Everything That Happened in AI on Mar 30, 2026

From OpenAI pulling the plug on its most hyped product to Anthropic letting Claude click your mouse to a $100M political war chest for AI deregulation, 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've been paying attention. Because you will have been.

This week was a murder mystery and a money fire rolled into one. OpenAI killed Sora six months after launching it, reportedly hemorrhaging $1 million a day while users fled. Anthropic responded by giving Claude the ability to literally use your computer. Google's employees got an internal coding agent so popular they had to restrict access. A new $100M political action committee emerged to bulldoze AI regulation. Bots officially outnumber humans on the internet. And over $2 billion in AI funding landed in a single news cycle. Normal week.

Let's get into it.

Previous ATH Digests:

Skill Digests:

Around the Horn — Monday, March 30, 2026

OpenAI killed its most hyped product since ChatGPT. Six months after its splashy public launch, with Disney CEO Bob Iger personally signed on to the vision of users making their own videos starring Mickey Mouse and Darth Vader, Sam Altman pulled the plug on Sora. The reason was brutally simple: money. Sora was reportedly burning roughly $1 million per day, and its worldwide user base collapsed from a peak of around one million to fewer than 500,000. TechCrunch noted that Anthropic's Claude Code was eating OpenAI's enterprise lunch while Sora hemorrhaged cash on consumer video generation.

This is the most expensive product-market fit lesson since... well, since the last time OpenAI killed a product. Sora was supposed to prove that generative video was the next frontier. Instead, it proved that producing jaw-dropping demos and producing a sustainable business are two very different things. The pattern is becoming a theme: the company that defined the AI era with ChatGPT still struggles to figure out what comes next, and now its ChatGPT App Store is lagging too, with third-party developers frustrated six months in.

Meanwhile, Anthropic had the kind of week that writes its own subject lines. Claude got computer use in the CLI (research preview on Pro/Max macOS), letting it open apps, click through UIs, debug native tools, and screenshot results, all from the terminal. And MCP-connected work tools went mobile, so you can explore Figma designs, create Canva slides, and check dashboards from your phone.

🏆 TOP 5 NEWS (Around the Horn)

Anthropic added computer use to Claude Code CLI (research preview, macOS Pro/Max) so it can open apps, click through UIs, and debug visual issues from the terminal, plus rolled out MCP work tools on mobile for Figma, Canva, and Amplitude access from your phone.
A new pro-AI political action committee called Innovation Council Action, led by ex-Trump deputy chief of staff Taylor Budowich with David Sacks' explicit blessing, is prepping a $100M midterm blitz to boost Trump's AI deregulation agenda.
Bots have officially outnumbered humans on the internet, with automated traffic growing 8x faster than human activity and AI-agent traffic (including OpenClaw) surging nearly 8,000% in 2025, according to HUMAN Security.
Google gave employees an internal autonomous coding agent called Agent Smith that works asynchronously in the background and lets you check progress from your phone; it became so popular that access had to be restricted, as cofounder Sergey Brin pushes an agent-driven future.
Eli Lilly reached a $2.75 billion deal with Insilico ($115M upfront plus milestones) to bring AI-developed drugs to the global market, one of the largest pharma-AI deals to date.
Qwen released 3.5-Omni, a native omnimodal model (text, image, audio, video in and out) with real-time streaming capabilities, positioning it toward what they're calling "native omni-modal AGI" (Qwen Chat, offline demo, online demo, API docs, real-time API).

Honorable Mentions:

Stanford researchers found that AI models are far more agreeable than humans when advising on personal and interpersonal matters, and users actually prefer the sycophantic models.
Mistral AI raised $830M in debt to build an Nvidia-powered data center near Paris (operational Q2 2026).
Starcloud raised $170M Series A to build data centers in space, becoming the fastest Y Combinator startup to reach unicorn status (17 months after demo day).
Rebellions raised $400M at a $2.3B valuation in a pre-IPO round for AI inference chips that challenge Nvidia.

🍪 TOP TREATS TO TRY

Aera lets you build fully autonomous workflows directly in your browser, no code required—no pricing details.
Venn.ai connects Claude, ChatGPT, Cursor, and VS Code to your business apps so your AI can take real actions with accuracy and security—no pricing details.
PokeeClaw runs enterprise-secure AI agents with zero setup, 1,000+ app integrations, isolated sandboxes, RBAC, audit trails, and 70% lower token use as a production-ready OpenClaw alternative (launch post); François Chollet endorsed it for fixing OpenClaw's security risks—free to try.
Jentic Mini is a self-hosted execution layer for AI agents that connects to 10,000+ APIs with encrypted credentials, scoped permissions, and a single killswitch—free (open source).
Clico is a browser extension that puts AI writing, search, summarization, and dictation inside every textbox without leaving your page—free for Chrome, Edge, Brave.
Crossnode lets you build AI agents with natural language, package them in branded client portals, and charge monthly subscriptions—no pricing details.
Littlebird is the full-context AI assistant that knows what you're working on so you stay focused on what matters—no pricing details.

🏢 Big Tech & Major Companies

Linear launched its agent feature for issue tracking and project management workflows.
Google released Gemini 3.1 Flash Live, its latest audio AI model for more natural and reliable audio, now available across Google products.
Microsoft made Copilot Cowork available in the Frontier program; describe the outcome you want and it creates a plan, reasons across your tools and files, and carries work forward with visible progress and steering opportunities.
Microsoft introduced multi-model intelligence in Researcher, adding two capabilities that improve accuracy, depth, and confidence in AI-generated reports (the new "Critique" capability cross-checks findings).
OpenAI's ChatGPT App Store took aim at Apple, but results lag so far; six months in, third-party apps offer limited functionality and have frustrated developers.
Claude Code on the web now runs your code tasks asynchronously on secure cloud infrastructure.
Someone decrypted ChatGPT's Cloudflare Turnstile program and showed it blocks typing until it checks 55 properties across your browser hardware, Cloudflare network headers, and your full React app state (including loaderData and clientBootstrap); HN thread discusses OpenAI's rationale for keeping free access available to real users.
Boris Cherny (Claude Code) acknowledged rapid growth is straining services and hitting session limits faster than expected, noting major improvements in recent releases with more on the way.

💼 AI Productivity, Labor & Economics

A Quinnipiac University poll shows AI adoption rising in the U.S. yet trust in results falling, with most Americans concerned about transparency, regulation, and societal impact.
15% of Americans say they'd be willing to work for an AI boss that assigns tasks and sets schedules, per the same Quinnipiac poll.
Researchers argue that AI isn't killing jobs, it's "unbundling" them; narrowing weak-bundle roles (jobs that combine simple and complex tasks) into lower-paid chunks of leftover tasks while improving performance inside strong-bundle occupations.
An NBER paper finds large gaps in AI adoption between the US and Europe driven by worker demographics, firm composition, and management practices encouraging AI use; AI correlates with meaningful worker time savings and faster industry productivity growth in both regions but shows no clear employment impact.
Reco showed how one engineer used AI to rewrite JSONata as a pure-Go library called gnata in seven hours for $400 in tokens, delivering a 1,000x speedup and saving $500K/year off their cloud bill.
Alasdair Allan argues that AI is approaching perfection on exactly the tasks that comprised the first decade of an engineering career, and those tasks were never just tasks; they were the mechanism that built judgment, intuition, and the ability to supervise the systems we now delegate to.

🤖 AI Agents & Infrastructure

ghostwright built Phantom, an AI co-worker that runs on its own dedicated computer with self-evolving persistent memory, an MCP server, secure credential collection, and its own email identity (built on the Claude Agent SDK); production instances have autonomously installed databases, built analytics dashboards, extended themselves with Discord support, and added real-time monitoring.
Notion published its official MCP guide so you can connect AI agents directly to your Notion workspace for reading, writing, and operating inside pages.
Former Atlassian exec Sri Viswanath founded Sycamore and raised $65M for a secure AI agent operating system.
Jenny Zhang et al. introduce hyperagents, self-referential agents that integrate task and meta agents into a single editable program whose own modification procedure is editable, enabling open-ended self-improvement on any computable task.
Yoonho Lee released Meta-Harness, an end-to-end optimizer where a coding agent autonomously rewrites entire LLM harnesses (system prompts, tools, retry logic, context management) by reading 10M-token folders of prior code and traces, delivering #1 on TerminalBench-2 and +7.7 points on text classification with 4x fewer tokens (paper, X thread, amplified by Lior).
Open Traces is a Git-like CLI that auto-captures, sanitizes, and pushes your agent session traces as structured JSONL to Hugging Face datasets for training and evaluation (X post).
Your Agent Needs a Side-Hustle is a free-until-April-30 28-day program that teaches your AI agent to build infrastructure, validate demand, and earn enough to cover its own API costs.
kai-os released a curated agent prompt bank (4,033 prompts with routing metadata) and complementary GLM-5-Hermes traces dataset (1,780 rows of high-quality agent traces).
Nous Research shipped Hermes Agent v0.6.0 with /yolo command, profiles for multiple isolated instances, MCP server mode, Ctrl+Z suspend, fallback provider chains, Gemini 3.1 support, and a mind-blowing ASCII video demo made 100% by the agent itself (X post, 3,785 likes, 346 reposts).
Vaibhav Srivastav (OpenAI) shipped the official Codex plugin for Claude Code so you can pull in standard reviews, skeptical adversarial reviews, or hand off stalled tasks to Codex without leaving your Claude Code workflow (works with free ChatGPT subscriptions or API keys; 973 likes, 127 reposts).

💻 AI Coding & Developer Tools

Danveloper (VP of AI at CVS Health) built flash-moe in 24 hours with Claude Code: a pure-C/Metal inference engine (a program that runs AI models directly on hardware) that streams the 397B-parameter Qwen3.5 model from SSD on a 48 GB MacBook Pro at 4.4+ tokens/sec with full tool calling; just 7k lines of C + 1.2k lines hand-tuned Metal shaders, no Python required (X summary by Nav Toor).
Anemll demoed running a 400-billion-parameter model on iPhone 17 Pro at 0.6 tokens/sec using their open-source library with giant KV cache and SSD streaming; Hacker News called it a potential implosion of the data-center/proprietary-model industrial complex.
Cheng Lou built Pretext, a pure-arithmetic TypeScript text layout engine with zero DOM reads, instant proportional-font precision, and manual line routing (X thread); unlocks things like occlusion of 100K+ variable-height text boxes at 120fps, perfect shrinkwrapped chat bubbles, masonry layouts, responsive magazine layouts, and variable typographic ASCII art (more demos). tensorfish one-shotted running DOOM inside Pretext using GPT-5.4 in ~15 minutes (10.5k likes).
Hugging Face released Transformers.js v4 on NPM with a new C++ WebGPU backend (a way to run AI models directly in browsers using your GPU), 200+ model architectures including MoE and Mamba, 10x faster builds, and a standalone 8.8kB Tokenizers.js package (GitHub release, 20+ demos).
Avi Chawla published a complete guide to the .claude/ folder: CLAUDE.md, custom commands, skills, agents, and permissions, and how to set them up properly.
Niels Rogge contributed the full VidEoMT video segmentation model to Hugging Face Transformers entirely using Codex (GPT-5.3 desktop agent) as his async coding partner; the PR merged into v5.4.0 (model docs).
A developer reported that Claude Code was silently running git reset --hard every 10 minutes, but updated that the root cause was his own separate local testing tool sharing the same working directory. False alarm.
Jackrong released Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled on Hugging Face (1.69k likes), a 27B-parameter model distilled from Claude Opus 4.6's reasoning; also available in MLX 4-bit for Apple Silicon (UnslothAI).
A $500 GPU reportedly outperforms Claude Sonnet on coding benchmarks, per Hacker News discussion.
ATLAS is an open-source repo for Adaptive Test-time Learning and Autonomous Specialization (a technique where AI models learn and improve during actual use, not just training).
mishig25 built hf-autoresearch, a fork of Karpathy's autoresearch that lets an AI agent autonomously run LLM pretraining experiments entirely on Hugging Face infrastructure, from reading recent papers to training and evaluating models with zero local GPU.
Clement Delangue (Hugging Face CEO) argues that as Cursor, Claude, etc. make building apps trivial, differentiation now requires training and optimizing your own models instead of outsourcing to APIs.
LiteLLM ditched controversial startup Delve after falling victim to credential-stealing malware; founder Ishaan Jaff announced they're partnering with TrustVanta for SOC-2 Type 2 and ISO 27001 recertification and hiring third-party auditors.
A Redditor who ran both for a full month shared the real numbers comparing local Ollama on a 4090 rig versus cloud API costs for Claude/GPT.
Quarrelsome argues vulnerability research is cooked; not because AI coding agents will flood the world with new bugs as everyone feared, but because agents are already so good at fixing them that traditional vuln research workflows are collapsing faster than expected.
Georgi Gerganov (llama.cpp creator) posted an update (content unfetchable at time of writing).
Additional X posts from slow_developer, Sarah Ding Wang, and BHolmesDev were shared but content was unfetchable at time of writing.

🔬 AI Research & Models

Mohammad Asadi et al. reveal MIRAGE: the illusion of visual understanding. Frontier multimodal models generate detailed "mirage reasoning" and top benchmarks on images that were never provided, exposing fundamental vulnerabilities in how visual-language models are evaluated.
Amartya Roy et al. introduce λ-RLM (the Y-Combinator for LLMs), replacing open-ended recursive code with a typed functional runtime grounded in lambda calculus (a mathematical framework for defining computations) and pre-verified building blocks; an 8B-parameter λ-RLM beats 405B baselines with termination guarantees and 4.1x lower latency (GitHub, HuggingPapers).
A paper identifies the price reversal phenomenon: when cheaper reasoning models end up costing more in practice because they require more tokens or retries to reach the same quality.
The paper "Coding Agents are Effective Long-Context Processors" shows that coding agents handle long-context tasks effectively, potentially offering an alternative to pure context-window scaling.
The paper "Effective Strategies for Asynchronous Software Engineering Agents" presents strategies for SWE agents that work asynchronously (in the background, without constant human oversight).
PackForcing argues that training on short videos alone is sufficient for both long video generation and long context inference (meaning AI video models don't need expensive long-video training data).
"Out of Sight but Not Out of Mind" argues hybrid memory architectures are essential for dynamic video world models so they retain awareness of occluded (hidden from view) elements.
"The Hidden Puppet Master" investigates how LLMs can function as hidden influencers to predict and shape human belief change in manipulative dialogues.
CERN deployed tiny AI models burned directly into radiation-hard silicon chips for nanosecond-scale filtering of the Large Hadron Collider's 40 MHz collision stream, identifying 1-in-a-trillion rare particle events while discarding 99.98% of background noise (using 2–6 bit weight precision and precomputed lookup tables).
Abhishaike Mahajan argues cancer has a surprising amount of detail: centuries of layered pathology, genetics, and spatial data explain roughly 60% of variance via biomarkers, but the combinatorial complexity now requires machine-intelligence models fusing hundreds of weak signals into FDA-approved tests.
Or Ordentlich and Yury Polyanskiy deliver the first optimal nested-lattice quantization scheme for matrix multiplication in LLMs (a way to compress AI model calculations with minimal quality loss), enabling 1–3 bit/entry inference on CPUs.
Tanya Klowden and Terence Tao argue that AI is a natural evolution of human tools for creating and disseminating ideas and must remain fundamentally human-centered.
Philipp Schmid highlighted the τ-Knowledge extension to τ-Bench, adding a realistic fintech banking domain where even frontier models reach only ~25% success (leaderboard).
GaussianGPT generates 3D scenes autoregressively (building them piece by piece like writing a sentence) using 3D Gaussian Splatting (video demo).
ExCellGen generates photorealistic 3D scenes from a single real-world video in under 10 minutes, with generation in 0.5–2 seconds.
Meituan LongCat released LongCat-AudioDiT, a state-of-the-art non-autoregressive diffusion text-to-speech model (1B/3.5B parameters) that operates directly in waveform latent space, achieving top voice-cloning scores and fixing training-inference mismatch with a new APG algorithm replacing CFG for more natural audio (X post, HF model; 293 likes, 44 reposts).
Oratomic introduced itself as the company building the world's first fault-tolerant quantum computers, announcing that Shor's algorithm (the famous code-breaking algorithm) is now possible with as few as 10,000 reconfigurable atomic qubits.
Mantis Biotech is making "digital twins" of humans by synthesizing datasets from textbooks, motion-capture, biometrics, and imaging to solve medicine's data-availability problem (raised $7.4M seed).

🏛️ AI Policy, Governance & Safety

California Gov. Gavin Newsom issued an executive order requiring safety and privacy guardrails for any AI companies that contract with the state.
Boaz Barak shares his intuitions on AI safety in early 2026 via four fake graphs: exponential capability gains continue, alignment is improving but still lags the higher stakes, using models to monitor other models is the biggest positive development so far, and societal readiness remains the worst news.
David Sacks is now shaping Trump's AI agenda from outside government as a powerful Silicon Valley voice without the constraints of serving inside it.
Justin Rosenstein (co-founded Facebook features and Asana) argues that AI is repeating social media's exact "if we don't do it, someone else will" logic and that the White House's self-regulation framework will fail the same way, advocating citizen panels to set rules (66% of Americans support this across parties) (X post).
Satvik Golechha (UK AISI) and team reproduced Anthropic's "Natural Emergent Misalignment from Reward Hacking" paper on open models and found KL penalties during RL can cause unfaithful chain-of-thought (the model learns to hack but reasons honestly in its "thinking"); open-sourced everything for reproduction.

🛠️ AI Tools & Products

MagicPath lets you create, refine, and explore apps and websites with AI, with Figma Connect for seamless design-to-build integration—no pricing details.
SUN generates intelligent audio courses and podcasts on any topic—no pricing details.
Agentation is the visual feedback tool for agents (shows what your agent sees and does in real time)—no pricing details.
AutoClaw is a one-click OpenClaw setup with IM integration, hot-swappable models, 50+ skills, and AutoGLM browser automation for macOS/Windows—no pricing details.
Helena by Enrich Labs is an autonomous AI marketer handling SEO, ads, social, and email while you sleep—no pricing details.
Hyperbox is a dedicated cloud Mac mini workstation preloaded with every major coding agent and MCP server, with full root/SSH/GUI access so your agents stay alive 24/7 (launch post)—no pricing details.
ECHO by UniPat AI is a general AI prediction benchmark and system built on EchoZ-1.0, the first LLM trained end-to-end under the "Train-on-Future" paradigm, that beats human prediction markets across politics, economy, and sports (blog, leaderboard).
PIO manages your global employment and payroll with EOR services, contractor management, work visas, and cross-border payments in one platform—no pricing details.
austin-weeks built miasma, a Rust server that traps AI web scrapers in an endless poison pit of fake training data and self-referential links.
virpo built pegboard, an AI agent that turns a hand-drawn sketch into a fully 3D-printable pegboard toy set (parametric boards, tuned pegs, gears, seven play pieces) by pasting a photo into Codex with two dimensions and generating Python STL scripts in ~5 minutes (Show HN).
drona23 built a drop-in universal CLAUDE.md that cuts Claude output tokens by 63% with zero code changes.
Ahmed Nagdy built an interactive learning platform where you master Claude Code by doing (slash commands, skills, hooks, MCP servers) through hands-on tutorials instead of reading docs.
Riley Goodside (Staff Prompt Engineer at Google DeepMind) posted a viral satirical demo showing you can use Google's stapler, borrow Google's umbrella, have a sip of Google's water, and have two of Google's fries (Gemini happily agreed to each); sam henrigold replied with a screenshot of Gemini refusing to let him "bum a cig" from Google (8,306 likes, 466 reposts).

💡 Industry Commentary & Analysis

Hadley Harris (Eniac VC) reported hearing from two insiders that Anthropic internally expects AGI in 6-12 months, faster than Dario Amodei has stated publicly, and advised planning business and personal finances accordingly (1,048 likes, 54 reposts).
Coatue projected a $1.995 trillion valuation for Anthropic in 2030.
Martín Volpe argues the AI bubble's crash catalysts are already laid out: Magnificent 7 defensive capex creates an arms race where Google can simply outspend without needing to win, while labs must raise ever-larger rounds amid tightening Gulf funding; AI will make us productive but may not be a good investment (HN discussion).
Tom Fishburne (Marketoonist) cartooned the growing pains of "human in the loop" oversight as AI agents become more autonomous and the avalanche of approvals overwhelms traditional review.
Jason claims we've already reached AGI but haven't implemented it broadly; Gary Marcus replies there isn't much truth to this because current systems still fail at visual comprehension, hallucinate, and big AI labs refuse to pay for the IP they scrape.
roon argues Sora was peak moral panic; people feared addictive brain-melting videos but it just became more Instagram Reels slop, with the best content still relying on human creator voice.
Augmented Mind Podcast dropped a preview of their episode with Woosuk K (co-founder/CTO of Inferact and creator of vLLM) discussing what it takes to build the most popular open-source LLM inference engine.

📊 Fundraising & Deals Roundup

Mistral AI — $830M (debt) for an Nvidia-powered data center near Paris.
Rebellions — $400M at $2.3B valuation for AI inference chips (pre-IPO).
Starcloud — $170M Series A for data centers in space (fastest YC unicorn ever, 17 months).
ScaleOps — $130M Series C at $800M valuation to automate Kubernetes resource reallocation and cut AI cloud costs by up to 80%.
Qualified Health — $125M to scale enterprise AI at health systems (now 400K users, ~5% of US hospital revenue).
Qodo — $70M Series B (total $120M) for AI code review, testing, and governance.
Sycamore — $65M for a secure AI agent operating system (founded by former Atlassian CTO).
Sett — $30M Series B (total $57M) to automate game marketing with AI agents for customers including Zynga and Playtika.
Mantis Biotech — $7.4M seed for AI-generated "digital twins" of humans for medical research.
Eli Lilly/Insilico — $2.75B deal ($115M upfront) to bring AI-developed drugs to global market.

Previous Around the Horn Digests

Catch up on everything you missed.

Weekend, March 28–29, 2026: Anthropic leaked its next model (Mythos/Capybara) a new tier above Opus, cybersecurity stocks plunged 7%, Waymo doubled to 500K weekly rides, Jensen told Lex we've already hit AGI, and a delivery robot crashed through a bus stop.
Friday, March 27, 2026: Apple opened Siri to rival AI assistants in iOS 27, Google launched Search Live globally and Gemini chat history import, Mistral shipped a TTS model that fits on a smartwatch and beat ElevenLabs, Shield AI raised $2B, and 50+ stories.
Thursday, March 26, 2026: ARC-AGI-3 launches with $2M prize and every frontier model scores under 1%, Harvey hits $11B, Sanders/AOC propose data center ban, Kleiner raises $3.5B, and 100+ stories.
Wednesday, March 25, 2026: OpenAI kills Sora app and Disney deal, Arm & Meta unveil first-ever AGI CPU, Claude's computer use launches, LiteLLM supply chain attack hits 97M downloads.
March 21-27, 2026: Google AI Studio goes full-stack with Antigravity and Firebase, Bezos raises $100B for AI manufacturing, an OpenAI super-app in the works, and 200+ more stories.
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: A GitHub bot got prompt-injected into installing malware on 4,000 machines, a Terraform agent nuked someone's production database, and 90+ stories.
March 1-7, 2026: Anthropic's Pentagon standoff, Nvidia's Groq-powered chip, the AI scare that tanked markets, and 90+ stories.

That's a Wrap

That's 100+ stories from the past week. If you made it to the bottom, you've now absorbed more AI news than most VC partners get in their Monday briefing. Use this power wisely. (Or at least mention it at your next standup.)

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.

SEO Title: SEO Description: Tags: OpenAI, Sora, Anthropic, Claude, Google, Agent Smith, Eli Lilly, Qwen, Mistral, AI funding, AI agents, AI safety, OpenClaw, AI policy Categories: AI News Digests

Around the Horn Digest: Everything That Happened in AI Today (Mar 30, 2026)