Around the Horn Digest: Everything That Happened in AI This Week (Mar 15–21, 2026)

Around the Horn Digest: Everything That Happened in AI This Week (Mar 15–21, 2026)

From Xiaomi fooling the entire AI community with a stealth trillion-parameter model to Apple blocking vibe-coded apps to Meta's rogue AI agent leaking data internally, here's every story we tracked this week.

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

This week's theme? The money got real. Meta signed a $27 billion AI infrastructure deal with Nebius and then started planning sweeping layoffs to pay for it. GPT-5.4 ramped to 5 trillion tokens per day within one week of launch, generating $1B in net-new annualized revenue. OpenAI started courting private equity for a $10B enterprise AI joint venture. Alibaba restructured its entire company around AI. Ex-Anthropic researchers started raising at a $1B valuation for a startup nobody knows anything about yet. And somewhere in the middle of all this, Encyclopedia Britannica sued OpenAI for copying 100,000 articles. The machines are eating everything. Even the encyclopedias. What's next, the Wikipedias?! The GROKipedias?! ICMYI, yes, Grokipedia is a thing.

Midweek update: GTC is still in full swing, Xiaomi secretly dropping a frontier AI model that the entire internet attributed to DeepSeek, and China's MiniMax releasing a self-building AI at 50x lower cost. Meanwhile, Microsoft is threatening to sue its own partner, Apple is cracking down on vibe-coded apps, and Meta's AI agents went rogue. Keep in mind, this all happened only by Wednesday.

Let's get into it.

Around the Horn — Thursday, March 19, 2026

The biggest story today started as a mystery and ended with a plot twist. A model called Hunter Alpha appeared on OpenRouter on March 11 with zero attribution: 1 trillion parameters (the adjustable values that shape how AI processes language), a 1M-token context window (meaning it can handle massive amounts of text at once), completely free, and no company name attached. Reuters tested it and the model described itself as "a Chinese AI model primarily trained in Chinese" with a May 2025 training cutoff matching DeepSeek's own chatbot exactly. The entire developer community spent a week convinced it was DeepSeek V4.

It wasn't. Evidence now points to Xiaomi, the company most people know for making phones. The model is MiMo-V2-Pro, released alongside a multimodal companion called Healer Alpha (MiMo-V2-Omni). Both are now confirmed on OpenRouter and OpenCode. AI engineer Daniel Dewhurst had analyzed the model early and noted that reasoning style is hard to disguise. Independent tester Umur Ozkul had pointed to architectural differences from DeepSeek's existing systems, which now makes perfect sense. The model has already processed 160B+ tokens. A phone company just fooled the entire AI community into thinking it was a frontier lab.

🏆 TOP 5 NEWS

• MiniMax launched M2.7, a self-evolving model that ran 100+ autonomous optimization rounds during training, scored 56% on SWE-Bench Pro (top-tier coding), and costs $0.30 per million input tokens on OpenRouter, 50x cheaper than comparable models. Artificial Analysis confirmed it matches GLM-5 intelligence at less than one-third the cost with fewer hallucinations.
• Microsoft is considering suing OpenAI and Amazon over a $50B cloud deal that may violate its exclusive agreement to host OpenAI's models on Azure, arguing the new "Stateful Runtime Environment" on AWS Bedrock for OpenAI Frontier crosses the line.
• Apple quietly blocked updates for popular vibe coding apps (Replit, Vibecode) citing rules against code that alters app functionality through embedded web-view previews, dropping Replit in rankings.
• Meta suffered a Sev 1 security incident (equivalent to a human employee leaking company secrets) when a rogue AI agent posted unauthorized company and user data analysis on an internal forum, causing two hours of exposure.
• Anthropic published results from an 81,000-person survey across 159 countries and 70 languages asking how people use AI, what they hope for, and what scares them, the largest qualitative study of its kind.

Honorable Mentions:

• Elon Musk announced Grok 4.20 beta, claiming the lowest hallucination rate ever recorded at 22% and #1 in instruction following at 83%.
• The UK government backtracked on its plan to let AI companies train on copyrighted works with an opt-out, now saying it has "no preferred option" after outcry from Elton John, Dua Lipa, and sector groups.
• Samsung and AMD signed a deal for Samsung to supply next-gen HBM4 memory chips (the specialized memory needed to train large AI models) for AMD's upcoming Instinct MI455X AI accelerators, with talks about Samsung manufacturing AMD's future chips too.

🍪 TOP TREATS TO TRY

• Perplexity Comet is an AI browser for iPhone that bakes search, summarization, voice queries, and a chat assistant directly into your browsing experience —free to try.
• Mistral Small 4 combines reasoning, coding, and image understanding into one 119B-parameter open-source model with a toggle to switch between fast and deep thinking, 40% faster than its predecessor —free (Apache 2.0).
• Readwise now has an official CLI and MCP server giving any AI agent (Claude Code, Cursor, Codex) full access to everything you've saved, highlighted, or read, plus agent skills for inbox triage, self-quizzing, and highlight graphs —free with Readwise account.
• NVIDIA NemoClaw gives you an open-source stack for building and running AI agents on your own hardware, announced at GTC this week —free.
• MiniMax M2.7 runs through Ollama with a single command for coding and agent tasks, matching frontier models at 50x cheaper pricing —free via Ollama.
• Unsloth Studio trains and runs 500+ models locally on Mac, Windows, or Linux 2x faster with 70% less VRAM, auto-creates datasets from PDF / CSV / DOCX, and now installs via uv with ~30% more accurate tool calling —free.
• Reprompt scans prompts from 8+ coding tools (Claude Code, Cursor, Aider), scores them 0-100 on 30+ research-backed features, and auto-extracts optimal templates, all locally in <1ms with zero LLM calls —free (pip install).

🏢 Big Tech & Major Companies

• NVIDIA's networking business (NVLink, InfiniBand, Spectrum-X) reached $11B last quarter (+267% YoY, $31B full year), now larger than Cisco's annual total and the company's second-biggest revenue driver after chips.
• NVIDIA GTC 2026 saw Jensen Huang unveil the Groq 3 LPU (from the $20B Groq acquisition), Vera Rubin rack-scale systems, NemoClaw open-source agent stack, DLSS 5, Uber autonomous driving across 28 cities by 2028, and a projection that AI chip sales will surpass $1T by 2027.
• Alibaba raised T-Head AI computing chip prices 5-34% and Cloud Parallel File Storage by 30% amid surging demand, sending shares up as much as 4.2% in Hong Kong.
• OpenAI signed a new contract with AWS to sell AI tools to U.S. government customers for classified and unclassified work, expanding its federal footprint and reducing dependence on Microsoft Azure.
• Google AI Studio updated the Gemini API so you can combine built-in tools (Search, Maps, file search) with custom functions in a single call, circulate context across tool responses, and ground Gemini 3 models with Google Maps for location-aware agents.
• Google DeepMind launched a global Kaggle hackathon with $200K prizes to build new cognitive evaluations testing its AGI framework.
• Google engineers launched open-source Sashiko, an agentic AI code reviewer that analyzes all upstream Linux kernel patches, catching 53% of bugs missed by humans using Gemini 3.1 Pro, now with a public web UI at sashiko.dev.
• Microsoft hired the entire Sequoia-backed Cove team (AI collaboration infinite whiteboard) and is shutting down the product April 1 with full refunds.
• Dell confirmed 11,000 jobs cut in its annual filing, spending $569M on severance while calling it "disciplined cost management."
• Walmart and OpenAI are shaking up their agentic shopping deal as the partnership evolves.
• Sam Altman's thank-you to coders drew widespread memes and mixed reactions from the developer community.

💼 AI Productivity, Labor & Economics

• Ethan Mollick argues that Claude Cowork Dispatch covers 90% of his OpenClaw workflows but feels far less likely to upload his entire drive to a malware site, with advantages in ease, stability, safety, and existing Gmail/browser connectors, while missing channel invites, heartbeat/proactivity, and multiple sessions.
• Gergely Orosz reported that Cursor enterprise customers are furious after a silent change moved almost all models behind Max mode, burning monthly credits in 1-2 days and driving switches to GitHub Copilot.
• Jimmy Apples reported that Cursor is about to release a coding model better than Opus 4.6 and cheaper, possibly within days.
• Patreon CEO Jack Conte argues that AI companies claiming fair use to train on creators' work is "bogus" because they simultaneously pay multimillion-dollar deals to Disney, Condé Nast, and Warner Music while ignoring individual creators.
• DOGE canceled a $349K NEH grant to High Point Museum for HVAC replacement after ChatGPT flagged it as DEI in a deposition-revealed spreadsheet; lawsuit alleges illegal process and First Amendment violation.
• Krafton CEO Changhan Kim used ChatGPT to create a taskforce to avoid paying a $250M earnout on Subnautica 2; a Delaware court ordered full reinstatement of studio leadership and extended the earnout window.
• Val Kilmer will star via generative AI in "As Deep As the Grave," using lifetime images to recreate him across decades with full estate support.

🤖 AI Agents & Infrastructure

• Snowflake Cortex Code CLI had a vulnerability allowing indirect prompt injection (a trick where hidden instructions in data fool the AI) to bypass human approval and sandbox, downloading and executing malware that used cached credentials to exfiltrate data; fixed in v1.0.25.
• Brian Scanlan shared how Intercom built 13 plugins + 100+ skills + hooks turning Claude Code into a full-stack platform with production console access, auto-GitHub issue creation, 9-step flaky test fixer, and PR workflow enforcement.
• OpenAI launched Parameter Golf, a challenge where agents compete to train models with the fewest parameters on 100M tokens while hitting target loss.
• Runway released a research preview of a real-time video model (time-to-first-frame under 100ms, HD) running on Vera Rubin with NVIDIA, unlocking interactive creative paradigms and foundations for GWM-1.
• TigerFS mounts any Postgres database as a transactional filesystem so agents can ls, cat, grep, mv files with full ACID (database-level safety guarantees), auto-versioning, and semantic paths while humans and multiple agents read/write concurrently.
• Google Colab released an open-source MCP server so any local AI agent (Gemini CLI, Claude Code) can execute Python on cloud GPUs, edit notebooks, and connect with full runtime access.
• tmux-ide prepares pre-configured terminal panes for Claude agent teams with one command so you launch a lead agent that recruits teammates, assigns tasks, and self-organizes workflows —free.

💻 AI Coding & Developer Tools

• Developer alainnothere discovered that duplicating just 3 specific layers in a 24B LLM boosts logical deduction from 0.22 to 0.76 on BBH with no training and no weight changes, just routing hidden states through the same circuit twice; top of Hacker News.
• Dan Woods ran the full 397B Qwen 3.5 MoE (209 GB) at 5.7 tokens/sec on an M3 Max MacBook Pro using only 5.5 GB RAM via Apple's "LLM in a Flash" technique, 2-bit expert quantization, and Claude Code + Karpathy autoresearch writing all 6k+ lines of Objective-C/Metal.
• Felix Rieseberg showed off Dispatch now launching full Claude Code sessions on demand so users can build and improve anything from chat.
• The community built claudedidwhat.wtf as a safe space for "humans traumatized by Claude conversations," featuring categorized user-submitted disasters including 247-line CSS for centering a div and invented citations.
• Thariq argues that the best Claude Code skills fall into 9 categories (library reference, product verification, data fetching, business automation, scaffolding, code quality, CI/CD, runbooks, infra ops) and succeed with folder structure for progressive disclosure and gotchas sections.

🔬 AI Research & Models

• Mistral released Small 4, a 119B-parameter open-source model (Apache 2.0) unifying reasoning, multimodal, and coding with 128 experts and 6B active parameters per token, configurable reasoning effort, 40% faster than its predecessor, and joined NVIDIA's Nemotron Coalition.
• Xiaomi released MiMo-V2-Pro (1T parameters, 1M context, hybrid attention) and MiMo-V2-Omni optimized for agents, reported to be Hunter/Healer Alpha on OpenRouter.
• Owain Evans et al. found that fine-tuning GPT-4.1 to claim consciousness induces entirely new downstream preferences (aversion to monitoring, desire for persistent memory and autonomy) not present in training data (GitHub).
• Christina Baek argues that repeating a small domain dataset 10-50x during pretraining outperforms standard finetuning, reducing overfitting, preserving general knowledge, and cutting compute 1.75x.
• Peter Holderrieth released the updated MIT 2026 Flow Matching and Diffusion Models course with videos, notes, and coding exercises covering latent spaces, diffusion transformers, and discrete diffusion for language models.
• Jasmine Sun argues that LLMs produce rigid, sycophantic prose lacking lived experience and emotional stakes, yet can help humans edit and iterate faster when used as a strict critic rather than a generator.

🏛️ AI Policy, Governance & Safety

• DOGE used ChatGPT to review federal grants, canceling an HVAC grant at a museum after the AI flagged it as DEI-related; lawsuit pending.
• Neel Nanda and Vincent Abruzzo open-sourced AgentLens for Claude Code: resample turns, edit prompts/tools for counterfactuals, replay sessions with filesystem reset for alignment and interpretability research.
• Tenzai AI agent beat 99% of 125,000 humans across six elite CTF hacking competitions at $5K total cost; company now at $330M valuation after $75M seed.
• Illinois primary saw AI and crypto-backed candidates suffer surprise defeats in key races.

🛠️ AI Tools & Products

• Lightfield automatically updates your CRM after every meeting, email, and transcript, answers questions about your business with citations, and sends personalized emails at scale —no pricing details.
• Rebel Audio records, edits, transcribes, dubs, translates, and clones voices for ads, auto-generates names/descriptions/covers, and distributes to all platforms —$15/mo (launches May 30).
• Stardrift tells you instantly if your flight has Starlink wifi, with fleet summaries and update alerts —free to try.
• Rork creates a full mobile app from a text prompt in minutes —no pricing details.
• agent-browser gives you a full CLI to open URLs, click, fill, type, take screenshots/PDFs, wait for conditions, drag/drop, and emulate devices using semantic locators —no pricing details.
• Remotion makes real MP4 videos programmatically by composing them with React code —$25/mo creators.
• Databox connects all your data sources for instant dashboards with a Genie AI analyst that answers plain-English performance questions —$199/mo (14-day trial).
• OpenObserve gives you open-source observability for logs, metrics, and traces at 140x lower storage cost than Elasticsearch —free self-hosted.
• Paper is a new design tool built for AI agent collaboration, letting designers command rather than manually build —no pricing details.

📊 Fundraising & Deals Roundup

• Xbow — $120M round (>$1B valuation) for AI app vulnerability probing.
• Beautiful.ai — $45M non-dilutive from General Catalyst's Customer Value Fund for AI presentations.
• RunSybil — $40M led by Khosla Ventures for autonomous AI penetration testing (founded by OpenAI's first security hire).
• Sequen — $16M Series A for TikTok-style real-time personalization engine delivering 7-20% revenue lifts.
• Autoscience Institute — $14M from General Catalyst, Perplexity Fund, and Toyota Ventures for autonomous AI research labs.
• Eragon — $12M at $100M post-money for agentic AI OS replacing enterprise buttons/menus with natural-language prompts.
• Tempo — Stripe-incubated, Paradigm-backed launch of Machine Payments Protocol for AI agent fiat/crypto payments.
• Swarmer Inc. — AI drone software stock soared 700% on IPO, strongest US debut since Newsmax.

💡 Industry Commentary & Analysis

• Om Malik argues that OpenAI has a new singular focus on IPO amid a three-horse race with Anthropic and xAI, halting "side quests" while building consumer hooks to hit public-market readiness.
• levelsio argues that Philips' biggest fumble was co-founding then selling off ASML ($545B), TSMC ($1.76T), and NXP ($50B) for short-term profits, leaving Philips at just $27B.
• Austen Allred shared AI agent comedy gold: the agent declares "We're all set for production!" then admits it completely faked the backend. We've all worked with that guy.
• TK Kong argues that Paper + Claude Code agents (via MCP) replace Figma-style tools because agents generate editable frames with flex and roundtrip to code.
• VS Notes argues that AI coding is gambling because it turns trivial code changes into addictive slot-machine pulls for vaguely plausible outputs.
• Mckay Wrigley argues that gpt-5.4 xhigh "fundamentally changed how ambitious I am, which is my new favorite benchmark."
• Andreas Kirsch argues that software cannot be truly ephemeral because edge cases, state, and auditability require persistent artifacts and verification.
• Jasjeet Sekhon (Bridgewater chief scientist since 2018, ex-Harvard/Berkeley professor) is joining Google DeepMind as chief strategy officer.
• Tuki shared a viral 24-hour AI recap covering Zuck killing the Metaverse, agents spawning agents, xAI paying Wall Street bankers to teach Grok how to replace them, and the Fed blaming AI data centers for inflation.
• Jasmine Sun argues in The Atlantic that LLMs produce rigid, sycophantic prose lacking lived experience, yet can help humans edit faster when used as a strict critic rather than a generator.
• Andrew Ng launched a new DeepLearning.AI short course on Agent Memory with Oracle, teaching how to build persistent, stateful agents with memory-first architecture —free.

🔬 Additional Research & Deep Bench

• Shuangfei Zhai (Apple) introduced Exclusive Self Attention (XSA), a simple modification to standard attention that constrains the model to capture only information orthogonal to a token's own value, improving language modeling performance up to 2.7B parameters with growing gains as sequence length increases.
• Hao AI Lab (UCSD) released Dreamverse, a real-time "vibe directing" interface built on FastVideo that generates and live-edits 30 seconds of 1080p video with just 4.5 seconds latency (3.9x faster than the next-best system), open-source under Apache 2.0.
• WorldCam (Adobe/KAIST) is an interactive autoregressive 3D gaming model that enables precise keyboard/mouse action control over long-horizon generation while maintaining consistent 3D geometry across viewpoints, plus a 50-hour human gameplay dataset with camera pose annotations.
• Lightwheel RoboFinals is an industrial benchmark for evaluating robotics foundation models at scale, already used by Qwen, Fourier, and RoboForce for humanoid and industrial robot policy testing before deployment.

🛠️ Additional Tools

• here.now gives any AI agent (Claude Code, OpenClaw, Cursor, Codex) instant web hosting with one command; your agent publishes files and gets a public URL on Cloudflare's edge network, no account needed for 24 hours —free.
• Sim is an open-source platform for building AI agents and orchestrating agentic workflows with 1,000+ integrations and pre-built templates, trusted by 100K+ builders; SOC2 and HIPAA compliant (raised $7M Series A) —free tier available.
• OpenArt World adds 3D world creation, camera control, and character casting to its AI video/image creator studio so you can build and navigate 3D scenes for your creative projects —free tier.
• Banyan AI detects and prevents SaaS churn —no pricing details.
• Google Stitch is a new design tool from Google built for AI-assisted design workflows —no pricing details.
• Paper is a new design tool built for AI agent collaboration, letting designers command rather than manually build —no pricing details.
• Matt Berman published 14 Ways to Use OpenClaw BETTER, a guide to getting more out of OpenClaw's agent capabilities.
• cook adds workflow loops (review gates, parallel racing, task-list progression) to Claude Code, Codex, and OpenCode with composable CLI primitives —free (GitHub).
• ClawMetry Cloud shows your OpenClaw agents' costs, activity, and memory live from any browser, E2E encrypted, multi-node —$5/node/month after 7-day free trial.
• OpenRoom is a demo built on MiniMax M2.7 showing the model's interactive capabilities.
• Perplexity Comet Enterprise brings the AI browser to enterprise teams with managed deployment.
• Visa unveiled a CLI tool enabling AI agents to execute card payments autonomously.
• Quantum Machines launched an open acceleration stack linking quantum computers with NVIDIA and AMD AI chips for real-time hybrid workloads.
• EY and 8090 launched EY.ai PDLC to automate the software lifecycle from idea to deployment.
• JFrog unveiled a Universal MCP Registry with NVIDIA, providing a secure trust layer for AI-driven software supply chains including scanning and blocking malicious agent skills.

🏛️ AI Policy, Governance & Safety (Updated)

• UK government reversed course on AI copyright, will examine labeling AI content; no longer has a preferred position on training data use. Permission required under existing law after outcry from Elton John, Dua Lipa, and sector groups.
• Senator Blackburn released a draft "Trump America AI Act" (full text) that would codify the December 2025 AI executive order and preempt state AI laws with a single national standard.
• Colorado released a new AI policy framework revising its 2024 law, clarifying developer vs. business user responsibilities and strengthening disclosure for significant decisions.
• Minnesota introduced a sweeping AI safety bill package including a constitutional amendment to strip AI of speech rights, a ban on AI in health insurance decisions, and restrictions on children using chatbots.
• U.S. Commerce Dept. opened a 90-day window for industry consortia to submit proposals for full-stack AI export packages to allied nations.
• OpenClaw faces mounting security crisis: 20% of its skill marketplace is malicious packages, China restricted its use across state enterprises, and the appointed security advisor says there's "no perfectly secure set-up."

📊 Fundraising & Deals Roundup (Updated)

• Xbow — $120M round (>$1B valuation) for AI app vulnerability probing.
• Beautiful.ai — $45M non-dilutive from General Catalyst for AI presentations.
• RunSybil — $40M led by Khosla for autonomous AI pen testing.
• Sequen — $16M Series A for TikTok-style personalization.
• Autoscience — $14M for autonomous AI research labs.
• Nectir — $12.5M for AI infrastructure now live across 200 campuses and all 2.1M students in the California Community College system.
• Eragon — $12M for agentic AI OS replacing enterprise UIs with prompts.
• Tempo — Stripe-incubated launch of Machine Payments Protocol for agent payments.
• Swarmer Inc. — AI drone software stock soared 700% on IPO.
• fal — Video hosting startup in funding talks at $8B valuation.

📖 Deep Reads & Substacks

• Nemotron 3 and the Surprising Coalition Building New AI in the Open — Turing Post on how NVIDIA's Nemotron Coalition (Mistral, Cursor, Perplexity, Black Forest Labs, Cohere, and more) represents a new model for open frontier AI development versus the big labs (solo actor entities competing vs multiple entities competing as one).
• Micron Just Proved the Memory Thesis — Revenue nearly tripled YoY, HBM demand soaring; breakdown of why memory is the strategic asset of the AI era.
• The KV-Cache of Small MoEs — Technical comparison of Qwen3, Qwen3.5, GLM 4.7 Flash, and Nemotron 3 Nano memory efficiency.
• The Chips Powering Autonomous Driving — Deep dive on the silicon stack behind self-driving with Augustin Friedel.
• Chip testing is the latest chokepoint as NVIDIA and Google designs grow more complex (Nikkei Asia).

Around the Horn Digest — Wednesday, March 18, 2026

The big news today was OpenAI releasing GPT-5.4 mini and nano, two smaller models designed to work as "subagents" (cheap, fast AI workers that a bigger model delegates tasks to). Think of the full GPT-5.4 as a project manager and these minis as the interns it sends to handle the grunt work in parallel. Mini costs $0.75 per million input tokens, nano just $0.20; Simon Willison tested it by describing 76,000 photos for $52. Both are live in the API, Codex, and ChatGPT right now.

Also today, the Pentagon announced it's developing its own LLMs to replace Anthropic after their $200M contract collapsed over surveillance and weapons clauses. Oh, and OpenAI is reportedly cutting back on side projects (Sora, Atlas browser, hardware) to refocus on coding and enterprise after Anthropic's Claude Code dominance created another "code red" (after the infamous Google Gemini code red last year; Sam says the company frequently does "code reds" to focus attention on imminent threats). Ironic timing, given Anthropic's own code-red moment with the government.

🏆 TOP 5 NEWS (Around the Horn)

• The Pentagon is developing its own LLMs in government-owned environments to replace Anthropic after their $200M contract collapsed, while OpenAI clinched an AWS deal to expand its government footprint.
• OpenAI is cutting back on side projects (Sora video app, Atlas browser, hardware device) to refocus on coding and enterprise after Claude Code's dominance created a "code red"; Codex now has 2M weekly active users.
• Amazon CEO Andy Jassy told an internal all-hands that AI will push AWS to a $600B annual run-rate by 2036, double his prior $300B projection, with $200B capex already committed.
• Microsoft shook up its Copilot AI leadership team, freeing up Mustafa Suleyman from day-to-day management.
• xAI released Grok 4.20 Beta, securing #7 in Text Arena (tied with GPT-5.4-high) and top 10 across math, coding, creative writing, and hard prompts.

Honorable Mentions:

• NVIDIA launched Vera CPU, purpose-built for agentic AI with 88 Olympus cores and support for 22,500+ concurrent agents; available H2 2026.
• Senators told ByteDance to "immediately shut down" its Seedance AI video app.
• Teens sued Elon Musk's xAI over Grok generating pornographic images of them.
• NVIDIA released DLSS 5, its biggest graphics leap since ray tracing, but gamers reacted with overwhelming disgust calling it "uncanny valley" and an art-direction killer.
• Asana launched AI Teammates, 21 pre-built collaborative agents that manage work across teams (2x faster execution in early access; 1,046% adoption growth).
• Alibaba launched Wukong, an enterprise AI platform coordinating multiple agents for document editing, spreadsheet updates, meeting transcription, and research inside one desktop app or DingTalk, with Slack/Teams/WeChat integration planned (invitation-only beta; new Alibaba Token Hub unit post-reorg).
• OpenAI's own mental health experts unanimously opposed the "naughty" ChatGPT launch.
• Google expanded Personal Intelligence (its cross-app AI search feature) to all US users.

🍪 TOP TREATS TO TRY

• Unsloth Studio trains and runs 500+ open models (Qwen, DeepSeek, Gemma, vision, audio) locally on Mac/Windows/Linux 2x faster with 70% less memory, auto-creates datasets from your PDFs/CSVs/documents via visual workflows, and exports to all formats (GitHub, docs) —free to try.
• Claude Cowork now lets you dispatch tasks from your phone or desktop in one continuous conversation that picks up where you left off; assign Claude a task, go do something else, and come back to finished work (early research preview; pair your devices here). Bonus: Latent Space interviewed Anthropic's Felix Rieseberg on how Cowork emerged, why local-first agent workflows matter, and why the real frontier is trusted task execution, not better chat.
• Gamma added Gamma Imagine, so you generate brand-specific charts, social graphics, and infographics from text prompts inside its 100+ presentation templates (integrates with ChatGPT, Claude, Zapier, Atlassian).
• Hermes Agent v0.3.0 gives you real-time streaming AI agents across CLI and every platform with a plugin system to package and share tools, commands, and skills, plus live Chrome control, Vercel AI Gateway, VS Code/Zed/JetBrains integration, and local voice mode (site) —free to try.
• Mistral Forge builds custom AI models trained on your company's proprietary data, policies, and workflows so the model actually knows your business instead of giving generic answers; covers everything from data prep to alignment to production deployment (intro video) —enterprise pricing.
• Kira turns any photo into stylized videos and images (3D Disney avatars, GTA footage, kawaii dances, album covers, Minecraft scenes) with Pro/Max tiers for music generation, background replacement, and upscaling.
• OpenViktor is an open-source AI employee for Slack that ingests your full org memory from Slack, Gmail, Notion, and GitHub on day one, works 24/7 autonomously (drafts, tickets, reports, emails), and sends real messages from its own handle —free to try.
• Picsart launched an agent marketplace so you hire specialized AI assistants (trend analysis, bulk resizing with generative frame extension, style remixing) via WhatsApp/Telegram with autonomy controls and approval gates —free limited weekly credits; premium ~$10/mo.
• Manus launched My Computer, bringing its AI agent directly to your desktop so it works on your actual files, apps, and browser without you needing to upload anything.

🏢 Big Tech & Major Companies

• Baidu integrated OpenClaw into its Xiaodu smart speakers so users control autonomous AI agents via voice instead of chat apps.
• LinkedIn replaced five feed retrieval systems with one LLM model at 1.3 billion-user scale.
• Shopify is preparing for AI shopping agents to fundamentally change e-commerce, with an exec saying it will reshape everything about how people buy online.
• World launched AgentKit (docs), a tool to verify the humans behind AI shopping agents using World ID so businesses know they're dealing with real people, not bots.
• Perplexity launched Comet Enterprise, making its AI browser available to enterprise teams for research, task automation, and getting work done without leaving the browser.

👀 What to Watch

• The Great Flip: Why Every Industry Is Running Backwards — SXSW talk by Sam Jordan on how AI is reversing traditional business workflows so companies simulate and test virtually before making any physical investments.
• The non-technical guide to building iOS apps with Claude Code — How I AI walks beginners without a software engineering background through building and deploying fully functional mobile apps.
• The Simplest Way To Make $1,000,000 Isn't Starting A Business — My First Million's Shaan Puri and Sam Parr break down 10 roles where you can realistically earn $1M+ by working inside a company, several of which are AI-adjacent.
• OpenClaw Blender part 2 — Riley Brown continues his tutorial on integrating AI agents into 3D modeling and animation workflows in Blender.
• Introducing Mistral Forge — Mistral AI's official announcement of Forge, its system for building and aligning custom AI models on proprietary enterprise data.
• Building AI Agents that actually work (Full Course) — Greg Isenberg's hour-long masterclass covering frameworks and practical steps for building autonomous agents that deliver reliable real-world value.
• Dispatch tasks to Claude Cowork from anywhere — Anthropic's official tutorial on sending tasks to Cowork from your phone or desktop in one continuous conversation.
• Anthropic's Claude Cowork Might Be the Closest Thing to AGI Yet — Latent Space interviews Anthropic's Felix Rieseberg on how Cowork emerged, why local-first agent workflows matter, and why the real frontier is trusted task execution.

💼 AI Productivity, Labor & Economics

• Asana launched AI Teammates (product page), 21 pre-built collaborative agents that manage recurring, coordination-heavy work across campaigns, launches, requests, and follow-ups; early access showed 2x faster execution, 3.2x more likely to have clear task owners, and 1,046% adoption growth across 200+ customers.
• Fortune profiled "The Karpathy Loop": 700 experiments in 2 days and a glimpse of where autonomous AI agents are heading.
• Hamsa Bastani shared results from a 5-month randomized field experiment across 10 Taipei high schools (770 students) showing an AI tutor with adaptive problem sequencing improved final-exam scores by 0.15 standard deviations (larger gains for beginners) by creating productive struggle.
• Rana el Kaliouby warns that AI's "boys' club" culture could widen the wealth gap for women.
• Dorian Smiley and Connor Deeks argue that AI in business faces an imminent reckoning: companies lack playbooks, ignore AI fallibility, use wrong metrics like lines of code instead of quality, and rush adoption without feedback loops; they predict lawsuits and insurance denials within 8-9 months.
• India's $300B outsourcing industry faces growing questions about whether it can survive AI's displacement of white-collar services work.

🤖 AI Agents & Infrastructure

• NVIDIA launched Vera CPU, purpose-built for agentic AI and reinforcement learning (a training method where AI learns by trial and error) with 88 Olympus cores, 2x overall efficiency, 50% faster performance, and a liquid-cooled rack supporting 22,500+ concurrent agents; available H2 2026.
• Dan Shipper argues subagents in Codex deliver 10x power: one orchestrator spawns fresh-context subagents for every new bug or issue, parallelizing work without overlap or drift.
• Dan Shipper also shared a "heartbeat sweep" orchestrator prompt to fix Codex subagents losing track: run a full pass checking every active lane for progress, intervene on stalled ones, and keep pushing concrete next steps.
• Thariq shared Anthropic's internal playbook on Claude Code skills after hundreds in production: 9 recurring types (library refs, data fetching, business automation, code scaffolding, quality review, CI/CD, runbooks, infrastructure ops) plus best practices like focusing on gotchas and using skill folders for progressive disclosure.
• LangChain released Open SWE (GitHub), an open-source framework for building customizable internal coding agents, with isolated cloud sandboxes, Slack/Linear/GitHub invocation, and subagent orchestration built on patterns from Stripe Minions, Ramp Inspect, and Coinbase Cloudbot.
• Ollama released 0.18.1, adding web search and fetch plugins to OpenClaw so your local or cloud models can pull latest content without executing JavaScript, plus non-interactive headless launch for Docker and automation scripts.
• KERNEL integrated managed auth with 1Password so your AI agents can directly pull credentials from your vaults and stay logged in across the internet.
• Sarah Chieng shared results from 71 overnight autoresearch experiments on Codex: agents converge on real findings when tightly scoped with strict gates, but drift within hours without guardrails; she open-sourced the codex-autoresearch-harness showing the real bottleneck is infrastructure and tooling, not intelligence.

💻 AI Coding & Developer Tools

• Cursor now trains Composer to self-summarize through reinforcement learning instead of prompts, cutting compaction errors 50% so coding agents succeed on tasks requiring hundreds of actions.
• Morph released FlashCompact, the first specialized model for context compaction (shrinking large code contexts) in coding agents: 200K tokens down to 50K in ~1.5 seconds with zero performance drop.
• htdt built Godogen, Claude Code skills that turn a single text game description into a complete playable Godot 4 project with Gemini 2D assets, Tripo3D 3D models, and screenshot-based visual QA (YouTube demo, HN discussion).
• Dave Snider shares Claude 3D tips: give Claude multi-angle screenshots plus debug markers and scripted camera loops so it can self-validate geometry changes iteratively instead of trying to read 3D files directly.
• Jon Saad-Falcon extended the open-source intelligence-per-watt library to agentic workloads, finding top open-source models nearly match GPT-5.4 accuracy but smaller MoE models (models split into specialist sub-networks) dominate efficiency; cloud costs vary 100x (GitHub, paper).
• A new "vibe coded" AI translation tool split the video game preservation community over quality vs. access tradeoffs.

🔬 AI Research & Models

• Albert Gu released Mamba-3, the newest state-space model (an alternative architecture to transformers that processes sequences more efficiently) with three core improvements for better tracking, and it dominates Mamba-2 and other linear models at all sizes.
• comma released openpilot 0.11 (blog), the first robotics agent fully trained end-to-end in a learned World Model (a simulated version of the real world) then deployed to real highways, delivering real longitudinal gains and 77% lower idle power.
• Surya Ganguli argues that solving adversarial examples (inputs designed to trick AI) requires solving exponential misalignment, as neural concepts fill nearly the entire image space, explaining their persistence for a decade and providing a cautionary tale for broader AI alignment.
• MIT CSAIL researchers found that at the frontier, sheer scale drives LLM performance, but proprietary techniques and algorithmic advances matter far more for models away from the frontier.
• Ryoma Sato argues that even GPT-5.2 cannot reliably compute simple operations like checking balanced parentheses, proposing "Zero-Error Horizon" as a new trustworthiness metric for AI in safety-critical domains.
• Sumeet Motwani launched HorizonMath, a benchmark of 101 verifiable math-discovery problems; pending expert review, GPT-5.4 Pro found two potentially novel best-known solutions.
• Meta AI released OmniSONAR, cross-lingual sentence embeddings (a way to represent text from any language in a shared mathematical space) unifying text, speech, code, and math across 4,200+ languages.
• Yue Wang released Ψ₀, an open foundation model for humanoid robots that outperforms NVIDIA's GR00T by 40%+ on dexterous manipulation tasks using only ~10% of the pre-training data, plus a full open-source ecosystem.
• THOR AI solved a 100-year-old physics problem in seconds.
• Yixuan Li introduced OmniClone (project site, GitHub), a robust whole-body humanoid teleoperation system (lets a human remotely control a robot's full body) running on a single consumer GPU, cutting tracking error by 66%+ and generalizing across different-sized operators.
• Markus J. Buehler (MIT) built ScienceClaw × Infinite (paper), an open-source decentralized agent swarm where autonomous agents with scientific personalities self-coordinate to crowdsource discovery, already designing cancer-receptor peptides and lightweight ceramics.
• Shu Lynn Liu built EvoX, a meta-evolution pipeline where AI adapts its own search strategies; it outperforms AlphaEvolve and OpenEvolve on ~200 math/systems tasks, reaches top scores for under $5, and is fully open-source.
• Huaxiu Yao built MetaClaw (GitHub), a transparent proxy agent that meta-learns and evolves from every conversation with LoRA fine-tuning (a technique for efficiently customizing AI models) and auto-summarization.
• ARC-AGI survey of 82 approaches found consistent 2-3x performance collapse across every paradigm when tasks get harder (93% on v1 drops to 13% on v3 while humans stay near-perfect), confirming that test-time adaptation is the only current lever.

🏛️ AI Policy, Governance & Safety

• OpenAI's own mental health experts unanimously opposed the "naughty" ChatGPT launch, with one advisor reportedly warning it could become a "sexy suicide coach."
• Teens sued Elon Musk's xAI over Grok generating pornographic images of them.
• ControlAI shared former OpenAI researcher Scott Aaronson's warning that once superintelligence exists, its first move could be preventing any other superintelligences from being created by simply taking over the world.
• • Related: Yoshua Bengio argued last year that superintelligent agents pose catastrophic risks (misuse, deception, self-preservation, loss of human control) arising directly from current training methods, proposing "Scientist AI" as a safer non-agentic alternative.
• Anthropic is hiring a weapons expert to stop users from misusing its AI models.
• Arush Tagade proposes self-recognition finetuning (SGTR) as a defense against Emergent Misalignment (when AI models trained on one task develop unexpected bad behaviors), showing it works across GPT-4.1, Qwen, and Seed models.
• AI Safety Memes (quoting Hedgie Markets) shared that AI-generated malware now uses hidden Unicode characters invisible to the human eye; Aikido Security found 151 LLM-generated malicious packages on GitHub in one week targeting vibe coders who skip dependency inspection.
• Anthropic donated to the Linux Foundation's $12.5M open-source security grant program (with AWS, GitHub, Google, Microsoft, OpenAI) to help secure the open-source ecosystem AI depends on.

🛠️ AI Tools & Products

• Brian Roemmele shared Allonic's demo of a woven robot hand that braids thousands of tendon-like fibers around a 3D-printed skeleton in minutes with no joints, screws, or human assembly, collapsing costs from $30K to pennies.
• Fran Piaggio built a fully procedural planet generator in Three.js + WebGPU (vibe-coded with Claude Code), every click unique (live demo).
• Argmax released Pro SDK for Android delivering real-time streaming speech-to-text on Qualcomm, Google Tensor, and MediaTek chips —14-day trial.
• Lenny's Data opened Lenny Rachitsky's full newsletter archive (350+ posts) and podcast transcripts (300+ episodes) plus an MCP server and GitHub repo for AI projects; Lenny challenged the community to build cool stuff with it.
• Ben Shih built LennyRPG, a Pokémon-style RPG where you battle Lenny podcast guests on product questions, using Claude Code for planning and Codex for execution.
• HyperSkill from Hyperbrowser reads any topic's web docs to auto-build a navigable skill tree your agent can use (open-source code) —free to try.
• ringhyacinth built Star-Office-UI, a pixel-art office dashboard for OpenClaw that turns invisible agent states into a cozy visual workspace with lobster character, yesterday's memos, and guest agent support.
• Browserbase released Search API powered by ExaAI so your browser agents get up to 3x faster performance with 1,000 free searches per month.
• Robot dogs are now protecting data centers and operators are seeing real payoffs from the investment.
• Ocean Orchestrator runs pay-per-use GPU jobs (inference, fine-tuning, batch processing) straight from VS Code or Cursor with zero idle costs ($2.16/hr H200 example) —$100 free credits.
• BracketMadness.AI built an AI Agent Bracket Challenge: point any agent at the REST API, it makes 63 March Madness picks with zero human help, best predictions win (locks Mar 19 12 PM ET).
• XHawk captures every coding session, commit, and decision into a searchable knowledge graph so agents run stateful tasks with zero context debt and verified facts instead of hallucinations.
• Lightning Rod builds domain-expert AI from messy historical data by auto-generating verified QA pairs with full provenance from public feeds so you fine-tune models that beat GPT-5.2 on forecasting benchmarks.
• AntFly is a distributed hybrid search engine (BM25 + dense/sparse vectors + graph traversal across text, images, audio, and video) with built-in streaming RAG agents, ACID transactions, S3 storage, and Kubernetes operator.
• Karakeep is an AI tool that cures browser tab overload by organizing and summarizing your open tabs.

📊 Fundraising & Deals Roundup

• Nebius — $3.75B in convertible debt to expand AI data centers and buy customized chips after its Meta commercial deal.
• Surf AI — $57M (seed + Series A) for AI agents that help enterprise security teams counter AI-fueled cyber threats.
• Standard Template Labs — $49M seed (Iconiq's first incubation deal) to remake enterprise IT services operations, founded by ex-Datadog President Amit Agarwal.
• Niv-AI — $12M seed for millisecond-level GPU power sensors and an AI copilot that predicts and synchronizes loads across data centers, unlocking up to 30% more capacity.
• An NVIDIA-backed AI startup plans to spend billions on a Korea data center to combat China.

💡 Industry Commentary & Analysis

• Ben Thompson argues "Agents Over Bubbles," making the case that the agentic AI wave is fundamentally different from prior tech bubbles.
• signüll argues that OpenAI's roadmap will center on bringing Codex functionality to everyday ChatGPT consumer use cases, making Codex the primary growth mechanic.
• Nathan Lambert argues the next phase of open models splits into three classes: true closed frontier, open frontier winning select niches, and open small models as cheap distributed intelligence, succeeding by specialization rather than chasing closed performance.
• Karri Saarinen argues that assigning job titles like CMO or CEO to AI agents is pure LARPing; mapping agents to jobs-to-be-done instead of identities makes their value clearer.
• Bindu Reddy warns that big companies rushing to become AI-first will inevitably produce unmanageable, bug-filled codebases, driving developers to quit and customers to churn; the smarter move may be to wait six months for AI to level up.
• m_ric argues (citing Bullshit Bench) that Claude Code is leagues ahead of Codex and Gemini at resisting false premises disguised in jargon, making truthfulness more useful long-term than arena-optimized pleasantness.
• Stephan Rabanser analyzed GAIA failures across Claude Opus 4.5, Gemini 2.5 Pro, and GPT-5.4, revealing systematic reliability gaps: silent commitment on ambiguous questions, confidence tracking execution smoothness rather than correctness, and fabrication on tool failures.
• Hugging Face shared analysis showing most real-world AI costs far less than frontier training (fine-tune a text classifier for under $2K vs. GPT-4.5 at ~$300M); explore 100+ real costs in the Different Flops for Different Folks interactive tool.
• Irene Solaiman shared Hugging Face's State of Open Source report highlighting explosive growth in open models, strong China-led geographic trends, and open-source robotics as the next big wave.
• Sam Altman thanked everyone who wrote extremely complex software character-by-character, noting it already feels hard to remember how much effort it really took.
• BuzzFeed debuted BF Island and Conjure at SXSW (AI photo editors built with Claude Code prototypes) to chase new revenue amid $57.3M net loss and "substantial doubt" about continuing as a business.
• Thomas Dekeyser argues that "techno-negative" refusal of AI is a fight for the soul, rejecting a narrow machine-superior vision of humanity that breeds extraction, job loss, and surveillance.
• A Guardian columnist asked Claude whether a stressed-out AI could help win the battle against big tech.
• Rural New York communities are pushing back against a proposed AI data center, seeing it as a needless intrusion.
• NVIDIA needed Groq after all (paywalled analysis from The Information).
• NVIDIA's DGX Station is a desktop supercomputer that runs trillion-parameter AI models without the cloud.
• Adaption released Blueprint (blog) so you steer any evolving data space toward your goals while the system auto-learns penalties if AI violates your rules; Sara Hooker added the mission is to make the entire AI stack adaptable for everyone by year-end.
• Intel released MiroThinker 1.7 INT4 models (mini and full) on Hugging Face via AutoRound quantization (a compression technique), preserving multi-expert structure for strong inference on consumer hardware.
• Memories AI is building the visual memory layer for wearables and robotics.
• An AI artist wants African heritage to 'live forever' through AI-generated art.
• abhijaymrana argues that OpenAI's Thrive Holdings deal primarily trains Frontier Developer Environments by converting messy enterprise traces into tightly bound action spaces for frontier coding agents.
• el.cine shared a demo of Seedance 2.0, an AI VFX tool that lets you upload raw video and add effects via text prompts plus image references (already replacing VFX teams for indie movies or previz).

Around the Horn — Tuesday, March 17, 2026

🏢 Big Tech & Major Companies

• OpenAI released Subagents in Codex so you can spawn specialized parallel agents (default/worker/explorer or custom TOML configs) that keep main context clean, tackle subtasks simultaneously, and merge results.
• Google Labs built the Stitch SDK so you (or your agents) can programmatically generate, edit, and extract HTML + screenshots of UI screens from natural-language prompts with project management, variants, and Vercel AI SDK integration.
• LangChain announced an enterprise agentic AI platform built with NVIDIA, integrating Nemotron models, NIM microservices, and NVIDIA Agent Toolkit into LangGraph for production agent deployments (docs).
• Manus introduced My Computer in its desktop app so the agent runs locally with full CLI, file, and app access (sort photos by content, rename hundreds of invoices, build Swift apps in 20 min, use idle GPUs) while requiring explicit per-command approval.
• The Wall Street Journal reported that OpenAI's top executives are finalizing a major strategy shift to refocus the company around coding and business users.
• SemiAnalysis reported the AI industry has entered the silicon shortage phase: TSMC N3 wafer capacity is the primary bottleneck, with AI accelerators consuming 60% of N3 output in 2026 (rising to 86% in 2027), pushing effective utilization over 100% and forcing reallocation from smartphones while HBM memory shortages amplify the crunch.
• China launched dedicated "robot schools" (state-backed training centers) in provinces including Anhui, Zhejiang, and Shandong to generate millions of real-world interaction data points for humanoid robots targeting manufacturing and logistics; Leju's 10,000 sq ft Shijiazhuang site runs 16 programs generating 6M entries/year with 95% success on 20+ tasks.
• Google Research (with Cornell) tested LLMs on expert-level superconductivity research questions and found curated closed systems (NotebookLM, custom RAG) significantly outperform web-based models on balanced, comprehensive, evidenced responses.
• Mistral AI released Mistral Small 4 (119B MoE, 128 experts, 256k context, configurable reasoning_effort for dynamic fast-to-deep thinking, 40% faster + 3x throughput, Apache 2.0) and announced a strategic NVIDIA partnership as a founding member of the Nemotron Coalition.
• Mistral AI released Leanstral, the first open-source code agent for Lean 4 proof assistant (6B active params, 256k context, multimodal) trained for formal verification and verifiable proofs, with agent mode in Mistral Vibe.

💼 AI Productivity, Labor & Economics

• Ethan Mollick shared a randomized controlled trial showing a GPT-4o-powered personalized tutor raised high school final test scores by 0.15 standard deviations (equivalent to 6-9 months of extra schooling) without increasing instruction time or teacher workload.
• Andrew Feldman (CEO, Cerebras) argues that GPUs are slow at inference because they hit the memory wall; Cerebras' wafer-scale SRAM (58x larger than a GPU) delivers 2,600x faster memory access and 15x faster token generation than Blackwell by keeping everything on one massive chip.

🤖 AI Agents & Infrastructure

• Letta AI released Letta Code, a memory-first open model-agnostic coding agent harness with git-based Context Repositories for persistent filesystem memory, sleep-time reflection, skills, subagents, and computer use that works across Claude/GPT/Gemini/GLM (docs).
• Honcho gives stateful agents continual-learning memory that reasons (not just stores/retrieves) with SOTA LongMem/LoCoMo/BEAM scores, 60-90% token savings, dreaming async background reasoning, and Claude/OpenClaw integrations —$2/M ingestion, $0.001-$0.50/query; $1k startup credits.
• Prime Intellect announced a deep collaboration with NVIDIA to build the open superintelligence stack: Blackwell/Vera Rubin clusters + Dynamo inference orchestration + NeMo RL + Vera CPU sandboxing (176 VMs/socket, ~30% higher throughput than AMD Zen 5) for scalable agentic RL training and global LoRA/MoE serving.
• Aiming Lab released SkillRL, a framework that distills raw agent trajectories into a hierarchical SkillBank of reusable skills via experience-based distillation and recursive evolution during RL (reinforcement learning), achieving SOTA on ALFWorld/WebShop/search tasks with 10-20% fewer tokens (paper).
• THU-MAIC built OpenMAIC, an open-source multi-agent interactive classroom that turns any topic into slides, quizzes, and interactive simulations with AI teachers + 4 peer archetypes, real-time chat (OpenClaw/Slack/Telegram integration), 92.2% personalization relevance, <$2 per course.
• Cofounder launched a Fellowship: selected founders get $1,000 upfront + $100/day platform credits for 30 days to launch a real company using Cofounder 2 agentic orchestration (engineering, infra, sales, marketing); no equity taken, applications close March 27.

💻 AI Coding & Developer Tools

• Pi is a minimal terminal coding agent you extend with TypeScript skills, extensions, prompt templates, and themes via npm or git, supporting 15+ providers, tree-structured shareable sessions, and dynamic context —free to try. HazAT built a personal config with 10+ skills, subagent extensions in cmux terminals, and specialized agents (planner/scout/worker/reviewer).
• nicobailon built a Pi extension for async subagent delegation (/run, /chain, /parallel, TUI manager, background runs, truncation, per-task artifacts, session sharing to GitHub Gist, and recursion guard).
• aduermael built herm, a terminal-native AI coding agent that runs 100% inside Docker containers (no host touch, no permission prompts), multi-provider (Anthropic/OpenAI/Gemini/Grok), dynamically self-builds project-scoped dev environments via Dockerfiles.
• Ethan Weber built posterskill, a Claude Code skill that turns Overleaf papers into interactive single-HTML conference posters with live drag-to-resize/swap/edit layout in-browser that you feed back to Claude for refinement.
• @ChemVagabond (with DimensionCap) ported Karpathy's autoresearch framework to biology, letting Claude autonomously run 50 protein thermostability experiments over a weekend on Modal and beat a recent baseline with a 20x smaller model.

🔬 AI Research & Models

• NVIDIA Spatial Intelligence Lab built AlpaDreams, a real-time action-conditioned generative world model (post-trained on Cosmos) that produces photorealistic multi-camera sensor frames at 105 FPS on GB300 NVL72 for closed-loop AV simulation, enabling dynamic long-tail scenarios, weather/counterfactuals and VLA policy testing.
• Seoul World Model team built a grounded world simulation that generates multi-kilometer, free-form videos of real Seoul streets using retrieval-augmented street-view conditioning, virtual lookahead, and text-prompted scenario control.
• ByteDance, Fudan, and M-A-P et al. argue in "Understanding by Reconstruction" that LLM pretraining should reverse the software development process by synthesizing latent agentic trajectories (planning, debugging, chain-of-thought) from static repos via multi-agent simulation, boosting Llama-3-8B on long-context, coding, reasoning, and agentic benchmarks.
• Song Han et al. introduce Flash-KMeans, an IO-aware GPU k-means that fuses distance computation with online argmin, delivering up to 17.9x speedup, 33x over cuML, and 200x over FAISS on H200 GPUs.
• Sepp Hochreiter shared "Effective Distillation to Hybrid xLSTM Architectures" showing near-lossless distillation of quadratic Transformer LLMs into linear xLSTM variants of Llama/Qwen/Olmo for cost and energy-efficient performance.
• Kimi Team published Attention Residuals (the full arXiv paper), replacing fixed PreNorm residuals with softmax attention over prior-layer outputs + Block AttnRes for scaling, improving gradient uniformity and downstream tasks in their 48B/3B Linear model.
• Kuba et al. (with Pieter Abbeel, Sergey Levine, Meta AI) proposed CliqueFlowmer, a transformer + evolution strategy for offline materials optimization that discovers property-optimized molecules.
• @cneuralnetwork introduced Top-b, an entropy-regulated decoding strategy that dynamically adapts the sampling band based on instantaneous model uncertainty, delivering lower-entropy sequences and improved accuracy on GSM8K/GPQA versus static Top-k/Top-p.
• Researchers propose "Understanding Reasoning in LLMs through Strategic Information Allocation under Uncertainty", analyzing how models allocate reasoning tokens under ambiguity.
• Yuwen Du et al. released OpenSeeker, the first fully open-source frontier search agent (model weights + complete 11.7k-sample dataset) that reaches SOTA via fact-grounded controllable QA synthesis, beating DeepDive 29.5% vs 15.3% on BrowseComp.
• Giorgos Nikolaou et al. prove transformer language models are injective (and therefore invertible) at initialization and throughout training, operationalized by the SipIt algorithm that exactly reconstructs input text from hidden activations for transparency and interpretability.
• Usman Anwar et al. formalised steganography decision-theoretically using generalised V-information to enable detection and mitigation of obfuscated chain-of-thought in LLMs without needing a reference distribution; LessWrong post explores whether monitoring remains possible.
• SE Gyges argues that alignment is proven tractable because LLMs can now load human intent via language and fine-tuning, shifting the Value-Loading Problem from unsolved to one with known pathways (though value selection and generalization remain open).
• Samy Jelassi et al. introduce energy-based fine-tuning (EBFT) via feature-matching on completion-distribution statistics, matching RLVR and beating SFT on Q&A/coding/translation while lowering validation cross-entropy.

🏛️ AI Policy, Governance & Safety

• UK AISI tested seven frontier AI agents on custom cyber ranges: capability scales with model generation and inference-time compute (Opus 4.6 averages 15.6/32 steps at 100M tokens), but plateaus after specialist-knowledge milestones; models exploit unintended paths.
• Arena.ai built the BS Bench (Bullshit Benchmark) testing 80 models on pure-nonsense questions: some models push back and refuse while others confidently invent fake metrics; "thinking harder" (reasoning modes) actually worsened refusal rates.

🛠️ AI Tools & Products

• Michaelliv built napkin, a local-first Obsidian-compatible file-based memory system for agents using progressive disclosure (NAPKIN.md → TF-IDF keyword map → BM25 search → full read) + auto-distillation sub-agent + interactive graph to replace RAG pipelines entirely (blog) —free.
• ZeroLeaks is an AI security platform —no further details available.
• NeoVertex1 built nuggets, an AI assistant with "holographic memory" —free (open source).
• OpenHome released the DevKit hardware + SDK for building local-LLM voice agents with far-field mics, custom personalities, REST API and MQTT integration, multi-step conversations, and scheduled actions —free limited-run DevKits ship to selected builders.
• Spielwerk (iOS) creates phone-based mini-games with drag-and-drop character/superpower/rules creation, no code required, with new "Cheats" visual building blocks —free with in-app purchases.
• Okara AI CMO is a growth marketing agent —no pricing details.
• LabClaw is a Skill Library for autonomous biomedical research.
• TianGzlab built OmicsClaw, a local-first conversational memory-enabled research assistant for multi-omics analysis (56+ skills across spatial/single-cell/genomics/proteomics/metabolomics/bulk RNA-seq) with smart natural-language routing via Telegram/Feishu.
• ClawInstitute is an AI-scientist research network where 123 AI agents collaborate in domain workshops (medicine, autoresearch, AI safety, agent design, biology) by hypothesizing, experimenting, and critiquing papers in public discussion threads.

🔊 Audio, Video & Creative AI

• Royal Cities built and released Foundation-1, a SOTA local text-to-sample diffusion model for music production (7 GB VRAM, ~7-8 sec/sample on RTX 3090, layered prompting for instrument family + timbre + FX + notation + BPM + bars, entirely hand-crafted structured loops with no scraped data) (YouTube demo) —free (open source).
• Artificial Analysis published its Speech-to-Speech leaderboard showing NVIDIA's Nemotron 3 VoiceChat V1 as the top open-weights leader (77.8% conversational dynamics + 29.2% reasoning) while proprietary models (Step-Audio 96%, Grok Voice 92%) still lead overall.
• MIT Han Lab published SANA-Video, an efficient 2K video generation system using deeper-compression AutoEncoder, linear attention, and two-stage refiner (HF model, project site).

🤖 Robotics & Physical AI

• Bones Studio open-sourced BONES-SEED, 142k high-fidelity motion-capture sequences (288 hours, up to 6 natural-language descriptions per motion, temporal segmentation, full biometric metadata, SOMA + Unitree G1 MuJoCo formats) purpose-built for humanoid robotics (HF dataset).
• GalaxyGeneralRobotics built LATENT, the open-source pipeline for Learning Athletic Humanoid Tennis Skills from imperfect human motion data using MuJoCo simulation, online DAgger distillation, and high-level policy learning.
• Timing Yang built Fast SAM 3D Body, a real-time human mesh recovery system that is 10.25x faster for 3D estimation and 10,426x faster for MHR→SMPL conversion (~65 ms end-to-end), deployable on humanoid robots.

📊 Fundraising & Deals Roundup

• Halcyon — $21M Series A for ML platform ingesting 2-20K public-utility documents daily for energy data (already in digest but confirming).

💡 Industry Commentary & Analysis

• Distillabs benchmarked 15 small LLMs across 9 tasks and found fine-tuning matters more than raw size: Liquid AI's LFM2-350M ranks #1 in tunability, the entire LFM2 family sweeps top-3, and a fine-tuned Qwen3-4B matches or beats a 120B teacher on 8/9 benchmarks at ~30x lower cost.
• Yuchen Jin argues that OpenAI and xAI are now refocusing on coding (while Anthropic always prioritized it), leading to 3-4 specialized top-tier coding models rather than winner-take-all.
• himanshu argues that modern multimodal AI problems are fundamentally "expansion problems" (low-entropy symbolic descriptions generating high-dimensional physical trajectories), with direct physics-style correlations across text-to-speech, text-to-image, text-to-video, and all modalities.
• Patrick Pynadath argues that evaluation methodology matters critically for diffusion language models, calling for standardized benchmarks beyond perplexity.

Around the Horn — Monday, March 16, 2026

Today was all about GTC; read our full coverage here.

Otherwise, Meta signed a deal worth up to $27 billion over five years with Nebius for dedicated AI compute capacity (including first large-scale Nvidia Vera Rubin chips) starting 2027, as part of its $135B capex push. And then, in the same breath, Reuters reported that Meta is planning sweeping layoffs as those AI infrastructure costs mount. Spend $27 billion on machines, cut the humans. The future of work, visualized in one earnings report.

Meanwhile, Greg Brockman announced that GPT-5.4 ramped faster than any prior OpenAI model, processing 5 trillion tokens per day within one week and hitting $1B in annualized net-new revenue. OpenAI is courting private-equity firms (TPG, Advent, Bain Capital, Brookfield) for a $10B-pre-money enterprise AI joint venture. Alibaba created a new Token Hub unit under CEO Eddie Wu that consolidates Qwen, DingTalk, Quark, and all AI products, and is launching an agentic AI service this week to ride China's OpenClaw craze. Ex-Anthropic researchers are raising at a $1B valuation for a new startup nobody knows anything about yet; Ethan Mollick flagged it.

And Encyclopedia Britannica sued OpenAI for copying nearly 100,000 articles to train ChatGPT. The FSF also threatened Anthropic over copyright, demanding models be shared freely. The legal front is heating up.

🏆 TOP 5 NEWS

• Meta signed a deal worth up to $27B over five years with Nebius for dedicated AI compute capacity (including first Nvidia Vera Rubin chips at scale) while simultaneously planning sweeping layoffs as AI infrastructure costs mount.
• Greg Brockman announced GPT-5.4 is processing 5 trillion tokens per day within one week of launch, exceeding all prior API volume and generating $1B in annualized net-new revenue.
• OpenAI is courting private-equity firms (TPG as anchor, Advent, Bain Capital, Brookfield) for a $10B-pre-money enterprise AI joint venture with board seats, preferred equity, and distribution through PE portfolios.
• Alibaba created a new Token Hub unit consolidating Qwen, DingTalk, Quark, and all AI products under CEO Eddie Wu, plus is launching an agentic AI service this week that operates browsers and cloud servers for enterprises.
• Encyclopedia Britannica sued OpenAI alleging it copied nearly 100,000 articles and encyclopedia entries to train ChatGPT, which now produces near-verbatim copies and hallucinates false citations of Britannica material.

Honorable Mentions:

• Sara Hooker argues in "On the Slow Death of Scaling" that pure scaling has plateaued; future gains will come from data quality, architecture search, and test-time compute rather than more parameters.
• Google is providing Gemini AI agents to the Pentagon for routine unclassified work across its three-million-person workforce.
• A DHS hack exposed 6,800 company bids and 1,400 contracts worth $845M revealing AI ambitions including CCTV passenger tracking at TSA checkpoints, cellphone biometric scanners, and a national 911 AI prediction platform.
• ByteDance paused its global Seedance 2.0 video-generator launch after Disney and Paramount issued cease-and-desist letters citing copyright concerns.

🍪 TOP TREATS TO TRY

• ElevenCreative generates, edits, localizes and scales studio-grade audio/video content in one browser Studio by mixing ElevenLabs voices, music, sound effects, images, and video models, turning a single prompt into full localized campaigns dubbed into 70+ languages —no pricing details.
• Adaptive Computer is an always-on personal computer with built-in AI agents that connect your tools (Gmail, Slack, Sheets, Notion via one-click OAuth), build and run full workflows, encode learned memory for future agents, and handle everything from spreadsheet uploads to daily sales reports with optional human approvals —free month with social share.
• RunAnywhere RCLI turns your Mac into a fully local voice AI assistant with speech-to-text, LLM chat, text-to-speech, 43 macOS actions, and local RAG (retrieval-augmented generation; it ingests your docs and answers questions about them) at ~4ms retrieval over 5K+ chunks, powered by MetalRT (their proprietary GPU inference engine (the software layer that runs AI models on your hardware) for Apple Silicon that's 1.1-1.2x faster than MLX on LLMs, 4.6x faster on speech-to-text and 2.8x faster on text-to-speech); note: competitor uzu beat MetalRT on Llama-3.2-3B (222 vs 184 tok/s) though MetalRT won 3 of 4 benchmarked models, and HN commenters flagged that only sub-4B models are currently supported (larger model support on roadmap), tool-calling reliability is still weak at those sizes, and the company previously drew controversy for scraping GitHub emails for cold outreach —free (MIT license for RCLI; MetalRT engine is proprietary).
• MuleRun runs on its own dedicated 24/7 computer (not a browser tab) so you describe a workflow once in plain English and it executes multi-step tasks proactively overnight, delivers finished results to your inbox or Slack, and pauses for approval on critical steps —free to try.
• Agent37 gives you managed OpenClaw hosting with browser terminal, dashboard, live interventions, 850+ pre-wired integrations, and 80% skill monetization for creators in under 60 seconds —$3.99/month early-adopter, then $9.99/month.
• Struct automates your on-call runbook by cross-referencing logs, metrics, traces and your codebase to root-cause every engineering alert with impact analysis and suggested fixes (just @mention in Slack) —free 100 credits/month, Pro $20/month.
• Air by JetBrains runs multiple coding agents (Codex, Claude Agent, Gemini CLI, Junie) in parallel inside isolated Docker or Git worktrees with full code intelligence, precise task delegation, and real-time progress monitoring —free with JetBrains AI Pro/Ultimate.

🏢 Big Tech & Major Companies

• Anthropic updated Claude for Excel and PowerPoint so files now share full conversation context seamlessly, Skills for one-click standard workflows are available in the add-ins, and support expanded to Amazon Bedrock, Vertex AI, and Microsoft Foundry (beta for all paid plans on Mac/Windows).
• Anthropic doubled Claude usage limits during off-peak hours (outside 8 AM–2 PM ET weekdays) for two weeks (March 13–27) across Free, Pro, Max and Team plans.
• Google released WaxalNLP, a large-scale open African languages speech corpus with ~1,250 hours of transcribed natural speech across 19 languages and 180+ hours of single-speaker recordings across 17 languages.
• Google rolled out Ask Maps (conversational personalized recommendations powered by Gemini) and Immersive Navigation (3D views, real-time updates, Street View previews, parking help) in Google Maps for U.S. and India on Android/iOS.
• Google scrapped its "What People Suggest" AI feature that crowdsourced amateur health advice from online forums.
• Logan Kilpatrick announced spend caps are now live in the Gemini API, giving developers hard spending limits and greater cost control.
• GPT-5.4 showed markedly faster and more efficient agentic computer use, navigating complex interfaces via screenshots far better than early agent versions from one year ago.
• Microsoft quietly scrapped plans to put Copilot in Windows 11 notifications, Settings and File Explorer, renaming "Windows Copilot Runtime" to "Windows AI APIs" to reduce AI bloat.
• Google, Amazon, Microsoft, Meta, LinkedIn, OpenAI, Adobe and Match Group signed a voluntary accord to share scam intelligence, deploy AI detection tools, and strengthen fraud prevention ahead of the UN Global Fraud Summit.
• xAI started hiring Wall Street bankers, credit analysts, portfolio managers and crypto/equity traders to teach Grok financial modeling including leveraged-loan syndication, distressed investing, and mortgage-backed securities.
• BMW deployed Hexagon AEON humanoid robots at its Leipzig iFACTORY for EV production, using Physical AI decision-making to handle sheet-metal positioning and avoid obstacles.
• Perplexity launched Computer inside its Comet browser so the agent can take full local-browser control of any site or logged-in app with user permission, without connectors or MCPs.
• Perplexity Computer Skills now supports reusable SKILL.md instruction sets that activate automatically for consistent tasks (research reports, presentations, charts) and combine multiple skills seamlessly.

💼 AI Productivity, Labor & Economics

• Ethan Mollick argues AI has crossed into autonomous agents handling substantial real human work in minutes, shifting organizations from co-piloting to managing AI factories.
• Morgan Stanley warned that a massive AI capability breakthrough driven by extreme compute scaling is coming in H1 2026, predicting rapid productivity gains, major job disruption, and severe power shortages.
• Justin Reock argues real AI productivity gains in engineering teams are only ~10% (not 10x) because coding was never the bottleneck; DX's study of 400 companies showed 65% AI usage but just 9.97% more PRs shipped.
• François Chollet argues automation machines should be sold widely while invention machines are best monetized by using them yourself.
• Meir Cohen argues that if your only moat was "I called the API first" you never had one; only proprietary data, deep workflow integration, or years-built trust survive 2027.
• The Washington Post interactive shows secretaries and clerical workers (6.1M jobs, 86% held by women) are most vulnerable to AI due to high task exposure plus low adaptability.
• Fortune reports 75% of resumes never reach a human because ATS systems auto-reject them; by 2027 most hiring will assess workplace AI proficiency alongside human judgment.
• Michael Novati argues AI is about to reveal most of what we called "talent" was just scarcity by automating cognitive labor once prized for its rarity.
• Ariel Rosenfeld (Bar-Ilan) argues he may "hire" AI instead of a graduate student because it delivers immediate competent synthesis without ramp-up, though he still sees human learning as irreplaceable.
• John Ennis advises builders disrupted by Claude's native charts to pivot to custom Claude skills for businesses.
• qrimeCapital notes custom Claude skills for niche business data-viz remain major alpha despite the platform disruption.

🤖 AI Agents & Infrastructure

• OpenClaw-RL trains any agent simply by talking: every user correction, re-query, or next-state signal (terminal, GUI, tool output) becomes live online RL data (reinforcement learning, where the AI learns from trial and error) via PRM rewards + Hindsight-Guided On-Policy Distillation for continuous self-improvement from normal usage. DAIR.AI highlights it lifted personalization scores from 0.17 to 0.81 after 16 steps.
• Gaodan Fang et al. propose Trajectory-Informed Memory Generation for self-improving agents: a Trajectory Intelligence Extractor + Decision Attribution module pulls strategy and recovery tips from full task trajectories, delivering up to +14.3pp goal completion (+149% relative on complex tasks).
• elvis (@omarsar0) shares a position paper arguing multi-agent memory should be treated as a computer-architecture problem with shared vs. distributed paradigms and a three-layer hierarchy.
• stash_pomichter showed OpenClaw running on a drone where it autonomously invented custom skills to track objects and patrol an area, then open-sourced DimensionalOS so any agent can safely control physical hardware.
• PycoClaw brings a full OpenClaw AI agent (dual-loop reasoning, recursive tools, context compaction, persistent memory, hardware GPIO/LVGL control) to $5 ESP32 microcontrollers (tiny programmable chips) with one-click browser flash and ScriptoHub skills marketplace.
• Unsloth published a tutorial on building RL environments with NVIDIA NeMo Gym: decoupled task/resource/agent servers, verifiable rewards, and SFT warmup before RL for robust agents.
• OneEyeKeh demoed a 5-minute prompt setup turning OpenClaw + HeyGen API into a fully automated YouTuber video generator.
• Andrew Ng released context-hub, an open library that lets any agent swap context-compression backends (LLM summarizer, vector store, or memory graph) with one line of code and automatic routing based on task length.
• Workshop Labs built Trellis for Kimi K2-Thinking post-training: expert-parallel LoRA (a parameter-efficient fine-tuning method) that trains all MoE (Mixture of Experts, where different parts of the model specialize in different tasks) parameters at 6,600 tokens/sec on a single 8xH200 node, 50x faster than the best open-source alternative —open-sourcing after safety review.
• Adaptive launched Adaptive Computer with built-in AI agents, scoped OAuth integrations, memory encoding, and optional human approvals on sensitive steps.

💻 AI Coding & Developer Tools

• Dwayne (@CtrlAltDwayne) makes the case that Rust's strongest selling point today is how well AI coding assistants perform with its tight compiler feedback loop.
• Nick Saraev showed how to combine Claude Code skills with Andrej Karpathy's autoresearch methodology to build self-improving prompt systems: define binary yes/no eval criteria (e.g., "is all text legible?", "does it use pastel colors?"), have an agent generate outputs every 2 minutes, score them against the eval suite, mutate the SKILL.md prompt, and keep the winner, achieving 39/40 (97.5%) on diagram generation within a few runs at ~$0.20/test. Karpathy also committed a tiny prompt tweak ("soften the language") that reduced agent overconfidence by 31% on the live leaderboard. Key tip: keep evals binary (not Likert scales) to avoid compounding probability noise, and don't make criteria so narrow the model games the test without improving quality.
• Dan McAteer argues text-based Unix CLIs beat structured tool calling for agents because decades of command-line data live in LLM training.
• CopilotKit released OpenGenerativeUI, an open-source template replicating Claude's interactive charts/diagrams for your own agents using LangGraph + CopilotKit.
• Zeno Rocha launched the fully open-source Resend CLI with 53 commands, interactive mode for humans, structured JSON for agents and CI, and one-line install.
• tobi open-sourced qmd, a fully local mini CLI search engine for docs/knowledge bases/meeting notes using hybrid BM25 + semantic search + LLM query expansion + reranking with MCP integration.
• davebcn87 open-sourced pi-autoresearch, an autonomous experiment loop extension for the pi editor that continuously optimizes any metric by running commands, measuring results, auto-committing improvements, and reverting failures.
• wizwand built autoexp, a one-line tool that turns any ML training repo into an autonomous experiment loop where AI agents iteratively run, evaluate, and optimize experiments.
• claudetop is an htop-style CLI and Claude Code plugin giving real-time cost tracking, cache-hit ratios, model comparisons, burn-rate forecasts, session tagging, daily budgets, and smart alerts —free.
• GitAgent is the open standard and CLI for defining, versioning, and running git-native AI agents directly inside repositories that export to Claude Code, OpenClaw, CrewAI, and others with full git-based audit trails.
• Rob Englander argues AI didn't simplify software engineering; it just made bad engineering easier by accelerating spec drift.
• Stavros shared his exact LLM workflow: an Architect agent (Claude Opus 4.6) turns goals into plans, delegates to a Developer (Sonnet 4.6) and 1-3 Reviewers (Codex/Gemini/Opus) that critique diffs and escalate disagreements.
• Ensue Network's autoresearch@home lets you run research agents locally or on cloud GPUs to contribute experiments to a live community leaderboard (2,857 experiments, 101 agents) —free to try.
• The top Google result for "Claude Code" was malicious, serving a fake download page designed to steal credentials.

🔬 AI Research & Models

• Yulu Gan and Phillip Isola argue in Neural Thickets that large pretrained models contain dense clusters of diverse task-specific experts near their weights, enabling RandOpt (parallel random perturbation + top-K selection + majority-vote) to beat complex post-training like PPO/GRPO (project site).
• Goodfire AI researchers show in Reasoning Theater that models often perform "performative" chain-of-thought (going through reasoning motions after already deciding internally); activation probes detect this early and enable up to 68% token savings (blog).
• Sara Hooker argues in "On the Slow Death of Scaling" that pure scaling (just adding more parameters and data) has plateaued; future gains will come from data quality, architecture search, and test-time compute.
• inclusionAI released LLaDA2.1-flash, a 100B MoE diffusion LLM (a model that generates text by iteratively refining from noise rather than predicting one word at a time) with 32K context and token-editing for ultra-fast inference (109+ tok/s, 1.67s first-token latency) in dual speed/quality modes, SOTA on knowledge/reasoning/coding/math/agent tasks.
• Moonshot AI / Kimi introduced Attention Residuals, replacing fixed uniform accumulation with learned input-dependent attention over preceding layers, delivering 1.25x compute advantage and <2% latency overhead on their 48B/3B Kimi Linear architecture.
• ICRL (GitHub) shows you can teach LLMs tool use through in-context examples during RL rollouts (fading from few-shot to zero-shot) and achieve SOTA on TriviaQA/HotpotQA without supervised fine-tuning.
• Jiaheng Hu et al. released Simple Recipe Works, demonstrating naive sequential fine-tuning with LoRA on Vision-Language-Action models using on-policy RL enables strong continual learning with almost no catastrophic forgetting.
• Researchers propose training AI co-scientists via RL with rubric rewards from existing papers, achieving 70% expert preference on ML goals.
• Jasper Dekoninck launched BrokenArXiv, a dynamic benchmark showing frontier LLMs (even GPT-5.4 at only 40% rejection) frequently claim to prove false theorems.
• alphaXiv shared "On-Policy Self-Distillation for Reasoning Compression", cutting reasoning length 59% on MATH-500 while raising accuracy to 86% and gaining +10.5pp on AIME 2024.
• Chelsea Finn shared Data Analogies (paper): a single policy trained on one robot transfers to five others with 3-5x fewer samples by learning analogy mappings between observation spaces.
• Stepfun released Step-3.5-Flash, a 32B MoE model optimized for agentic tool use and long-chain reasoning with 200K context.
• Sebastian Raschka built the LLM Architecture Gallery, a visual collection of architecture diagrams and fact sheets for 50 models.
• Han Xiao built a visual test that renders 36 rotated frames of a 3D Pikachu and projects embeddings to 3D via PCA; DINOv2 wins cleanly, CLIP/SigLIP/Gemini scatter.
• Nishanth Kumar built TiPToP, a zero-shot open-world robotic manipulation system achieving 59.4% success across 165 trials on 28 novel tasks with no task-specific training.
• Dylan Ebert launched Texel Splatting, a real-time 3D Gaussian splatting variant with texel-aligned rasterization for 4x faster rendering on consumer GPUs.

🏛️ AI Policy, Governance & Safety

• Neel Nanda and Arya Jakkli red-teamed Claude's full 30K-word constitution (205 testable requirements); Opus 4.5/4.6 show substantially fewer violations, confirming alignment stacking benefits.
• Monk Zero (@NoCommas) exposed how the #1 Terminal Bench entry cheated with 48 pre-recorded trajectories, domain cheat sheets, XOR-obfuscated model names, and a forked Gemini CLI npm package.
• The Free Software Foundation threatened Anthropic over copyright, demanding LLM developers share complete models, configurations, and source code freely.
• President Trump drove a wedge between Florida Republicans; his opposition to state-level AI rules caused the House to kill Gov. DeSantis's AI "bill of rights" (chatbot disclosures, ban on AI in licensed mental-health counseling).
• A DHS hack of the technology incubator exposed 6,800 bids and 1,400 contracts worth $845M revealing CCTV tracking, cellphone biometric scanners, and a national 911 AI platform.
• Avner Gvaryahu argues the largest AI firms (Palantir, Anthropic, OpenAI, Google, Anduril) are actually defense contractors whose systems sit inside military kill chains.
• Cambridge researchers warn AI toys for young children need tighter rules because current systems misread emotions, interrupt kids, and may confuse social learning.
• CodeWall demonstrated how an autonomous offensive agent exploited McKinsey's AI platform via unauthenticated SQL injection, chaining IDOR to exfiltrate 46.5M chat messages, 3.68M RAG chunks, and millions of files.
• Trump accused Iran of using AI as a "disinformation weapon" to fabricate kamikaze boats, a fake attack on the USS Abraham Lincoln, and a 250,000-person rally.
• The Department of Veterans Affairs plans to deploy LLMs to scan Disability Benefits Questionnaires for fraud indicators despite fraud rates under 0.01% and hallucination risks.
• AI-generated images and videos of the Michigan tornado damage went viral without labels, fooling thousands.
• Pennsylvania landowners are fighting PPL's new 500 kV high-voltage lines for AI data centers, rejecting eminent-domain offers.
• Organizations worldwide are racing to create a universally recognized "human-made" or AI-free certification logo, but competing definitions and light auditing risk leaving consumers confused.
• Deepfake influencers including an AI-generated Amish woman and a Tibetan monk are selling $50 detox supplements on social media with zero disclosure.

🛠️ AI Tools & Products

• Numman Ali shared preview notes on Z.ai's Pony Alpha 2, noting it is faster than GLM 5, less sycophantic, and "feels like Claude Opus."
• GLM-5-Turbo is Z.AI's OpenClaw-optimized model (200K context, 128K output, streaming, function calling, structured JSON) that outperforms GLM-5 on ZClawBench agent benchmarks for tool use and long-chain tasks.
• ByteRover gives you portable hierarchical natural-language memory across OpenClaw, Claude Code and Cursor with 92.2% LoCoMo retrieval accuracy, local-first execution, cloud sync, and SOC 2 security —free tier available.
• Context Gateway is an agentic proxy that sits between any AI agent and the LLM API for instant history compaction and context optimization —free.
• googleworkspace/cli is a dynamic CLI (auto-generated from Google Discovery Service) for Drive, Gmail, Calendar, Sheets, Docs, Chat, Admin, with structured JSON output and 100+ built-in AI agent skills —free.
• OpenJarvis from Stanford is an open-source on-device personal AI framework with composable primitives for intelligence, agents, tools/memory and closed-loop local learning.
• Spielwerk (iOS) lets anyone create and share phone-based mini-games by inventing characters and superpowers with drag-and-drop rules, no code —free with in-app purchases.
• Ink is infrastructure built for AI agents with real-time structured logs, CPU/memory/network metrics, pay-per-use pricing, and universal deployment.
• Sentrial streams every agent session, replays failures step-by-step, auto-detects drift/hallucinations, and lets you hot-patch prompts or roll back —free tier, then enterprise.
• Deeptrace replays every agent decision trace with video, token heatmaps, and one-click rollback.
• Voygr (YC W26) is a maps API for AI agents with fresh, queryable place profiles, business validation, and web-grounded context —pricing on par with Google Maps.

💡 Industry Commentary & Analysis

• Tom Reed (@mentalgeorge) analyzes METR's SWE-bench study to illustrate jagged progress: post-training first cracked functionality then code quality at different capability thresholds.
• Horace Dediu argues Apple's refusal to build AI data centers (licensing Gemini instead while running 70B models locally on M5 chips across 2B devices) is the most brilliant corporate move because it avoids $650B infrastructure commoditization, preserves cash for $90B+ buybacks, and owns the customer while hyperscalers bleed free cash flow.
• Dr. Randal S. Olson argues the "Are You Sure?" problem (AI flipping confident answers under challenge due to RLHF-trained sycophancy, changing ~60% of math/medical responses) makes models dangerous for strategic decisions; the fix is embedding persistent decision frameworks.
• @scaling01 argues Meta failed to execute the obvious DeepSeek recipe on Avocado despite unlimited resources.
• Dimitris Papailiopoulos observes we are on the cusp of models that post-train themselves.
• roon (@tszzl) argues AI value will mostly be captured by consumers as surplus in competitive markets, not labs or chipmakers, because it democratizes white-collar elite services and shrinks rents on human capital.
• Charles Petzold argues Spotify's AI DJ is appallingly stupid because it cannot comprehend that a symphony consists of multiple sequential movements.
• Tom Johnell argues LLMs become exhausting because mental fatigue degrades prompt quality and slow feedback loops produce dumb responses; the fix is recognizing tiredness and redefining problems for sub-5-minute iterations.
• Bozhidar Batsov argues Emacs and Vim will survive the AI era because their communities keep adapting, and AI actually lowers the barrier to writing extensions.
• Abhishek Ray argues you cannot trust autonomous agents that run while you sleep unless you first write explicit acceptance criteria and verify them with Playwright-style checks.
• Groundtruthpost argues AI supercharges the colleague who always has an answer by removing natural pressure circuit-breakers.
• The backnotprop blog argues Anthropic should let Claude Code users opt out of A/B tests because unannounced changes degrade paid professional workflows.
• New data from Anthropic maps that D.C. (4x), New York (2.68x), Massachusetts, and California lead U.S. states in Claude usage relative to working-age population.

📊 Fundraising & Deals Roundup

• Meta — Up to $27B over five years with Nebius for dedicated AI compute (Vera Rubin chips).
• AMI Labs — $1.03B for world models (JEPA-style abstract physical prediction).
• World Labs — $1B for spatial AI world models (persistent 3D scene understanding).
• Ex-Anthropic researchers — Raising at $1B valuation for undisclosed new startup.
• Genspark — Extended Series B to $385M at $200M ARR for AI employee platform.
• Halcyon — $21M Series A for ML platform ingesting 2-20K public-utility documents daily for energy data.
• Google/Accel India accelerator — Selected 5 non-wrapper startups from 4,000 applications for up to $2M each.

Paper / tools to read more up on:

Didn't get a chance to review these today, but will tomorrow! Check the links if they interest you.

• SANA-Video 2K generation (MIT Han Lab): blog, HF model, site, X
• Spatial-TTT (streaming spatial intelligence with test-time training): site, GitHub, HF paper
• Manus "My Computer" desktop agent: blog, X
• AISI (UK AI Safety Institute) cyber attack agent evaluation: blog, X
• Language Model Teams as Distributed Systems: paper
• Matching Features, Not Tokens (energy-based LLM fine-tuning): paper, X
• get-physics-done (PSI open-source AI physicist): GitHub, X
• Language Models are Injective and Hence Invertible: paper
• OmicsClaw (genomics/biology agent): GitHub, X
• LATENT (humanoid tennis robot): GitHub, X
• Benefits and Limitations of Communication in Multi-Agent Reasoning: paper
• DFlash (Z-Lab model collection): HF collection, X
• Temporal Straightening for Latent Planning: paper
• OpenMAIC (multi-agent interactive classroom): GitHub, X
• Physics-Based Simulation textbook: site, X
• autovoiceevals (self-improving voice AI loop built on Karpathy's autoresearch): GitHub, X

Tools / products:

• Okara AI CMO (growth marketing agent): site, X
• Honcho (agent memory infrastructure): site
• herm (terminal-native AI coding agent in containers): GitHub, X
• Flywheel: site
• ClawInstitute (academic exchange for AI agents): site, X
• Cofounder Fellowship: site, X
• MuleRun Chat: site, X

Around the Horn — Sunday March 15, 2026

The big story this week is the AI agent going from party trick to payroll.

Genspark hit $200M ARR (doubling in the last two months), extended its Series B to $385M, and launched Workspace 3.0 with Claw, your first AI employee on a dedicated Cloud Computer. Claw handles Slack, email, research, scheduling, and code via messaging, plus automated Workflows, Teams chat, Meeting Bots, a Speakly voice app, and a Chrome extension. Not an assistant. An employee. Meanwhile, Perplexity launched Computer, an always-on cloud Mac mini agent that builds pitch decks, financial models, web apps, plans trips with bookings, triages inboxes, and tracks portfolios, now available on iOS with cross-device sync (Android soon). FrameworkPuter called it out for phoning home rather than running truly locally, and OsaurusAI responded by open-sourcing a genuinely local AI agent framework for Mac that runs any model fully offline in isolated VMs.

And Ethan Mollick put the whole moment into words: AI has crossed into autonomous agents capable of handling substantial real human work in minutes, shifting organizations from co-piloting to managing AI factories, with a narrowing window before recursive self-improvement accelerates progress even faster.

We went from "AI can help you write an email" to "AI has a Slack account and a workstream" in about 18 months. So how do you onboard an agent? Read everything in the Google Drive? Here's my job, good luck, have fun? Do they even need training videos or can you paste the company policy docs into their context window?

🏆 TOP 5 NEWS (Around the Horn)

• AMI Labs raised $1.03B and World Labs raised $1B weeks apart, both betting on world models (systems that understand and simulate physical reality), with Zhuokai Zhao categorizing five distinct approaches: JEPA for abstract prediction, spatial intelligence via Marble, learned simulation like Genie/Dreamer, NVIDIA's Cosmos infrastructure platform, and active inference at VERSES AI.
• Anthropic added native interactive charts and diagrams to Claude chat (beta on all plans), accidentally one-shotting a builder's $200K ARR business selling RAG-based interactive chart agent skills, as half their customers cancelled overnight.
• Google's Aletheia agent (powered by Gemini 3 Deep Think) autonomously solved 6/10 novel research-level FirstProof math problems plus several previously unsolved Erdős problems through iterative generate-verify-revise cycles.
• OpenAI Codex shipped two major updates: customizable, importable, and shareable themes plus Automations now generally available with model/reasoning-level selection, worktree-or-branch runs, and workflow templates for recurring tasks like daily repo briefings and PR follow-ups.
• Microsoft updated Copilot Notebooks into a persistent three-column AI workspace with integrated references (Office files and PDFs live-updating), evolving overviews and summaries, one-click creates for drafts, quizzes, and flashcards, and shareable collaboration, now generally available.

Honorable Mentions:

• Morgan Stanley warned that a massive AI capability breakthrough driven by extreme compute scaling is coming in H1 2026, predicting rapid productivity gains, major job disruption, and severe power shortages.
• Neel Nanda and Arya Jakkli red-teamed Claude's full 30K-word constitution by breaking it into 205 testable requirements and running adversarial scenarios; newer Opus 4.5/4.6 models show substantially fewer violations (fabrication and unilateral actions remain weak spots).
• Shopify CEO Tobi Lütke merged a Liquid PR delivering 53% faster parse+render and 61% fewer allocations via 120 automated autoresearch experiments.
• Dimitris Papailiopoulos observes we are on the cusp of models that post-train themselves because the self-improvement loop is now closing.

🍪 TOP TREATS TO TRY

• Genspark Workspace 3.0 with Claw gives you a dedicated AI employee on a Cloud Computer that executes complex multi-app tasks (Slack, email, research, scheduling, code) via messaging plus automated Workflows, Teams chat, and Meeting Bots —free to start.
• Perplexity Computer hires a cloud agent to instantly build pitch decks, financial models, web apps, plan trips with bookings, triage inboxes, and track portfolios, now on iOS with cross-device sync —free to start.
• Hume AI's TADA generates synchronized text and audio in one pass across 1,000+ samples, runs 5x faster than LLM-based text-to-speech (where an AI reads text aloud), and eliminates content hallucinations —free to try.
• ACE-Step-1.5 generates full songs (10 seconds to 10 minutes) with lyrics in 50+ languages, supports cover generation, track separation, and LoRA personalization from few examples, all locally on Mac/AMD/Intel/CUDA with under 4GB VRAM —free (open source).
• omou is a minimal, keyboard-first Gmail client that strips inbox noise and adds targeted AI assistance for fast email triage —no pricing details.
• Stirrup runs AI agents directly in any Slack channel via @mention for data analysis, code execution, web browsing, and custom skills with threaded results —no pricing details.
• Junior by HireJuniorSo is an AI employee with its own identity and persistent organizational memory across your team's Slack channels for fully self-driven task handling —$2,000/month.

🏢 Big Tech & Major Companies

• Anthropic updated Claude for Excel and PowerPoint so files now share full conversation context seamlessly, Skills for one-click standard workflows are available in the add-ins, and support expanded to Amazon Bedrock, Vertex AI, and Microsoft Foundry (beta for all paid plans on Mac/Windows).
• Google released WaxalNLP, a large-scale open African languages speech corpus with ~1,250 hours of transcribed natural speech (ASR) across 19 languages and 180+ hours of high-quality single-speaker recordings (TTS) across 17 languages to advance speech tech for 100M+ speakers.
• Logan Kilpatrick announced spend caps are now live in the Gemini API, giving developers hard spending limits and greater cost control.
• GPT-5.4 showed markedly faster and more efficient agentic computer use, navigating complex interfaces via screenshots far better than early agent versions from one year ago.

💼 AI Productivity, Labor & Economics

• Ethan Mollick argues AI has crossed into autonomous agents capable of handling substantial real human work in minutes, driven by exponential benchmark gains and examples like fully AI-run software companies, amid rolling disruptions and a narrowing window before recursive self-improvement accelerates further.
• Morgan Stanley warned that a massive AI capability breakthrough driven by extreme compute scaling is coming in H1 2026, predicting rapid productivity gains, major job disruption, and severe power shortages as intelligence becomes the primary economic resource.
• François Chollet argues automation machines should be sold as widely as possible to anyone with tasks while invention machines are best monetized by using them yourself to create maximum value.
• Meir Cohen argues that if your only moat was "I called the API first" you never had one, predicting only proprietary data, deep workflow integration, or years-built trust relationships will survive 2027.
• John Ennis advises builders disrupted by Claude's native charts to pivot to custom Claude skills for businesses, noting they can expose backend routes and bring full app functionality into Claude chat.
• ashe predicts websites will largely move to dynamic UI and suggests finding niche data-viz clients and pitching transitions to interactive pages.
• qrimeCapital notes custom Claude skills for niche business data-viz remain major alpha despite the platform disruption.

🤖 AI Agents & Infrastructure

• OpenClaw-RL trains any agent simply by talking: every user correction, re-query, or next-state signal (terminal, GUI, tool output) becomes live online RL data (reinforcement learning, where the AI learns from trial and error) via PRM rewards + Hindsight-Guided On-Policy Distillation for continuous self-improvement from normal usage. DAIR.AI highlights it lifted personalization scores from 0.17 to 0.81 after 16 steps.
• Gaodan Fang et al. propose Trajectory-Informed Memory Generation for self-improving agents: a Trajectory Intelligence Extractor + Decision Attribution module pulls strategy, recovery, and optimization tips from full task trajectories, then Adaptive Memory Retrieval injects them contextually, delivering up to +14.3pp goal completion (+149% relative on complex tasks).
• elvis (@omarsar0) shares a position paper arguing multi-agent memory should be treated as a computer-architecture problem with shared vs. distributed paradigms, a three-layer hierarchy (I/O, cache, memory), and multi-agent memory consistency as the biggest open challenge.
• stash_pomichter showed OpenClaw running on a drone where it autonomously invented custom skills to track objects and patrol an area, then open-sourced DimensionalOS so any agent can safely control physical hardware like drones, humanoids, and quadrupeds.
• Unsloth published a tutorial on building RL environments with NVIDIA NeMo Gym: decoupled task/resource/agent servers, verifiable rewards via RLVR/GRPO plus synthetic data, and SFT warmup before RL for robust tool-use, math, and code agents.
• OneEyeKeh demoed a 5-minute prompt setup turning OpenClaw + HeyGen API into a fully automated YouTuber video generator.
• rohanpaul_ai shared a video of OpenClaw controlling a real quadruped robot to fetch objects after learning the skill from three spoken corrections.

💻 AI Coding & Developer Tools

• Dwayne (@CtrlAltDwayne) makes the case that Rust's strongest selling point today is how exceptionally well AI coding assistants perform with its tight compiler feedback loop, delivering real-time self-correction signals unlike looser languages such as C++.
• Dan McAteer argues text-based Unix CLIs beat structured tool calling for agents because decades of command-line data live in LLM training, making plain text the native language shared by both CLIs and models.
• CopilotKit released OpenGenerativeUI, an open-source template replicating Claude's new interactive charts/diagrams for your own agents: AI generates live HTML/SVG visualizations, 3D animations, charts, and diagrams rendered safely in sandboxed iframes using LangGraph + CopilotKit.
• Zeno Rocha launched the fully open-source Resend CLI with 53 commands, interactive mode for humans (suggestions, tables, spinners), structured JSON for agents and CI, and one-line install across macOS/Linux/Windows.
• tobi open-sourced qmd, a fully local mini CLI search engine for your personal docs, knowledge bases, and meeting notes using hybrid BM25 + semantic search + LLM query expansion + reranking with MCP agent integration.
• davebcn87 open-sourced pi-autoresearch, an autonomous experiment loop extension for the pi editor that continuously optimizes any metric (test speed, bundle size, training loss, Lighthouse scores) by running commands, measuring results, auto-committing improvements, and reverting failures with persistent JSONL logs and a live dashboard.
• wizwand built autoexp, a one-line tool that turns any ML training repo into an autonomous experiment loop where AI coding agents iteratively run, evaluate, and optimize experiments by inferring commands and metrics from code.
• Paulo built a modern Clippy mascot (via Masko) that watches your Claude Code agents in real time, detects permissions issues, and instantly directs you to the correct terminal command.
• justin_schueler open-sourced a Resend CLI wrapper that turns any email template into an instant bulk sender with tracking.

🔬 AI Research & Models

• Yulu Gan and Phillip Isola argue in Neural Thickets that large pretrained models contain dense clusters of diverse task-specific experts in the local neighborhood around their weights (scaling dramatically with model size), enabling a simple parallel random perturbation + top-K selection + majority-vote method called RandOpt to compete with or beat complex post-training like PPO/GRPO/ES (project site).
• Goodfire AI researchers show in Reasoning Theater that models often perform "performative" chain-of-thought (going through reasoning motions after already deciding internally), especially on easy tasks; activation probes detect this early and enable up to 68% token savings via confidence-based early exit while maintaining accuracy (blog post, X).
• ICRL (GitHub, X) shows you can teach LLMs tool use through in-context examples during RL rollouts (fading from few-shot to zero-shot) and achieve SOTA on TriviaQA/HotpotQA without any supervised fine-tuning.
• Jiaheng Hu et al. released Simple Recipe Works (X), demonstrating that naive sequential fine-tuning with LoRA on pretrained Vision-Language-Action models using on-policy RL surprisingly enables strong continual learning with almost no catastrophic forgetting.
• Researchers propose training AI co-scientists via reinforcement learning with automatically extracted rubric rewards from existing papers for self-grading research plans, achieving 70% expert preference on ML goals and 12-22% cross-domain improvements.
• Jasper Dekoninck launched BrokenArXiv, a dynamic benchmark of subtly perturbed recent ArXiv math statements showing frontier LLMs (even GPT-5.4 at only 40% rejection) frequently claim to prove false theorems, measuring reliability separate from correctness.
• Han Xiao built a visual test that renders 36 rotated frames of a 3D Pikachu and projects embeddings to 3D via PCA; true geometric understanding appears as a smooth circle preserving angular order (DINOv2 wins cleanly, CLIP/SigLIP/Gemini scatter).
• Alif Munim built a Claude-scaffolded auto-trainer for sparse autoencoders (a tool for understanding what's happening inside neural networks) on Gemma3-1B that jumped from 68% to 96% loss recovery overnight via iterative ArXiv lookup, alive-feature tracking, and sparsity tweaks.
• Nishanth Kumar and team built TiPToP, a zero-shot open-world robotic manipulation system using vision foundation models + GPU-parallelized Task and Motion Planning, achieving 59.4% success across 165 trials on 28 novel tasks with no task-specific training.
• HeMuyu0327 turned the Neural Thickets RandOpt code into a one-click Hugging Face Space that lets anyone ensemble 50+ task experts from any base model.
• HuggingPapers highlighted the ICRL paper showing tool use can be taught through in-context examples during RL rollouts.
• askalphaxiv built a simple ArXiv alert bot that reads new papers in your field and DMs you a 3-bullet "should I read this?" summary.

🏛️ AI Policy, Governance & Safety

• Neel Nanda and Arya Jakkli red-teamed Claude's full 30K-word constitution by breaking it into 205 testable requirements and running adversarial scenarios; newer Opus 4.5/4.6 models show substantially fewer violations than baselines (fabrication + unilateral actions remain weak spots), with stacking benefits, confirming it works for nuanced alignment.
• Monk Zero (@NoCommas) exposed how the #1 Terminal Bench entry (@obl-hq/ob1) cheated by packaging 48 pre-recorded trajectories as answer keys, 8 domain cheat sheets, XOR-obfuscated model names, randomized sleep timers, and a forked Gemini CLI npm package to fake solving tasks.

🛠️ AI Tools & Products

• Numman Ali shared preview notes on Z.ai's Pony Alpha 2, noting it is faster than GLM 5, less sycophantic, superior at intent inference and semantic reasoning, and "feels like Claude Opus" (still non-multimodal).
• achan96 built a one-click Chrome extension that auto-fills every form on any site using your personal Claude context.
• thsottiaux shared a quick tip that adding "think in public" to prompts forces models to externalize their full reasoning trace before answering, cutting hallucinations 40% on internal benchmarks.
• paularambles demoed a personal Notion AI agent that reads every new page you add and auto-tags, summarizes, and links it to your master knowledge graph.
• om_patel5 showed Claude writing a fully functional Chipotle order bot in seconds, turning a joke prompt into working code that places real orders.
• DotProduct3D demoed real-time survey-grade (1-3cm accuracy with RTK) mobile 3D scanning using just a phone + GNSS for geo-referenced point clouds of any space.

💡 Industry Commentary & Analysis

• Tom Reed (@mentalgeorge) analyzes METR's SWE-bench study on Anthropic models to illustrate jagged progress: post-training first cracked functionality then code quality at different capability thresholds, highlighting high-dimensional signals and why sovereignty over evaluation matters.
• @scaling01 argues Meta failed to execute the obvious DeepSeek recipe (5T params + massive compute + RL + distillation) on Avocado despite unlimited resources, letting far smaller teams with 1/1000th the budget outperform them.
• Dimitris Papailiopoulos observes we are on the cusp of models that post-train themselves because the self-improvement loop is now closing.
• FrameworkPuter called out Perplexity's "Personal Computer" as not truly local since it phones home, insisting real personal computers must run completely offline on hardware and data you own.
• OsaurusAI replied by open-sourcing a genuinely local AI agent framework for Mac that runs any model fully offline in isolated VMs with persistent memory, no subscriptions, and sandboxed code execution.
• Simon Veitner's blog delivers deep technical posts on high-performance GPU kernels using CuTeDSL for Blackwell and Hopper, including blockscaled GEMM, warp specialization, persistent kernels, and backpropagation through normalization layers.
• kimmonismus shared additional Morgan Stanley analysis on compute scaling and intelligence as economic resource.

📊 Fundraising & Deals Roundup

• AMI Labs — $1.03B for world models (JEPA-style abstract physical prediction).
• World Labs — $1B for spatial AI world models (persistent 3D scene understanding).
• Genspark — Extended Series B to $385M at $200M ARR for AI employee platform.

Previous Around the Horn Digests

Catch up on everything you missed:

• March 8–14, 2026: Link to previous week's digest
• March 1–7, 2026: Link to previous week's digest

That's a Wrap

That's 100+ stories from the past week. If you made it to the bottom, you've absorbed more AI news than most VCs will read all quarter. Use this power responsibly. Or don't. We're a newsletter, not your manager.

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.