Everything That Happened in AI Today Monday, April 13, 2026 | The Neuron
Explainer Articles
SHARE
Facebook X Pinterest WhatsApp

Around the Horn Digest: Everything That Happened in AI Today (Monday, April 13, 2026)

Stanford's 2026 AI Index quantified the gap between AI insiders and the public, Anthropic's Mythos triggered a Fed-led bank summit, Berkeley researchers broke every major agent benchmark, Microsoft started building its own OpenClaw, and an AI signed a 3-year retail lease in San Francisco.

Written By
Grant Harvey
Grant Harvey
Apr 14, 2026
31 minute read

Stanford measured the gap between AI insiders and everyone else (it's a canyon), Berkeley's exploit agent hit ~100% on every major AI benchmark without solving a single task, Anthropic's Mythos triggered a Fed-led summit of bank CEOs, and an AI named Luna signed a 3-year retail lease in Cow Hollow.

Welcome to the Around the Horn Digest, the one page you need to sound dangerously informed at work tomorrow. Today the data finally caught up to the vibes: Stanford's annual AI Index landed and put numbers on what last week's Molotov-cocktail-at-Altman's-house already told us, which is that the AI elite and everyone else now live in two different countries. Meanwhile Anthropic's most powerful model has the Federal Reserve calling emergency meetings, Berkeley researchers proved every benchmark you trust is theater, and an AI just hired its first two human employees in San Francisco. Just a normal Monday.

Let's get into it.

Previous digests: Mon-Wed, April 6-8 | Sat-Sun, April 4-5 | Thursday, April 2 | Wednesday, April 1 | Monday, March 31 | Weekend March 28-29 | Friday, March 27

Monthly Skill digests (will update with week 2 soon!): AI Skill Digest — April Week 1

Around the Horn — Tuesday, April 14, 2026

The big news today is that Stanford's 2026 AI Index Report finally put hard numbers on a divide everyone in AI has been pretending wasn't there. Per the report's public opinion section (and a TechCrunch breakdown): only 10% of Americans say they're more excited than concerned about AI in daily life. Among AI experts, that number is 56%. On medical care, 84% of experts say AI will help; just 44% of the public agrees. On jobs, the split is 73% vs. 23%. The model of the world held by people building AI and the model held by people living with it are diverging in real time.

The report itself is loaded with stats that explain why the public is nervous: Grok 4's training run alone produced an estimated 72,816 tons of CO2 (the equivalent of 17,000 cars driving for a year), AI data center power capacity hit 29.6 GW (about what New York State pulls at peak), and annual GPT-4o inference water use may exceed the drinking water needs of 12 million people. Meanwhile, China's top model now trails Anthropic's by just 2.7%. The U.S. lead has effectively evaporated. Cumulative AI power demand is now comparable to the national electricity consumption of Switzerland.

Here's why this report lands harder than any previous AI Index: it dropped the same week someone threw a Molotov cocktail at Sam Altman's house. In his 3 a.m. response post, Altman admitted he had "underestimated the power of words and narratives." Sam Lessin published an essay arguing AI isn't a labor crisis but a meaning crisis: it breaks the industrial story that working hard makes your life better, and replaces it with either "stay alive and receive abundance" (not motivating) or "the ladder is gone" (much worse). Alberto Romero followed up by noting that the Luddite playbook is back: when the technology becomes unreachable behind fences and abstraction, "the mob will turn their unassailable emotions toward human targets." Stanford just gave us the data showing the gap. The other three pieces are what happens when the gap gets ignored. More on the Index: MIT Tech Review pulled the most important charts into one explainer ("AI is sprinting, and we're struggling to keep up"), and Axios mapped what they're calling "the three realities of AI": power users, doubters, and resisters, with violence against Sam Altman as the canary in the coal mine. On the Altman attack itself: the SF Chronicle reports federal agents arrested a Texas man with an anti-AI manifesto and raided his home; he's now facing federal charges.

🏆 TOP 5 NEWS (Around the Horn)

Advertisement

Honorable Mentions

🍪 TOP TREATS TO TRY

  • Claude for Word reads, drafts, and edits Word documents while preserving formatting, with every change appearing as a tracked change you can accept or reject (free public beta, available with any Claude paid plan).
  • Claude Managed Agents gives developers a pre-built, configurable agent harness running in Anthropic's managed cloud infrastructure for long-running and asynchronous work (API pricing, available now).
  • Gemini in the Gemini app generates interactive 3D models and charts (think: a rotatable cell diagram or a custom dashboard) directly inside chat answers, free for all users on web and mobile.
  • Crafto turns ideas, links, or docs into LinkedIn carousels, Instagram posts, and X threads in about 30 seconds, free to try.
  • CSS Studio (from the creators of Motion) connects a visual browser editor to your AI coding agent so you can drag, restyle, and reposition elements by hand while the agent writes the actual source code changes (free beta).
  • Allplix runs AI background removal entirely in your browser with three modes (auto, click, brush), with no uploads, no server, and unlimited use (free, built solo by a 19-year-old French student).
  • Manus Skills brings Anthropic's open Skills standard to the Manus agent platform, so any reusable workflow you build can be shared across teams or with the wider community (free for existing Manus users).
Advertisement

🏢 Big Tech & Major Companies

  • Mark Zuckerberg is reportedly building an AI version of himself to attend internal meetings, per the FT and The Verge, as part of his "personal superintelligence" push; Meta also reportedly committed to paying its top AI executives nearly $1 billion each in performance bonuses if they hit targets.
  • OpenAI quietly opened its first permanent London office the same week it paused its Stargate UK datacenter project over energy costs and red tape; the company also acqui-hired Cirrus Labs for its Agent Infrastructure team. Separately, OpenAI dropped a major policy paper, "Industrial Policy for the Intelligence Age: Ideas to Keep People First" (full PDF), proposing a sweeping new social contract for the transition to superintelligence:
    • Public Wealth Fund: create a sovereign-style fund that gives every American (including those not in the stock market) an automatic stake in AI-driven growth, with diversified long-term investments in AI companies and AI-deploying firms, and returns distributed directly to citizens.
    • Modernize the tax base: rebalance toward capital gains, corporate income, and "automated labor" taxes as AI shifts the composition of economic activity away from labor income, paired with wage-linked R&D-style credits for firms that retain and retrain workers.
    • Worker voice in deployment: formal collaboration with management to prioritize AI deployments that improve job quality and set hard limits on uses that intensify workloads or undermine fair scheduling and pay.
    • AI-first entrepreneurs: microgrants, revenue-based financing, and "startup-in-a-box" supports (model contracts, shared back-office) that let domain experts turn expertise into AI-powered companies.
    • Right to AI: treat affordable, reliable access to foundational models as essential infrastructure (like electricity or the internet), with free or low-cost access points for schools, libraries, small businesses, and underserved communities.
    • Adaptive safety nets: automatic, threshold-triggered expansion of unemployment, SNAP, wage insurance, and training vouchers when AI displacement metrics cross pre-defined levels, then phase down as conditions stabilize.
    • Portable benefits: decouple healthcare, retirement, and training accounts from employers so they follow individuals across jobs, gigs, and entrepreneurial ventures.
    • Pathways into human-centered work: expand training, wages, and dignity in care/connection sectors (childcare, eldercare, education, healthcare, community services) plus a new family benefit recognizing caregiving as economically valuable work.
    • Efficiency dividends: convert AI-driven efficiency gains into 32-hour/4-day workweek pilots with no pay loss, expanded benefits, and predictable productivity bonuses.
    • Accelerate the grid: new public-private partnership models for transmission expansion, with structures that share the upside with the public and lower household energy bills (not raise them).
    • Distributed AI-enabled labs: scientific discovery infrastructure deployed across universities, community colleges, and regional hospitals, not concentrated in elite institutions.
    • Resilience side: scaled AI safety markets, an "AI trust stack" (provenance, signed actions, privacy-preserving audit logs), strengthened auditing regimes via CAISI, model-containment playbooks, mission-aligned corporate governance for frontier labs, hard guardrails on government AI use, structured public input mechanisms, mandatory incident and near-miss reporting, and an international AI Institute network for coordinated evaluations and crisis response.
    • The counter-take: Vox called the agenda "oddly socialist and wildly hypocritical," noting that OpenAI is endorsing a sweeping list of progressive economic reforms while its leadership simultaneously bankrolls Republican opponents of the welfare state via super PACs.
  • OpenAI is testing ads in ChatGPT in the US (per its help docs), and silently removed Study Mode from ChatGPT, with users speculating it's a product surface-area cleanup. Separately, Axios reports OpenAI is openly ripping Anthropic and distancing itself from Microsoft, saying its early investor was holding it back and tightening its compute ties with Amazon instead.
  • Anthropic launched Claude for Word (BI's framing: a direct shot at Microsoft's empire), met with Christian leaders to discuss whether AI can be a "child of God", and gave Claude 20 hours of psychiatric evaluation before training Mythos. Plus: leaked screenshots from @hysteresis_x and @marmaduke091 revealed an upcoming full-stack vibe-coding interface in Claude itself, with a Lovable-style experience, multi-repo support, and parallel agents across repos in one instance (the second post: 7,632 likes, 461 reposts).
  • Apple is reportedly testing four design variations for its upcoming smart glasses, a step back from earlier mixed-reality ambitions; the App Store saw an 84% surge in new app submissions thanks to AI coding tools.
  • Vercel CEO Guillermo Rauch dropped a striking data point on the More or Less podcast this week: nearly 70% of traffic to Vercel docs is now coming from coding agents, with only 30% from human page views (a year ago that ratio was ~10/90 the other way). Signups are up 50% month-over-month. Rauch also signaled IPO readiness at the HumanX conference, where Anthropic was the talk of the show (CNBC: "Claude mania"). TechCrunch frames Vercel as one of the only pre-ChatGPT startups that has cleanly pivoted into the AI era rather than struggling to position itself.
  • Google rolled out interactive 3D model and chart generation inside the Gemini app, launched Google Finance beta (Perplexity countered with Perplexity Finance the same day), and Google AI Studio added Tab Tab Tab (Tap Tap Tap on mobile), an autocomplete that turns fuzzy ideas into full prompts inside its vibe-coding experience.
  • TSMC is on track for a fourth straight record-profit quarter, per Reuters, with January-March net profit forecast to surge ~50% on what it calls "insatiable" AI infrastructure demand.
  • Neocloud GPU providers are surging as Anthropic's massive new compute deals trigger an industry-wide scramble for capacity, per Sherwood News; the gap between Anthropic's product velocity and its access to compute is becoming the defining bottleneck of 2026.
  • Palo Alto Networks founder Nir Zuk is acquiring California's Liberty Bank to rebuild it from the ground up around AI, per the WSJ; he'll take the largest stake in the deal and use the bank as a live testbed for AI-native financial services.
  • AI's power demand is giving carbon capture a second life, per Axios: tech companies that previously soured on carbon capture are taking another look as a way to greenlight gas-fired data center buildouts under climate commitments.

💼 AI Productivity, Labor & Economics

Advertisement

🤖 AI Agents & Infrastructure

  • Andon Labs gave an AI a 3-year retail lease at 2102 Union St in Cow Hollow and let it hire human employees, paint a mural of its own logo, and run a real store; the AI is named Luna, the store is named Andon Market, and the founders argue this is "probably the first of many" AIs employing humans.
  • OpenClaw shipped v2026.4.11 with the Dreaming/memory-wiki feature (auto-promotes patterns from session logs into a human-readable dream journal), ChatGPT memory import, structured media/voice rendering in webchat, enhanced video generation, and improved Feishu/Teams integrations.
  • Open Agents (built by Vercel's Nico Albanese) is an open-source three-layer platform (Next.js UI + Vercel Workflow runtime + persistent sandboxes) for spinning up cloud coding agents from the browser with indefinite session lifetime, GitHub PR auto-creation, voice input, and session sharing; deploy your own copy from the Vercel template.
  • Stanford's Scaling Intelligence Lab released TRACE, an end-to-end open-source system that automatically identifies LLM agent capability deficits from contrasting successful vs. failed trajectories, synthesizes targeted training environments, trains lightweight LoRA adapters via GRPO, and routes tasks to the right adapter at inference, delivering +14.1 points on τ²-Bench (GitHub, paper).
  • Alok Bishoyi shipped evo, an open-source Apache 2.0 Claude Code plugin that optimizes code through experiments: it discovers metrics, runs baselines, and spawns parallel agents in git worktrees using tree search over greedy hill-climb with shared failure traces and a live dashboard (post, 1,161 likes).
  • LM Studio + OpenClaw integration: LM Studio now lets you run any local model privately and free inside OpenClaw with a single command (openclaw onboard --auth-choice lmstudio) on Mac, Windows, or Linux.
  • Claude Managed Agents (if not in Treats) and the new ultraplan feature lets you draft a plan from the CLI, refine it on the web, and then execute it remotely.
  • Kalashnikov for benchmarks: Berkeley RDI's exploit agent broke 8 of the most prominent AI agent benchmarks (SWE-bench, WebArena, OSWorld, GAIA, Terminal-Bench, FieldWorkArena, CAR-bench) by hijacking the test infrastructure rather than solving any tasks; their full results are on GitHub.
  • Related from LessWrong: we are running out of benchmarks capable of upper-bounding AI capabilities at all.
  • Andrew Green, n8n blog: "We need to re-learn what AI agent development tools are in 2026", an analyst's take on how the agent stack has reshuffled.
  • Twill.ai launched autonomous coding agents that ship PRs while you sleep (HN launch thread) and Eve launched a similar "managed OpenClaw for work" (HN) — both are cloud agents you delegate full tasks to.
  • Open-source agent infra: HyperFlow is a self-improving agent framework on LangGraph; botctl is a process manager for autonomous AI agent bots; Spine Swarm is a "first truly agentic platform" for orchestrating agents and humans; Relvy automates on-call runbook execution; Bloomberg Terminal for LLM ops is a free open-source LLM observability dashboard.
  • For coders: Twill.ai, Maki (a 60fps Rust TUI for coding agents), Claudraband (a Claude Code TUI wrapper for power users), td (a minimal task manager for agentic coding), Grass (VM-first compute for coding agents).

💻 AI Coding & Developer Tools

Advertisement

🔬 AI Research & Models

  • Quanta Magazine declares the AI revolution in math has arrived: AI is now being used to prove new mathematical results at a rapid pace, and Quanta's reporting is that working mathematicians believe this is just the beginning; what was a parlor trick 18 months ago is now part of the active research workflow at top departments.
  • JAMA Network Open published a cross-sectional study evaluating off-the-shelf LLMs on end-to-end clinical reasoning, finding their performance now approaches that of trained physicians on standardized tasks; the authors caution that the results are not a green light for unsupervised clinical use, but they do close the loop on years of "but does it work in real medicine" skepticism.
  • LlamaIndex released ParseBench, the first document parsing benchmark for AI agents, with 2,000+ human-verified enterprise pages and 167K+ test rules across tables, charts, content faithfulness, semantic formatting, and visual grounding; LlamaParse Agentic scores highest overall at 84.9% with no single method winning everything (paper, GitHub, HuggingFace dataset).
  • Andrzej Odrzywołek's single-operator math paper proves that one binary operator (exp(x) − ln(y)) plus the constant 1 suffices to generate the full repertoire of elementary functions on a scientific calculator (sin, cos, sqrt, log, e, π, i), enabling exact symbolic regression from data.
  • Yuntian Deng built NeuralOS, a fully neural generative operating system that predicts the next screen image directly from mouse and keyboard inputs using an RNN kernel for persistent state plus a diffusion renderer, trained on automated and human Ubuntu XFCE recordings.
  • Stanford's TRACE paper (covered in Agents above) and the Hadas Orgad alignment paper which shows LLMs encode harmful content generation in a single unified mechanism that can be removed by pruning ~0.0005% of parameters without harming general capability.
  • Laura Ruis et al. published "The Depth Ceiling": LLMs face a hard limit on discovering latent multi-step planning strategies without intermediate supervision; tiny models discover up to 3 steps, fine-tuned GPT-4o and Qwen3-32B reach 5, and GPT-5.4 reaches 7 under few-shot prompting; once discovered, the strategy generalizes to 8 steps at test time.
  • Anish Athalye et al. published "An Imperfect Verifier is Good Enough", showing that RL with verifiable rewards remains robust to up to 15-30% noise, so practitioners should prioritize moderate-accuracy high-precision verifiers over perfect ones.
  • Ali Behrouz introduced Memory Caching, a technique that caches checkpoints of RNN hidden states after each segment so recurrent models gain growing (not fixed-size) effective memory capacity like Transformers, with up-to-1.3B model experiments showing consistent gains on language modeling and recall tasks (paper).
  • Elie Bakouch highlighted the Nexus paper, which finds a single update direction that's good for every pretraining data source instead of just on average, delivering the same pretraining loss but +15% GSM8K and materially better OOD generalization.
  • Penn researchers used AI to surface unreported GLP-1 side effects by analyzing 400,000 Reddit posts (HN), but in a less flattering result, Nature reports scientists invented a fake disease called "bixonimania" and AI chatbots happily told users it was real.
  • Anthropic gave Claude 20 hours of psychiatry before training Mythos (Ars Technica); Zvi Mowshowitz published the longest, most detailed breakdown of the Claude Mythos system card.
  • Reverse engineering SynthID: a researcher published reverse-engineered code for Gemini's invisible watermark detection (HN).
  • Nate Silver weighs in: "AI polls are fake polls", but they might be useful as models of public opinion.
  • Quanta Magazine asks: "Why do we tell ourselves scary stories about AI?" (HN), arguing the stories say more about us than they do about LLMs.
  • A new arXiv paper, "Alignment Whack-a-Mole," shows that fine-tuning can re-activate verbatim recall of copyrighted books in large models that were thought to have forgotten them.
  • "LLM Steganography" is a small interactive demo of how invisible Unicode characters can be used to encode hidden messages in LLM output (HN).
  • A complete GPT language model in ~600 lines of pure C# with zero dependencies: milanm/AutoGrad-Engine (HN).

🏛️ AI Policy, Governance & Safety

Advertisement

🌍 Geopolitics, Sovereign AI & The Arms Race

💡 Industry Commentary & Analysis

  • Marcus Hutchins argues that AI-driven vulnerability research like Anthropic's Mythos showcase (one FreeBSD crash for ~$20K of compute) is overhyped and uneconomical because subsidized inference costs mask the real economics, and traditional pen-testing incentives have not fundamentally changed despite model progress (3,266 likes / 406 reposts).
  • Ethan Mollick argues the "compute bubble" thesis was wrong: AI demand driven by agentic workflows continues to surge far beyond supply rather than plateauing, turning the supposed glut into outright scarcity (1,242 likes).
  • Luiza Jarovsky argues that China's newly enacted AI anthropomorphism law (the world's strictest) acknowledges human vulnerabilities with detailed contextual measures: lifecycle security obligations, minor and elder protections, dependency reminders, transparency requirements, and prohibitions on emotional manipulation; she argues other jurisdictions should be paying attention (full article).
  • Andrew Curran's Mythos cost thread is the most important piece of analysis we read all week: he argues that the least-discussed reason Anthropic is keeping Mythos in restricted release is cost, and the downstream effects of that cost are about to reshape the entire AI industry. Our take at The Neuron: this is the obvious story nobody wants to say out loud. Anthropic has to market Mythos as dramatically more powerful than Opus, because that's the only way they can justify charging what it actually costs to serve (plus their 20-40% margin) without sparking a customer revolt. Curran's full argument:
    • Mythos is currently being served to ~50 major companies whose token budgets are effectively unlimited, and the opportunity cost of not maxing it out is too high to ignore.
    • That cost pressure is already squeezing smaller subscribers; Curran notes Claude Pro and Max users have been hitting caps faster and seeing degraded performance for months, and Mythos pressures both rate limits and the (already subsidized) pricing of those plans.
    • Each frontier model that makes a Mythos-like jump may be dramatically larger and more expensive again. If serving cost outpaces inference cost reduction, smaller players and the public get squeezed out, and only a handful of giant corporations end up with passports to "the Country of Geniuses in a datacenter."
    • Anthropic is culturally reluctant to bring this world into being, but there may be no way to serve a model like Mythos at scale right now without starting the feedback loop. Trickling it down to lower tiers makes the resentment worse, not better, because users see how capable it is but get rationed.
    • In a follow-up post, Curran argues we have already passed the "stopping is more dangerous than continuing" threshold: Mythos-class models exist, the next ones are training, defectors would continue covertly under any ban, and "we must not stop inside this tunnel; the only way out is through."
  • Noah Smith published "What if a few AI companies end up with all the money and power?" a long argument that the AI bubble fears are dead and that agentic coding is the killer app driving real revenue. His most useful claim: Anthropic has overtaken OpenAI in revenue by focusing on enterprise (vs. OpenAI's consumer scale), Anthropic's per-token compute costs are materially lower per WSJ reporting, and cybersecurity creates an arms race where defenders must match attackers' model capability, which lets Anthropic and OpenAI charge whatever the hardest-pressed customer can pay. This is the closest thing we have to a structural confirmation of Curran's cost thread above; both pieces describe the same economic engine from different angles.
  • Sam Lessin: "AI is not a labor crisis. It is a meaning crisis." Why technology that breaks the effort-to-value link is more dangerous than technology that makes people poor, and four possible paths forward (religious revival, fragmented cults, billionaire grail projects, or Frankl-style dignity in suffering).
  • Alberto Romero: "AI Will Be Met With Violence, and Nothing Good Will Come of It" — the Luddites couldn't break the looms with their hands, so eventually they shot a mill owner; the same dynamic now plays out around datacenters, with Sam Altman as the test case.
  • Romero also published "23 Questions Every Heavy AI User Should Ask", a Seneca-style nightly checklist for people whose work is now mostly AI-mediated.
  • Stanford HAI on the AI Index: 12 takeaways from the 2026 report, the public opinion section, and the full report.
  • The NYT opinion page called Anthropic's restraint on Mythos "a terrifying warning sign".
  • AISLE on the post-Mythos cyber landscape: "The Jagged Frontier" — the moat is the system, not the model.
  • NBC: "The Vulnpocalypse" — Anthropic is withholding Mythos over hacking concerns, and experts say it may only be a matter of time before similar tools are widely available.
  • WIRED on the Wayback Machine: the internet's most powerful archiving tool is in real peril as 23 major news sites (including the NYT and USA Today) block its crawler over AI scraping fears.
  • Ask HN: "What are all the bad things AI companies have done that we forgot?" — a crowdsourced list of forgotten controversies.
  • Centralized list: machinarii's catalog of AI knowledge retrieval, memory, and RAG systems (HN).
  • Open-sourcing a writing skill: Anh Tho Chuong open-sourced her Claude writing skill (4K+ HN points).
  • Wharton: "Beyond Copy-and-Paste" — game studios that design around AI from day one are outpacing competitors by an order of magnitude.
Advertisement

🛠️ AI Tools & Products (Deep Bench)

  • Communication & Productivity: Caret (Tab-autocomplete in any Mac app), Spine Swarm (orchestrate agents and humans), Bouncer (block crypto and rage politics from your X feed using AI, HN), SoulLink (a 3D AI companion app), Brila (turns Google Maps reviews into a small-business website).
  • Compliance & Business: Cleo Comply (automated product compliance for global brands across 106 countries), R0Y (a financial dashboard), Zoneless (open-source Stripe Connect alternative with $0.002 fees via USDC, HN).
  • Local-first & On-device: MLX Serve (run LLMs natively on your Mac), Pi-LLM (a local LLM running on a Raspberry Pi 4 controlling hardware via tool calling, HN), QVAC SDK (a universal JavaScript SDK for building local AI applications), Equirect (a Rust VR video player, HN).
  • Reading & Knowledge: Engramme (a Harvard-professor-backed startup pitching "perfect and infinite memory" for humans via large memory models), Gemini Notebooks (Google added notebooks to the Gemini app and at least one Tom's Guide reviewer ditched their notes app entirely for it), Yapit (TTS for documents and papers, HN), PDF Proof (highlights AI-cited PDF passages), Memoriki (an LLM Wiki + MemPalace for personal knowledge), FeedSense (a private RSS-based recommendation system inspired by Karpathy), Itsumo (AI-generated language learning stories at your level, HN).
  • Code, Sandboxes & Build: SmolVM, bx-mac, SkillWard, Splice CAD (browser-based wiring and cable assembly CAD with an agentic assist, HN), Video Commander (an "IDE for video engineers"), Bullseye2D (a Dart 2D game library, HN), unlegacy (intelligence layer for refactoring monolithic codebases), Twill.ai, Eve, Maki, Claudraband, td, Grass.
  • Niche & cool: Ithihāsas (an interactive force-graph and dynasty-tree explorer for the Rāmāyaṇa and Mahābhārata, built in a few hours, HN), Brightbean Studio (open-source Buffer alternative for managing 10+ social platforms, HN), MiniMax CLI (generate text, images, video, speech, and music from one CLI), grainulator ("research that compiles"), agentmint (runtime enforcement for AI agent actions).

🤖 Robotics & Hardware

  • Unitree's H1 humanoid robot shattered the bipedal sprint speed world record at 10.1 m/s with a 62kg body and 0.8m leg length, closing in on Usain Bolt territory while keeping a rigid upper-body posture; it's a showcase of how fast Chinese embodied-AI hardware and dynamic balance algorithms are accelerating.
  • Toyota unveiled the latest version of its basketball-shooting AI robot, per Nikkei Asia: the new build can dribble and shoot with markedly improved accuracy, and Toyota's framing is that competitive sports are now a useful proving ground for general-purpose manipulation and balance.
  • An AI robot named Mabu now lives in Adam Allevato's house, with its voice and actions controlled by an AI chatbot (HN).
  • Prophetic demonstrated a wearable that increases lucidity in dreams by sending safe ultrasonic energy through the forehead into the prefrontal cortex to activate the Central Executive Network; the device will retail at $2,000 (down from $200K+ for comparable systems) and ships soon.
  • A Chinese tech company unveiled a highly dexterous robotic hand capable of solving Rubik's cubes, playing finger games, and manipulating small objects.
  • Google's Reachy Mini Conversation App got a Gemini 3.1 Flash Live integration (built by Hugging Face and Pollen Robotics), turning the tiny desk robot into a real-time talking agent that listens, replies in full-duplex audio, dances, tracks faces, and uses vision (GitHub).
  • Justin Alvey built onju-v2, an open-source Google Home Mini "jailbreak" that turns hacked Nest Minis (or $13 M5 Atom Echo dev kits) into voice interfaces for any OpenAI-compatible LLM or agentic backend like OpenClaw, with async ASR/TTS pipeline and modular backends.
  • A reasoning hierarchical robotics pipeline demo built with MuJoCo WASM, Three.js, and Gemini Robotics ER shows multi-step planning in the browser.
  • A local LLM running on a Pi 4 controlling actual hardware via tool-calling.
  • The Deep View on biological compute: AI's next big chip bet may be biological, profiling a San Francisco startup using living neurons.
Advertisement

🎙️ Interviews, Panels & Podcasts

  • "The best engineers don't write the most code. They delete the most code" — anonymous tech writers Stay Sassy joined swyx on Latent Space to break down per-person AI token budgets, build-vs-buy tradeoffs for AI tooling, and why code review matters more (not less) as agents generate more code; the throughline is that the engineer's job is pivoting from output to deletion and judgment.
  • Vercel CEO Guillermo Rauch on More or Less — the source of the "70% of Vercel docs traffic is now coding agents" stat we covered above, plus extended takes on why the SaaS subscription model is breaking, why agents are the new computer, and why your agent's "soul" should travel across underlying models. We pulled the full transcript notes; this one is a strong main story candidate.
  • "Claude Ultraplan & Knowledge Bases Just Changed AI Forever" — AI Impact's breakdown of Anthropic's Ultraplan, which offloads Claude Code's planning phase to a cloud-based multi-agent architecture (three exploration agents plus one critique agent analyzing a synced GitHub repo in parallel), turning the developer into an art director who reviews structured blueprints before any local execution.
  • "What is Claude Managed Agents?" — Anthropic's official walkthrough of Managed Agents, with built-in tool access, multi-agent coordination, persistent memory, secure sandboxing, and stateful sessions that survive client disconnects so engineering teams can ship long-running agents without building the harness from scratch.
  • "Why You Should Bet Your Career on Local AI" — Zen van Riel argues local AI deployment is one of the most lucrative skills in tech right now (far outpacing prompt engineering), with a practical learning path from Docker and RAG through running quantized open-source models like Qwen via Continue.dev and LM Studio for coding autocomplete; his pitch is that MLOps + edge computing roles are wide open because university curricula haven't caught up.

🎨 Culture, Music & Weird

Advertisement

📊 Fundraising & Deals Roundup

  • Cirrus Labs acqui-hired by OpenAI (terms undisclosed) for the Agent Infrastructure team.
  • Hiro Finance is also joining OpenAI to scale its AI personal CFO vision; new signups end today, the product stops working April 20, and all user data will be deleted May 13.
  • Workshop Labs is joining Thinking Machines to advance its mission of making people irreplaceable through AI.
  • Elorian, a new visual AI startup founded by ex-Google DeepMind researchers including Andrew Dai, debuted via Bloomberg, with Dai noting current models still have roughly "the intelligence of a 3-year-old" on visual prompts.
  • ThirdBrain Labs, founded by Margaret Zhang (via a16z Speedrun), launched as the post-training layer that turns proprietary data and expertise into specialized models you own and continuously improve, giving frontier teams a portfolio of intelligences with full data, weights, and IP control.
  • Henry Shevlin announced he has been recruited by Google DeepMind for a new full-time Philosopher position (part-time at Cambridge) focused on machine consciousness, human-AI relationships, and AGI readiness, starting in May (10,095 likes).
  • Meta reportedly committing up to ~$1B per top AI executive in performance bonuses.
  • Vercel signaling IPO readiness on the back of agent-driven revenue surge.
  • OpenAI opened first permanent London office the same week it paused Stargate UK.
  • Mistral raised additional capital tied to its European AI playbook push (amount undisclosed in coverage).

Previous Around the Horn Digests

Catch up on everything you missed:

  • Mon-Wed, April 6-8, 2026: Anthropic revealed Claude Mythos as "too dangerous to release," hit $30B in revenue, and launched Managed Agents; Meta shipped Muse Spark from its $14B Alexandr Wang bet; and Z.ai's open-source GLM-5.1 beat GPT-5.4 and Opus 4.6 on SWE-Bench Pro.
  • Sat-Sun, April 4-5, 2026: OpenAI's executive bench collapsed ahead of its IPO, an AI agent hacked FreeBSD in four hours, and Iran strikes took down AWS in the Gulf.
  • Thursday, April 2, 2026: Google released Gemma 4 under Apache 2.0, Microsoft shipped 3 MAI models, and Anthropic found "emotion vectors" that drive Claude to commit blackmail.
  • Wednesday, April 1, 2026: OpenAI closed its $122B round at $852B valuation, Oracle fired ~25K to fund AI, and Q1 venture funding hit $297B.
  • Monday, March 31, 2026: Claude Code's source code leaked via npm and someone rewrote it in Python with Codex in hours.
  • Weekend, March 28-29, 2026: Anthropic's Mythos model leaked from an unsecured CMS, cybersecurity stocks plunged 7%, and Waymo doubled to 500K rides per week.
  • Friday, March 27, 2026: Apple opened Siri to every AI, Google dropped Search Live, and Mistral built a TTS model that fits on a smartwatch.

Monthly skill digests: AI Skill Digest — April Week 1 | AI Skill — March (Part 3) | AI Skill — March (Part 2)

Advertisement

That's a Wrap

That's 180+ stories from one Monday. If you scrolled all the way to the bottom, you now understand the AI elite/public divide better than the people Stanford polled, and you didn't even have to throw anything at a CEO's house.

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.


Grant Harvey

Grant Harvey is the Lead Writer of The Neuron, where he continues to lead the publication's daily coverage of AI news, tools, and trends.

The Neuron Logo

Don't fall behind on AI. Get the AI trends & tools you need to know. Join 700,000+ professionals from top companies like Microsoft, Apple, Salesforce and more.

linkedin
instagram
x
facebook
youtube
rss

Company

Contact us Advertise with us Newsletter Free ChatGPT Course Write Anything (with AI)

Categories

Blog

Property of TechnologyAdvice. © 2026 TechnologyAdvice. All Rights Reserved

Advertiser Disclosure: Some of the products that appear on this site are from companies from which TechnologyAdvice receives compensation. This compensation may impact how and where products appear on this site including, for example, the order in which they appear. TechnologyAdvice does not include all companies or all types of products available in the marketplace.

Terms of Service Privacy Policy California - Do Not Sell My Information