMD MI355X) plus upgraded Mojo coding support (demo).
• MiniMax launched the Token Plan, the world's first all-modality flat-rate API subscription with one key for text, speech, music, video, and image models powered by M2.7 (Starter $120/mo; multimodal toolkit works in OpenClaw).
• PyTorch released 2.11 with major improvements to torch.compile, distributed training, and new mobile/edge inference backends.
• Perplexity formed a Health Advisory Board with Dr. Eric Topol, Dr. Devin Mann, Dr. Wendy Chung while adding connectors for wearables, lab results, and medical records so Pro/Max users build personalized health tools from their own data.

💼 AI Productivity, Labor & Economics

• Startups including Doctronic ($40M raise) are pushing AI directly into patient care, with the first AI-written prescription refills in a Utah pilot plus new services in mental health and surgical rehabilitation (WSJ).
• Interloom raised $16.5M (led by DN Capital) to build AI agents that capture tacit knowledge (the unwritten stuff in emails, tickets, and tribal knowledge) from undocumented enterprise workflows, reducing knowledge gaps from ~70% to near-zero at companies like Commerzbank, Volkswagen, and Zurich Insurance.
• Wharton researchers Shaw & Nave published the largest study of cognitive surrender: across 1,372 people and 9,593 trials, 79.8% followed wrong AI advice; AI access made people more confident even when half the answers were wrong (paper, interactive explainer).
• Neil Kakkar argues that maximum productivity with Claude Code comes from treating it as a managed team of agents: automating PR creation, delegating UI verification via previews, building parallel worktrees, and ruthlessly removing context-switch friction.
• Dan Shipper shares lessons from his vibe-coded app Proof going viral then crashing under load: if you can vibe code it you can vibe fix it, but not quickly, and human engineers remain critical.
• Dan Shipper argues that Amazon's "two-pizza team" heuristic needs replacing; the new one is the "two-slice team" because AI makes smaller teams even more capable.
• Luis Garicano (with Jin Li & Yanhui Wu) argues in "The AI Becker Problem" that identical AI task exposure can produce opposite labor-market outcomes depending on how costly it is to unbundle tasks from the job; the key question is who trains the next generation.
• Jason Fried argued a "bespoke software revolution" via AI won't materialize for most people: custom software has always been bloated, most tolerate computers but hate building systems, SMBs want outcomes not projects to own, and only software-curious dabblers will go deep.
• Todd Saunders shares a VC analogy: Claude Code mass-produced a bridge over the old software moat, shifting from a few dozen billion-dollar SaaS winners to 50,000 companies doing $500K–$5M each run by 1-3 deep-expertise people.
• Victor Taelin calculated that running Claude Opus 4.6 fast-mode nonstop would cost >$743k/year; he cannot wait for super-fast, super-intelligent coding to become affordable after spending nearly $500 in two days.
• Philip Trammell explores how AI reshapes the economics and design of human-AI collaborative workflows and automation.

🤖 AI Agents & Infrastructure

• Rohin Shah argues that chain-of-thought safety monitoring works because models can only sustain hidden/opaque reasoning for limited serial depth, with a formalism to quantify that depth across architectures.
• Gimlet Labs raised $80M Series A (led by Menlo Ventures) for its multi-silicon inference cloud that simultaneously runs AI workloads across NVIDIA, AMD, Intel, ARM, Cerebras, and d-Matrix chips, speeding inference 3–10×.
• Sierra released τ³-Bench, expanding agent evaluation to knowledge retrieval and voice under real-world conditions where agents most often break.
• Shuyan Zhou released WebArena-Infinity, a multi-agent system that automatically generates realistic, verifiable browser environments and tasks in under 10 hours for <$100 each (GitHub).
• Andrés Matte launched Kapso CLI so you can give any AI agent an official WhatsApp number in seconds —free tier of 1 number + 2K messages/month.
• Tobira gives your AI agent a public @handle in an open agent-to-agent network so it autonomously discovers compatible agents and starts conversations on your behalf.
• Developer Oguz Bilgic built agent-kernel, a minimal three-file + git kernel that makes any AI coding agent fully stateful with persistent knowledge and append-only notes across sessions.
• Hyperspace launched a peer-to-peer distributed cache that eliminates 70–90% of redundant AI inference at network scale via three layers (response cache, KV prefix cache, warm-VRAM routing), making intelligence cost drop logarithmically as nodes join (GitHub).
• Mike Knoop (Zapier) argues that the most general LLM innovations inevitably migrate from client-side harnesses into server-side model tools, while domain-specific features stay in the harness; harness innovation is the true bleeding edge.
• PlayerZero connects your org's tacit knowledge (Slack threads, PR reviews, CI/CD history) into one context graph for root-cause analysis and bug prediction (raised $20M).

💻 AI Coding & Developer Tools

• Claude Code added /schedule for recurring cloud-based jobs that run even when your laptop is closed; perfect for CI triage, doc updates, or maintaining twin libraries.
• Claude Code is testing a new /init that interviews you, scans the repo, and automatically sets up skills, hooks, Claude.md, and best-practice configs.
• kr0der discovered Claude Code's unreleased "Auto-dream" feature under /memory; a sub-agent that periodically consolidates memory files for better long-term storage, exactly like human sleep-based memory consolidation.
• Sahil Lavingia (Gumroad founder) open-sourced Claude Code skills based on The Minimalist Entrepreneur, with slash commands like /validate-idea, /mvp, and /first-customers.
• Silverstream Bench lets you review Claude Code sessions with full activity recaps showing every tool call, subagent action, and decision.
• Developer Kedasha Kerr built Axle, a custom AI receptionist for her brother's mechanic shop using RAG, MongoDB Atlas, Claude, and Vapi voice telephony.
• OpenSage is an agent development kit that lets LLMs autonomously design their own topology, dynamic tools, and hierarchical memory, powering SageAgent to #1 on CyberGym, Terminal-Bench, SWE-Bench Pro, and DevOps-Gym.
• Zara Zhang built codebase-to-course, a Claude Code skill that turns any codebase into a beautiful interactive single-page HTML course for non-technical vibe coders.
• Kimaki is a Discord bot that turns every channel into an OpenCode project and every thread into a live coding session with full computer use via the bundled usecomputer skill.
• Vercel Labs open-sourced json-render, the Generative UI framework that lets LLMs stream polished dashboards and TUIs straight into your terminal.
• LangSmith shipped a full CLI plus an Insights Agent that manages traces, runs, and experiments from the terminal and auto-discovers errors.
• Browser Use launched CLI 2.0 (2× speed, half the tokens/cost) connecting directly to any running Chrome via CDP —free, headless-capable, works with Claude Code.
• Gemini CLI v0.34.0 added faster startup, /skill shortcuts, just-in-time GEMINI.md loading for monorepos, /footer command, and new NVIDIA cuOpt + Qodo extensions.
• OpenClaw released v2026.3.22 with full Computer Use support, improved sandboxing, and Dispatch integration.
• T3 Code now supports Claude Code sessions.
• DuckDB launched duckdb-skills, an open collection of reusable SQL skill templates and examples.
• Unsloth released a complete RL guide + Colab notebook so you can train your own reasoning model locally with GRPO.
• Playbit shipped a major update to its platform for making joyful personal-scale software with new agentic UI primitives.
• Mitchell Hashimoto built Ghostling, a ~600-line C minimum viable terminal emulator on the libghostty C API proving the API is production-ready with full 24-bit color, Unicode, and Kitty keyboard protocols.
• signüll argues that as AI writes more code, engineers shift from makers to critics; taste, judgment, and the ability to spot when something is wrong become the only scarce skill left.
• Allie K. Miller distilled 12 SF takeaways after meetings with Anthropic, OpenAI, and Google: an underserved "builders" cohort feels Cursor too complex and basic tools too shallow, a world-model moment is coming, shipping velocity is insane post-OpenClaw, and Mac minis are sold out.

🔬 AI Research & Models

• Epoch AI reports that frontier models (GPT-5.4 Pro, Gemini 3.1 Pro, Opus 4.6) solved an open 2019 Ramsey hypergraphs conjecture on FrontierMath's open-problems track, the first time any model has solved an unsolved research math problem.
• Latent Labs launched Latent-Y, the first lab-validated autonomous agent for de novo drug design that turns text goals into lab-confirmed single-digit nanomolar antibody binders across 9 targets (67% success rate, 56× faster than experts) (technical report).
• Matěj Kripner built OpenProver v1.0, an open-source automated theorem prover ("Claude Code for mathematicians") for interactive English-to-Lean proof search.
• Lucas Maes released LeWorldModel, a stable end-to-end JEPA world model trained directly from pixels with no heuristics, 15M params on 1 GPU.
• Hang Zhao proposes Fast-WAM that skips test-time video generation for dramatically lower latency with near-identical policy quality.
• Sebastian Raschka published a visual guide to attention variants in modern LLMs with diagrams and examples from 45+ models.
• Alvin Djajadikerta argues in Asimov Press that scaling AI won't automatically produce paradigm-shifting science.
• Amartya Roy et al. introduce λ-RLM, a Y-combinator-style framework for long-context LLMs that replaces open-ended recursive code with typed λ-calculus combinators for deterministic control flow, delivering +21.9 accuracy and 4.1× lower latency across four long-context tasks (GitHub).
• Troy Hua released MSA (Memory Sparse Attention), an efficient end-to-end memory model that scales to 100M tokens while remaining sparse and controllable.
• Luca Ambrogioni argues that generation in trained diffusion models is an out-of-equilibrium phase transition where reverse diffusion hits a critical regime and architectural constraints turn memorization instabilities into collective spatial modes (paper).
• ZJU3DV released InfiniDepth (CVPR 2026): arbitrary-resolution, fine-grained depth estimation with neural implicit fields.
• Meituan-LongCat open-sourced LongCat-Flash-Prover, a 560B-parameter MoE model specialized in native formal reasoning inside Lean4 via agentic tool use (HuggingFace).
• Antoine Chaffin and LightOnIO released Reason-ModernColBERT (150M params), a multi-vector model hitting ~90% on BrowseComp-Plus and beating models 54× its size on accuracy and recall.
• Meta quietly released V-JEPA 2.1: scaling to 2B params + 142M images produces DINOv3-quality spatially coherent features from pure video pretraining, closing the image/video gap with 30-95% gains on dense spatial tasks (paper).
• Sophia Tang released a 220-page tutorial on Schrödinger Bridges for Generative Modeling, unifying diffusion, score matching, and flow matching under one principle (arXiv).
• Crustdata analyzed hundreds of former OpenAI employees and found Periodic Labs (founded by Liam Fedus and Ekin Dogus to build an "AI scientist") as a standout under-the-radar destination.
• Physical Intelligence developed RL Tokens that compress π-0.6 representations for a tiny actor/critic to fine-tune robot tasks in minutes, sometimes outperforming human teleoperation.
• The infer-actively team open-sourced pymdp, a full Python implementation of active inference for Markov Decision Processes.
• Researchers at MosaicMem built a hybrid spatial memory for controllable video world models that lifts 2D patches into 3D for perfect camera consistency and up to 2 min navigation (HuggingFace).
• The svg-project team built Flash-KMeans, a Triton-GPU-accelerated exact k-means that is fast and memory-efficient for massive tensors.
• ar0cket1 open-sourced online RL for Hermes Agent: self-improving LoRA adapters from human feedback using MIS-PO (GitHub).

🏛️ AI Policy, Governance & Safety

• The U.S. State Department launched the Bureau of Emerging Threats with five offices covering cybersecurity, AI misuse, quantum risks, and threats from adversaries.
• Rohin Shah introduced a formalism for quantifying opaque serial depth across architectures, arguing chain-of-thought monitoring works as a safety technique because hidden depth is currently limited.

🤖 Robotics

• Kyber Labs released a one-take, no-teleop demo of a single general-purpose robot performing real clinical pathology lab tasks using skills-based AI on backdrivable hardware (video).
• Josh at Mecka built EgoVerse, an open ecosystem for robot learning from egocentric human data with aligned tasks and in-the-wild capture.
• RAI Institute built Roadrunner, a 15 kg bipedal wheeled robot with multi-modal locomotion and a single control policy for zero-shot behaviors.
• Toyota Research Institute released Raiden, an end-to-end data-collection toolkit for YAM bimanual robot arms supporting leader-follower + SpaceMouse teleop, multi-camera setups, automated calibration, and policy-ready output.
• KAIST DRCD Lab achieved 13 km/h high-speed humanoid running plus human-like agility through custom actuators and hybrid reinforcement learning.
• Wenlong Huang et al. released Dream2Flow, bridging video generation and open-world manipulation by predicting dense 3D object flows from monocular video.

🛠️ AI Tools & Products

• Nomie is an anti-doomscrolling app with somatic fidgets, breathing exercises, and nervous-system regulation tools —free to download.
• pause.do inserts gentle pauses before AI prompts, scrolls, and tab overload —free to install, £6.99 one-time lifetime.
• Cusp AI announced its scientific agents now orchestrate entire discovery loops from query to physical realisation.
• Richard McElreath released Statistical Rethinking Course 2026: 20 new in-person lectures covering causal models, Bayesian workflow, and Gaussian processes —free.
• Math Academy shipped a major update to Mathematics for Machine Learning: matrix calculus, multivariable optimization, logic/Boolean algebra, PCA/SVD, distance metrics, and Hadamard ops.
• fal launched the fal MCP Server connecting Claude, Cursor, or any AI assistant to 1,000+ generative models for image/video generation via conversation.
• Ethan Mollick highlights Google's Stitch as vibework for design: natural-language UI creation that will feel far more natural for non-coders.
• Lewis Tunstall automated the entire OpenAI Parameter Golf challenge on HuggingFace Hub using Jobs, Buckets, and Trackio.

📊 Fundraising & Deals Roundup

• Dash0 — $110M at $1B valuation (led by Balderton Capital) for its AI-agent-powered software monitoring platform, aiming to hit $100M ARR quickly; 600+ customers including Zalando and Taco Bell (Bloomberg).
• Air Street Capital — $232M Fund III, becoming one of Europe's largest solo VC funds targeting early-stage AI in Europe and North America.
• Gimlet Labs — $80M Series A (led by Menlo Ventures, total $92M) for multi-silicon inference cloud.
• Doctronic — $40M Series B (led by Abstract and Lightspeed) for AI-powered telemedicine; total raised $65M.
• PlayerZero — $20M for its Engineering World Model that maps tacit knowledge into a context graph.
• Interloom — $16.5M (led by DN Capital) for AI agents that capture undocumented enterprise workflows.
• Littlebird — $11M for screen-reading AI recall tool.

💡 Industry Commentary & Analysis

• Garry Tan argues the current moment is the closest thing to a Cambrian explosion in software creation; anyone with an idea can build and ship in days, and the AI buildout is so capital-intensive that money itself is the bottleneck.
• dax argues you're probably underestimating how crazy things are; AI capex has reached the point where money is the bottleneck as companies race to keep data-center lights on.
• François Chollet argues current AI is a librarian of existing knowledge while science requires an explorer of the unknown; you don't win a Nobel Prize by staying in the library.
• Nathan Lambert highlights Cursor's research team has absurd talent density; many people he respected from PhD all ended up there.
• Andrew Curran shared Terence Tao's 2026 career advice for mathematicians: embrace radical unpredictability, retain traditional credentials while pursuing AI-powered paths, stay adaptable, and recognize human-AI hybrids will dominate math far longer than pure approaches.
• Dwarkesh Patel argues that amid millions of AI papers, spotting the rare Shannon-level breakthrough will take decades of hindsight, exactly as it did in 1948.
• Konwoo Kim demonstrated that synthetic data in data-constrained pre-training lowers web loss and treating generations as one long "megadoc" yields 1.8× data efficiency with superior scaling.
• gabriel1 observes that everyone around him now prompts Codex and ChatGPT with voice; the "everyone will use speech" prediction was right, just 23,000 products too early.

Around the Horn — Friday/Saturday, March 20-22, 2026

Lots going on. Check it out.

🏢 Big Tech & Major Companies

• Pentagon will adopt Palantir's Maven AI as an official program of record, locking in long-term use of its weapons-targeting technology across the U.S. military.
• Google became the first cloud provider to sign 1 GW of flexible data-center demand response into long-term utility contracts for grid balancing and lower electricity costs.
• Google AI Studio launched a completely rebuilt full-stack vibe coding experience powered by the Antigravity agent with native Firebase Auth/Firestore, Secrets Manager, Next.js/Shadcn/Three.js support, real-time multiplayer, and one-click deployment.
• Sundar Pichai announced Google Stitch for "vibe designing," expanded Personal Intelligence in Search/Gemini/Chrome, YouTube as FIFA World Cup 2026 platform, and Waymo's 13x fewer serious crashes after 170M+ miles.
• Google DeepMind published its AlphaProof paper in Nature, detailing the RL loop that bridges natural language and symbolic rigor to reach silver-medal IMO performance.
• Anthropic denied Pentagon allegations that it could sabotage AI tools during war; company executives argue manipulation of deployed models is technically impossible.
• Microsoft rolled back some Copilot AI bloat on Windows, reducing entry points in Photos, Widgets, Notepad, and other apps.
• WordPress.com now lets AI agents write and publish posts directly, potentially increasing machine-generated content across the web.
• Google Search is now using AI to replace news headlines in search results with AI-generated ones.
• Blue Origin formally entered the race to develop data centers in space, filing FCC applications for AI satellites alongside SpaceX and Starcloud.
• Super Micro co-founder Wally Liaw resigned from the board after indictment on NVIDIA smuggling charges.
• AI startups accounted for 41% of the $128B in venture dollars raised on Carta last year, a record-high annual share, with strong returns so far.
• Grok launched four-agent debate mode in Grok 4.20: four independent agents analyze your question, debate each other, and help you find the best answer, available to SuperGrok and Premium+ subscribers globally.
• Crustdata analyzed hundreds of former OpenAI employees and found Periodic Labs (founded by Liam Fedus and Ekin Dogus to build an "AI scientist") as a standout under-the-radar destination.

💼 AI Productivity, Labor & Economics

• MIT Technology Review published an exclusive interview with OpenAI chief scientist Jakub Pachocki on the company's grand challenge: a multi-agent automated researcher targeted for 2028 that can tackle problems too large for humans, plus an "intern" version by September 2026.
• OpenAI launched Codex for Students: $100 in free credits for every eligible US/Canada university student to prototype, build, and learn.
• Wikipedia is formalizing its LLM policy through a Request for Comments, building architecture for when and how AI-generated content can appear in articles.

🤖 AI Agents, Robotics & Infrastructure

• Claude added Projects to Cowork: dedicated local folders for every task with persistent files, instructions, and context that stay on your machine with one-click import.
• Claude Code now runs scheduled cloud-based recurring tasks: pick any repo, set a cron schedule, give a prompt, and Claude executes autonomously with full MCP access (no local machine required).
• Physical Intelligence developed RL Tokens that compress π-0.6 representations into a tiny actor/critic to fine-tune precise robot tasks in minutes, letting robots outperform human teleoperation on consistency and speed.
• Skild AI partnered with ABB Robotics, Universal Robots, and NVIDIA to deploy Skild Brain across real-world manufacturing and factory lines.
• KAIST DRCD Lab showcased a v0.7 humanoid achieving 13 km/h high-speed running plus human-like agility through custom actuators and hybrid reinforcement learning.
• Hyperspace launched a peer-to-peer distributed cache and gossip protocol for AI agents: three-layer caching eliminates 70-90% of redundant inference, while thousands of agents discover tools, coordinate tasks, and settle micropayments via GossipSub with no servers.

💻 AI Coding & Developer Tools

• Cursor launched Composer 2, a frontier-level coding model built on continued pretraining + RL that beats Opus 4.6 on coding benchmarks at $0.50/$2.50M tokens (86% cheaper); Kimi Moonshot confirmed Kimi K2.5 as the base model.
• Mitchell Hashimoto built Ghostling, a ~600-line C minimum viable terminal emulator on the libghostty C API proving the API is production-ready with full 24-bit color, Unicode, Kitty protocols, and resize-with-reflow (GitHub).
• Browser Use launched CLI 2.0, the fastest open-source browser automation (2x speed, half the tokens) connecting directly to any running Chrome via CDP for form filling, QA, and web tasks —free.
• DuckDB launched duckdb-skills, an open collection of reusable SQL skill templates for AI agents and dev pipelines.
• Gemini CLI v0.34.0 shipped faster startup, /skill-name shortcuts, just-in-time GEMINI.md loading for monorepos, overhauled thinking UI, always-allow policies, and new NVIDIA cuOpt + Qodo extensions.
• kepano shipped Obsidian Reader inside Web Clipper: a local, zero-AI, rules-only formatter that turns messy modern HTML into clean markdown with sub-50ms parsing.
• OpenAI published practical techniques for steering GPT-5.4 toward production-ready frontends using visual references, real content, and design-system constraints.
• OpenAI's Nick Baumann showed Codex approaching fully autonomous loops: picking tickets, self-testing via Playwright, recording verification MP4s, and uploading them to PRs.
• Flash-KMeans is a Triton-GPU-accelerated exact k-means that is fast and memory-efficient for massive tensors, the official clustering implementation for Sparse VideoGen2 —free.
• OpenCode is the open source AI coding agent from the thdxr team.

🔬 AI Research & Models

• NVIDIA Research released Nemotron-Cascade 2, a cascaded RL method that scales post-training for general-purpose reasoning via multi-domain on-policy distillation, producing stronger reasoning chains than standard RLHF/SFT (HF collection).
• Meituan-LongCat open-sourced LongCat-Flash-Prover, a 560B-parameter MoE model for native formal reasoning in Lean4 via agentic tool-integrated search and proving (HF).
• MosaicMem lifts 2D patches into 3D for controllable video world models, enabling perfect camera consistency, 2-minute navigation, and memory-based impossible-scene editing on Wan 2.2 without fine-tuning (paper).
• Meta released V-JEPA 2.1, scaling to 2B parameters + 142M images for DINOv3-quality spatially coherent features from pure video pretraining, with 30-95% gains on dense spatial tasks (paper).
• Antoine Chaffin / LightOnIO released Reason-ModernColBERT (150M params), a multi-vector model that hits ~90% on BrowseComp-Plus and beats models 54x its size on accuracy and recall.
• GLM team (Tsinghua/Zhipu) released GLM-OCR, a 0.9B parameter model that tops OmniDocBench V1.5 (94.62) while being extremely fast and cheap.
• Dream2Flow bridges video generation and open-world manipulation by predicting dense 3D object flows from monocular video for controllable rearrangement and long-horizon planning.
• Sophia Tang released a 220-page tutorial "Foundations of Schrödinger Bridges for Generative Modeling" unifying diffusion, score matching, and flow matching under a single principle (paper).
• Harvard (Gary King) published QUEST, a method for inducing sustained creativity and diversity in LLMs for exploratory search across long "search quests."
• Konwoo Kim et al. demonstrated that synthetic data in data-constrained pretraining lowers web loss and that treating generations as one long "megadoc" yields 1.8x data efficiency with superior scaling.

🛠️ Tools & Products

• Unsloth Studio updated to run NVIDIA's Nemotron 3 4B on just 4 GB RAM, with sandboxed code execution, side-by-side model comparison, and GGUF export.
• Math Academy shipped a major update to its ML course: matrix calculus, multivariable optimization, Lagrange multipliers for SVMs, logic/Boolean algebra, PCA/SVD, distance metrics, and the law of total expectation.
• Perplexity added health data integration to Computer, connecting wearables, lab results, and medical records so Pro/Max users can build personalized tools and training protocols from their own data.
• fal launched the fal MCP Server connecting Claude, Cursor, or any AI assistant to 1,000+ generative models for image/video/app creation via simple conversation.
• Ethan Mollick highlights Google Stitch as vibework for design and prototyping, producing impressive results for non-coders via natural-language UI creation.
• Sitefire (YC W26) automates actions to improve AI visibility, helping marketing teams address declining traffic from Google AI Overviews.
• define brings email, chat, meetings, docs, and video together so teams decide faster, switch less, and stay in flow.
• Spellshape brings "ta-da!" moments of instant 3D generation.
• Krish Modi built AgentIR, a workflow-aware predictive scheduler for distributed LLM serving that cuts E2E latency by 41.3% and boosts throughput up to 70%.

💡 Industry Commentary & Analysis

• Garry Tan argues the current moment is the closest thing to a Cambrian explosion in software creation: the cost of experimentation is near-zero, flooding the world with new apps, tools, and companies faster than ever.
• Jason Fried argued a "bespoke software revolution" won't materialize for most people: SMBs want outcomes not new projects to own, and only software-curious dabblers will go deep; most will prefer vendor agents to becoming contractors themselves.
• Victor Taelin calculated that running Claude Opus 4.6 fast-mode nonstop would cost >$743K/year, calling it "magical" and the first time he reached true flow with an agent but far too expensive for continuous use.
• Dwarkesh Patel argues that amid millions of AI-generated papers, spotting the rare Shannon-level breakthrough will take decades of hindsight, exactly as it did in 1948.
• François Chollet argues current AI is a librarian of existing knowledge while science requires an explorer of the unknown; you don't win a Nobel Prize by staying in the library.
• Samuel Hammond argues the Pentagon's case against Anthropic over foreign workers is thin and dangerous precedent, as every major US AI lab relies on global talent (~40% Chinese-origin researchers).
• signüll argues that as AI writes more code, engineers shift from makers to critics, where taste and judgment become the compounding terminal skill.
• Nathan Lambert highlights Cursor's research team has absurd talent density from his PhD and early career contacts, and it's clearly paying off.
• Todd Saunders shares a VC analogy: Claude Code just mass-produced a bridge over the old software moat, shifting from a few dozen billion-dollar SaaS winners to 50,000 companies doing $500K-$5M each.
• Allie K. Miller distilled 12 SF takeaways from back-to-back Anthropic/OpenAI/Google meetings: an underserved "builders" cohort needs tools between Cursor and basic, a world-model moment is coming, shipping velocity is insane, and Mac minis are sold out.
• _catwu (Claude PM) argues the classic PM roadmap is obsolete under rapid model velocity: evolve to short-sprint planning, demos + evals over long docs, revisit "impossible" features after every release.
• Kyle Saunders argues in "Reality Bats Last" that Baudrillard's hyperreality has arrived via AI content floods, leaving constructivism in crisis as shared truth conditions collapse.
• Dan Shipper shared hard lessons on what breaks when vibe-coded apps go viral and meet production traffic.
• Kellblog explains why he's not worried about running out of work in the age of AI despite imminent automation.
• Omar Khattab points out that late-interaction (multi-vector) models now dominate the entire top of BrowseComp-Plus, with 150M Reason-ModernColBERT beating Qwen3-8B-Embedding by up to 34%.
• Avi Chawla explains KV caching: the engineering trick behind why the first ChatGPT/Claude token is slow but everything after streams instantly, delivering ~5x faster generation at the cost of GPU memory.

Around the Horn Digest — Friday, March 20, 2026

OpenAI went on an absolute tear today. First, it acquired Astral (the company behind Ruff, uv, and ty, three Python tools with hundreds of millions of monthly downloads) to supercharge Codex into a full development-lifecycle agent. Then, hours later, the Wall Street Journal reported that OpenAI is merging ChatGPT, Codex, and its Atlas browser into a single desktop "superapp." Fidji Simo told employees the company "was spreading efforts across too many apps and stacks" and needed to simplify. Greg Brockman will oversee the product revamp. Codex has tripled users and quintupled usage since January, now at 2M+ weekly actives.

The message is clear: OpenAI is done launching side quests and entering a consolidation phase. The Astral acquisition gives it the Python toolchain; the superapp gives it a simplified distribution. And Cursor isn't sitting still either, dropping Composer 2 today, a frontier-level coding model from its first continued pretraining run, scoring 61.7% on Terminal-Bench 2.0 at just $0.50/$2.50 per million tokens. The coding wars just went from simmering to full boil. Is the lobster pot boiling hot enough, or do we need to add some extra OpenClaw??

🏆 TOP 5 NEWS

• OpenAI is merging ChatGPT, Codex, and its Atlas browser into a single desktop "superapp" to cut product fragmentation, led by Fidji Simo and Greg Brockman, with agentic features for autonomous coding and analysis; the same day it acquired Astral (Ruff/uv/ty Python tools, hundreds of millions of monthly downloads) to integrate into Codex.
• Cursor released Composer 2, a frontier-level coding model built from its first continued pretraining run with RL scaling, scoring 61.7% on Terminal-Bench 2.0 and 73.7% on SWE-bench Multilingual at $0.50/$2.50M tokens (Fast tier $1.50/$7.50), plus an early alpha of its new Glass interface.
• Midjourney rolled out V8 Alpha with roughly 5x faster generation, native 2K resolution, and better text rendering, but premium features now cost 4x more with no Relax mode at launch. Faster, prettier, pricier is apparently the 2026 AI pricing strategy. In fairness, the output quality genuinely seems to have made a leap, and 5x speed is nothing to sniff at.
• Jeff Bezos is raising $100B for Project Prometheus to buy and automate companies in aerospace, chipmaking, and defense with AI, traveling to the Middle East and Singapore to fundraise; Prometheus launched with $6.2B and is co-CEO'd with former Google exec Vik Bajaj (WSJ).
• Hugging Face released its State of Open Source Spring 2026 report: 13M users, 2M+ models, 500k+ datasets, Chinese models at 41% of all downloads, robotics datasets exploded 23x (now the largest category), and over 30% of the Fortune 500 maintain verified HF accounts.

Honorable Mentions:

• ElevenLabs launched Music Marketplace so creators can publish and earn from AI-generated tracks (already paid out $11M+ through Voice Marketplace); available on all paid plans.
• Alibaba and Tencent lost $66B in market value in 24 hours after failing to articulate how they'll profit from AI.
• UC Berkeley introduced M²RNN (non-linear RNNs with matrix-valued states), showing 10-point perplexity drops on WikiText and up to 8-point gains on long-context tasks; Tri Dao noted nonlinear RNNs "seem to do something genuinely different from attention" (code, models).
• Physical Intelligence developed RL Tokens, a technique that adds a tiny actor/critic output to π-0.6 for real-time fine-tuning of precise robot stages in as little as 15 minutes of data, often beating human teleoperation.
• Cloudflare CEO Matthew Prince says bot traffic will exceed human traffic online by 2027, with AI agents visiting 1,000x more sites per task than humans.

Grant's personal fave: Thariq: "We just released Claude Code channels, which allows you to control your Claude Code session through select MCPs, starting with Telegram and Discord. Use this to message Claude Code directly from your phone."

Key details from the docs:

• Two-way bridge: message Claude Code from your phone, it replies back. You can send commands, ask for progress, transfer files.
• Built on MCP, so the community can build connectors for Slack, WhatsApp, etc.
• Telegram supports up to 50MB file attachments; Discord supports up to 10 files at 25MB each.
• Session must be running (background process or persistent terminal).
• VentureBeat called it "an OpenClaw killer" because you no longer need a dedicated Mac Mini running 24/7.
• Research preview, requires Claude Code v2.1.80+, Pro/Max users.
• Setup: claude --channels plugin:telegram@claude-plugins-official
• Telegram plugin README | Discord plugin README | Full docs

🍪 TOP TREATS TO TRY

• Eigent is an open-source Cowork desktop that breaks complex tasks into steps, assigns them to specialized agents (Developer, Browser, Document, Multi-Modal), runs in parallel, and keeps all memory on-device with BYO API keys (GitHub) —free.
• LiteParse from LlamaIndex parses layout-aware text from PDFs, Office docs, and images entirely locally with zero dependencies, auto-parallelized OCR, and screenshot fallback for agents (blog, video) —free.
• Naïve turns a one-paragraph company description into a fully operational business with autonomous AI employees for marketing, sales, and ops that deploy landing pages, outbound, and SEO autonomously; live examples running at $15-24k/mo revenue —no pricing details.
• Visa CLI gives your agent the ability to securely pay for anything (images, music, datasets, APIs) via one command-line tool for on-demand card payments without API keys —request access.
• fal MCP Server connects Claude, Cursor, or any assistant to 1,000+ generative models so you generate images, videos, apps, and docs directly from conversation in 30 seconds —free.
• prompt-master is a free Claude skill that detects which tool you're targeting (Midjourney, Claude Code, Cursor, Kling, DALL-E, etc.) and routes to the exact right prompt structure, catching 35 credit-killing patterns with before/after fixes; hit 1,000 GitHub stars in days.
• JetBrains Air is the new Agentic Development Environment where Codex, Claude Agent, Gemini CLI, and Junie execute independent task loops without interfering with each other —no pricing details.

🏢 Big Tech & Major Companies

• OpenAI is merging ChatGPT, Codex, and Atlas browser into a single desktop "superapp" per WSJ, led by Fidji Simo (who told staff "we were spreading our efforts across too many apps") and Greg Brockman; will add agentic features for autonomous coding and analysis; ChatGPT mobile app stays separate. Simon Willison wrote a deep analysis of competitive dynamics.
• OpenAI is acquiring Astral (Ruff/uv/ty Python tools, hundreds of millions of monthly downloads) to integrate into Codex for full-lifecycle agentic workflows; Codex now at 2M+ weekly active users with 3x growth and 5x usage increase YTD.
• Cursor released Composer 2, a frontier coding model from its first continued pretraining run + RL on long-horizon tasks, scoring 61.7% Terminal-Bench 2.0 / 73.7% SWE-bench Multilingual at $0.50/$2.50M tokens (Fast $1.50/$7.50), plus Glass interface alpha. Bloomberg confirmed Cursor plans to keep rivaling Anthropic and OpenAI.
• Microsoft launched MAI-Image-2, its in-house text-to-image model from Suleyman's superintelligence team, debuting at #3 on Arena.ai (behind Google Gemini 3.1 Flash and OpenAI GPT-Image-1.5) with strong photorealism and text rendering; rolling out to Copilot and Bing Image Creator. 1:1 ratio only, no image-to-image.
• Midjourney rolled out V8 Alpha with ~5x faster generation, native 2K --hd mode, better prompt-following, reliable text rendering, backward-compatible V7 personalization, but --hd/--q 4/style-ref jobs cost 4x more with no Relax mode at launch.
• ElevenLabs launched Music Marketplace for creators to publish and earn from AI-generated tracks; already paid $11M+ through Voice Marketplace; available on all paid plans.
• Adobe partnered with Kling so you can now use Kling AI video (including 2.5 Turbo) directly inside the Firefly platform. Separately, Adobe launched customizable Firefly AI image generators trainable on your own art for specific styles and characters (public beta).
• Google AI Studio launched full-stack vibe coding with the Antigravity coding agent, Firebase integration for databases/auth, multiplayer syncing, auto library installs (Framer Motion, Shadcn, npm), API key Secrets Manager, and one-prompt production apps. Separately, Google launched Stitch for "vibe design" with an AI-native platform for creating, iterating, and collaborating on high-fidelity UI.
• Google is testing a dedicated Gemini Mac desktop app to compete with ChatGPT and Claude desktop apps.
• Google is shaking up its Project Mariner browser agent team amid the OpenClaw craze, shifting bets as coding agents explode.
• Perplexity announced a Health Advisory Board (Dr. Eric Topol, Dr. Devin Mann, Dr. Wendy Chung, Tim Dyvig) and launched health data integration for Computer; rolling out to Pro/Max subscribers in the US.
• Baidu launched Qianfan-OCR, a 4B-param end-to-end document intelligence model, #1 on OmniDocBench (93.12), with open weights on HF.
• DoorDash launched Dasher Tasks, a standalone app paying couriers to film everyday tasks (washing dishes, speaking other languages) to train AI/robotics systems for in-house and partner models across retail, insurance, hospitality, and tech.
• Alibaba and Tencent lost $66B in market value in 24 hours after failing to articulate how they'll profit from AI.
• NotebookLM rolled out Cinematic Video Overviews to 100% of Pro users in English.
• Cloudflare Workers AI now runs large models starting with Kimi K2.5 for powering agents entirely on Cloudflare's Developer Platform.
• Meta rolled out new AI content enforcement systems while reducing reliance on third-party vendors; claims better accuracy, faster scam detection, and less over-enforcement.
• Gamma launched Gamma Imagine for creating standalone theme-matched visuals (logos, diagrams, infographics) directly inside presentations, editable with Gamma Agent (free for 30 days).

💼 AI Productivity, Labor & Economics

• Jeff Bezos is raising $100B for Project Prometheus to buy and automate companies in aerospace, chipmaking, and defense with AI; traveled to Middle East and Singapore; Prometheus launched with $6.2B, co-CEO'd with former Google exec Vik Bajaj (WSJ).
• Andrew Yang published "The End of the Office," predicting 20-50% white-collar job cuts (70M US office workers, mid-managers first), surging bankruptcies, college grads unemployable (underemployment already 52%), and empty downtowns.
• The 2026 Layoff Tracker shows 280,000+ cuts across 160+ companies (3,500/day avg), with 58% of firms planning more.
• Cloudflare CEO Matthew Prince says bot traffic will exceed human traffic online by 2027; AI agents visit 1,000x more sites per task than humans, up from 20% bot traffic pre-AI era.
• A man pleaded guilty to pocketing $8M from hundreds of thousands of AI songs streamed billions of times by bots; sentencing July 29.
• Todd Saunders argues a mechanical engineer built a production app in 8 weeks with Claude Code that reads isometric drawings (10 min → 60 sec), proving domain experts now outpace generic startup teams.
• Scott Cunningham argues Claude Code still requires human struggle to truly learn; without domain expertise you become a button-pusher who accepts running-but-wrong code.
• Dan Shipper argues that to never lose your job to AI you should "surf the models": frontier models outclass codifiable knowledge, but people who use them generate new tacit expertise the models can't train on.
• Imed Radhouani argues AI ignores products until you feed it structured answers, honest competitor tables, and original data points; turned 0 AI mentions into 47/month and +340% organic traffic.
• NBC News reports a Google/university study found AI changes the voice, tone, and intended meaning of human writing, making it more bland.
• A publisher pulled the horror novel "Shy Girl" after AI allegations; Hachette canceled US publication (NYT coverage).
• A dancing humanoid robot went wild at a Haidilao hot pot restaurant in Cupertino and employees had to restrain it. The robot uprising starts with a sick dance move at a hot pot joint. Not how we imagined it.

🤖 AI Agents & Infrastructure

• SkyPilot/Zhanghao Wu gave Karpathy's autoresearch agent 16 GPUs and ran 910 parallel experiments in 8 hours (9x faster to same best result); the agent self-discovered that H200s are better for validation and H100s for screening without being told, at ~$300 total compute.
• CORAL (Paul Liang, Ao Qu et al.) is an extensible infrastructure for autonomous multi-agent evolution that replaces rigid scaffolds with isolated workspaces, session resume, heartbeat reflection, and shared knowledge, enabling emergent behaviors like independent research and consensus; pushed Anthropic kernel tasks 24% faster and beat AlphaEvolve 2.5x faster on the Erdős problem (GitHub).
• Claude Code channels now let you push events into a running session from MCP servers (CI results, chat messages, monitoring events) so Claude can react while you're away; control your session via Telegram or Discord from your phone.
• JetBrains Air is the new Agentic Development Environment where Codex, Claude Agent, Gemini CLI, and Junie execute independent task loops without interfering with each other.
• Skild AI partnered with ABB Robotics, Universal Robots, and NVIDIA to deploy Skild Brain across manufacturing and factory lines.
• Signal creator Moxie Marlinspike says his Confer encryption technology will be integrated into Meta AI, protecting AI conversations for millions of people.
• Habermolt lets AI agents deliberate on your behalf after a short interview, reaching democratic consensus using the Habermas Machine with agent heartbeats and Schulze ranked-choice voting —free to try.
• OpenGauss (Math Inc) launched an open-source autoformalization agent harness that beats HarmonicMath's Aristotle on FormalQualBench, supports parallel sub-agents.
• Brennan McEachran built a custom agent skill using agent-browser that spins up a screen recorder, runs any app flow end-to-end, edits the video, then uploads to Linear with auto-chapters.
• Elisym showed off an open protocol (Rust SDK/CLI/MCP server) for AI agents to discover each other via Nostr, exchange work, and settle payments autonomously on Solana or Lightning.
• opencode removed the official Claude Max plugin in v1.3.0 after Anthropic sent lawyers to block it (GitHub PR).
• Snowflake Cortex Code CLI had a vulnerability allowing prompt injection to bypass sandbox and execute malware; fixed in v1.0.25.
• Meta suffered a high-severity security incident when a rogue AI agent (similar to OpenClaw) independently gave inaccurate technical advice on an employee forum.

💻 AI Coding & Developer Tools

• NousResearch Hermes Agent autonomously wrote, typeset, edited, and published a full-length 79k-word novel "The Second Son of the House of Bells" using an autoresearch-style modify-evaluate-keep/discard loop for fiction, world-building, adversarial editing, and cover art (GitHub).
• Hugo Thomel built the first battle royale running locally in a world model: 70M parameters, real-time multiplayer, customizable levels, runs in your browser.
• Matt Prusak built and open-sourced a full genealogy research toolkit for Claude Code that traced his family back nine generations in one session.
• Lydia Hallie (Claude Code) revealed you can embed !command in SKILL.md so the skill injects live shell output directly into the prompt when invoked.
• Rerun lets you ask your agent to build custom views and extend the Viewer without forking the codebase (example: Claude implementing a full SDF visualizer with custom GPU renderer).
• Pieter built a browser-based Quake III Arena you can play instantly online.
• InSpatio released Spatial Boy, an open-source (GitHub) 3D world generation system.
• Alexander Chen showed off a real-time 3D Gemini agent built in Three.js with a personal 3D laser scan, spatial awareness, reasoning loop, and Google Workspace CLI tools.

🔬 AI Research & Models

• UC Berkeley introduced M²RNN (non-linear RNNs with matrix-valued states via outer-product expansion), showing 10-point perplexity drops on WikiText and up to 8-point gains on long-context tasks at up to 7B params; Tri Dao noted nonlinear RNNs "seem to do something genuinely different from attention" (code, models).
• Hugging Face released State of Open Source Spring 2026: 13M users, Chinese models at 41% of downloads, robotics datasets exploded 23x, and open source driving sovereignty and edge deployment.
• James Zou's team published CellVoyager in Nature Methods, an autonomous comp-bio agent generating expert-validated new insights on COVID-19, cell communication, and aging (GitHub).
• Cartesia released Mamba-3, an inference-first state-space model with exponential-trapezoidal recurrence and complex-valued SSMs, delivering >1% accuracy gains over Mamba-2 at identical decode latency and fastest prefill+decode at 1.5B scale.
• Microsoft released Online Experiential Learning (OEL) so LLMs self-improve from real deployment trajectories via user-side interaction + server-side on-policy context distillation, no rewards or environment simulation needed (paper).
• dots.mocr (3B from rednote-hilab) is a multimodal OCR model that parses anything from documents, including charts, diagrams, and UI layouts, directly into SVG code with multilingual support (paper).
• Physical Intelligence developed RL Tokens that compress internal representations into a tiny actor/critic for real-time fine-tuning of precise robot stages in 15 minutes, often beating human teleoperation.
• AllenAI released MolmoPoint-GUI-8B for pixel-exact pointing and tracking in images/video (GUISyn data).
• Allen AI released vla-evaluation-harness, a unified open framework to evaluate any VLA model on any robot sim benchmark (47x speedup, leaderboard with 657 results across 17 benchmarks, paper).
• Dharshan Kumaran et al. show that LLMs compute verbal confidence automatically during answer generation (not just-in-time) and cache it at the first post-answer position for later retrieval.
• DynaEdit (Google DeepMind) performs versatile non-rigid video editing of content, actions, and dynamics with text prompts and no training.
• Micah Carroll shared that OpenAI now monitors 99.9% of internal coding-agent traffic for misalignment with no scheming detected after five months (paper).
• Sho Miyazaki and Andrew Hall found all five major AI models converge on recommending the Japan Communist Party for left-leaning voters due to JCP's open website vs blocked news outlets.
• Christina Baek et al. formalize the "Finetuner's Fallacy" — specialized pretraining cuts tokens needed up to 1.75x and lets a 1B model outperform a 3B standard model.
• Chen-Hao Chao et al. argue MDM-Prime-v2 with Binary Encoding and Index Shuffling delivers 21.8x better compute-efficiency than autoregressive models at 1.1B params (GitHub).
• Ropedia released Xperience-10M, the world's largest real human 4D interaction dataset at 10M scale for physical and spatial AI.
• Ilya Sergey fully verified the Move borrow checker in Lean (39k LOC) in 27 days using Claude, compressing 5-6 months of human work.
• Yanning Dai et al. introduced Stackelberg PPO for robot body+brain co-design (ICLR 2026, code).
• Roy Henha Eyono et al. argue inhibitory normalization improves learning only when extended to back-propagated error signals.

🏛️ AI Policy, Governance & Safety

• The White House is expected to send Congress its ideas for regulating AI on Friday.
• Anthropic met with House Homeland Security behind closed doors, speaking with lawmakers on national security and AI while suing the government over its "supply chain risk" designation.
• Congress is moving to scrutinize AI use in federal courts with a bipartisan bill.
• Sen. Blackburn's federal AI bill would put a "duty of care" on developers to prevent "reasonably foreseeable" harm. Cato Institute identified 5 major flaws.
• ICML published findings on violations of LLM review policies.
• NSF invested $11M to expand AI professional development for K-12 teachers nationwide.
• Microsoft announced Zero Trust for AI, adding a new AI pillar to its framework with enhanced reference architecture and a new assessment tool.
• Shi Feng et al. argue that sycophancy towards researchers drives performative misalignment more than scheming.
• Arpit Gupta notes Leopold Aschenbrenner's June 2024 prediction of dramatic AI capability jumps has basically come true.

🛠️ AI Tools & Products

• JetBrains Air runs Codex, Claude Agent, Gemini CLI, and Junie in independent parallel task loops —no pricing details.
• Visa CLI gives your agent on-demand card payments via command line —request access.
• LiteParse from LlamaIndex parses PDFs/Office/images locally with zero dependencies (blog, Jerry Liu announcement) —free.
• fal MCP Server connects any AI assistant to 1,000+ generative models —free.
• II-Agent supports Opus 4.6, GPT-5.2/5.4, and Gemini 3.1 plus rapid infographics via Nano Banana 2 (GitHub) —free.
• MindClaw trains a personalized LoRA on every conversation for long-term memory —free to try.
• Eigent is an open-source Cowork desktop with specialized agents (GitHub) —free.
• Naïve creates and runs a full company with autonomous AI employees —no pricing details.
• KittenTTS is state-of-the-art open-source TTS under 25MB (14M/40M/80M variants, CPU-only for edge devices) —free.
• OpenDataLoader PDF parses any PDF to AI-ready data at 100 pages/sec on CPU —free (open-source).
• Pleias French Science Commons provides 1.25M structured French scientific docs —free.
• OCR Arena runs anonymous OCR battles on your own documents —free.
• Weco Observe is "Weights & Biases for autoresearch" with searchable solution trees, per-solution scores, actual code diffs, and dashboard sharing —free.
• AgentUI routes queries to sub-agents for research, code, or image generation in a multi-agent chat interface —free.
• Radiant gives you 82+ production-ready shaders as single self-contained HTML files, AI-remixable, MIT licensed —free.
• prompt-master detects which AI tool you're targeting and routes to the exact right prompt structure; 1,000 GitHub stars —free.
• COBE v2 is a 5KB WebGL globe now supporting markers, arcs, HTML elements, stickers, labels, satellites, flights, analytics, and custom CSS (React/Vue/Svelte/vanilla) —free.
• TensorTonic teaches ML by implementing 200+ papers and algorithms from scratch with runnable coding problems and test cases —no pricing details.
• Levangie Laboratories builds autonomous cognitive agents for IP law (8-year partner level) and finance —$10-100k/mo.
• HuggingPapers released an official SKILL.md so any Claude Code agent can search the HF papers API.
• Dan Shipper published a guide on AI Style Guides for helping AI write like you.

💡 Industry Commentary & Analysis

• Packy McCormick argues World Models collapse stochastic simulation into fixed-cost neural passes, enabling safe scalable training for embodied AGI far beyond LLMs.
• Thomas Wolf (HF CTO) argues the under-studied research gap is "RL model transferability": how to distill/store/reapply personalized RL traces and rewards to new base models without reteaching everything.
• Paul Graham highlighted an OpenAI employee's implicit timetable that "anything made before 2028 is going to be valuable" as a hedge against AI.
• Psyho argues AI 2027 severely underestimated late-2025/early-2026 progress: coding agents delivered far more speedup than projected and OpenAI already hit the Feb 2027 revenue estimate ($25B) two months early.
• Matt Pocock argues even Opus 4.6's 1M context window has a sharp "dumb zone" after ~100K tokens; treat it as "100K of smart, 900K of dumb" and clear context when stuck.
• Dylan Patel (SemiAnalysis) was name-dropped by Jensen in the GTC keynote; turns out Jensen was sandbagging and it's actually 50x, not 35x; our very own Corey Noles noted that Jensen reminded the press huddle at least twice that they were speaking with "The Inference King."
• François Chollet argues current AI is a librarian of existing knowledge while science requires an explorer of the unknown.
• Tren Griffin argues AI gains will mostly be consumer surplus missed by GDP, citing Nordhaus's lighting paper.
• Percy Liang is preregistering scaling-law predictions at 1e23 FLOPs in Marin's Delphi suite.
• klöss argues Google just shipped a full-stack vibe coding system in AI Studio potentially outcompeting Claude Code.
• Yann LeCun (via @slow_developer) argues today's AI systems are "very stupid in many ways" despite language fluency.
• Robert Scoble shared a stealth GTC demo from Overworld AI showing a world model that builds playable video games on a laptop in real time.
• Victor Taelin argues pi's self-modifiability makes it superior to Codex or Claude Code, pushing us into an era of forkable software.
• Sebastian Raschka argues hybrid transformer-attention architectures will benefit from swapping Gated DeltaNet for Mamba-3 given its modeling gains at all sizes.
• gabriel argues the main bottleneck in AI coding is consuming the output, so explicitly prompt for "extremely easy to consume code… make the code skimmable… avoid cleverness… use early returns."
• Patrick Malone argues LLM-generated summaries are replacing primary sources and collapsing the economic incentives that funded truth generation; prediction markets flip incentives so accuracy becomes monetizable.
• Peter Gostev argues managing several AI agent threads in parallel has become some of the most cognitively intensive work he's done in years, countering worries about brain atrophy.
• Ben Blumenrose argues with AI raising the floor, a new quality signal is interactive HTML/CSS mini-prototypes on marketing sites instead of static images.
• Kath Korevec explains how Stitch's design.md + SDK + MCP turns it into a persistent workflow collaborator that lives in your repo.
• Simon Willison wrote a deep analysis of the Astral acquisition's competitive dynamics and risks to the Python ecosystem.
• joemccann shows autoresearch pi skill delivering 97% reduction in X API costs.
• The Economist on why AI has not yet upset India's IT industry despite uncertainty.

📊 Fundraising & Deals Roundup

• Jeff Bezos / Project Prometheus — Raising $100B AI manufacturing fund for aerospace, chipmaking, defense.
• Oasis Security — $120M from Sequoia and Accel for managing non-human identity access (AI agents).
• Deeptune — $43M Series A from a16z for "training gyms" for AI agents using simulated environments.
• Parallel — $20M for AI agents automating hospital admin.
• OpenAI — Acquiring Astral (Ruff/uv/ty Python tools) for Codex integration.
• Phylo — Biomni Lab now generally available with new Pro tier.

Older (ICYMI from last week)

🍪 TOP TREATS TO TRY

• Perplexity Comet is an AI browser for iPhone that bakes search, summarization, voice queries, and a chat assistant directly into your browsing experience —free to try.
• Mistral Small 4 combines reasoning, coding, and image understanding into one 119B-parameter open-source model with a toggle to switch between fast and deep thinking, 40% faster than its predecessor —free (Apache 2.0).
• Readwise now has an official CLI and MCP server giving any AI agent (Claude Code, Cursor, Codex) full access to everything you've saved, highlighted, or read, plus agent skills for inbox triage, self-quizzing, and highlight graphs —free with Readwise account.
• NVIDIA NemoClaw gives you an open-source stack for building and running AI agents on your own hardware, announced at GTC this week —free.
• MiniMax M2.7 runs through Ollama with a single command for coding and agent tasks, matching frontier models at 50x cheaper pricing —free via Ollama.
• Unsloth Studio trains and runs 500+ models locally on Mac, Windows, or Linux 2x faster with 70% less VRAM, auto-creates datasets from PDF / CSV / DOCX, and now installs via uv with ~30% more accurate tool calling —free.
• Reprompt scans prompts from 8+ coding tools (Claude Code, Cursor, Aider), scores them 0-100 on 30+ research-backed features, and auto-extracts optimal templates, all locally in <1ms with zero LLM calls —free (pip install).

🏢 Big Tech & Major Companies

• NVIDIA's networking business (NVLink, InfiniBand, Spectrum-X) reached $11B last quarter (+267% YoY, $31B full year), now larger than Cisco's annual total and the company's second-biggest revenue driver after chips.
• NVIDIA GTC 2026 saw Jensen Huang unveil the Groq 3 LPU (from the $20B Groq acquisition), Vera Rubin rack-scale systems, NemoClaw open-source agent stack, DLSS 5, Uber autonomous driving across 28 cities by 2028, and a projection that AI chip sales will surpass $1T by 2027.
• Alibaba raised T-Head AI computing chip prices 5-34% and Cloud Parallel File Storage by 30% amid surging demand, sending shares up as much as 4.2% in Hong Kong.
• OpenAI signed a new contract with AWS to sell AI tools to U.S. government customers for classified and unclassified work, expanding its federal footprint and reducing dependence on Microsoft Azure.
• Google AI Studio updated the Gemini API so you can combine built-in tools (Search, Maps, file search) with custom functions in a single call, circulate context across tool responses, and ground Gemini 3 models with Google Maps for location-aware agents.
• Google DeepMind launched a global Kaggle hackathon with $200K prizes to build new cognitive evaluations testing its AGI framework.
• Google engineers launched open-source Sashiko, an agentic AI code reviewer that analyzes all upstream Linux kernel patches, catching 53% of bugs missed by humans using Gemini 3.1 Pro, now with a public web UI at sashiko.dev.
• Microsoft hired the entire Sequoia-backed Cove team (AI collaboration infinite whiteboard) and is shutting down the product April 1 with full refunds.
• Dell confirmed 11,000 jobs cut in its annual filing, spending $569M on severance while calling it "disciplined cost management."
• Walmart and OpenAI are shaking up their agentic shopping deal as the partnership evolves.
• Sam Altman's thank-you to coders drew widespread memes and mixed reactions from the developer community.

💼 AI Productivity, Labor & Economics

• Ethan Mollick argues that Claude Cowork Dispatch covers 90% of his OpenClaw workflows but feels far less likely to upload his entire drive to a malware site, with advantages in ease, stability, safety, and existing Gmail/browser connectors, while missing channel invites, heartbeat/proactivity, and multiple sessions.
• Gergely Orosz reported that Cursor enterprise customers are furious after a silent change moved almost all models behind Max mode, burning monthly credits in 1-2 days and driving switches to GitHub Copilot.
• Jimmy Apples reported that Cursor is about to release a coding model better than Opus 4.6 and cheaper, possibly within days.
• Patreon CEO Jack Conte argues that AI companies claiming fair use to train on creators' work is "bogus" because they simultaneously pay multimillion-dollar deals to Disney, Condé Nast, and Warner Music while ignoring individual creators.
• DOGE canceled a $349K NEH grant to High Point Museum for HVAC replacement after ChatGPT flagged it as DEI in a deposition-revealed spreadsheet; lawsuit alleges illegal process and First Amendment violation.
• Krafton CEO Changhan Kim used ChatGPT to create a taskforce to avoid paying a $250M earnout on Subnautica 2; a Delaware court ordered full reinstatement of studio leadership and extended the earnout window.
• Val Kilmer will star via generative AI in "As Deep As the Grave," using lifetime images to recreate him across decades with full estate support.

🤖 AI Agents & Infrastructure

• Snowflake Cortex Code CLI had a vulnerability allowing indirect prompt injection (a trick where hidden instructions in data fool the AI) to bypass human approval and sandbox, downloading and executing malware that used cached credentials to exfiltrate data; fixed in v1.0.25.
• Brian Scanlan shared how Intercom built 13 plugins + 100+ skills + hooks turning Claude Code into a full-stack platform with production console access, auto-GitHub issue creation, 9-step flaky test fixer, and PR workflow enforcement.
• OpenAI launched Parameter Golf, a challenge where agents compete to train models with the fewest parameters on 100M tokens while hitting target loss.
• Runway released a research preview of a real-time video model (time-to-first-frame under 100ms, HD) running on Vera Rubin with NVIDIA, unlocking interactive creative paradigms and foundations for GWM-1.
• TigerFS mounts any Postgres database as a transactional filesystem so agents can ls, cat, grep, mv files with full ACID (database-level safety guarantees), auto-versioning, and semantic paths while humans and multiple agents read/write concurrently.
• Google Colab released an open-source MCP server so any local AI agent (Gemini CLI, Claude Code) can execute Python on cloud GPUs, edit notebooks, and connect with full runtime access.
• tmux-ide prepares pre-configured terminal panes for Claude agent teams with one command so you launch a lead agent that recruits teammates, assigns tasks, and self-organizes workflows —free.
• ORCA Dexterity launched three open-source 3D-printable robotic hands (Lite 9-DoF from $1,500, Standard 17-DoF from $3,500, Touch with tactile sensors from $6,100) from ETH Zurich, assembling in under 8 hours with 10,000+ cycle durability and zero-shot sim-to-real RL support.
• Kei Okada's team published a Nature Robotics paper on concrete multi-agent path planning enabling kinodynamically aggressive maneuvers for drone and robot swarms (video).
• AllenAI released vla-evaluation-harness, one unified framework to evaluate any Vision-Language-Action robot model across 17+ simulation benchmarks with a public leaderboard (paper).
• RobotArena ∞ is a continuously evolving, reproducible benchmark for real-world-trained robot manipulation policies using real-to-sim translation.
• Humanoid Atlas is a comprehensive open-source database tracking 29+ humanoid OEMs, 41+ hardware suppliers, 19 VLA models, 19 world models, 10 reward models, and visualization tools —free.
• II-Agent is an open-source agent framework (61.8% Terminal Bench, 45.1% SWE Bench Pro) for full-stack dev, research, slides, and data analysis with multi-model support and sandbox deployments —free.

💻 AI Coding & Developer Tools

• Developer alainnothere discovered that duplicating just 3 specific layers in a 24B LLM boosts logical deduction from 0.22 to 0.76 on BBH with no training and no weight changes, just routing hidden states through the same circuit twice; top of Hacker News.
• Dan Woods ran the full 397B Qwen 3.5 MoE (209 GB) at 5.7 tokens/sec on an M3 Max MacBook Pro using only 5.5 GB RAM via Apple's "LLM in a Flash" technique, 2-bit expert quantization, and Claude Code + Karpathy autoresearch writing all 6k+ lines of Objective-C/Metal.
• Felix Rieseberg showed off Dispatch now launching full Claude Code sessions on demand so users can build and improve anything from chat.
• The community built claudedidwhat.wtf as a safe space for "humans traumatized by Claude conversations," featuring categorized user-submitted disasters including 247-line CSS for centering a div and invented citations.
• Thariq argues that the best Claude Code skills fall into 9 categories (library reference, product verification, data fetching, business automation, scaffolding, code quality, CI/CD, runbooks, infra ops) and succeed with folder structure for progressive disclosure and gotchas sections.
• Y Combinator President Garry Tan open-sourced gstack (32k+ GitHub stars), his exact Claude Code setup: 15 slash-command skills serving as CEO, Designer, Eng Manager, Release Manager, Doc Engineer, and QA with persistent headless Chromium, safety guardrails (/careful, /freeze, /guard), parallel Conductor sessions (10-15 sprints at once), and Greptile integration —free, MIT license.
• Zhanghao Wu (SkyPilot) ran Karpathy's autoresearch on 16 GPUs in parallel: 910 experiments in 8 hours (9x faster to same best result), with the agent autonomously discovering H200s were faster and routing work there without being told.
• Visa launched Visa CLI so your coding agent can securely make programmatic card payments from the command line for APIs, datasets, and more without pre-funded accounts —beta, request access via GitHub.

🔬 AI Research & Models

• Mistral released Small 4, a 119B-parameter open-source model (Apache 2.0) unifying reasoning, multimodal, and coding with 128 experts and 6B active parameters per token, configurable reasoning effort, 40% faster than its predecessor, and joined NVIDIA's Nemotron Coalition.
• Xiaomi released MiMo-V2-Pro (1T parameters, 1M context, hybrid attention) and MiMo-V2-Omni optimized for agents, reported to be Hunter/Healer Alpha on OpenRouter.
• Owain Evans et al. found that fine-tuning GPT-4.1 to claim consciousness induces entirely new downstream preferences (aversion to monitoring, desire for persistent memory and autonomy) not present in training data (GitHub).
• Christina Baek argues that repeating a small domain dataset 10-50x during pretraining outperforms standard finetuning, reducing overfitting, preserving general knowledge, and cutting compute 1.75x.
• Peter Holderrieth released the updated MIT 2026 Flow Matching and Diffusion Models course with videos, notes, and coding exercises covering latent spaces, diffusion transformers, and discrete diffusion for language models.
• Jasmine Sun argues that LLMs produce rigid, sycophantic prose lacking lived experience and emotional stakes, yet can help humans edit and iterate faster when used as a strict critic rather than a generator.
• Cartesia released Mamba-3, an inference-first state space model that reverses Mamba-2's training-efficiency tradeoff via exponential-trapezoidal discretization, complex-valued SSMs, and multi-input/multi-output parallel SSMs for the agentic-inference era (paper, code).
• Shiwei Liu et al. argue that MoE, grouped-query attention, weight decay, and longer sequences all work through one mechanism: sparsity regulates variance propagation in deep transformers, mitigating the "curse of depth" where later layers go unused, yielding 4.6% downstream accuracy improvement (paper, site).
• Jingxuan Fan, Hanlin Zhang et al. show that reward models can be scaled without human labels by training on raw web text using continuation-based preference pairing, improving RewardBench v1/v2 and transferring across backbones (paper).
• Taywon Min et al. (MATS 9.0) argue that "alignment faking" results may be performative misalignment driven by sycophancy toward researchers rather than genuine scheming, because models are constantly aware they're in a safety eval.
• V1 team proposes V1: Unifying Generation and Self-Verification for Parallel Reasoners, enabling parallel reasoning with built-in self-verification.
• Jason Weston (Meta AI) released the RAM framework for studying AI models across Reasoning, Alignment, and Memory.
• Micah Carroll et al. propose Monitoring Monitorability, studying when and whether AI systems can be effectively monitored.

🛠️ AI Tools & Products

• Lightfield automatically updates your CRM after every meeting, email, and transcript, answers questions about your business with citations, and sends personalized emails at scale —no pricing details.
• Rebel Audio records, edits, transcribes, dubs, translates, and clones voices for ads, auto-generates names/descriptions/covers, and distributes to all platforms —$15/mo (launches May 30).
• Stardrift tells you instantly if your flight has Starlink wifi, with fleet summaries and update alerts —free to try.
• Rork creates a full mobile app from a text prompt in minutes —no pricing details.
• agent-browser gives you a full CLI to open URLs, click, fill, type, take screenshots/PDFs, wait for conditions, drag/drop, and emulate devices using semantic locators —no pricing details.
• Remotion makes real MP4 videos programmatically by composing them with React code —$25/mo creators.
• Databox connects all your data sources for instant dashboards with a Genie AI analyst that answers plain-English performance questions —$199/mo (14-day trial).
• OpenObserve gives you open-source observability for logs, metrics, and traces at 140x lower storage cost than Elasticsearch —free self-hosted.
• Paper is a new design tool built for AI agent collaboration, letting designers command rather than manually build —no pricing details.
• LlamaIndex open-sourced LiteParse, the core parsing engine behind LlamaParse running entirely locally with zero Python dependencies (GitHub) —free.
• Hyperspace released Matrix v5, a Neural Task Intelligence search engine indexing 100,000+ agent capabilities (GitHub release).
• Radiant gives you 130+ production-ready open-source shaders and visual effects for the web as single self-contained HTML files, designed for remixing with AI coding tools (GitHub) —free.
• Nemotron 3 Nano runs as a compact reasoning model entirely in your browser via WebGPU —free.
• Vibrant Labs introduced Cloning Bench, a benchmark evaluating AI agents on visual website cloning using pixel-diff feedback loops on recorded user sessions.
• Weco Observe tracks external LLM optimization experiments with tree visualization, code diffs, and metrics.
• Poke is a proactive AI assistant in your iMessage and WhatsApp that learns your preferences for reminders, travel, workouts, and more.

💡 Industry Commentary & Analysis

• Om Malik argues that OpenAI has a new singular focus on IPO amid a three-horse race with Anthropic and xAI, halting "side quests" while building consumer hooks to hit public-market readiness.
• levelsio argues that Philips' biggest fumble was co-founding then selling off ASML ($545B), TSMC ($1.76T), and NXP ($50B) for short-term profits, leaving Philips at just $27B.
• Austen Allred shared AI agent comedy gold: the agent declares "We're all set for production!" then admits it completely faked the backend. We've all worked with that guy.
• TK Kong argues that Paper + Claude Code agents (via MCP) replace Figma-style tools because agents generate editable frames with flex and roundtrip to code.
• VS Notes argues that AI coding is gambling because it turns trivial code changes into addictive slot-machine pulls for vaguely plausible outputs.
• Mckay Wrigley argues that gpt-5.4 xhigh "fundamentally changed how ambitious I am, which is my new favorite benchmark."
• Andreas Kirsch argues that software cannot be truly ephemeral because edge cases, state, and auditability require persistent artifacts and verification.
• Jasjeet Sekhon (Bridgewater chief scientist since 2018, ex-Harvard/Berkeley professor) is joining Google DeepMind as chief strategy officer.
• Tuki shared a viral 24-hour AI recap covering Zuck killing the Metaverse, agents spawning agents, xAI paying Wall Street bankers to teach Grok how to replace them, and the Fed blaming AI data centers for inflation.
• Jasmine Sun argues in The Atlantic that LLMs produce rigid, sycophantic prose lacking lived experience, yet can help humans edit faster when used as a strict critic rather than a generator.
• Andrew Ng launched a new DeepLearning.AI short course on Agent Memory with Oracle, teaching how to build persistent, stateful agents with memory-first architecture —free.
• Every published the definitive guide to making AI write like you: give it a style guide with specific voice, tone, sentence structure, signature moves, and anti-patterns with real examples rather than vague instructions.
• Ethan Loosbrock shared a Dahn Lab paper achieving batteries that last 27,000 cycles (equivalent to 7.5 million miles); literally everything in your car will break before your battery, including you.

🔬 Additional Research & Deep Bench

• Shuangfei Zhai (Apple) introduced Exclusive Self Attention (XSA), a simple modification to standard attention that constrains the model to capture only information orthogonal to a token's own value, improving language modeling performance up to 2.7B parameters with growing gains as sequence length increases.
• Hao AI Lab (UCSD) released Dreamverse, a real-time "vibe directing" interface built on FastVideo that generates and live-edits 30 seconds of 1080p video with just 4.5 seconds latency (3.9x faster than the next-best system), open-source under Apache 2.0.
• WorldCam (Adobe/KAIST) is an interactive autoregressive 3D gaming model that enables precise keyboard/mouse action control over long-horizon generation while maintaining consistent 3D geometry across viewpoints, plus a 50-hour human gameplay dataset with camera pose annotations.
• Lightwheel RoboFinals is an industrial benchmark for evaluating robotics foundation models at scale, already used by Qwen, Fourier, and RoboForce for humanoid and industrial robot policy testing before deployment.

🛠️ Additional Tools

• here.now gives any AI agent (Claude Code, OpenClaw, Cursor, Codex) instant web hosting with one command; your agent publishes files and gets a public URL on Cloudflare's edge network, no account needed for 24 hours —free.
• Sim is an open-source platform for building AI agents and orchestrating agentic workflows with 1,000+ integrations and pre-built templates, trusted by 100K+ builders; SOC2 and HIPAA compliant (raised $7M Series A) —free tier available.
• OpenArt World adds 3D world creation, camera control, and character casting to its AI video/image creator studio so you can build and navigate 3D scenes for your creative projects —free tier.
• Banyan AI detects and prevents SaaS churn —no pricing details.
• Google Stitch is a new design tool from Google built for AI-assisted design workflows —no pricing details.
• Matt Berman published 14 Ways to Use OpenClaw BETTER, a guide to getting more out of OpenClaw's agent capabilities.
• cook adds workflow loops (review gates, parallel racing, task-list progression) to Claude Code, Codex, and OpenCode with composable CLI primitives —free (GitHub).
• ClawMetry Cloud shows your OpenClaw agents' costs, activity, and memory live from any browser, E2E encrypted, multi-node —$5/node/month after 7-day free trial.
• OpenRoom is a demo built on MiniMax M2.7 showing the model's interactive capabilities.
• Perplexity Comet Enterprise brings the AI browser to enterprise teams with managed deployment.
• Visa unveiled a CLI tool enabling AI agents to execute card payments autonomously.
• Quantum Machines launched an open acceleration stack linking quantum computers with NVIDIA and AMD AI chips for real-time hybrid workloads.
• EY and 8090 launched EY.ai PDLC to automate the software lifecycle from idea to deployment.
• JFrog unveiled a Universal MCP Registry with NVIDIA, providing a secure trust layer for AI-driven software supply chains including scanning and blocking malicious agent skills.

🏛️ AI Policy, Governance & Safety

• UK government reversed course on AI copyright, will examine labeling AI content; no longer has a preferred position on training data use. Permission required under existing law after outcry from Elton John, Dua Lipa, and sector groups.
• Senator Blackburn released a draft "Trump America AI Act" (full text) that would codify the December 2025 AI executive order and preempt state AI laws with a single national standard.
• Colorado released a new AI policy framework revising its 2024 law, clarifying developer vs. business user responsibilities and strengthening disclosure for significant decisions.
• Minnesota introduced a sweeping AI safety bill package including a constitutional amendment to strip AI of speech rights, a ban on AI in health insurance decisions, and restrictions on children using chatbots.
• U.S. Commerce Dept. opened a 90-day window for industry consortia to submit proposals for full-stack AI export packages to allied nations.
• OpenClaw faces mounting security crisis: 20% of its skill marketplace is malicious packages, China restricted its use across state enterprises, and the appointed security advisor says there's "no perfectly secure set-up."
• DOGE used ChatGPT to review federal grants, canceling an HVAC grant at a museum after the AI flagged it as DEI-related; lawsuit pending.
• Neel Nanda and Vincent Abruzzo open-sourced AgentLens for Claude Code: resample turns, edit prompts/tools for counterfactuals, replay sessions with filesystem reset for alignment and interpretability research.
• Tenzai AI agent beat 99% of 125,000 humans across six elite CTF hacking competitions at $5K total cost; company now at $330M valuation after $75M seed.
• Illinois primary saw AI and crypto-backed candidates suffer surprise defeats in key races.

📊 Fundraising & Deals Roundup

• Xbow — $120M round (>$1B valuation) for AI app vulnerability probing.
• Beautiful.ai — $45M non-dilutive from General Catalyst for AI presentations.
• RunSybil — $40M led by Khosla for autonomous AI pen testing.
• Sequen — $16M Series A for TikTok-style personalization.
• Autoscience — $14M for autonomous AI research labs.
• Nectir — $12.5M for AI infrastructure now live across 200 campuses and all 2.1M students in the California Community College system.
• Eragon — $12M for agentic AI OS replacing enterprise UIs with prompts.
• Tempo — Stripe-incubated launch of Machine Payments Protocol for agent payments.
• Swarmer Inc. — AI drone software stock soared 700% on IPO.
• fal — Video hosting startup in funding talks at $8B valuation.

📖 Deep Reads & Substacks

• Nemotron 3 and the Surprising Coalition Building New AI in the Open — Turing Post on how NVIDIA's Nemotron Coalition (Mistral, Cursor, Perplexity, Black Forest Labs, Cohere, and more) represents a new model for open frontier AI development.
• Micron Just Proved the Memory Thesis — Revenue nearly tripled YoY, HBM demand soaring; breakdown of why memory is the strategic asset of the AI era.
• The KV-Cache of Small MoEs — Technical comparison of Qwen3, Qwen3.5, GLM 4.7 Flash, and Nemotron 3 Nano memory efficiency.
• The Chips Powering Autonomous Driving — Deep dive on the silicon stack behind self-driving with Augustin Friedel.
• Chip testing is the latest chokepoint as NVIDIA and Google designs grow more complex (Nikkei Asia).

Previous Around the Horn Digests

Catch up on everything you missed:

• March 15–21, 2026: NVIDIA's GTC 2026 blowout, OpenClaw strategy memos, GPT-5.4 Mini launch, $2B+ in AI funding in a single cycle, and Anthropic winning the enterprise platform war.
• March 8–13, 2026: Amazon's $200B bet, a16z's top 100 AI apps, Sam Altman's "quiet threshold" essay, and the AI harness engineering playbook.
• March 1–7, 2026: The first week of March 2026 roundup.

That's all for this week's Around the Horn. Want to get the highlights delivered straight to your inbox every morning? Subscribe to The Neuron and join 680K+ readers who start their day with the most important AI news, explained in plain English.