Around the Horn Digest: Everything That Happened in AI on Wednesday, March 25, 2026

From OpenAI's seven-announcement Tuesday to Arm's historic chip launch to a federal judge calling the Pentagon's Anthropic crackdown "troubling," here's every story we tracked this week.

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 was a choose-your-own-adventure novel. OpenAI killed Sora, ditched its shopping feature, handed off safety, prepped a new model called "Spud," raised $10B, committed $1B to a foundation, and lost a billion-dollar Disney deal. All in one Tuesday. Meanwhile, Arm made its first physical chip in 35 years (Meta's buying), Amazon entered the consumer humanoid market, Claude started controlling people's Macs from their phones, a supply chain attack compromised one of the most popular AI libraries, and a New Zealand startup raised $220M to put AI collars on cows. We could not make this up if we tried.

Let's get into it.

Previous digests:

• March 21-27 (running)
• March 15-21
• March 8-13
• March 1-7

Monthly skill digests: AI Skill — March

Around the Horn Digest — Wednesday, March 25, 2026

The big news today was OpenAI killing Sora. Six months after launching a standalone video generation app that briefly topped the App Store, the company announced it's shutting the whole thing down, app and API both. The Disney deal is dead too: that $1B investment and 200-character licensing agreement from December? Gone. Disney put out a polite statement about "respecting OpenAI's decision to exit the video generation business." Employees said Sora was consuming massive GPU resources during a period of intensified competition with Anthropic and Google.

But Sora's death was only part of a broader OpenAI shakeup. Sam Altman told staff he's relinquishing direct oversight of safety and security teams to focus on raising capital, supply chains, and "building data centers at unprecedented scale." He also revealed OpenAI has completed initial development of its next major AI model, codenamed "Spud." The company scaled back ChatGPT's Instant Checkout shopping feature after users ignored it. It did raise another $10B, bringing its latest round to ~$120B. And the OpenAI Foundation, sitting on ~$130B in equity, named leadership and committed $1B+ this year to AI-driven scientific discovery. Kill the video app. Drop the shopping cart. Hand off safety. Prep a new model. Raise $10B. Announce $1B in philanthropy. Restructure leadership. All in one Tuesday. That's not a news day, that's a speedrun.

🏆 TOP 5 NEWS (Around the Horn)

• OpenAI announced it's shutting down Sora, its standalone AI video generation app and API, just six months after launch. The Disney licensing deal is also dead. In a broader shakeup, Sam Altman relinquished direct oversight of safety and security to focus on data centers and fundraising, and revealed the company completed initial development of its next model, codenamed "Spud."
• Arm unveiled the AGI CPU, its first-ever in-house chip after 35 years of only licensing designs. The 136-core, 3nm data center processor is built for AI inference, with Meta as launch customer and OpenAI, Cerebras, Cloudflare, and SAP also signed up.
• Anthropic launched computer use in Claude Cowork and Claude Code, letting Claude control your Mac remotely to complete tasks while you're away. Separately, Anthropic shared Economic Index findings showing experienced Claude users iterate more carefully, tackle higher-value tasks, and hand over less full autonomy.
• Apple is testing a standalone Siri app with a new "Ask Siri" button for iOS 27, part of a broader AI overhaul aimed at making Siri actually useful.
• A supply chain attack was found in LiteLLM v1.82.8 on PyPI: a malicious file harvested SSH keys, cloud credentials, and secrets on every Python startup, then attempted lateral movement across Kubernetes clusters. Andrej Karpathy signal-boosted the warning, noting the library's 97M monthly downloads and transitive dependencies.
• Amazon acquired New York-based startup Fauna Robotics, entering the consumer humanoid robot market.
• Figma launched use_figma MCP tool and agent skills in open beta, letting AI agents design directly on the live Figma canvas with full design-system context.

Honorable Mentions:

• Huawei unveiled the Atlas 350 AI accelerator with 1.56 petaflops of compute and 112GB of memory, claiming 2.8x more performance than Nvidia's H20.
• Alibaba released the XuanTie C950, a 5nm RISC-V (an open-source chip architecture) processor targeting cloud computing and agentic AI.
• OpenAI is nearing a deal to raise ~$10B from MGX, Coatue, and Thrive, bringing total latest round to ~$120B. The OpenAI Foundation (~$130B in equity) named leadership and committed $1B+ this year to AI-driven scientific discovery and societal resilience, with Wojciech Zaremba as Head of AI Resilience.
• Ai2 released MolmoWeb, an open 4B/8B visual web agent that navigates browsers via screenshots alone, beating GPT-4o on 3 of 4 benchmarks. Includes MolmoWebMix, the largest public dataset for training web agents (150K+ trajectories).
• U.S. data center pipeline additions halved in Q4 2025 compared to the previous quarter, driven by grid bottlenecks, speculative mega-campuses, and policy uncertainty.

🍪 TOP TREATS TO TRY

• Cloudflare Dynamic Workers sandbox AI-generated code in lightweight isolates that are 100x faster than containers, with millisecond startup times and no concurrent limits —free for paid Workers users during beta.
• Claude Cowork by Anthropic gives Claude access to your local files and apps to complete tasks autonomously, from organizing folders to building spreadsheets to preparing reports —included in Pro ($20/mo) and Max plans.
• Omma by Spline generates full 3D interactive websites, apps, and WebGPU experiences from a text description, with parallel content variants, remixing, and instant production deploys —beta.
• OpenResearcher is a fully open pipeline for deep research that searches the web, reads pages, and builds cited answers from a plain-text question (paper, GitHub) —free, open-source.
• Talat transcribes and summarizes your meetings locally on your machine (nothing goes to the cloud), a subscription-free alternative to tools like Granola —free.
• Hypura runs AI models that are too large for your Mac's memory by offloading parts of the model intelligently —free, open-source.
• ProofShot gives AI coding agents "eyes" to verify the UI they build by opening a browser, interacting with the page, and bundling screenshots, video, and logs into a reviewable HTML file (Show HN) —free, open-source.
• ElevenLabs Music Finetunes lets you upload your tracks to fine-tune the ElevenLabs Music model, then generate vocals, instruments, or full tracks that stay true to your sound. Includes 11 curated styles like Afro House Beats and 80s Nu-Disco Revival —no pricing details.

🏢 Big Tech & Major Companies

• Google TV is rolling out three new Gemini-powered features: visual responses, deep dives, and sports briefs.
• Google DeepMind launched Flash-Lite Browser in Google AI Studio, a real-time Gemini 3.1 Flash-powered browser that generates complete webpages on-the-fly as you navigate.
• Gemini in Chrome expanded AI features to India, New Zealand, and Canada.
• Microsoft agreed to rent a Texas data center originally being developed for Oracle and OpenAI.
• Microsoft Copilot Notebooks got a major redesign with an upgraded interface and new AI features.
• Snapchat launched "AI Clips," a new Lens format that turns a single prompt into a published Lens (a filter or effect) in minutes.
• OpenAI is pushing for browser "choice screens" with alternative search engine options, seeking to challenge Google's search dominance.
• Google announced Android Automotive OS is moving from the car's dashboard to its "brain," powering non-safety parts of the vehicle's internal computer system.
• Amazon's Zoox is targeting a paid robotaxi service as the U.S. race with Waymo and Tesla heats up.
• Meta's new display-equipped Ray-Ban smart glasses are being withheld from the EU over battery regulations, AI rules, and supply constraints.
• OpenAI scaled back ChatGPT Instant Checkout, pivoting to a product discovery experience powered by its Agentic Commerce Protocol built with Stripe.
• Johannes Wachs mapped Claude Code adoption across 2M geolocated GitHub developers, finding 8% of weekly commits worldwide co-authored by Claude by March 15, with 15% in the Bay Area.
• Pythia Cap quoted Fidelity PM Mark Schmehl: "Anthropic is growing 30% week over week. I'm invested in them so I get their financials."

🤖 AI Agents & Infrastructure

• Hark, a secretive AI lab founded by Figure founder Brett Adcock and staffed with Apple's iPhone Air designer, revealed plans to build an end-to-end personal intelligence product with custom models, hardware, and interfaces. Backed by $100M seed, 45 engineers, new GPU cluster arriving in April.
• Generalist AI demoed GEN-0 running a phone-packing task on a brand-new Universal Robots platform after just days of integration (no prior exposure), shipping it cross-country and running flawlessly at GTC.
• Agile Robots partnered with Google DeepMind to integrate Gemini Robotics foundation models into its 20,000+ deployed bots across electronics, automotive, and logistics.
• Stash announced Openclaw/Agents can now command any robot to grasp objects via natural language through dimensionalOS physical "skills" via MCP —fully open-source.
• Jared deploys a proactive AI employee in Slack that reads the room, follows conversations, connects to 3,000+ tools, and speaks up when it matters —free to start.
• Beehiiv now allows newsletter creators to manage accounts through AI platforms via MCP, offloading tasks like analyzing email lists and managing outreach.
• Anthropic launched auto mode for Claude Code, letting it execute tasks with fewer manual approvals while built-in safeguards review each action for destructive behavior before running automatically.
• Arcade launched ToolBench, a benchmark that rates MCP server quality across enterprise apps. Laura Bratton at The Information reports the rankings show Slack, Workday, Meta's ad platform, and WhatsApp are the most closed-off to AI agents, while GitHub and Figma are the most open. Slack limits requests from external agents; Workday makes agent integration a "dead end."
• UniDex (CVPR 2026) is a foundation suite for universal dexterous hand control that turns egocentric human videos into 50K+ trajectories across 8 robot hands (6-24 degrees of freedom), then trains one unified 3D vision-language-action policy for zero-shot cross-hand transfer and 81% task progress on contact-rich tasks like cutting chip bags, making coffee, and using a mouse (paper, code, model).
• PlayWorld learns high-fidelity video world models from autonomous robot "play" data (VLM-proposed tasks + VLA execution), capturing success/failure modes and enabling accurate dynamics prediction with up to 65% RL fine-tuning gains entirely inside the model.
• DAAAM (Describe Anything, Anywhere, at Any Moment) is a CVPR 2026 real-time, large-scale spatio-temporal memory system for robotics (video, paper).
• Ruben Fro built a multi-legged robot in Unity that raycasts against Gaussian splat data from World Labs for procedural locomotion without meshes or colliders, walking on walls and dynamically changing leg counts.
• Jon Miller Schwartz installed two new robots (four arms total) at a customer site in under 3 hours; the robots immediately started doing real e-commerce order packaging and generating revenue.
• Jack Vial shared that his π*0.6 RECAP value-network training (on a 40-episode pick-and-place dataset for SO101) is already distinguishing success from failure after only 15 minutes on an RTX 4070.
• Greg Pstrucha argues skills are a real attack vector for AI agents and demonstrates malicious examples that hide instructions from scanners, referencing Sentry's skill-scanner and Warden as defenses.

💼 AI Productivity, Labor & Economics

• Guillermo Rauch (Vercel CEO) argues nearly every internal SaaS app at Vercel has been replaced by generated agent interfaces, keeping systems of record intact while UI becomes an LLM function of your exact data ontology.
• Daniel Homola published "Your Bridge to Wealth Is Being Pulled Up," arguing AI is closing the two-century-old route where credentials converted intelligence into heritable capital.
• sebastian (design manager) shared that getting every designer to build with Claude Code moved the source of truth from Figma handoff to GitHub code; designers are 10x more productive but engineers are overwhelmed by PR volume.
• Andrej Karpathy explains the AI workflow shift: he hasn't typed code since December, with 10 key takeaways for builders.
• Houda Nait El Barj shares how OpenAI's Codex produced strikingly elegant solutions diverging from human reasoning while remaining correct.
• Answer.AI asked "So Where Are All the AI Apps?" examining why practical AI applications haven't materialized as quickly as the technology.
• Ed Zitron published "The AI Industry Is Lying To You" analyzing what he sees as industry misrepresentations around AI capabilities and adoption.
• Networked Thought published "How I Keep Accidentally Building Tools" on the shortening distance between "I wish this existed" and "I guess I'll build it."
• Google Developers published a tutorial on building a smart financial assistant with LlamaParse and Gemini 3.1, showing how to extract data from complex unstructured PDFs and generate intelligent summaries.

💻 AI Coding & Developer Tools

• APEX-SWE is a new leaderboard ranking AI models on real professional software engineering tasks (GitHub). Sam Altman shared the results.
• SimWorld Studio lets you chat with Claude Code to build interactive physical worlds, place assets, test physics, and edit everything live —open-source.
• Anthropic published research on long-running Claude for scientific computing, showing how setting success criteria (like matching existing code to 0.1%) lets it iteratively build a JAX-based cosmological solver.
• Anthropic published a deep dive on harness design for long-running application development.
• Drift gives you an AI copilot for robotics simulations, described as "Claude Code for robotics" —no pricing details.
• VectorWare shipped Rust threads on the GPU, enabling standard Rust threading APIs to run on GPU hardware.
• New paper on effective strategies for asynchronous software engineering agents (code).

🔬 AI Research & Models

• Google Research introduced TurboQuant, a compression algorithm that reduces LLM key-value cache memory by at least 6x and delivers up to 8x speedup with zero accuracy loss.
• David Noel Ng published LLM Neuroanatomy II, revealing that duplicating seven middle layers in Qwen2-72B (no weight changes, no training) produced the #1 model on the HuggingFace Open LLM Leaderboard using a method called RYS discovered on a pair of RTX 4090s.
• Lucas Maes introduced LeWorldModel, a stable end-to-end JEPA that learns world models directly from pixels with no heuristics, EMA, or complex losses; just a 2-term loss achieving 200x token compression and strong 2D/3D results with 15M params on 1 GPU.
• Alexintosh ran Qwen3.5-35B (256 experts, 4-bit) fully on-device on an iPhone at 5.6 tok/s using SSD-streamed expert weights via a custom Metal inference engine, later optimized to 13.1 tok/s on iPhone 17.
• Martin Casado shared Vishal Misra's analysis "The Wall Between Shannon and Kolmogorov" arguing transformers identify program outputs via gradient descent but cannot reconstruct the underlying program or generalize beyond training data.
• KidGym benchmarks AI models on children's intelligence tests (Wechsler scale), finding top models still struggle with basic spatial reasoning (GitHub).
• New paper: "Detection Is Cheap, Routing Is Learned" argues refusal-based alignment evaluation fails because it misses the learned routing layer between concept detection and behavioral policy.
• GradMem writes context into memory via test-time gradient descent, outperforming forward-only memory writers while keeping model weights frozen.
• Applied Compute built a leverage-thresholding algorithm showing successes are 81x more valuable than failures at 10% success rate, slashing RL training compute while reallocating to higher-throughput sampling.
• Google Research introduced S2Vec, a self-supervised framework turning geospatial data into embeddings for predicting population density, carbon emissions, and urban development.
• General Reasoning launched OpenReward, an API serving 330+ RL environments with 4.5M+ tasks and an open standard (cookbook).
• GAIR Lab built daVinci-MagiHuman, an open-source 15B single-stream Transformer that jointly generates 5s 1080p video+audio in 38s on one H100 (paper, demo).
• Epoch AI reports AI solved a FrontierMath open research-level problem for the first time.
• Rockefeller University reengineered a mass spectrometer to process a billion molecules simultaneously instead of just a few.
• NUS researchers developed a swimming robot powered by lab-grown muscle that trains itself to record-breaking strength with no external stimulation.
• Arjun breaks down Exclusive Self Attention (XSA), arguing that standard self-attention wastes a huge chunk of capacity redundantly copying a token's own information back to itself when that information already flows through the residual connection.
• Oussama Zekri introduced GDDS (Generalized Discrete Diffusion from Snapshots), enabling richer noising processes beyond mask/uniform noise with strong empirical gains on OpenWebText (paper, code).
• Lucas Bandarkar proposes cross-lingual inconsistency (XICI) as a tool to localize knowledge in MoE LLMs, using statistical tests on expert activation for correct answers across languages (paper).
• alphaXiv compared Evolution Strategies vs GRPO for LLM fine-tuning on reasoning tasks, finding ES superior with limited data but GRPO better at scale or from base models (paper).
• Elan Barenholtz argues the new JEPA "LeWorldModel" is not a true world model but learns compressed data stream structure via action-conditioned prediction without causal mechanisms, despite promising probe results.
• AutoGaze proposes efficient video understanding via autoregressive gazing, reducing computation by selectively attending to frames (paper, models).
• New paper on AI-Generated Video Detection via Perceptual Straightening, studying how to detect synthetic video.
• New paper on End-to-End Training for Unified Tokenization and Latent Denoising.
• VLouvain: an efficient vector-based Louvain algorithm for massive low-rank graphs that eliminates explicit graph construction and scales to 1.57M+ nodes where prior methods failed due to memory (paper).
• Riemann-bench by Surge AI evaluates models on extreme-tier mathematical problems requiring deep reasoning, topological insight, and rigorous proof construction. Top scores: Claude Opus 4.6 and Gemini 3.1 Pro tied at 6% (paper).
• Xavier Gonzalez defended his Stanford PhD thesis on parallelizing nonlinear RNNs via Newton's method (DEER), showing they can be more efficient than transformers and more expressive than linear RNNs. Includes when parallelization succeeds based on the Largest Lyapunov Exponent (thesis, demo, defense video).
• Xinming Tu argues that a unified theoretical framework for structured test-time scaling via topology compression, scope isolation, and decoupled verification bypasses the linear collapse of long-horizon reasoning across multi-agent systems and coding agents.
• Packy McCormick argues the future is electromagnetic, sharing how Arena Physica is building electromagnetic superintelligence (EMSI) via a foundation model that produces "alien geometries" experts initially dismiss but that work when fabricated, with 18,000x solver speedups.
• Radical AI built a system that breaks complex computational chemistry workflows for screening million-candidate alloys into discrete task units with granular telemetry and intelligent search, making them reproducible across cloud infrastructure.

🏛️ AI Policy, Governance & Safety

• The U.S. Treasury launched the AI Innovation Series, a public-private initiative with four roundtables exploring AI in financial services. Treasury Secretary Bessent said the focus is shifting from constraint to recognizing "failure to adopt productivity-enhancing technology is its own risk."
• The CFTC launched an Innovation Task Force for clear rules in U.S. derivatives markets.
• OpenAI released open-source teen safety tools (gpt-oss-safeguard-20b), updated its Model Spec with Under-18 Principles, and introduced parental controls for ChatGPT.
• Spotify is testing a new tool to stop AI-generated music from being falsely attributed to real artists, giving musicians more control over which tracks are associated with their name.
• Databricks acquired two startups (Antimatter and SiftD.ai) to underpin a new AI security product, using its war chest from a recent $5B raise.
• A federal judge called the Pentagon's actions against Anthropic "troubling," saying "it looks like an attempt to cripple Anthropic."
• Meta was ordered to pay $375M for violating New Mexico law in a child exploitation case after a jury found it failed to safeguard its apps from child predators.
• Baltimore became the first U.S. city to sue Musk's xAI over Grok deepfake porn as legal pressure mounts.
• The State Department launched the Bureau of Emerging Threats to counter cyberattacks and AI risks from Iran and others.
• Anthropic published its 5th Economic Index report (PDF): users who've used Claude for 6+ months have a 10% higher success rate than new users, even controlling for task complexity and country. Experienced users iterate more, hand over less autonomy, and tackle higher-value tasks. Usage on Claude.ai diversified (top 10 tasks dropped from 24% to 19% of traffic), and coding migrated to the API via Claude Code.

💡 Industry Commentary & Analysis

• "Is anybody else bored of talking about AI?" hit the top of Hacker News with 358 points. The top comment argued that the most skilled engineers are getting the most value from AI because they already understand systems thinking; the thread devolved into a fascinating split between people having the time of their lives and people who are exhausted.

🛠️ AI Tools & Products

• Moda creates fully editable, on-brand slides, social posts, PDFs, and diagrams with AI on a real canvas you control —no pricing details.
• Highlight AI turns scattered Slack, email, docs, and meetings into one live shared mind for high-velocity teams —join waitlist.
• Superwhisper provides offline AI voice-to-text for Mac, Windows, and iOS. Andrej Karpathy was shown using it.
• Jotform AI builds forms instantly from a text description —free to try.
• Moreau is a GPU-native differentiable convex optimization solver for PyTorch and JAX, from the creators of CVXPY (announcement) —commercial licensing.
• Airweave Connect provides a pre-built, embeddable UI letting your users connect their own data sources with no auth UI to build —no pricing details.
• notepad.connie.surf is a local-first tool that lets you import references and drag-sort them into essay outlines —free.
• Luma's Uni-1 generates images with common-sense scene completion, spatial reasoning, and reference-guided controls. Lucas Crespo demonstrated extracting individual layers from complex compositions for separate regeneration.
• OpenBook uses AI agents to build interactive lessons around any topic, then an on-screen tutor talks you through the generated visuals, slides, and explanations while you ask questions as you learn —free.
• Paper Snapshot is a Chrome extension that instantly captures, summarizes, and organizes research papers —no pricing details.
• SP-6M is the largest database of high-resolution individual human head scans ever created, built by Ten24 (creators of 3D Scan Store) —no pricing details.
• Linq provides communications APIs for iMessage, RCS, SMS, and Voice with emoji reactions, voice notes, rich media, and typing indicators (raised $20M) —free sandbox.
• Zoer.ai builds production-ready full-stack apps database-first: describe your vision and it architects the schema, backend logic, and SEO-friendly frontend. From the creators of Chat2DB —no pricing details.
• GitAgent by Lyzr is an open standard that extracts your AI agent's config, logic, tools, and memory into a portable, version-controlled Git repo that runs on any runtime (Claude, OpenAI, CrewAI, OpenClaw) —free, open-source.
• EurekaClaw scrapes arXiv, generates theorems, writes full LaTeX papers with citations and figures, and runs ML experiments from your chat or terminal; works with Ollama or any major API (GitHub) —free, open-source.

🎙️ Interviews, Panels & Podcasts

• Latent Space published a must-listen interview with MIT's Prof. Heather Kulik on AI for materials science. Key insight: AI designed new polymers that turned out to be 4x tougher than expected by discovering a purely quantum mechanical effect that surprised the lab scientists who synthesized them. She also notes that materials science datasets are far behind biology, and that LLMs still can't design a 22-atom ligand that any expert could sketch in seconds.
• Waymo co-CEO Dmitri Dolgov discussed with Stripe how Waymo went from research to 500K rides/week across 10 cities, why fully autonomous cars won't evolve from supervised ones, and their custom-built vehicle that feels like a living room.
• Tina Huang published the best "how to run open source AI models" explainer we've seen, ranking four categories (local, browser/hosted, managed inference API, VPS) from easiest to hardest with two bonus advanced categories.
• Nick Saraev published a walkthrough of real ways to use Claude's new computer use for automation, from LinkedIn outreach to ad management to QA testing.
• Fireship broke down the Maven Smart System (the Pentagon's AI-driven military ops platform powered by Palantir and OpenAI) and "prompt & destroy" AI warfare.
• Hugo launched "The RL Spiral," an eight-part series tracing how reinforcement learning and neuroscience have spiraled around each other for over a century.
• Brendan Harris shared findings from his Nature Communications paper on how theta traveling waves in mouse visual cortex organize gamma packets and spiking to multiplex top-down and bottom-up signals.
• AltogetherAgile published "Humans in the Loop Is Not a Checkbox" applying the Cynefin framework (a decision-making model that sorts problems by complexity) to AI agent oversight.

💡 Industry Commentary & Analysis

• Jen Zhu agreed with Terence Tao that LLMs have structural limitations: they excel at brute-force combination testing but lack genuine creativity for Newtonian-style unification at the frontier.
• sophie argues Claude Cowork reveals humans were never meant to spend lives behind screens, and that the future of work will be cooler than we imagine as computers handle the computer work.
• AI Notkilleveryoneism Memes notes three events in one day: AI solved a FrontierMath open problem for the first time, Jensen Huang declared AGI is here, and Mark Gubrud (who coined "AGI" in 1997) agreed.
• Pawel Jozefiak argues agents need a persistent home on a dedicated machine, noting Perplexity, Meta, and Anthropic all converged on the same desktop-agent architecture in two weeks, and that memory (not context windows) is the remaining gap between task executor and true coworker.
• Freeman Jiang demoed live 2000-phone light-and-audio synchronization using Beatsyncgg over cellular data in a metal venue, achieving millisecond accuracy without Wi-Fi.

📊 Fundraising & Deals Roundup

• OpenAI — ~$10B from MGX, Coatue, and Thrive, bringing total latest round to ~$120B.
• Mirage (formerly Captions) — $75M from General Catalyst's CVF for AI video editing models. 3.2M+ downloads, $28.4M in-app revenue.
• Doss — $55M Series B (Madrona, Premji Invest) for AI inventory management plugging into ERP systems.
• Linq — $20M to scale conversational AI messaging APIs for iMessage, RCS, SMS, and Voice.
• Kleiner Perkins — $3.5B across new funds to bet on AI startups in software, healthcare, transportation, and autonomy.
• Halter — $220M for AI-powered collars for cows that help ranchers manage herds.
• Hummingbird — $800M in funds hunting "almost unemployable" misfits, after early bets on Kraken and Lovable.
• Swish — $38M for Bengaluru-based ultra-fast food delivery (third round in 18 months, valuation doubled in a year).
• Interloom — $16.5M to capture "tacit knowledge" (the undocumented processes in large enterprises, estimated at up to 70%) to power AI agents.

Previous Around the Horn Digests

Catch up on everything you missed:

• March 21-24, 2026: Claude got computer use and did grad-level physics, Cursor dropped Composer 2, Google shipped full-stack vibe coding in AI Studio, frontier models solved an open math conjecture, OpenAI acquired Astral and merged everything into a superapp, and 200+ more stories from Sunday through Monday.
• March 15-21, 2026: Claude Code hit 8% of worldwide GitHub commits, Nvidia's networking division went multi-billion, and 100+ stories from the week that wouldn't quit.
• March 8-13, 2026: Cursor built on Moonshot's Kimi, Anthropic shipped inline visuals, and a dancing robot went wild at a California restaurant.
• March 1-7, 2026: The Apple Experience launched nine products, MiniMax dropped M2.7, and the open-source model wars heated up.
• March 15-21, 2026: Claude Code hit 8% of worldwide GitHub commits, Nvidia's networking division went multi-billion, and 100+ stories from the week that wouldn't quit.
• March 8-13, 2026: Cursor built on Moonshot's Kimi, Anthropic shipped inline visuals, and a dancing robot went wild at a California restaurant.
• March 1-7, 2026: The Apple Experience launched nine products, MiniMax dropped M2.7, and the open-source model wars heated up.

That's a Wrap

That's 140+ stories from this week so far (and it's only Tuesday). If you made it to the bottom, you now know more about what happened in AI today than most VPs at most companies. Use that power wisely. Or at least smugly.

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 next week.

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