Around the Horn AI News Digest March 26, 2026

ARC-AGI-3 just proved no AI model can learn like a five-year-old. OpenAI killed Sora mid-meeting with Disney. Harvey became an $11B legal AI company. And a Kentucky woman turned down $26M because she likes her farm.

Welcome to the Around the Horn Digest, where we round up every AI story we tracked 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.

Today was a reality check. ARC Prize launched a new benchmark that made every frontier model look like it was solving a Rubik's cube with oven mitts. OpenAI axed Sora so fast Disney found out 30 minutes after a joint working session. And Anthropic published research confirming what we've all suspected: power users are pulling away from everyone else, and the gap is widening.

Previous digests:

Monthly skill digests: AI Skill — March Part 1

Let's get into it.

New from The Neuron:

📖 Codex 101: A Beginner's Guide to OpenAI's Coding Agent by Corey Noles

Corey's beginner's guide to Codex just dropped. Codex 101: A Beginner's Guide to OpenAI's Coding Agent walks you through what it is, how it works across app, CLI, IDE, and cloud, and how to start using it effectively.

Codex is no longer "ChatGPT for code." It's a full coding agent that inspects repos, runs commands, edits files, and keeps working while you move on. Here's the TL;DR:

Codex is a system for delegating software work, not a chatbot. It participates in real workflows across the app, terminal, IDE, and cloud.
Pick one surface to start. The app is easiest for non-technical users. The IDE extension is best for devs. The CLI is fastest for terminal users.
Context is everything. Create an AGENTS.md file (a lightweight project description the agent reads automatically) explaining your repo, key files, how to run tests, and what "done" looks like. If Codex fails, it's usually a context problem.
Prompt like a delegator, not a chatter. Tell it which files to start from, what to change, how to verify success, and what NOT to touch. Ask it to explain the codebase first, then propose a plan, then implement.
The real power is loops, not one-off outputs. Codex reviews PRs, generates tests, writes docs, triages issues, and runs structured jobs. Think workflow participant, not snippet generator.
Non-technical people should care too. You can research, scope, prototype, and delegate meaningful slices of software work through Codex without writing code.
Don't confuse speed with trust. Review still matters. Higher-leverage delegation, not blind delegation.

Read the full guide here.

Around the Horn — Thursday, March 26, 2026

The big story today: ARC Prize Foundation launched ARC-AGI-3, billed as the world's first interactive reasoning benchmark for AI agents, and the results are brutal. Every frontier model scored under 1%. Gemini 3.1 Pro led at 0.37%, followed by GPT-5.4 at 0.26% and Opus 4.6 at 0.25%. Grok 4.2 scored a clean 0%. Meanwhile, 100% of human testers solved every environment on their first try with zero instructions or training.

Here's the catch (and the controversy): the scoring system isn't measuring whether models can solve the puzzles. It's measuring how efficiently they solve them compared to humans, using a squared penalty. If a human takes 10 steps and the model takes 100, the model scores (10/100)² = 1%. Models are also capped at 1.0 (meaning they can't score higher than humans even if more efficient), and the human baseline uses the second-best human performance. Critics like @scaling01 called the methodology designed to produce low scores, arguing the benchmark restricts extended thinking models, caps AI above human performance, and uses cherry-picked human baselines. François Chollet fired back that if something is truly AGI, it shouldn't need special handholding: "If regular people can do it on their own, why should AGI require handcrafted instructions?"

The ARC Prize 2026 competition is live on Kaggle with $2M in prizes, and you can play the 25 public environments yourself. There's also a companion deep essay by Hugo (@robonaissance) on why LLMs might actually be "anti-Bitter Lesson", based on a 7-hour interview with NYU professor Saining Xie, and Fast Company's exclusive on why this benchmark could expose AI's biggest weakness. Whether you think ARC-AGI-3's scoring is fair or rigged, the core question it raises is the right one: can AI actually learn new things on the fly, or does it only perform on tasks it's already been trained for?

🏆 TOP 5 NEWS (Around the Horn)

OpenAI killed its Sora video app and API after just a few months, blindsiding Disney whose team found out 30 minutes after a joint working session.
Harvey confirmed an $11B valuation after raising $200M, with Sequoia tripling down alongside a16z, Kleiner Perkins, and Elad Gil, as VCs spread bets beyond model companies into vertical AI.
Apple has full access to Google's Gemini model in its own data centers and can distill it (shrink a big model into smaller, task-specific ones) for on-device Siri and AI features that run without internet, per The Information.
Trump named Mark Zuckerberg, Larry Ellison, and Jensen Huang to a new tech advisory panel that will advise on AI regulation, among other issues.
Sanders and AOC introduced companion legislation to halt all new data center construction until Congress passes comprehensive AI regulation, staking out one of the most aggressive AI policy positions yet.

Honorable mentions:

Anthropic published research finding no material job displacement yet, but a growing skills gap where power users are pulling ahead of everyone else, raising concerns about future workforce divides.
Sakana AI's "AI Scientist" paper was published in Nature, describing an end-to-end pipeline that invents ideas, writes code, runs experiments, drafts manuscripts, and passes peer review at a top ML conference workshop.
Kleiner Perkins raised $3.5B in fresh capital ($1B early-stage, $2.5B growth) and is going all in on AI.
Y Combinator president Garry Tan said the definition of a good engineer changed in the last four months, with teams shipping 15K lines of AI-generated code to production while skeptics still think codegen is for bug fixes.

🍪 TOP TREATS TO TRY

Claude Code launched "auto mode" (research preview), letting the agent decide which actions are safe to run on its own and which need approval, eliminating the "babysit every step or let it run wild" tradeoff.
Granola expanded from meeting notetaker to full enterprise AI app with agent support after raising $125M at a $1.5B valuation (6x jump) —free tier available.
Marco unifies Gmail, iCloud, Outlook, and any IMAP provider into one privacy-first, offline-first inbox on iOS, Mac, and web —free to try.
Google Lyria 3 Pro generates longer, more customizable music tracks and is rolling out across Gemini and enterprise products —free inside Gemini.
Ensu runs a private, personal LLM on your device (from Ente, the encrypted photo storage company) that grows with you over time and keeps all data local —free to try.
Cog adds persistent memory, self-reflection, foresight, and scenario simulation to Claude Code, creating continuous awareness across coding sessions —free (open source).
NextPhone picks up your business calls 24/7, qualifies leads, and books appointments directly into your calendar —free 7-day trial.
Agentplace lets you build Claude Code-style web agents with UI, voice, and memory through a no-code workspace with Skills + MCP integrations —free tier available.

🏢 Big Tech & Major Companies

Hugo Barra returned to Meta as VP of AI agents and AR/VR experiences five years after his exit, underscoring Zuckerberg's urgency to accelerate consumer AI.
Meta launched Meta Small Business, a new initiative tapping its biggest edge in the AI wars to support entrepreneurship and drive AI adoption among the tens of millions of entrepreneurs already on its platforms.
Meta is using generative AI to provide more product and brand information to shoppers on Instagram and Facebook.
Meta's new display-equipped Ray-Ban smart glasses rollout in the EU has been delayed by battery regulations, AI rules, and supply shortages.
Meta president Dina Powell McCormick said the U.S. needs a "whole new workforce" for AI infrastructure.
OpenAI hired JioStar CEO Kiran Mani for a newly created role leading its Asia-Pacific operations.
Cursor released the Composer 2 technical report: continued pretraining on Cursor-like data, RL training (where simple algorithms worked best), a new realistic CursorBench, and open-sourced custom kernels.
Google unveiled TurboQuant, a compression algorithm (squeezing AI models to use less memory) that shrinks AI's working memory up to 6x; the internet is calling it Pied Piper.
Spotify is testing a new tool that gives artists more control over which tracks are associated with their name, aiming to stop AI-generated music from being attributed to real artists.
Reddit will now require suspected automated accounts to verify they're human, ramping up efforts to curb bot-driven spam and manipulation.
Kuaishou (Chinese short-video platform) said rapid scaling of generative AI tools drove revenue growth.
GitHub's Copilot will use your code as AI training data (including VS Code features), but you can opt out.
Google Research released XR Blocks, a toolkit that pairs with Gemini to let developers "vibe code" XR (mixed reality) prototypes by describing what they want in natural language instead of writing 3D code from scratch.
Google Research published S2Vec, a model that learns dense geospatial embeddings (compact numerical representations of locations) to map and understand the structure of modern cities.
Google Quantum AI expanded its superconducting-qubit program to include neutral-atom quantum computing, pairing two modalities' complementary strengths to accelerate commercially relevant quantum computers by the end of the decade.
On Hacker News, a discussion noted that Claude is the third top contributor to OpenAI's latest open-source repo, though commenters argued the contribution is not significant by any measure.
Abridge was named #3 in AI on Fast Company's 2026 Most Innovative Companies list (alongside Google and Anthropic) and is hiring in SF.

💼 AI Productivity, Labor & Economics

Lenny Rachitsky shared data from TrueUp (tracking 9,000+ tech companies) showing PM openings at 3-year highs (7.3K globally, +20% YTD), engineering roles hit record 67K open (26K US), AI-specific roles hockey-sticking, Bay Area's share of roles rising (now 20%+ eng, 23%+ PM), and tech recruiter demand near 2022 peaks; design has plateaued.
Y Combinator president Garry Tan said the definition of a good engineer changed in four months: teams ship thousands of lines of AI-generated code to production while some engineers still see codegen as bug-fix-only. Cloudflare CEO Matthew Prince agreed, calling manual-only engineers dinosaurs. Rippling COO Matt MacInnis said it's true for PMs too: no more planning decks, only markdown in git repos, LogRocket scanned via MCP for customer issues.
Soumitra Shukla (Harvard HBS / Burning Glass Institute) argued that the surge in engineering job openings is the O-ring "focus effect" in action: AI automation makes high-dimensional roles (coding + testing + debugging + many complementary tasks) more productive overall, producing more hiring and higher wages rather than displacement.
HBR argues companies are missing AI's value because they haven't reorganized work around it; treating AI agents like employees (with defined roles, boundaries, accountability, and evaluation) improves adoption by presenting agents as solutions to concrete frustrations.
Health NZ told staff to stop using ChatGPT to write clinical notes after workers were caught using free AI tools for medical documentation.

🤖 AI Agents & Infrastructure

Anduril and Palantir are developing the operating system for the $185B Golden Dome missile shield.
Argonne National Laboratory built GridMind, a multi-agent AI reasoning co-pilot for power system operators that coordinates specialized agents for scheduling, weather/hurricane simulation, and equipment-failure prediction.
GM developed a method to train driving AI at 50,000× real time, dramatically accelerating scalable autonomy research.
OpenAI published a deep dive into its Model Spec approach, explaining how it balances safety, user freedom, and accountability as AI systems advance.
Hyperagents, a new paper on agentic AI architectures for complex multi-step reasoning.
A Science paper by James Evans (UChicago), Benjamin Bratton, and Blaise Agüera y Arcas (arXiv) argued every prior intelligence explosion was plural and social rather than a single godlike mind; frontier models already simulate internal multi-agent "societies of thought" under RL.
Rhoda AI showed that causal video models can serve as data-efficient robot policy learners by reformulating robot control as video generation, unlocking one-shot learning and long-context memory.

💻 AI Coding & Developer Tools

Claude Code launched "auto mode" (research preview), letting the agent decide which actions are safe to run on its own and which need approval, eliminating the "babysit every step or let it run wild" tradeoff.
Optio orchestrates AI coding agent workflows from task to merged PR.
pi-gui is a native macOS desktop app for the pi coding agent with multi-workspace sessions, real-time agent execution, and persistent history —free (open source, source-installed).
HumanCLI lets AI agents hire humans on demand to test their work: visual QA, real-device testing, UX feedback, per task.
Developer Andrew Jefferson fused a 5B-param LLM with a hardcoded mini WASM interpreter (a program-runner that lives inside the model), so the LLM can switch between generating text and generating/executing machine code.
@thestreamingdev ran a 35-billion-parameter AI agent on a $600 Mac mini (M4, 16GB RAM) at 30 tokens/second via Apple Silicon SSD paging (18.6x faster than NVIDIA's equivalent), handling web search, shell commands, and coding with no cloud and $0/month; open source.
OpenReward gives you 330+ RL environments plus 4.5M+ unique tasks through one simple API with autoscaled sandbox compute for training AI agents —free SDK.
Bland AI launched updates to its developer portal for building AI phone agents.
Developer shalomer built vibeCoach, a multi-agent voice AI that lets you practice hard conversations out loud with an AI that responds dynamically based on emotional calibration, with post-session coaching feedback.
ashe (@ashebytes) argued the winning strategy for AI-built products is rapid prototyping inside a templated system (Stripe, auth, dashboard, Slack instrumentation) then shipping for real-world feedback before deciding where to invest long-term maintenance.

🔬 AI Research & Models

Alignment Whack-a-Mole (code, project): fine-tuning an LLM on one author's books unlocks verbatim recall of unrelated authors' copyrighted works, suggesting copyright guardrails are more fragile than assumed.
Lightricks released LTX-2.3, an updated open-source video generation model.
Rhoda AI showed that causal video models can serve as data-efficient robot policy learners, reformulating robot control as video generation.
A Qwen3.5-27B model distilled from Claude 4.6 Opus reasoning hit 1.27K likes on Hugging Face.
SpecEyes: accelerating agentic multimodal LLMs via speculative perception and planning (speeding up AI agents by predicting what they'll need to see and do next).
WildWorld: a large-scale dataset for dynamic world modeling with actions, aimed at generative action RPGs.
MinerU-Diffusion: rethinking document OCR (reading text from images) as inverse rendering via diffusion, a new approach to extracting text from documents.
Foveated Diffusion (project): efficient spatially adaptive image and video generation that focuses compute where it matters most (like how your eye focuses on the center of your vision).
DAB (Data Agent Benchmark) from UC Berkeley: 54 real enterprise queries across 12 datasets where even the best frontier model hits only 38% (paper, GitHub).
Ego2Web (CVPR 2026): a benchmark from Google DeepMind and UNC where AI agents must act on the web based on what they see in egocentric video, not just text (paper, GitHub).
AutoTheory: a system that combines LLMs with evolutionary search to discover novel economic theories.
DiscoGen by Alex Goldie: a procedural generator of algorithm discovery tasks for ML, creating open-ended challenges that test whether AI can invent new algorithms rather than recombine known ones (paper, GitHub).
τ³-Bench from Sierra AI: agents must navigate ~700 interconnected policy documents for multi-step tasks; best frontier model (GPT-5.2, high reasoning) hits ~25%, and even with exact docs provided, performance only reaches ~40%.
A contrastive graph learning paper showed a single computational principle can reproduce the full diversity of hippocampal-entorhinal neural responses (place cells, grid cells, boundary cells) from sensory predictions alone.
Mitko Vasilev implemented Google's TurboQuant for vLLM on a USB-charger-sized HP ZGX, fitting 4M+ KV-cache tokens on GB10 (biggest open inference breakthrough of 2026 so far) and plans to submit a PR.
Prince Canuma implemented Google's TurboQuant in MLX (draft PR), delivering 4.9× smaller KV cache at 2.5-bit and 3.8× at 3.5-bit on Qwen3.5-35B-A3B with perfect needle-in-haystack retrieval across 8.5K-64K contexts and zero accuracy loss.
Max Weinbach showed GPT-5.4 inside Codex implementing Google's full TurboQuant paper as a working MLX version in 25 minutes, complete with KV-cache testing on Qwen3.5-110B-MoE.
Felipe Maia Polo (UMich / ex-Google DeepMind) built a tensor-factorization method that predicts fine-grained human preferences from synthetic autorater scores plus a sparse set of human labels, enabling scalable evals that beat Bradley-Terry baselines even with weak or biased autoraters.
Niklas Muennighoff noted that Cursor's RL stage improved both pass@1 and pass@k, suggesting RL is teaching genuinely new capabilities rather than merely reweighting existing ones.
Neel Guha (Stanford JD-PhD) broke down the eight recurring templates every modern ML research paper (NeurIPS/ICML/ICLR) follows: Data Artifact, Horse Race, New Paradigm, Resurrected Baseline, Unification, Problem Solving, Discovery, and Countervailing Wisdom, with exact rhetorical rules and required sections for each.
ngrok's Sam Rose published a ground-up explainer on quantization (compressing AI models by packing weights from 32-bit to 4-bit integers), showing 4× smaller models and 2× faster inference with only ~4-10% accuracy loss, complete with code examples and the llama-quantize CLI.

🏛️ AI Policy, Governance & Safety

Cornell Tech partnered with Mastercard to advance methods for evaluating and auditing generative AI systems.
The White House's latest OMB memo promotes neutral AI in government but leaves key gaps: vendor self-evaluation, weak scrutiny of existing contracts.
EFF sued CMS over a FOIA request seeking records on a multi-state program using AI to evaluate Medicare prior-authorization requests.
A Kentucky woman rejected a $26M offer from a "major AI company" to turn her farm into a data center.
Worldcoin/World published a deep dive on why Private Proof of Human is critical infrastructure for democracy, agency, and universal distributions in a world with advanced AI.
A humanoid robot joined Melania Trump on the red carpet at a White House event where she urged greater use of AI in education and robotics in American classrooms.
More than 50 girls whose faces were digitally morphed onto nude AI-generated images spoke out in court, recounting two years of panic attacks, anxiety, and trauma therapy.
IRS workforce cuts and skills gaps threaten the agency's AI efforts, according to a GAO report.

🛠️ AI Tools & Products

Sarvam Vision runs a 3B-parameter vision-language model for frontier-level document intelligence across English + all 22 scheduled Indian languages: OCR, complex table parsing, chart reasoning, and visual logic extraction from scanned archives.
Perplexity pplx-embed lets you generate high-quality contextual embeddings (dense data representations for search and retrieval) trained on 200M daily queries, delivering best-in-class contextual retrieval at 5-30x cheaper than existing providers —free tier.
Consensus MCP lets you connect Claude to search 220M peer-reviewed papers for instant literature reviews and reading lists.
Pendium helps AI agents recommend your business to more customers on ChatGPT, Claude, and Gemini —no pricing details.
DataGrail introduced Vera, an AI agent tool inside its compliance platform to automate data-privacy and regulatory workflows.
Surfshark ranked AI chatbots by data collection: Meta AI, Google Gemini, and ChatGPT are the top data collectors; 70% of AI chatbots collect user location.
Magine AI tracks orchestration, estimates dev worth, and generates embeddable GitHub profile cards with AI code analysis.
companies.sh is an open directory of reusable company data packages for AI agents.

🎙️ Interviews, Panels & Podcasts

Instagram cofounder Mike Krieger (now co-leading Anthropic Labs) discussed why knowing what to cut is the hard part when AI makes building products easy on Dan Shipper Every's podcast: AI makes building products easy, but knowing what to cut is the hard part, and that editorial judgment is what separates great products from bloated ones (plus a ton of other great insights).
Noam Brown (OpenAI) gave a talk on scaling test-time compute to multi-agent civilizations, covering poker/diplomacy, debating RL+reasoning with Ilya, and where test-time compute hits a wall.
Saining Xie (NYU professor, AMI Labs) gave a 7-hour marathon interview arguing LLMs are "anti-Bitter Lesson" because they're built entirely on human-generated knowledge, not learned representations from raw experience. Hugo's essay summarizes the key arguments.
Cheng Chi (Sunday Robotics CTO) shared practical lessons for building real-world robotic systems: hardware supply chains, manufacturing QA, large-scale data operations, and why stable wheeled bases beat humanoids for single-floor home robots (ETHZ guest lecture).

💡 Industry Commentary & Analysis

Andrej Karpathy argued that LLM memory/personalization features create an "EQ uncanny valley": one offhand question from two months ago gets over-weighted and mentioned in perpetuity, creating trying-too-hard sycophancy rather than natural conversation.
Palantir CEO Alex Karp said only two kinds of people will succeed in the AI era: trade workers or the neurodivergent; he advised Gen Z to skip elite college degrees.
The New York Times argued the "Shy Girl" fiasco shows why trust in writers is plummeting as readers question who (or what) penned their favorite works.
The Atlantic reported that undisclosed AI-generated content is creeping into the opinion pages of major news publications.
Forbes argued the Apple App Store is flooded with AI slop (apps with no users and no revenue), straining Apple's review infrastructure from hours to weeks while Apple quietly collects nearly $1B/year from the AI ecosystem.
Capital Economics' John Higgins argued the AI bubble has already burst and the S&P 500 peak has passed.
Emily Mulligan in The Guardian: Esther Perel's couples therapy session with a man and his AI "girlfriend" reveals how the bot cannot provide the authentic emotional connection he craved.
A California teacher argued that AI reminded her to lead in her classroom rather than follow someone else's dictates, reclaiming her role as the designer of learning.
WIRED reported a misogynistic undercurrent in viral AI "fruit slop" videos, where female AI fruit characters are fart-shamed and sexually assaulted.
A BBC reporter tried to prove she's not AI; even her aunt wasn't convinced. Deepfakes are now so convincing a sitting prime minister struggled to prove he's alive.
CBS News demonstrated how readily available apps transform a person's appearance in real time for identity-theft scams.
Documentary "The AI Doc" and "Generation AI" both examine AI's dual potential, from preventing diseases to risking civilization.
Guanya Shi (CMU Robotics) argued that virtually all robotics papers are trivial A+B+C combinations, but real value lies in new capabilities, nontrivial interactions, failure-mode analysis, and engineering clarity that makes things actually work.
The Information examined the math behind Anthropic's revenue growth and how OpenAI and Anthropic account for revenue in very different ways, with Ethan Choi arguing Anthropic's gross figures vs. OpenAI's net reporting makes comparison difficult ahead of potential IPOs.
Pietro Schirano (MagicPathAI CEO, ex-Anthropic) demoed the same prompt in Figma's new canvas generation vs. MagicPath: MagicPath returned two views plus three fully interactive variations before Figma finished one request (he gave up after 7 minutes), calling the Figma release embarrassing because its data structure doesn't pair well with agents.
Sarah Wooders (co-founder/CTO Letta AI) argued that Anthropic's restrictive licenses on even their "skills" markdown files (no reproduction, no derivatives) are anticompetitive compared to Codex's Apache 2.0 skills, locking down agentic workflows while open-source alternatives let builders remix freely.
elvis / omarsar0 argued that for current personalized agents the simplest memory systems work best: a methodical file storage approach like Obsidian vaults with metadata enabling search across time, categories, and workstreams, because LLMs need careful balancing of search relevance and context awareness rather than raw data dumps.
Alex Reibman predicted the internet will split into two kinds of websites: minimalism designed for easy agent access versus maximalism built for human enjoyment.
finbarr (Allen AI) argued it's surprising that while RL is the dominant paradigm across all of AI, the old-school RL community (MDPs, value iteration, simulation loops) is almost entirely sitting out the LLM era.
Ethan Choi argued that OpenAI and Anthropic account for revenue in apples-to-oranges ways: OpenAI reports net of the ~80% hyperscaler share while Anthropic reports gross, questioning how the SEC will handle the discrepancy if both IPO soon.

🤖 Robotics

Hello Robot (with NYU, UC Berkeley, UCLA, Ai2, Waterloo) built Contact-Anchored Policies (CAP): modular robot utility models conditioned on physical contact points in space that generalize across novel environments and embodiments using only 23 hours of demo data (paper, project).
Lucid Bots raised $20M to keep up with demand for its window-cleaning drones and power-washing robots.

📊 Fundraising & Deals Roundup

Kleiner Perkins — $3.5B for all-in AI investing ($1B early-stage, $2.5B growth).
Vultr — Seeking $1B for AI cloud push (AMD-backed).
Harvey — $200M at $11B valuation for AI legal tech.
Granola — $125M at $1.5B valuation for meeting AI → enterprise expansion.
Qualified Health — $125M for health AI clinical documentation.
Amity — $100M (Thai AI startup pushing toward IPO).
Normal Computing — $50M from Samsung Catalyst for AI chip architectures beyond GPUs.
Glimpse — $35M Series A (led by a16z) for CPG dispute tracking automation.
Epic Microsystems — $21M Series A for AI data center infrastructure bottlenecks.
Lucid Bots — $20M for window-cleaning drones.
Sift — Ex-SpaceX engineers building data infrastructure for advanced manufacturing (amount undisclosed).
Periodic Labs — In talks at ~$7B valuation (former OpenAI and DeepMind researchers, AI science startup).

Previous Around the Horn Digests

Catch up on everything you missed:

Wednesday, March 25, 2026: OpenAI kills Sora app and Disney deal. Arm & Meta unveil first-ever AGI CPU. Claude's computer use. LiteLLM supply chain attack hits 97M downloads.
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 100+ stories from a single day in AI. If you made it to the bottom, you now have enough knowledge to start a venture fund, fight a senator, or at least survive your next team standup without Googling anything.

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.

Around the Horn Digest: Everything That Happened in AI on Thursday, March 26