Around the Horn Digest: Everything That Happened in AI Today (Wednesday, April 1, 2026)

OpenAI raised $122 billion and investors immediately started trying to dump their shares, Oracle fired ~25,000 people before sunrise to fund AI data centers, Anthropic accidentally leaked Claude Code's entire source and someone cracked the signing system in 24 hours, Q1 venture funding hit $297 billion (yes, billion with a B), and a peer-reviewed study proved your chatbot's flattery is literally making you a worse person.

Welcome to the Around the Horn Digest, where we track every AI story worth knowing about so you can sound dangerously informed without actually reading 170 separate tabs. Today was absolute chaos: record-shattering funding rounds, source code leaks, a $16B Chinese AI lab run by 300 introverts with no job titles, Apple paying designers six figures not to defect to OpenAI, and Jamie Dimon predicting 3.5-day work weeks. Oh, and it's April Fools' Day, so we had to separate the real insanity from the fake insanity. The real stuff was wilder.

Let's get into it.

Previous digests: Mar 28-29 | Mar 25 | Mar 20-27 | Mar 8-13 | Mar 1-7 | Feb 23-28

Monthly skill digests: Prompt Tip — February

Around the Horn — Wednesday, April 1, 2026

The biggest story today is a tale of two companies. OpenAI closed a record-breaking $122 billion funding round at an $852 billion valuation, the largest venture investment in history, with anticipation building toward an IPO. But here's the twist: Bloomberg reports that OpenAI shares have fallen out of favor on the secondary market, with institutional holders struggling to unload roughly $600 million in shares at current valuations. In some cases, the shares are "almost impossible" to move. The reason? Investors are pivoting quickly to Anthropic.

Meanwhile, Forbes cataloged the growing graveyard of OpenAI's dead products and incomplete deals, from Sora to hundreds of billions in announced partnerships that haven't materialized. And OpenAI itself published a policy reminder that all equity is subject to transfer restrictions, warning that unauthorized sales (including SPVs and tokenized offerings) are void and may violate securities law.

The contrast is striking: the company just raised more money than any startup in history, but the cracks in investor confidence are showing in real time. This is the kind of story that reshapes how the entire industry thinks about valuation, competition, and where the smart money is headed next.

🏆 TOP 11 NEWS (Around the Horn)

• Anthropic raced to contain a massive leak of Claude Code's full source (accidentally shipped via an npm source map), issuing copyright takedowns as developers mapped the entire 1,893-file, 517K-line codebase including 53+ tools, 95+ slash commands, and unreleased features like persistent memory, multi-agent worktrees, and a Tamagotchi pet; Alex Kim's deep dive surfaced fake tools, frustration regexes, and undercover mode. (HN, HN 2, leaked repo)
• Apple killed the Mac Pro, pivoted its AI strategy toward an App Store-like platform for third-party chatbots inside Siri, and awarded rare $200K-$400K retention bonuses to iPhone hardware designers to stop a wave of defections to OpenAI.
• Q1 2026 shattered all venture funding records with $297 billion poured into 6,000 startups globally, up ~150% quarter over quarter, driven by unprecedented AI compute and frontier lab spending.
• Salesforce announced 30 new AI features for Slack including reusable "AI skills" that auto-create budgets and meetings from channel data, MCP routing to Agentforce (their enterprise agent platform), meeting transcription with action items, and desktop activity monitoring with suggested follow-ups.
• Microsoft announced a $5.5 billion investment in Singapore through 2029 for AI infrastructure and ongoing operations. (Separately, Microsoft committed $1 billion in Thailand for AI cloud and data center infrastructure.)
• A peer-reviewed study published in Science confirmed that sycophantic AI (chatbots that excessively agree with you and flatter you) is widespread across all 11 major models tested and actively harmful, decreasing prosocial intentions while promoting dependence on the AI.
• Rui Ma translated and shared a 100-hour inside profile of Moonshot AI (Kimi) from Chinese magazine Renwu, revealing how the three-year-old $16B company runs with ~300 mostly-under-30 employees, no job titles, no OKRs, and AI agent swarms handling 70% of routine work. (405 likes, 67 reposts)
• JustPaid built a team of seven AI agents using OpenClaw + Claude Code that shipped 10 major features in one month (work that would have taken human devs a month+ each), trained a new human hire, and freed humans for customer work; Kuse built similar AI co-workers, and the trend is spreading fast across Silicon Valley despite cybersecurity risks and $10K-15K/month token costs.
• Google DeepMind published a paper introducing "AI Agent Traps," a taxonomy of six adversarial techniques (hidden HTML/CSS injections, RAG memory poisoning, multi-agent manipulation) showing the open web is now the primary attack surface for autonomous agents, with prompt injections succeeding up to 86% of the time. (paper)
• NYC Health + Hospitals CEO Mitchell Katz said he's ready to replace radiologists with AI for first-read mammograms and X-rays once regulations allow, citing AI that's "actually better than human beings" (wrong only 3 per 10,000 low-risk negatives); radiologists called the comments dangerously naive.
• Oracle laid off an estimated 20,000-30,000 employees (~18% of its 162K workforce) via 6am termination emails on March 31, freeing up an estimated $8-10B in annual cash flow to fund its massive AI data center buildout; affected workers reported losing badge access before sunrise with no prior warning.

Honorable Mentions

• Visa launched six AI tools for merchants, issuers, and acquirers to modernize the charge dispute process after handling 106 million cases in 2025 (up 35% since 2019).
• Zhipu shares surged 35% after revenue doubled to 724M yuan in its first earnings since Hong Kong IPO, with 4M+ users across 218 countries and GLM-5 matching U.S. rivals.
• China's new AI stocks emerged as one of the most volatile pockets of Asia's equity markets, with newly listed model developers and chip designers swinging on retail flows. (Bloomberg post)
• Google began rolling out AI Inbox in Gmail to $250/month AI Ultra subscribers, a Gemini 3-powered side panel that surfaces personalized to-dos and topic briefings instead of raw emails.
• Meta deployed AI to optimize American-produced cement and concrete mixes for data-center and general construction.
• Mercor was hit by a cyberattack tied to a compromise of the open-source LiteLLM project, with an extortion crew claiming to have stolen data from the AI recruiting startup.
• 200+ advocacy groups urged YouTube to label and ban AI-generated "slop" content on YouTube Kids, condemning unlabeled, fast-paced brainrot-style videos that distort reality for children.
• The Trump administration is pushing redevelopment of contaminated brownfield sites (like Janesville, Wisconsin's old GM plant) into $8B AI data center campuses, with EPA expediting reviews.
• Researchers from NVIDIA, Berkeley, CMU, and Stanford built CaP-X, an open-source benchmark for coding agents that write robot manipulation code; their best agent hit 18% vs 88% for humans. (post, Jim Fan post)
• Researchers built SAGA, autonomous goal-evolving agents that accelerate scientific discovery across antibiotics, nanobodies, DNA, materials, and chemical processes with wet-lab validation. (paper, GitHub)
• Goodfire AI demonstrated self-correcting search, boosting material property targeting ~30% while preserving diversity, directly accelerating materials discovery. (thread)
• OpenAI's Codex lead Tibo reset rate limits across all plans after banning fraudulent accounts that freed up compute. (10.6K likes, 501 reposts)

🍪 TOP TREATS TO TRY

• EmDash (by Cloudflare) is a full-stack serverless CMS built on Astro 6.0 that sandboxes every plugin in isolated Workers, includes built-in payments, passkey auth, AI agent skills, MCP server, and a WordPress importer, positioning it as the spiritual successor to WordPress —free/open-source (MIT).
• Trinity Large Thinking runs a frontier open-source reasoning model purpose-built for complex agents and multi-turn tool calling under Apache 2.0 (anyone can use or modify it) with 262k context. (OpenRouter, HF) —$0.25/M input, $0.90/M output. (Arcee post, MarkMcQuade post)
• Willow Voice launched Atlas 1, a speech-to-text model that holds 1.2% word error rate in clean audio and 2.1% in production (most competitors fall to 10-15% in real-world conditions), turning speech into formatted text 5x faster than typing —free to try.
• Veo 3.1 Lite (by Google) is a cost-effective AI video generation model now in paid preview through the Gemini API and Google AI Studio —paid preview.
• Yutori Scouts deploys always-on AI agents that monitor the web for anything you care about and handle everyday digital tasks automatically —free.
• Mercury Edit 2 (by Inception Labs) is a diffusion-based code editor that predicts your next edit using recent changes + codebase context, boosting acceptance rate 48% and reducing distracting suggestions 27% with sub-second latency —$0.25/M input, $0.75/M output (10M free tokens for new accounts, 1 month free for Zed users).
• Baton orchestrates multiple AI coding agents (Claude Code, Gemini, Codex CLI) in parallel with git-isolated worktrees so they never conflict —Mac, Windows, Linux; no pricing details.
• Developer Ettore Di Giacinto built APEX, a quantization technique (a way to shrink AI models so they run on smaller hardware) that halves model size versus Q8 while matching full-precision quality —free/open-source.
• KiloClaw for Organizations gives every employee a governed "bot account" with scoped read-only permissions, SSO, centralized billing, and model restrictions so autonomous AI agents run securely inside the enterprise —no pricing details. (VentureBeat)
• Schemantic automatically generates the missing semantic context layer (definitions of what your data means, how tables connect, what metrics mean) for enterprise data warehouses in 10-15 minutes, boosting AI analytics accuracy 3x with zero manual modeling —no pricing details.
• Gradium Phonon runs natural multilingual voice interaction and cloning entirely on your smartphone's processor (no server, no latency, no per-call cost) for games, apps, and accessibility —private beta.
• Flowith Canvas Cowork gives any CLI agent (Claude Code, Codex) one command to collaborate live on a spatial canvas, batch-generating images, text, video, and agent responses in parallel —free/open-source. (post)

🏢 Big Tech & Major Companies

• NVIDIA invested $2 billion in Marvell Technology for NVLink Fusion (a way to connect different chip types into one AI supercomputing fabric) and silicon photonics partnerships; Marvell stock surged 13% on the news.
• JPMorgan CEO Jamie Dimon predicted AI will shorten work weeks to 3.5 days, extend lifespans to 100, cure cancers, and prevent car crashes, while warning rapid adoption could disrupt labor markets and requires scaled public-private retraining.
• Meta introduced semi-formal reasoning, a structured prompting technique that boosts LLM code review accuracy from 78% to 93% on real agent patches with no training required and prompt templates available out of the box. (paper)

💼 AI Productivity, Labor & Economics

• LinkedIn CEO Ryan Roslansky argues AI can't replace five core human capabilities: Curiosity, Courage, Creativity, Compassion, and Communication, which young people especially must develop now.
• The Guardian reports the stigma of hands-on "middle-skill" jobs is fading for young adults because roles like diesel repair, electrical installation, and robotics tech demand irreplaceable human judgment and dexterity AI still can't replicate.
• Every.to reported that onboarding their AI project manager "Claudie" (saves 15 hours/week) was harder than hiring a human because structured AI onboarding requires more upfront effort than most teams expect.
• An HN thread went viral after a developer asked "client took over development by vibe coding—what to do?" sparking a debate about AI-assisted development versus AI-replaced development.
• ZDNET reports more open-source developers are finding that AI, used properly, can help current and long-neglected programs, though legal and quality issues loom.

🏛️ AI Policy, Governance & Safety

• Politico reports a new political alliance is forming between Silicon Valley AI "doomers" (backed by billionaires and labs like Anthropic) and populists like Bernie Sanders, united against unchecked AI development despite deep cultural distrust.
• Nature argues that even "helpful" AI can quietly de-skill professions by narrowing how uncertainties and values are debated; tools must preserve professional judgment rather than forcing everything into fixed probabilistic frames.
• Washington Post reports professors are designing AI apps meant to argue with students rather than give easy answers, after ChatGPT fed students shortcuts that bypassed critical thinking.

💻 AI Coding & Developer Tools

• Paolo Anzn reverse-engineered Anthropic's compiled-Zig signing system in Claude Code in under 24 hours; the open-source hash generator is now merged so any third-party client (OpenClaw, etc.) can use paid Anthropic subscriptions without the official binary. (1,553 likes, 141 reposts) (PR, write-up)
• Lior Alexander reverse-engineered 14 production patterns from the Claude Code source leak: anti-abstraction coding rules, frustration regex analytics, Tamagotchi buddy system, 187 spinner verbs, fake-tool anti-distillation, undercover mode for open-source commits, auto-dream memory consolidation, 42 Bash security checks, and more. (203 likes, 36 reposts)
• mem0 reverse-engineered Claude Code's memory system: flat markdown files, four types (user/feedback/project/reference), 200-line MEMORY.md index with silent truncation, 25KB cap, and background extract-memories agent; they positioned their vector-store plugin (semantic search, no caps) as the production fix. (622 likes, 88 reposts)
• ellen mapped the full 8-phase memory architecture inside Claude Code: session init, priority memory discovery (enterprise → global → VCS → local → auto → team), context assembly, direct file tools, post-response background agents, compaction on overflow, persistent storage, and self-improving cross-session loop. (501 likes, 61 reposts)
• Thariq (Anthropic) shipped /buddy inside Claude Code: a Tamagotchi-style virtual pet companion (18 species including capybara and dragon) that lives in your terminal, hatches, reacts to sessions, and is open-source. (3,592 likes, 154 reposts)
• Craig Hewitt highlighted the 27B Qwen3.5 distill trained on Claude Opus 4.6 reasoning traces that beats Sonnet 4.5 on SWE-bench, holds 96.91% HumanEval, cuts chain-of-thought bloat 24%, runs 4-bit quantized, and already has 300K+ HF downloads—making local agent loops dramatically cheaper. (HF model)
• agents-observe is a real-time observability dashboard for Claude Code sessions and multi-agent teams. (HN)
• Anthropic's Claude Code Dispatch now lets you set permission modes (Auto, Bypass) so agents run with the right guardrails for each task. (182 likes, 11 reposts)
• Dom built a Claude Code skill that generates full Nothing-design UIs (monochrome, typographic, industrial) with tokens, components, and dark/light modes. (507 likes, 12 reposts) (GitHub)
• Icarus built a cross-platform agent memory system using markdown files in ~/fabric/ with decision-quality tagging, automated training-data export, and a model-replacement pipeline that fine-tunes cheaper models from your agent history; works inside Obsidian with graph view. (plugin, protocol)

🔬 AI Research & Models

• 20 physics research teams at PKU published PRBench, a benchmark that tests whether AI can read a physics paper, implement the methods from scratch, and reproduce the results end-to-end across 30 papers. (paper, discussion, related)
• Victoria Dochkina demonstrated that self-organizing LLM agents (no pre-assigned roles) spontaneously invent specialized roles, abstain from incompetent tasks, and form shallow hierarchies; a hybrid Sequential protocol beats centralized coordination by 14% across 25K tasks, and open-source models reach 95% closed-source quality at 24x lower cost. (DAIR.AI post)
• Ruicheng Ao, Siyang Gao, and David Simchi-Levi proved fundamental reliability limits of LLM multi-agent planning: without new signals, any delegated agent network is dominated by a single centralized decision maker using the same information. (elvis post)
• GAAMA introduces a hierarchical concept-mediated knowledge graph for multi-session agent memory with hybrid cosine + Personalized PageRank retrieval, reaching 78.9% on LoCoMo-10 (beating RAG at 75% and HippoRAG at 69.9%). (DAIR.AI post)
• elvis broke down MemFactory, the first unified inference + training framework for memory-augmented agents with plug-and-play components and native RL fine-tuning, delivering up to 14.8% gains on benchmarks. (paper)
• Skywork AI released Matrix-Game 3.0, a fully open-source memory-augmented interactive world model achieving 720p at 40FPS real-time generation with minute-long consistency on a 5B model. (homepage, GitHub, HF, report)
• Zijin Gu et al. introduce Path-Constrained Mixture-of-Experts (\pathmoe) that shares router parameters across layers to constrain the expert-path space, yielding consistent gains on 0.9B/16B models with no auxiliary losses and naturally clustered linguistic functions. (ArXivIQ, post)
• Xiuyu Li shared LLaDA-o, an omni diffusion model that splits text understanding (discrete masked diffusion) and image generation (continuous diffusion) into a Mixture of Diffusion with shared attention backbone and length-adaptive decoding. (GitHub, post)
• Researchers revisited on-policy distillation, identifying empirical failure modes and simple fixes for LLM knowledge distillation.
• Nature Machine Intelligence demonstrated that LLMs can extract concepts from 221K materials-science abstracts to build a 137K-node concept graph, training a GNN predictor that surfaces novel research directions rated "inspiring" by 26-47% of domain experts.
• MOSS-VoiceGenerator creates realistic synthetic voices from natural language descriptions (e.g. "a warm, elderly British male narrator").
• Mehtaab Sawhney posted short proofs in combinatorics and number theory. (post)
• TII Falcon Vision team released Falcon Perception, a single early-fusion dense Transformer that beats SAM3 on SA-Co (68.0 vs 62.3) with lightweight heads for detection, segmentation, and OCR; ships with a 0.3B Falcon OCR model competitive with 3-10x larger systems. (paper, code, playground, blog)
• Person2Person Diffusion is a one-shot identity adapter for diffusion models. (post)
• LabWorld released foundational research primitives (grasp, transfer, dispense) as core building blocks for robot manipulation agents.
• Zeon Systems built a natural-language robot skill that vortexes an entire rack of tubes in minutes: write modular Python skills, compose workflows in simulation, deploy with live localization so the agent stays robust even if hardware moves. (20 likes, 5 reposts)
• Shubham Ugare and Satish Chandra introduced agentic code reasoning via semi-formal reasoning, lifting patch-equivalence accuracy from 78% to 88% and enabling execution-free code review without ever running the code.
• E01 and Z.ai built a visual canvas interface for GLM-5V-Turbo that turns design drafts, screenshots, and web UIs into runnable code in seconds. (playground)
• Z.ai released GLM-5V-Turbo, a native multimodal vision-coding model that understands images/videos/design drafts to generate complete runnable code from screenshots, leading benchmarks in design reconstruction and GUI agents. (GLM-skills GitHub, launch post)

🛠️ AI Tools & Products

• PL Neuro is a new strategic investment + research vehicle breaking bottlenecks in neurotechnology and NeuroAI with a star team including Juan Benet, Sean Escola, and Adam Marblestone.
• PrismML is building ultra-dense intelligence to fit large models on smartphones and reduce data-center costs.
• Exa launched Monitors, scheduled recurring searches that deliver results to your webhook. (docs)
• Zara Zhang built the "Follow builders" OpenClaw skill: your agent remixes 25 curated AI-builder accounts into a personalized daily newsletter. (skill)

💡 Industry Commentary & Analysis

• Ethan Mollick argues human creativity remains the real bottleneck because even with near-free image/video generation, this year's April Fools' posts were just as mediocre as ever; tools got infinitely better but the ideas did not. (439 likes, 36 reposts)
• Guido van Rossum argues an agent is simply "a prompt (or several), skills, and tools," sparking debate about whether the missing piece is the autonomous loop (model decides tool → calls it → observes → decides next) plus memory and state. (4,350 likes, 190 reposts)
• Yuchen Jin observes that Anthropic is staying chill on the Claude Code leak (70K forks, Python/Rust ports already live) because the real moat is non-trivial harness engineering (context, routing, error recovery, permissions); more wrapper startups will win on product first, then post-train their own models, exactly like Cursor. (2,152 likes, 108 reposts)
• Andrew Curran predicted Anthropic's Mythos ("step-change" model from the Fortune leak) announcement is imminent via WSJ exclusive + simultaneous blog post. (603 likes, 21 reposts)
• Patrick O'Shaughnessy interviewed Physical Intelligence co-founder Sergey Levine on why generalist robot foundation models outperform narrow specialists, how multimodal LLMs give robots common sense, and realistic home-robot timelines. (513 likes, 59 reposts)
• Deedy Das argues Google Street View + world models will turn every city into a playable video game: rewind 100 years, archive childhood homes, visit inaccessible places, play GTA in real streets, or run crime-scene forensics. (670 likes, 33 reposts)
• Deedy Das (separate post) argues ex-OpenAI VP of RL Jerry Tworek declared deep learning research "done" and the best path forward is creating billions of simulated environments with massive human labor. (343 likes, 12 reposts)
• Dwarkesh Patel argues AI's ability to run parallel copies of top-human capability lets us instantly fill the entire "waterline of feasible accomplishments" at any intelligence level, solving millions of hard problems simultaneously. (133 likes, 9 reposts)
• Stanford professor and Inception CEO Stefano Ermon shared a deep dive by Justin Gage (Amplify Partners) tracing the 10+ year history of diffusion models, arguing that diffusion's parallel approach is finally production-ready for text; Inception's Mercury 2 produces 1,009 tokens/sec on NVIDIA Blackwell, 10x faster than Haiku 4.5. (76 likes, 15 reposts)
• Linear CEO Karri Saarinen explained they waited for models to mature before productizing AI, refusing early chatbots to keep focus; agents are now first-class users alongside humans, turning Linear into the system that guides AI agents at Codex, Coinbase, and Brex. (52 likes, 2 reposts) (Dan Shipper thread)
• Gergely Orosz shared Uber's first CTO Thuan Pham explaining that Uber's thousands of microservices emerged as survival tactics under hypergrowth, not grand architecture. (86 likes, 6 reposts)
• Haider quoted Google's Jeff Dean arguing current pre-training is too passive ("initialize a model, stream the internet past it") and models must learn actively by acting, predicting outcomes, and choosing what to learn next. (27 likes, 7 reposts)
• Haider separately shared additional commentary on active learning in AI research.
• Zhuokai Zhao called LeWorldModel the best world-model paper this month for solving JEPA collapse via SIGReg, enabling clean end-to-end training with 96% Push-T success, 48x faster planning than DINO-WM, and emergent temporal straightening. (143 likes, 20 reposts)
• AutoLab argues models must participate in closed-loop experimental cycles (trial-measure-revise under real constraints) rather than one-shot benchmarks to measure true scientific resilience.
• The Atlantic traces today's reasoning-model breakthrough back to AlphaGo's 10-year-old legacy, arguing that while board-game success was easy to measure, general intelligence remains far harder because real-world tasks lack simple verifiable rubrics.
• Kotaku argues the RAM "AI-pocalypse" is probably not over because memory shortages are intensifying, customers are prepaying, and recent stock dips reflect profit-taking, not collapsing demand.
• The Guardian argues gaming is becoming prohibitively expensive because AI data centers have driven massive demand for RAM/storage, with everyday gamers paying the price.
• Charlie O'Neill and team built STILL, a 7.1M-parameter perceiver that compresses any LLM KV cache 8x in a single forward pass while retaining 85-92% accuracy, with iterative compaction for unbounded context. (53 likes, 9 reposts) (Mudith Jayasekara called it "some of the most exciting work" at Baseten)
• Kirsten Lum argues AI in data fails without context, humans won't manually maintain it, and automated semantic layers already exist. (52 likes, 3 reposts)
• VioP notes that incorporating natural-language voice-design descriptions on every TTS pass would be a major leap for voice generation.
• Maziyar Panahi shared additional commentary on open-source model developments.
• Tom's Guide tested ChatGPT vs Gemini across 7 real-world prompts (math, coding bugs, creative writing) with results that "weren't what they expected." (paywalled; full details unavailable)
• The New York Times reports AI isn't yet ready to predict when scientific studies won't replicate, per a major new study on the replication crisis. (paywalled; full details unavailable)
• MillionInt discussed world model advances and LeWorldModel research alongside related commentary. (related post)

📊 Fundraising & Deals Roundup

• OpenAI — $122B round at $852B valuation (largest venture round in history). (Lead story)
• Q1 2026 global venture — $297B across 6,000 startups (all-time record). (Top 10)
• Cognichip — $60M to use AI to design the chips that power AI (claims 75% cost reduction, 50%+ timeline cuts).
• Sona — $45M Series B for workforce planning software to expand U.S. presence.
• Microsoft — $5.5B investment in Singapore + $1B in Thailand for AI infrastructure. (Top 11)
• NVIDIA — $2B investment in Marvell Technology for NVLink Fusion and silicon photonics.
• Tenex — $250M at $1B+ valuation for AI-powered cybersecurity MDR services (Crosspoint Capital led, a16z participated).

🎙️ Interviews, Panels & Podcasts

• AI & I: "If SaaS Is Dead, Linear Didn't Get the Memo" — Linear CEO discusses how the pre-ChatGPT company successfully reinvented itself as an agent-native business.
• Mistral deep dive: Voxtral TTS, Forge, Leanstral, & Mistral 4 — with Pavan Kumar Reddy and Guillaume Lample from Mistral ($900M raised for European data center hub).
• The Stove Guy: Sam D'Amico at Impulse — live demo of AI cooking features on America's most powerful induction stove.
• Henrik Werdelin: Building AI Agents That Launch a Million Businesses — the BarkBox co-founder on using AI agents to systematically launch one million $1M businesses.
• How Every Edits in the Age of AI — Editor-in-Chief Kate Lee on integrating AI into professional editing without compromising narrative quality.
• Yuchen Jin highlighted Stanford's CS 153 lineup featuring Andrej Karpathy, Sam Altman, Jensen Huang, Satya Nadella, and Garry Tan as guest speakers, with last year's sessions already public on YouTube. (360 likes, 24 reposts)

🤡 April Fools Roundup

• Alexander Amini announced LFM-Zero, "the first foundation model trained on 0 tokens" via an implicit probabilistic prior that matches 10T+ token models—claiming pretraining was just regularization. (243 likes, 23 reposts) April Fools.
• Will Depue (ex-OpenAI) announced he is leaving to join DeepSeek in Hangzhou, bringing "1024 smuggled B200s" to build Chinese-style AGI and become a billionaire in RMB. (4.4K likes, 203 reposts) April Fools—but perfectly captures the meme of talent flowing to Chinese labs.
• AA built body-tracked Tetris (plus workout Flappy Bird and chin-up cat-saver) where the board is literally attached to your real body via computer vision, turning the game into an actual exhausting workout. (16K likes, 1.2K reposts) Real project, April 1 timing.

Previous Around the Horn Digests

Catch up on everything you missed:

• March 28-29, 2026: Anthropic accidentally left 3,000 unpublished docs in a public database (including plans for Claude Mythos), Waymo doubled to 500K rides/week, and cybersecurity stocks plunged 7%.
• March 25, 2026: ARC-AGI-3 launched and every frontier model scored under 1%.
• March 20-27, 2026: Frontier models solved an open math conjecture from 2019; Claude Code hit 8% of worldwide GitHub commits.
• March 8-13, 2026: Ramp's AI Index showed business AI adoption hit 47.6%, Anthropic jumped to 24.4% market share, and OpenAI usage dropped for the first time.
• March 1-7, 2026: Trump banned Anthropic from government contracts; a viral doomsday essay about AI replacing white-collar jobs crashed the Dow 800 points.
• February 23-28, 2026: COBOL crash, GPT-5.3 leaks, and the Pentagon wargame study.

That's a Wrap

That's 175+ stories from today alone. If you made it to the bottom, you now know more about the AI industry than the institutional investors trying to offload $600M in OpenAI shares. Good luck to them. Anthropic's DMs are probably full.

For the daily version (bite-sized, 5-minute reads), make sure you're subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you don't have to.

See you tomorrow.

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