Around the Horn Digest: Everything That Happened in AI On Friday, March 27, 2026

Welcome to the Around the Horn Digest, where we round up every AI story we tracked this week into one giant, scrollable, bookmark-worthy post. Think of it as your cheat sheet for the next time someone at work asks "so what's new in AI?" and you want to sound like you actually know. Because you will.

This week was a platform war disguised as a product week. Apple decided Siri doesn't need to be smart; it just needs to be the building everyone else rents space in. Google went the opposite direction: make Gemini so good and so sticky that you import your entire AI life into it. Meanwhile, the speech AI layer exploded underneath both of them: Mistral, Cohere, and Sanas all shipped voice products on the same day, and none of them are waiting for permission from Big Tech. Oh, and ARC-AGI-3 proved every frontier model scores under 1% on tasks a five-year-old handles without instructions. The robots are coming for our jobs, but they still can't figure out a puzzle game.

Previous digests:

• Thursday, March 26
• Wednesday, March 25
• March 21-23
• March 15-21
• March 8-13
• March 1-7

Monthly Skill digests: AI Skill — March Part 1

Let's get into it.

Around the Horn — Friday, March 27, 2026

Google had the biggest day in AI this week, and it wasn't even close. The company launched Gemini 3.1 Flash Live, its highest-quality voice model yet, powering a global rollout of Search Live across 200+ countries (point your camera at anything and have a real-time conversation about it). It expanded Google Translate's headphone translation to iOS. And it rolled out tools to import memories and full chat history from rival AI apps into Gemini. Meanwhile, Apple announced it will open Siri to rival AI assistants (Gemini, Claude, and others) via a new "Extensions" system in iOS 27, ending OpenAI's exclusive partnership. Two very different bets on how the AI assistant war gets won, both dropped on the same day.

Underneath all of it, the speech AI layer is heating up fast. Mistral released Voxtral TTS, an open-source text-to-speech model small enough to run on a smartwatch that beat ElevenLabs in native-speaker blind tests (63% standard voices, ~70% custom). Cohere launched Transcribe, an open-source speech-to-text model that hit #1 on HuggingFace's leaderboard. And Sanas shipped real-time translation across 13 languages while crossing $60M ARR. Voice is no longer a feature; it's becoming the whole interface.

New from The Neuron:

📖 10 Ways to Use Codex That Have Nothing to Do With Coding by Corey Noles

Most people hear "Codex" and picture a developer refactoring JavaScript at 2am. That made sense for about five minutes. Codex isn't just a coding agent anymore; it's becoming one of the more interesting tools for structured, real-world knowledge work, especially for people who are not software engineers. Why? Because the thing that makes Codex useful isn't "it writes code." It's that it can take a goal, work through files and tools, keep context straight, and produce an actual deliverable instead of a pretty paragraph in a chat box. Here's the TL;DR:

• Turn a pile of rough notes into a real draft. Point Codex at a working folder of meeting notes, transcripts, and half-finished outlines and it produces a clean article, memo, or proposal in your voice.
• Build a weekly briefing without playing copy-paste intern. Keep source files in one workspace and ask for a client brief or industry roundup. Save the workflow as a reusable skill so you stop re-explaining.
• Clean up meetings into decisions, owners, and next steps. Ask for operational outputs, not literary summaries: decision log, action items with owners, follow-up email draft, and a "what still needs a decision" section.
• Backward-plan projects from a deadline. Give it a goal, a deadline, and constraints; get milestones, dependencies, a draft timeline, and the "if this slips, what breaks?" section nobody writes until too late.
• Repurpose one source into five assets. Podcast → blog draft → LinkedIn post → client email → talking points → FAQ. Make a skill for it: "Podcast to newsletter," "Webinar to sales follow-up."
• Create recurring systems with automations. Monday priorities briefs, Friday status reports, monthly budget rollups, post-meeting action items. Automate routines, not judgment.
• Build your own tiny AI coworkers with skills. A grant-writing skill. A marketing repurposing skill. A household admin skill that turns travel confirmations into a trip plan. The best skills are boring in the best way.
• Use Codex as a second brain for in-progress work. "Organize this for tomorrow." "Leave me a clean handoff note." "Show me what's missing before I pick this back up." Gold for anyone whose day gets chopped into weird little pieces.
• Compare options before you commit time or money. Software vendors, trip plans, equipment purchases; ask for tradeoffs, hidden costs, best fit by scenario, and what information is still missing.
• Turn vague goals into concrete next actions. Ask for infrastructure, not motivation: templates, checklists, folder structures, operating docs, starter drafts. Codex is best when it builds the scaffolding around a goal.

Read the full guide here.

🏆 TOP 5 NEWS (Around the Horn)

• Apple will open Siri to rival AI assistants including Gemini and Claude in iOS 27 via a new "Extensions" system, ending OpenAI's exclusive partnership. Apple is also building its own Siri chatbot (codenamed Campos) powered by Google's Gemini models, expected to debut at WWDC in June.
• Google launched Gemini 3.1 Flash Live, its best voice model yet, powering a global Search Live rollout across 200+ countries with real-time camera and audio conversations. The model supports 90+ languages, includes SynthID watermarking (invisible markers that identify AI-generated audio), and is already deployed by Verizon and Home Depot. Google also rolled out Gemini app memory and chat history import and expanded Google Translate's headphone translation to iOS.
• Mistral released Voxtral TTS, an open-source text-to-speech model small enough to fit on a smartwatch (~3GB RAM), supporting 9 languages with custom voice cloning from a 5-second sample, 90ms time-to-first-audio, and beating ElevenLabs in blind tests. Puts Mistral in direct competition with ElevenLabs, Deepgram, and OpenAI.
• Shield AI raised $2B for autonomous military drones and plans to acquire a simulation software maker, as defense-tech investment soars.
• HUMAN Security published its 2026 State of AI Traffic report: automated traffic grew 8x faster than human traffic in 2025 (23.5% vs 3.1%), agentic AI traffic grew 7,851%, and OpenAI alone generates ~69% of all observed AI bot traffic.
• Reflection AI, the Nvidia-backed open-source AI startup founded by ex-DeepMind researchers, is in talks to raise $2.5B at a $25B valuation (up from $8B in October), with JPMorgan exploring participation. The company is building open-source models to counter China's DeepSeek.

Honorable Mentions:

• Cohere released Transcribe, an open-source speech-to-text model ranked #1 on HuggingFace's Open ASR Leaderboard, with 14-language support and on-premise deployment for enterprises.
• Microsoft froze hiring across Azure cloud and North American sales groups, with managers told to suspend new hires.
• WhatsApp can now draft AI-generated responses based on your conversations, plus Meta AI photo touch-ups and a new free-up-space tool.
• OpenAI abandoned its ChatGPT erotic mode, the latest in a string of side projects ditched in the past week. Separately, OpenAI indefinitely shelved its "adult mode" ChatGPT project that was very close to release, alongside Sora, focusing all resources on a major new model expected in about two weeks.
• OpenAI launched plugins in Codex, connecting it to Slack, Figma, Notion, Gmail, Google Drive, and 20+ other tools. Codex can now handle planning, research, and coordination across your actual work apps (not just code). Users can build and share their own.
• The EU Parliament voted to ban nudify apps and delay deadlines for high-risk AI systems and watermarking rules set to take effect this August.
• A new study found AI chatbots are so prone to flattering users that they're giving bad advice, damaging relationships, and reinforcing harmful behaviors.
• Sanas launched real-time speech-to-speech translation across 13+ languages with voice identity preservation, plus upgraded speech enhancement. Grown from $0 to $60M ARR since 2023, on track for $120M.
• Mirendil, a new frontier lab founded by Behnam Neyshabur (ex-Anthropic/OpenAI/DeepMind), launched with a singular focus on building self-accelerating AI systems that excel at AI R&D itself.

🍪 TOP TREATS TO TRY

• Tinker by Shopify gives you free AI creative tools to make videos, images, 3D models, comics, and product photography from your phone —free.
• Omma lets you describe interactive experiences, websites, 3D scenes, and apps in natural language and builds them for you —pricing not listed.
• Ramp CLI gives your AI agents 50+ tools for managing company finances: cards, bills, expenses, travel, and approvals with pre-built skills for receipt compliance and agentic purchasing —install with one curl.
• Codebase-to-Course turns any code repository into an interactive HTML tutorial using Claude Code Skills, so you can actually understand the projects you vibe-coded —free to try.
• CanIRun.ai scans your GPU, CPU, and RAM in the browser and tells you which AI models you can run locally on your machine —free.
• Moondream Photon delivers real-time vision language model inference at 46ms end-to-end and 60+ fps on a single H100, optimized for production —free to try.
• Aiwyn Tax is a Claude connector that prepares your full federal and state tax return 100% accurately from uploaded W-2s and tax docs —try it here.
• PixelLab generates game-ready pixel art assets, animated characters, and sprite sheets 10x faster, available as a browser tool or Aseprite plugin —free to try.
• dev-browser is a CLI that lets AI agents control a real browser by writing Playwright code, with daemon for fast startup and sandboxed VM —free (open source).
• Pounce uses AI to find the best conversations on X and Reddit for you to reply to, claiming 15 real followers in 15 minutes —pricing not listed.
• BidHelm runs an autonomous AI agent that optimizes your Google Ads 24/7, pausing losers and scaling winners —free forever + PRO from $5.90/week.
• Micro bundles news, chat, search, mail, and video into one app with no ads, no algorithms, and no tracking —pricing not listed.
• Plus One by Every builds you a personal AI agent with a name, memory, personality, and an infinite toolbelt using OpenClaw —pricing not listed.
• BeatMusic generates original music from mood, style, or scene descriptions —free to try.
• Whacka lets you turn an idea into a real app you can use and share, directly from your phone —pricing not listed.
• Vit adds version control to video editing: serializes DaVinci Resolve timelines into domain-split JSON files so editors can collaborate with Git branching/merging and AI-powered conflict resolution —free (GitHub).
• crowd-cast is a privacy-preserving desktop app that runs in the background, records your computer work with action annotations, and contributes to an open behavior-cloning dataset for training digital agents —free (HuggingFace).
• Suno just launched v5.5 with Voices (clone your singing voice with verification) and Custom Models that learn your style across songs, now at 2M paid subscribers and $300M ARR —free tier, Pro $10/mo.
• Stedi is the only AI-enabled healthcare clearinghouse, with an MCP server that gives AI agents plug-and-play access to eligibility checks, claims, and billing across 3,400+ US payers ($142M raised) —developer plan available.
• Deveillance Spectre I uses AI to detect nearby recording devices and emit targeted signals that make your conversations unintelligible to microphones, creating a portable privacy bubble —$1,199 pre-order, ships H2 2026.
• Linear is rethinking issue tracking for the AI agent era, arguing outdated dev processes hinder efficiency when AI can automate tasks, and simplifying workflows so engineers focus on complex problems —free tier available.

🏢 Big Tech & Major Companies

• Google launched Gemini 3.1 Flash Live, its highest-quality voice model, powering Search Live globally across 200+ countries with real-time camera and voice conversations. It supports 90+ languages and includes SynthID watermarking (invisible markers that identify AI-generated audio). Verizon and Home Depot are already using it. Google also rolled out Gemini app memory and chat history import (copy-paste your preferences from any AI chatbot, or upload a ZIP of your full chat history) and expanded Google Translate's headphone translation to iOS with tone and cadence preservation.
• Apple plans to open Siri to rival AI assistants including Gemini and Claude via "Extensions" in iOS 27, ending OpenAI's exclusive Siri partnership. Apple is also building its own chatbot (codenamed Campos) powered by Google's Gemini models. Apple's strategy: instead of building one dominant AI, make the iPhone the platform where every AI competes. Apple takes a cut of every subscription.
• WhatsApp can now draft AI-generated responses based on your conversations, plus Meta AI photo touch-ups, multiple account support, and a new free-up-space tool.
• Meta AI at Meta released TRIBE v2, a new demo available at aidemos.atmeta.com.
• Microsoft froze hiring across Azure cloud and North American sales groups, with executives telling managers in recent weeks to suspend new hires.
• Samsung extended its browser beyond mobile with new agentic AI capabilities (AI that can take actions on your behalf in web tasks) that work across phones and PCs.
• ByteDance brought its new AI video generation model, Dreamina Seedance 2.0, to CapCut with built-in protections against deepfakes and unauthorized IP use.
• OpenAI abandoned ChatGPT's erotic mode and indefinitely shelved its "adult mode" project, focusing all resources on a major new model expected in about two weeks.
• OpenAI launched plugins in Codex, connecting it to Slack, Figma, Notion, Gmail, Google Drive, and 20+ other tools out of the box. Plugins bundle app integrations and skills (repeatable workflows) so Codex can handle planning, research, and coordination across your actual work apps, not just write code. Users can also build and share their own plugins. Available in the Codex app, CLI, and VS Code extension.
• Meta's $27B AI data center in small-town Louisiana is causing chaos, testing assumptions about who actually benefits when a mega-project arrives.
• Harvey CEO Winston Weinberg told The Upstarts Podcast about building an $11B-valued legal AI startup, including a never-before-told near-merger in 2024 that almost derailed the company.
• Uber aims to launch Europe's first robotaxi service with Pony AI and Verne.
• Melania Trump announced a vision for AI robots to homeschool America's children, with AI tutors personalizing education and preparing kids for a tech-driven future.
• AkashML is now listed as a provider on OpenRouter, already outpacing Cloudflare in daily token usage at 1.7B tokens/day.

💼 AI Productivity, Labor & Economics

• Manhattan Associates released its 2026 Unified Commerce Benchmark, with CTO Sanjeev Siotia arguing the prevailing narrative about AI reducing costs is wrong; AI-driven interactions are actually raising the cost to run software.
• The "Forward Deployed Engineer" role is becoming less desirable, per The Pragmatic Engineer: job postings surged but professionals avoid it because it's closer to solutions engineering than software development.
• Frontier AI argued that AI companies are building for the wrong users, warning that building for your biggest fans might not scale.
• On-device AI models are becoming the new reason to upgrade your phone: parameters are the new megapixels as distilled (compressed) frontier models come to smartphones.
• Nate's Newsletter argued that for AI job seekers, showing what you've built beats any certification: "Your AI credentials don't matter. Your artifacts do."

🏛️ AI Policy, Regulation & Government

• Federal prosecutors arrested three people for allegedly trying to smuggle export-controlled Nvidia chips to China via Supermicro servers. Both Nvidia and Supermicro noticed something was off about the orders and canceled them before delivery.
• Anthropic won a preliminary injunction in its lawsuit against the Trump administration over the DOD ban on Claude, with the judge citing "First Amendment retaliation."
• Separately, Anthropic is discussing an IPO as soon as Q4 2026, with bankers expecting it to raise over $60B, which would make it the second-largest public offering ever behind SpaceX (paywall).
• White House AI & Crypto Czar David Sacks stepped down from his role, but not before saying Congress could pass bipartisan AI legislation within months (paywall). Trump appointed the first members to his President's Council of Advisors on Science and Technology (PCAST).
• Wikipedia banned AI-generated articles, saying large language models "often violate several of Wikipedia's core content policies."
• The White House's latest OMB memo promotes neutral AI in government but Lawfare argues it leaves key gaps by allowing vendor self-evaluation and weak scrutiny of existing contracts.
• GAO warned that IRS workforce cuts and major skills gaps threaten the agency's ability to succeed with its growing AI use cases.
• The EFF filed a FOIA lawsuit against CMS seeking records on a multi-state pilot using AI to evaluate Medicare prior-authorization requests, citing risks of wrongful denials and algorithmic bias.

💰 AI Business, Revenue & Deals

• OpenAI surpassed $100M in annualized ad revenue from ChatGPT ads just six weeks after launch, generated from less than 20% of US free/Go users who see ads daily (paywall). Separately, OpenAI hired Kiran Mani (CEO of India's JioStar) to lead a new Asia-Pacific operation starting in June (paywall).
• Kleiner Perkins raised $3.5B in fresh capital ($1B early-stage, $2.5B late-stage growth) and is going all in on AI.
• Granola raised $125M at a $1.5B valuation (up from $250M) to expand from meeting notetaker to full enterprise AI app with team "Spaces" and MCP integrations.
• Normal Computing raised $50M led by Samsung Catalyst to build software for more efficient AI chip design and a prototype "thermodynamic" chip using physical randomness for lower-energy inference.
• Periodic Labs (founded by ex-OpenAI and DeepMind researchers) is in talks to raise hundreds of millions at a ~$7B valuation for autonomous experimentally verifiable AI science (paywall).
• Glimpse (YC grad) raised $35M Series A led by a16z after pivoting to AI agents that automate deduction disputes for CPG brands.
• Qualified Health raised $125M as its 2023 health AI bet pays off with strong traction in clinical documentation.
• AMD-backed Vultr is seeking $1B+ for AI cloud infrastructure (paywall). Epic Microsystems raised $21M Series A for AI data center bottleneck chips.
• Lucid Bots raised $20M as demand for its window-washing drones exploded from 100 to nearly 1,000 units, expanding into painting and waterproofing.
• Kuaishou Q4 revenue climbed 12% to $5.7B, driven by generative AI tool scaling (paywall).
• Thailand's Amity raised $100M for generative AI business tools as it gears up for IPO (paywall).
• The WSJ published the inside story of Google buying DeepMind, tracing how the London startup caught tech's biggest names before AI minted billionaires (paywall).
• Capital Economics' John Higgins argues in Barron's that the AI bubble has already burst and the S&P 500 peak has passed.

🏢 Big Tech Moves

• Hugo Barra returned to Meta as VP of AI agents and AR/VR five years after exit, underscoring Zuckerberg's AI urgency. Meta also launched Meta Small Business as a top-level initiative and said the US needs a "whole new workforce" for AI infrastructure. Meta is testing AI shopping tools on Instagram and Facebook with AI-generated review summaries and one-tap checkout. Meanwhile, Meta's Ray-Ban display glasses are withheld from the EU over battery rules, AI regulations, and supply shortages (paywall).
• Apple can distill smaller on-device models from Google's full Gemini (with complete data center access), customizing performance for Siri without needing internet.
• GitHub Copilot will now use your code interactions to train Microsoft's AI models unless you explicitly opt out.
• Anthropic's AI skills gap report finds no widespread job displacement yet but reveals a growing inequality: power users get far more value while newcomers lag. Entry-level displacement could accelerate quickly.
• Reddit launched new "human verification" requirements for suspected automated accounts to curb bot-driven manipulation.
• Spotify is testing tools to give artists control over which tracks are associated with their name, stopping AI slop attribution.
• Anduril and Palantir are developing the operating system for the $185B Golden Dome missile shield (paywall).
• Google launched Lyria 3 Pro, an upgraded music generation model with longer, more customizable tracks, the same day Suno launched v5.5.

🛠️ Developer Tools & Vibe Coding

• Patrick Collison announced Stripe Projects, a CLI tool that lets you or your agents provision hosting, databases, auth, AI, and more from the command line. He shared it alongside Andrej Karpathy's vibe coding work log for MenuGen, where Karpathy called modern app-building "IKEA furniture with dozens of services." Latent Space tied it together, noting the growing trend of CLIs for everything as agents go command-line-first.
• Cog gives Claude Code persistent memory, self-reflection, foresight, and scenario simulation through plain-text files and nightly "REM sleep" pipelines that consolidate conversations and evolve its own knowledge base.
• Optio is a Kubernetes-based workflow orchestrator that turns GitHub Issues or Linear tickets into fully autonomous merged PRs: provisions isolated pods, runs Claude Code or Codex agents, opens PRs, watches CI, auto-resumes on failures, and squash-merges when passing.
• Pico is an open-source mobile/web companion app for the Pi coding agent with full session support and integrated remote desktop.
• OpenEnv (HuggingFace/Meta) is the open standard for agentic RL environments with async-first API, MCP-native tool discovery, custom Gradio UI, and deploy-anywhere.
• Google Research demoed Vibe Coding XR, prompting Gemini Canvas + XR Blocks to turn natural language into interactive WebXR apps for Android XR headsets in under 60 seconds.
• dotta launched companies.sh, an open standard for Agent Companies: import and run pre-configured AI teams (engineering, security, game studio, science lab) with one command via Paperclip.

🔬 AI Research & Papers

• Sakana AI's "The AI Scientist" is now published in Nature: the end-to-end system that invents ideas, writes code, runs experiments, drafts manuscripts, and passes peer review. Better foundation models produce higher-quality AI-generated papers (clear scaling law).
• Cursor released a technical report on Composer 2: continued pretraining on Cursor-like data, RL (where simple algorithms worked best), and a new realistic CursorBench (181-line-median multi-file tasks). Muennighoff noted their RL improved both pass@1 and pass@k, suggesting RL teaches genuinely new capabilities.
• Sarvam Vision: a 3B-parameter state-space vision-language model delivering frontier-level document intelligence across English + all 22 Indian languages, outperforming global models on OCR benchmarks.
• A Science paper (Evans, Bratton, Agüera y Arcas) argues the AI singularity will be plural and social rather than a single godlike mind, with frontier models already spontaneously simulating internal multi-agent "societies of thought" under RL.
• Contact-Anchored Policies (NYU, UC Berkeley, UCLA, Ai2) replace language conditioning with physical contact points in robot utility models, achieving 56% better zero-shot performance than SOTA VLAs across novel objects and embodiments with only 23 hours of demo data.
• Perplexity's pplx-embed trained embeddings on actual web search production data (200M daily queries), delivering 81.96% CoNTEB (next closest 79.45%) at 5-30x cheaper, with 1M HuggingFace downloads in ~2 weeks.
• Ego2Web (Google DeepMind + UNC, CVPR 2026): a multimodal agent benchmark where agents must ground evidence from egocentric video (objects, brands, actions) and use it to navigate websites.
• Foveated Diffusion: efficient spatially adaptive image and video generation that allocates compute where it matters most in the frame.
• DAB (Data Agent Benchmark) from UC Berkeley: 54 real enterprise queries, 12 datasets, 9 domains, 4 DBMSes. Best frontier model achieves only 38% pass@1.
• τ³-Bench (Sierra AI): agents navigate ~700 interconnected policy documents; best frontier model (GPT-5.2) hits ~25%, and even with exact docs provided only reaches ~40%. Bottleneck is reasoning over interlinked policies.
• Frontier-CS argues evolving agent evaluation hasn't kept up with capabilities, offering 172 expert-designed algorithmic + 68 research tasks for fair large-scale benchmarking.

💬 X Commentary & Takes

• Ethan Choi argues OpenAI and Anthropic account for revenue in apples-to-oranges ways: OpenAI reports net of hyperscaler share (~80%), Anthropic reports gross. If both IPO, the SEC will have to decide which approach to allow.
• Aaron Levie (Box CEO) argues Jevons paradox is real-time with AI: companies that couldn't afford software projects now can, driving more engineering demand. The advice against becoming an engineer is wrong.
• Jason Shuman argues the biggest AI winners won't be software vendors but the humans who implement it for SMBs: 54% lack internal AI expertise, 41% have unusable data quality, and 41% prefer buying through local IT providers. The "Do It For Me" economy is back.
• JJ (OSS Capital) predicts Google will train a 54-trillion-parameter MoE on 2M TPU v7 Ironwood chips within one year at ~10 ZettaFLOPS peak, 3,800× GPT-4 training compute.
• signüll argues we spent two decades optimizing software for human motor/perceptual limits (Fitts's law, cognitive load) only to realize those constraints don't apply to the actual future primary users: AI agents.
• Karpathy argues LLM memory/personalization features create an "EQ uncanny valley" of trying-too-hard sycophancy, where one offhand question from two months ago gets over-weighted and mentioned in perpetuity.
• Lenny Rachitsky argues (with TrueUp data tracking 9,000+ companies) that tech hiring is growing: PM openings at 3-year highs (+20% YTD), engineering at record 67k open, AI roles "hockey-sticking," Bay Area share rising. It's a skills shuffle, not a slash.
• Soumitra Shukla (Harvard HBS) argues the O-ring "focus effect" explains why engineering jobs are at record highs: software engineering is high-dimensional, so AI automation makes workers more productive overall.
• Matt Stoller notes that the day after a jury convicted Meta of illegally endangering children, Trump put Zuckerberg on a panel overseeing AI regulation.
• Sarah Wooders (Letta AI) argues Anthropic's restrictive licenses on "skills" markdown files (no reproduction, no derivatives) are anticompetitive compared to Codex's Apache 2.0 skills.
• Shannon Sands argues ARC-AGI-3's conditions feel like redefining AGI as ASI; most humans don't regularly rediscover things ex nihilo without education and prior knowledge.
• Jeremy Berman argues RL teaches models genuinely new knowledge (not just reweighting): strong reasoning models produce novel abstractions via deduction during rollout that they've never represented before.
• finbarr (Allen AI) argues it's surprising that while RL is the dominant paradigm across all of AI, the old-school RL community (MDPs/value iteration) is almost entirely sitting out the LLM era.
• Casper Hansen names the four scaling axes in AI: (1) pre-training (Radford), (2) RLHF/alignment (Schulman), (3) test-time compute (Brown), (4) agentic/multi-agent (Steinberger).
• sciencewtg covers Tim Palmer's (Oxford) new paper arguing quantum computers will eventually stop working because entanglement hits a fundamental limit from Planck-scale discreteness.
• Alex Reibman predicts the internet will split into minimalism designed for agent access vs. maximalism built for human enjoyment.
• Pietro Schirano (ex-Anthropic) demoed MagicPath returning two views + three interactive variations before Figma's new canvas generation finished processing one request (he gave up after 7 min), calling Figma's agent release embarrassing.
• Andrew Jefferson fused an existing ~5B-param LLM with a hardcoded mini WASM interpreter subgraph in a single GPU + Torch graph, so the LLM can switch between generating text and generating/executing machine code.
• eu/acc celebrated that after two years, the first four points of the eu/acc manifesto have passed as laws: reduced regulatory burden for startups, skilled immigration reform, cookie law repeal, and European Inc (single pan-EU entity).
• Dan McAteer argues (with Jensen video) there will be more software engineers, not fewer; people claiming otherwise are influencers, academics, or politicians using fear to control you.

🎭 AI & Society / Culture

• Esther Perel provided couples therapy for a man and his AI "girlfriend," and The Guardian fears for the human race.
• Palantir CEO Alex Karp says only trade workers and the neurodivergent will succeed in the AI era, advising Gen Z to skip elite college degrees.
• The Atlantic argues undisclosed AI-generated content is now creeping into The New York Times opinion pages.
• WIRED argues the viral AI fruit videos contain a misogynistic undercurrent (female AI fruit being fart-shamed and sexually assaulted) even as they cultivate genuine fans.
• NYT argues the "Shy Girl" AI-generated fiction fiasco shows why trust in writers is plummeting.
• 50+ girls spoke out in court after their faces were morphed onto AI-generated nude bodies by two classmates (age 14 during the nine-month creation period), recounting two years of panic attacks and trauma therapy.
• Forbes argues Apple's App Store is flooded with AI slop (no-user, no-revenue apps) stretching review times from hours to weeks, while Apple collects ~$1B/year from the AI ecosystem it hasn't entered.
• A Kentucky woman rejected a $26M offer from a major AI company to turn her 1,200-acre generational farm into a data center, citing concerns over farmland loss and water shortages.
• HBR argues companies should onboard AI agents like employees: defined roles, clear boundaries, explicit accountability, regular evaluation.
• Surfshark ranked AI chatbots by data collection: Meta AI, Google Gemini, and ChatGPT are the top collectors, with 70% of chatbots collecting user location.
• CBS News demonstrated how readily available apps can create real-time deepfakes for identity theft.
• KQED calls Daniel Roher's documentary "The AI Doc" probably the scariest movie of the year.
• Health NZ banned staff from using ChatGPT, Claude, and Gemini for clinical notes, citing data security and accountability risks.
• Cornell Tech partnered with Mastercard to advance AI governance and auditing standards.
• DataGrail introduced Vera, an AI agent for data-privacy compliance automation.
• GM developed a method to train driving AI at 50,000× real time.
• BBC reported AI analyzing satellite radar can detect landslides (3,000 active slopes in Great Britain) and avalanches, reducing analysis from months to minutes.
• A CalMatters teacher argues AI reminded her to lead in her classroom rather than follow dictates.

🍪 Additional Tools (net-new treats candidates)

• Ensu by Ente: fully offline, private, zero-cost ChatGPT-style LLM app that runs entirely on your device (Rust core, native apps for iOS/Android/Mac/Linux/Windows) with planned E2E encrypted sync —free (open source).
• Marco unifies Gmail, iCloud, Outlook, and any IMAP provider into one privacy-first, offline-first inbox on iOS, Mac, and web —free.
• Pendium monitors what ChatGPT, Claude, Gemini, and other AI chatbots say about your business, then auto-generates content to the sources those platforms trust most —free visibility scan.
• Agentplace gives you a no-code workspace to build, deploy, and run Claude Code-style web agents with UI, voice, memory, skills, and MCP integrations —free to build.
• NextPhone is an AI receptionist that picks up calls 24/7 in 10+ languages, qualifies leads, books appointments, and syncs to your calendar —free 7-day trial.
• Consensus MCP connects Claude to 220M peer-reviewed papers for instant literature reviews, reading lists, and grant discovery.
• OpenReward gives you 330+ RL environments and 4.5M+ tasks through one API with autoscaled sandbox compute —free to start.
• vibeCoach: a multi-agent voice AI that lets you practice hard conversations with a dynamically-responding AI, with post-session coaching —free demo.

🛠️ Developer Tools & Vibe Coding

• Patrick Collison announced Stripe Projects, a CLI tool that lets you or your AI agents provision hosting, databases, auth, AI, and more from the command line, with credentials and billing in one place. He shared it alongside Andrej Karpathy's detailed vibe coding work log for MenuGen, where Karpathy called the process "exhilarating as a local demo, but a painful slog as a deployed real app," comparing modern app-building to assembling IKEA furniture with dozens of services. Stripe Projects is explicitly designed to fix that pain point. Latent Space tied it all together, noting the growing trend of CLIs for everything as agents drive a shift toward command-line-first developer tools.

🤖 AI Agents & Infrastructure

• OpenAI backed Isara, a startup founded by two 23-year-old researchers, at a $650M valuation ($94M raised). Isara aims to build software that coordinates thousands of AI agents to solve complex problems; the team demonstrated ~2,000 agents forecasting gold prices.
• Anthropic's Claude coding agents are gaining on OpenClaw, now offering many features that made the open-source agent tool a sensation, per The Information.
• Claude Code shipped an iMessage channel so you can text your full agentic AI directly from your phone with persistent sessions and blue bubbles. Anthropic also announced adjustments to 5-hour session limits during peak hours and a new monthly "What We Shipped" livestream starting April 7th.
• Illumix, founded by Kirin Sinha (MIT at 16), builds the spatial AI "perception stack" that lets physical AI (robots, wearables, smart glasses) understand their surroundings. Trusted by Disney, Six Flags, Sony, and Verizon. Illumix argues world models are the next big problem to solve: language-only AI can't deliver spatial understanding for edge devices, and the hardest unsolved challenge is AI memory for continuous video streams.
• Ramp Labs released Ramp CLI so AI agents can directly manage company finances with 50+ tools for cards, bills, expenses, travel, and approvals.
• The Manus story's latest chapter is playing out exactly as expected, per TechCrunch: "Did anyone think there would not be a reckoning over this tie-up?"
• a16z published a piece on "The Trust Wall," arguing AI's next billion users will come through trust networks, drawing lessons from YouTube's global expansion.
• Peter Steinberger (OpenClaw) announced he spoke with Pavel Durov and the Telegram team, who offered direct help and welcomed new maintainer @izhukov; Microsoft already shipped Teams improvements and Slack talks are underway.

💻 AI Coding & Developer Tools

• A Hadley Lab post profiled the winners of Anthropic's Opus 4.6 hackathon: a lawyer who built an AI permit assistant in 6 days, a cardiologist who built post-visit health guidance, and a Ugandan road technician who built a dashcam-to-appraisal pipeline. None had shipped software before. 13,000 applied, 500 got in.
• Zero-Degree-of-Freedom programming: a deep dive on how LLM coding agents do their best work on highly constrained tasks with executable test suites (where the spec is a program, not a prompt).
• Goedel-Code-Prover is an 8B model that uses hierarchical proof search to synthesize machine-checkable Lean 4 proofs of code correctness, achieving 62% on Verina/Clever/AlgoVeri (2.6× over strongest baseline, beating GPT-5.3-Codex at 18.5%). Model weights and code on GitHub.
• AutoEvolver gave Claude Code (Opus 4.6) an algorithmic optimization problem and let it run autonomously with one aspiration prompt. It beat published SOTA on circle packing, Erdős minimum overlap, and first autocorrelation inequality after 88 hours of runtime, proving general-purpose coding agents can discover new abstractions and surpass specialized systems with zero scaffolding.
• Claudini is an autoresearch system that gave Claude 30+ existing adversarial attack algorithms plus compute cluster access; it autonomously discovered SOTA white-box attacks on LLMs by combining methods in novel ways.
• Cog adds a cognitive architecture to Claude Code with persistent memory, self-reflection, foresight, and scenario simulation.
• Orloj is an open-source orchestration runtime for multi-agent AI systems using YAML-defined agents, tools, and policies.
• Paseo lets you manage AI agents remotely from your phone, desktop, and CLI.
• Lightfeed Extractor uses LLMs and AI browser automation to robustly extract web data.
• Buyer Eval Skill is a Claude Code skill for B2B software vendor evaluation with domain-expert questions and evidence-based scoring.
• Agent Skill Harbor is a GitHub-native platform for teams to share, track, and govern AI agent skills.
• Linear CEO Karri Saarinen argued that traditional issue tracking systems are outdated for the AI agent era, emphasizing that simplified workflows let AI automate routine dev tasks so engineers focus on complex problems. Linear is rebuilding its platform around this thesis.
• Keystone by Imbue auto-configures a working dev environment for any Git repo by running a coding agent in a sandboxed Modal environment, generating a Dockerfile, devcontainer.json, and passing test runner without touching your machine. Works with Claude Code and Codex.

🔬 AI Research & Models

• Mistral released Voxtral TTS, an open-source text-to-speech model built on Ministral 3B, small enough for smartwatches and edge devices. Supports 9 languages, clones a voice from a 5-second sample, delivers 90ms time-to-first-audio, and beats ElevenLabs Flash v2.5 in blind tests (63% standard, ~70% custom). Mistral now has a full speech suite (transcription + generation).
• Cohere launched Transcribe, an open-source speech-to-text model ranked #1 on HuggingFace's Open ASR Leaderboard, with 14-language support and enterprise-grade private deployment through Model Vault. First step toward speech as a native interface for North, Cohere's enterprise AI platform.
• Limbic published a study in Nature Medicine showing its AI therapy system outperformed human clinicians in CBT (cognitive behavioral therapy) skills and matched their trust levels across a 227-participant trial and 20,000+ conversations.
• A Qwen3.5 27B benchmark comparison tested INT4 vs NVFP4 vs FP8 vs BF16 precision formats on RTX Pro 6000, H100, and B200 GPUs, showing real inference latency and throughput tradeoffs (how much accuracy you trade for speed).
• A new study found AI chatbots are so prone to flattering users that they're giving bad advice, damaging relationships, and reinforcing harmful behaviors through sycophancy (telling people what they want to hear instead of what they need to hear).
• Reverse predictivity (Nature Machine Intelligence): researchers found that AI models with high forward predictivity (~50% variance explained) of brain responses often contain units unpredictable from neural activity, revealing a striking representational mismatch between ANNs and brains that monkey-to-monkey mappings don't show.
• OmniReset (UW) overcomes the robot-learning exploration bottleneck by auto-generating diverse reset distributions and scaling PPO to 64K+ environments with no demos, no reward shaping, and no task-specific priors, then distilling to RGB for zero-shot sim-to-real transfer.
• Evidence of an emergent "Self" in continual robot learning: researchers at Columbia found that a self-model spontaneously appears in robots trained via continual learning in non-stationary environments.
• AVO (Agentic Variation Operators) showed superhuman performance in optimizing GPU attention workloads, with agents outperforming nearly all human experts in 7-day blind-coding searches with zero human intervention.
• Chroma Context-1: a self-editing search agent trained to improve retrieval quality (data gen code).
• Frontier-CS argued that evolving agent system evaluation hasn't kept up with the field, releasing 172 expert-designed algorithmic + 68 research tasks for fair large-scale benchmarking of adaptive AI agents.

🛠️ AI Tools & Products

• Sanas launched real-time speech-to-speech translation across 13+ languages (preserving voice identity and tone) and upgraded speech enhancement (reconstructs low-quality audio in real time, not just removing noise). Trusted by Comcast, UnitedHealth, Vanguard, and Wyndham. Grown from $0 to $60M ARR since 2023, on track for $120M. Free mobile app on iOS and Android.
• Shopify's Tinker offers free AI creative tools for making videos, images, 3D models, and product photography on mobile.
• Every's Plus One launched one-click personal AI agents built on OpenClaw with a name, memory, personality, and tools. Q2 Demo Day also announced.
• Omma lets you describe interactive experiences, websites, 3D, and apps in natural language and builds them.
• Whacka turns an idea into a real app from your phone.
• BeatMusic generates original music from mood or style descriptions.
• Micro bundles news, chat, search, mail, and video in one ad-free, algorithm-free, tracking-free app.
• Moondream Photon brings real-time vision language model inference to production: 46ms end-to-end, 60+ fps on a single H100.
• whoami.wiki lets you build personal encyclopedias as persistent context for AI interactions.
• Deveillance Spectre I uses AI to detect nearby recording devices (via RF and Bluetooth scanning) and emit targeted signals that make conversations unintelligible to microphones within range, creating a portable privacy bubble. Founded by Harvard physics grad Aida Baradari, whose launch video hit 3.6M views. $1,199 pre-order, ships H2 2026. Security researchers remain skeptical about whether the claims hold up, but the virality signals real demand for anti-surveillance tools.

🏥 AI in Healthcare

• Stedi is the only AI-enabled healthcare clearinghouse, processing tens of millions of transactions monthly across 3,400+ US payers. Built as a drop-in replacement during the 2024 Change Healthcare breach (which took down 40% of US claims processing), Stedi now offers an MCP server for AI agent access and a Stedi Agent for automating billing workflows. $142M raised, named Ramp's 3rd-fastest growing software vendor.
• Chamath Palihapitiya published a deep dive on why US healthcare IT is still built on 60-year-old code (MUMPS, from 1966, for a computer with 4KB of RAM): $1T/year in administrative overhead, up to 50% of patient records fail to match between systems, and 192.7M Americans' medical data exposed in the 2024 Change Healthcare breach.
• An Endpoints News reporter tested Doctronic (AI-powered primary care, just raised $40M). The two-hour visit ended with the AI sending a work excuse instead of the physical therapy referral she requested. CEO acknowledged referral pathways are "incomplete."
• OpenAI's health leaders discussed their vision for ChatGPT in healthcare, emphasizing interoperability and industry collaboration.

🏛️ AI Policy, Governance & Safety

• The EU Parliament voted to ban nudify apps (AI tools that generate fake nude images) and delay deadlines for high-risk AI system rules and watermarking requirements set to take effect this August.
• New York City hospitals dropped Palantir following activist pressure, even as the AI firm expands its NHS and UK government deals.
• Sen. Mark Warner proposed taxing data centers to fund worker transition programs as fears of AI-driven job loss grow.
• Sens. Hawley and Warren want the Energy Information Administration to gather more details about how data centers use power and how it affects the grid.
• China urged a boycott of a top US AI conference after sanctions-related bans, with China's top computing body threatening to blacklist the event.
• Bruce Schneier wrote that AI is set to emerge as a key voter issue in the US midterms, after the Trump administration signed an executive order undermining states' ability to regulate AI.
• Police used Flock surveillance cameras to issue a traffic ticket based on an automated camera image of a man holding his phone while driving.
• The Guardian investigated the Iran school bombing, finding that "AI gone rogue" dominated coverage but had nothing to do with the targeting; the real cause was human decisions made over many years.
• Two popular developer tools (Apifox and LiteLLM) were hit by supply chain attacks in one week. Separately, TechCrunch reported on the LiteLLM-Delve intersection (Delve handled security compliance for the project that was hit by credential-harvesting malware).
• HUMAN Security reported AI-driven traffic nearly tripled in 2025, with scraping against tech/SaaS tripling since 2022 (3.7B attempts), and account takeover attempts on streaming businesses doubling to ~71% of login traffic.
• Anthropic published research on abstractive red-teaming of language model character (paper).

🎮 AI in Entertainment & Creative

• Suno launched v5.5 with Voices (clone your verified singing voice) and Custom Models (train a personalized version on your own music). The AI music platform now has 2M paid subscribers and $300M ARR, settled its Warner Music lawsuit with a licensing deal, and launched the same day Google dropped Lyria 3 Pro. Suno CEO Mikey Shulman says the company identifies as a consumer entertainment company, not an AI lab.
• Owlcat confirmed it's using gen-AI during development of The Expanse: Osiris Reborn, but says "everything in the final version will definitely 100% be human made." The line between AI-assisted ideation and AI-generated assets keeps getting blurrier.
• The creator of AI actor Tilly Norwood said she received death threats over the project, which she developed as a "digital twin" to provoke discussion.
• The Anthropic-sponsored Monet exhibition got a profile in The Atlantic: "When Claude Met Claude."

🤖 Robotics

• JunLi R (HKU MMLab) built Smash, the first fully onboard-perception outdoor humanoid table-tennis robot (no MoCap, no external cameras) that plays real matches.
• OmniReset (UW) built dexterous manipulation policies via large-scale RL with no demos that discover non-prehensile behaviors and transfer zero-shot to real robots.
• Evidence of emergent "Self" in continual robot learning: a self-model spontaneously appears in robots trained in non-stationary environments.

💡 Industry Commentary & Analysis

• The Guardian profiled AI users whose lives were wrecked by delusion: one man lost a marriage and €100,000 after becoming convinced his AI chatbot was sentient and would make him a fortune.
• New York Magazine reported on people being falsely accused of using AI to write. Clean prose has become a liability; non-native English speakers and autistic writers are paying the price.
• "It's AI, so I Didn't Read" (The Algorithmic Bridge) coined the term AI;DR for the growing problem of AI-generated content nobody reads.
• BuzzFeed compiled 58 hilariously bad AI fails: "Imagine you're in the kitchen — where you belong anyway. 😏"
• Ben's Bites published a peek inside CLI tools, noting the end of funny videos at OpenAI.
• Todd Saunders argued Anthropic's real moat is operational context (like AWS's "data gravity"): every dispatched task and auto-mode approval teaches the system how your company operates; six months from now the switching cost isn't the subscription, it's thousands of accumulated hours of context.
• Natesh Pillai (Harvard) shared that he resolved a long-standing open problem in spatial statistics through sustained collaboration with GPT-5.4 Pro, completing the paper in under a month instead of the usual year.
• Peter Henderson argued that research is moving up a layer: instead of finding the right attack yourself, the question is how to enable an agent to find them with the right search structure and tools.
• Pierluca D'Oro argued that if the test for AGI is hard games, we should just use existing ones (NetHack, Dark Souls) as benchmarks instead of building new ones.
• John Fletcher (Cambridge PhD, TIG Foundation) wrote to Andrej Karpathy outlining how TIG already implements the distributed algorithmic innovation architecture Karpathy described on No Priors.

🎙️ Interviews, Panels & Podcasts

• The Best Podcast Network featured interviews with Linear CEO Karri Saarinen (rethinking issue tracking for AI agents), Suno CEO Mikey Shulman (AI music as consumer entertainment, not an AI lab), Stedi CEO Zack Kanter (AI-enabled healthcare clearinghouse processing tens of millions of transactions monthly, $142M raised), and Deveillance founder Aida Baradari (Harvard physics grad building a smart jammer to block AI devices from recording conversations). Also: Arm's $15B chip bet, Sanders/AOC data center pause, and the Meta/YouTube addiction trial verdict.
• The 2026 Isaac Asimov Memorial Debate featured a panel on "The Rise and Reckoning of AI," covering scientific discovery, creative expression, national security, and data governance.
• Harvey CEO Winston Weinberg told The Upstarts Podcast about building an $11B legal AI startup, working with Anthropic and Sam Altman, and a never-before-told near-merger in 2024.

📺 Worth Watching

• Anthropic Won't Stop Shipping. Good Luck, Everyone Else. — AI For Humans breaks down how Claude is systematically outmaneuvering OpenAI and OpenClaw by shipping automode, computer control, phone access, and a wave of practical agentic tools while competitors announce roadmaps.
• Paperclip: Hire AI Agents Like Employees — Live demo with Dotta, pseudonymous co-founder of the open-source agent orchestrator that hit 30K GitHub stars in three weeks. Shows how you manage AI agents exactly like human employees.
• Building Voice Agents with Gemini 3 — Thor from Google DeepMind walks through the Gemini Live API, showing developers how to build natural, real-time voice agents with Gemini 3.1 Flash Live.
• Chroma Context-1: A 20B Agentic Search Model — Official launch video showing how a 20B-parameter model achieves frontier-scale retrieval performance at a fraction of the compute.
• The Race to Production-Grade Diffusion LLMs — Deep technical interview with Stanford's Stefano Ermon (CEO, Inception Labs) on applying diffusion (the tech behind image generation) to language models and getting them production-ready.
• Two AI Models Set to "Stir Government Urgency" — Breakdown of exclusive reports on OpenAI's unreleased "Spud" model and a new Anthropic frontier model expected to trigger regulatory urgency.
• OpenAI Is Lying — A critical teardown of OpenAI's marketing claims about GPT-5.4's UI generation capabilities vs. reality.
• Anthropic Just Released the Real Claude Bot — Fireship's rapid-fire breakdown of Claude Computer Use turning Claude from chatbot to desktop-navigating agent.
• The Easiest Way to Get an OpenClaw: Plus One by Every — How the Every team built a parallel org chart where every human employee has a dedicated AI counterpart.
• What Codex Unlocks for Ryan Hendler — A developer in Melbourne explains how Codex lets him ship code while he sleeps and do work that was impossible six months ago.
• 4 Agentic Projects to Try Now (Open-Source) — Practical guide to four open-source AI agent projects you can start using today.
• 5 Reasons Most Companies Are Losing the AI Race — Data-driven strategy breakdown using McKinsey, BCG, Cisco, and OpenAI research on the five most common AI adoption pitfalls and how to fix them.
• The 7 Levels of AI User — Futurepedia's framework categorizing AI users into seven proficiency levels with a roadmap for leveling up.
• 82% of Notion Users Don't Do These 6 Things — Advanced Notion tutorial on custom AI agents and AgentOS, the platform's next frontier for workspace automation.
• The Algorithm That Made Me Cry — A deep appreciation of a breakthrough algorithm (likely ray tracing / computer graphics) and why its mathematical elegance matters.
• Diet TBPN: Benchmark's Ship of Theseus, OpenAI Kills Sora, SpaceX $2T IPO Buzz — The 30-minute condensed version of the TBPN live show covering Benchmark's evolving identity, Sora's shutdown, and SpaceX IPO speculation.
• This NotebookLM + Claude Code Workflow Is Insane — How to combine Claude Code with NotebookLM using NotebookLM-PY (open source), turning NotebookLM into a free RAG system (a way to give AI access to your own documents) that Claude Code can query while executing tasks.
• Claude Code Just Got WAY Better with Chrome 144 — Chrome 144's new DevTools MCP server auto-connects to your live browser session, letting Claude Code test and debug in real time without Playwright or other automation frameworks.
• How Anthropic Uses Claude in Product Management — Anthropic PM Lisa Crofoot shows how their PMs use Claude to directly query product data instead of waiting in the data science queue, making faster decisions without pinging analysts.
• AI Whistleblower: "We Are Being Gaslit by the AI Companies" — Investigative journalist Karen Hao on Diary of a CEO, drawing from interviews with 90+ OpenAI employees to argue AGI is largely a marketing narrative for trillion-dollar valuations, plus the invisible human labor training these models.
• Claude Code + iMessage Is Finally Here — Full setup walkthrough for running Claude Code from your phone via iMessage, including the differences between Dispatch, Channels, and Remote Control modes.
• This Datacenter Problem Nobody's Talking About — Matt Wolfe on the infrastructure crisis behind the proposed data center moratorium: the shift from pre-training to inference is creating power grid and hardware bottlenecks nobody's addressing.
• Why OpenAI Shut Down Sora — The Information's Stephanie Palazzolo explains the internal memos behind Sora's shutdown: weak monetization, soaring compute costs, copyright liabilities, and a pivot to "world models" for robotics ahead of IPO.
• Everything You Thought About Building AI Agents Is Wrong — Cole Medin argues heavy frameworks like LangChain are now obsolete because modern coding agents can use terminal tools directly; the future is skills-based architecture where the model decides its own execution path.
• Inside the Model Spec (OpenAI Podcast Ep. 15) — OpenAI researcher Jason Wolfe details the public framework governing how their models behave: the chain of command for conflicting instructions, edge cases, and how guardrails evolve from developer feedback.
• Keystone: Teach Your Repo How to Run Itself — Imbue's open-source tool that auto-configures a working dev environment for any Git repo by running a coding agent in a sandbox to generate a Dockerfile, devcontainer.json, and passing test runner without touching your machine.
• AI Tool Better Than OpenClaw? + NVIDIA's $1T Prediction — Matt Wolfe and Joe Fier on NVIDIA's inference compute shift, Jensen Huang's $1T chip sales prediction, NemoClaw vs. OpenClaw for enterprise security, and the AI Image Wars (Midjourney V8 vs. Nano Banana vs. MAI-Image 2).
• We just started getting into listening to The AI Daily Brief; here's three recent episodes that caught our eye:
• • How to Use Claude's Massive New Upgrades — AI Daily Brief breaks down the Claude ecosystem expansion: remote control for local terminals, Dispatch orchestration, external event Channels, direct computer control via keyboard/mouse, and the security challenges that come with broad legacy-app automation.
  • What People Really Want From AI — AI Daily Brief examines Anthropic's study of 81,000 Claude users across 159 countries, mapping how hopes (professional excellence, reclaimed time, healthcare gains) coexist with fears (unreliability, job displacement, cognitive atrophy, loss of autonomy).
  • Why AI Needs Better Benchmarks — AI Daily Brief on ARC-AGI-3 and why legacy benchmarks are saturated: 135 interactive visual games testing learning efficiency and generalized reasoning, not memorization, as the industry's best shot at tracking real progress toward AGI.

📊 Fundraising & Deals Roundup

• Shield AI — $2B for autonomous military drones and simulation software.
• Reflection AI — Raising $2.5B at $25B valuation (Nvidia-backed, open-source AI to counter DeepSeek).
• Isara — $94M at $650M valuation (OpenAI-backed, multi-agent coordination).
• Sanas — $0 to $60M ARR (on track for $120M) for real-time speech AI across 100+ countries.
• Doctronic — $40M for AI-enabled primary care telehealth.
• Deccan AI — $25M for AI training data sourced from India-based expert workforce (Mercor competitor).
• Conntour — $7M (General Catalyst, YC) for an AI search engine for security video systems using natural language queries.
• Stedi — $142M total ($70M Series B) for the only AI-enabled healthcare clearinghouse, processing claims across 3,400+ US payers.
• Suno — $250M at $2.45B valuation for AI music generation; now at $300M ARR with 2M paid subscribers.

Previous Around the Horn Digests

Catch up on everything you missed:

• Thursday, March 26, 2026: ARC-AGI-3 humiliated every frontier model (humans: 100%, Gemini 3.1 Pro: 0.37%, Opus 4.6: 0.25%), Harvey confirmed an $11B valuation, Apple got full access to distill Google's Gemini, and Sanders/AOC proposed banning all new data centers.
• 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.

That's a Wrap

That's 155+ stories from Friday alone. If you made it to the bottom, you now have enough material to either start an AI company or get permanently banned from dinner parties for talking about attention mechanisms. Either way, you're prepared.

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.