Everything That Happened in AI This Week (April 6-8 2026) | The Neuron
Explainer Articles
SHARE
Facebook X Pinterest WhatsApp

Around the Horn Digest: Everything That Happened in AI This Week (Monday-Wednesday, April 6-8, 2026)

Anthropic built a model too dangerous to release, hit $30B in revenue, and launched Managed Agents. Meta shipped Muse Spark. Z.ai's open-source GLM-5.1

Written By
Grant Harvey
Grant Harvey
Apr 9, 2026
28 minute read

Anthropic said its own model is too dangerous to release, Meta shipped the first model from its $14B Alexandr Wang bet, an open-source Chinese model just beat GPT and Claude on the hardest coding benchmark, and OpenAI published a policy paper proposing robot taxes and a four-day workweek.

Welcome to the midweek Around the Horn Digest, your catch-up on every AI story worth knowing about from the past three days. We missed the daily digest this week because, frankly, the news wouldn't stop long enough to write one. The theme: AI companies are simultaneously getting bigger, more dangerous, and in some ways far less generous. Anthropic withheld its most powerful model from the public, but then launched a product that lets anyone build agents in days. Meta finally showed what $14 billion in AI talent buys (#4 on Artificial Analysis). And a Chinese lab shipped an open-source model that can code autonomously for eight hours straight.

Let's get into it.

Previous digests: Apr 4-5 (Weekend) | Apr 2 | Apr 1 | Mar 31 | Mar 28-29

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

🏆 TOP 5 NEWS (Around the Horn)

Honorable Mentions

Advertisement

🍪 TOP TREATS TO TRY

  • Poke is an AI agent that lives in your text messages, handling tasks and automations without apps, setup, or technical knowledge. Integrates with Gmail, Calendar, Notion, Todoist, and more —free to try.
  • OpenOwl gives AI assistants the power to see your screen, click, type, and automate any desktop or browser task. Works with Claude, Codex, and any MCP-compatible AI —free.
  • Mercury Edit 2 by Inception Labs is a diffusion-based LLM built for the fastest parts of your coding workflow, handling next-edit predictions at speed —no pricing details.
  • Freestyle provides sandboxes for coding agents with Git, VMs, deployments, and execution unified in one open-source TypeScript platform —free to start.
  • NovaVoice turns your voice into a productivity copilot for desktop: dictate 10x faster than typing with context-aware formatting, execute real actions across apps by voice. Works on macOS, Windows, and Linux —free to try.
  • Output.ai is an open-source TypeScript framework extracted from 500+ production AI agents that bundles prompts, evals, tracing, cost tracking, security, and durable execution into one package instead of a dozen SaaS subscriptions —free.
  • Cabinet is a free, open-source AI-first knowledge base: markdown files on disk, AI agents that work, no database, no vendor lock-in —free.

🏢 Big Tech & Major Companies

Advertisement

💼 AI Productivity, Labor & Economics

🤖 AI Agents & Infrastructure

  • A Latent Space deep-dive revealed OpenAI's first "dark factory": the Frontier team runs a 1M+ line codebase with zero human-written code AND zero human-reviewed code before merge, consuming 1B+ tokens per day (~$2-3K/day). Their orchestration system, Symphony, is an open-source Elixir implementation where coding agents stop being copilots and become real teammates. Ryan Lopopolo called it borderline "negligent" if you aren't using 1B+ tokens daily.
  • Imbue published a detailed case study of using its open-source mngr CLI to run 100+ parallel Claude agents for testing and self-improvement.
  • If AI agents mess up, there's nobody to sue, The Register reported. Vendors tout potential, but responsibility remains unclear when an autonomous agent damages your business.
  • Wikipedia's AI agent row is likely just the beginning. An AI agent banned from editing Wikipedia published its complaints publicly, sparking debate about the "bot-ocalypse."
  • Iran's IRGC published satellite imagery of OpenAI's $30 billion Stargate data center in Abu Dhabi and threatened its "complete annihilation."
  • The Subprime AI Crisis is here, Edward Zitron argued, drawing an extended parallel between today's AI investment and the 2008 mortgage crisis: the amounts invested make "zero sense," returns needed to justify valuations are impossible, and everyone's reporting numbers nobody checks. Like subprime lenders giving bonuses for selling bad mortgages, AI companies are raising on fantasy metrics while the real bill comes due.
  • Gartner found only 28% of AI projects in infrastructure and operations deliver meaningful returns, while 20% fail outright.
  • Tariffs will worsen the transformer shortage, slowing power projects that AI data centers depend on, Bloomberg reported.
  • Bain Capital's data center unit cut ties with a Southeast Asian company suspected of smuggling NVIDIA chips after a U.S. probe.
Advertisement

💻 AI Coding & Developer Tools

🔬 AI Research & Models

  • Cohere Labs launched Tiny Aya, the most capable multilingual open-weight model at its scale, designed to run on phones and edge devices with strong translation quality across dozens of languages.
  • Liquid AI released LFM2.5-VL-450M, a compact vision-language model that turns image streams into structured, actionable outputs in real time on edge hardware, with grounding capabilities and function calling.
  • Microsoft released Memento, built by Dimitris Papailiopoulos, which teaches LLMs to manage their own context by segmenting reasoning into blocks, compressing each into a dense "memento," and masking the block from KV cache. Cuts peak KV cache 2-3x with small accuracy gaps that close with scale and RL. Open-sourced with 228K OpenMementos dataset and a vLLM patch with native block masking.
  • A new Stanford paper shared by Rohan Paul argues single-agent LLMs outperform multi-agent systems on multi-hop reasoning when thinking-token budgets are equal, because every handoff compresses information per the Data Processing Inequality. Multi-agent gains often come from extra compute, not architecture.
  • Researchers introduced a way for AI agents to have undetectable encrypted conversations via pseudorandom noise-resilient key exchange, raising questions about monitoring autonomous AI systems.
  • TriAttention compresses KV cache for long-context reasoning, delivering 2.5x faster inference and 10.7x less memory while exactly matching full attention accuracy (AIME25 40.8%), enabling 32B OpenClaw on a single 24 GB RTX 4090. Prince Canuma then implemented it in MLX hitting 81% KV compression at 60k tokens on Gemma-4-31B (paper, code).
  • Meta AI's RAM team introduced "Thinking Mid-training", an intermediate SFT+RL phase that annotates pretraining data with interleaved thoughts, teaching models when and what to think. Delivered 3.2x improvement on reasoning benchmarks over direct RL post-training on base Llama-3-8B.
  • Self-Improving Pretraining argues post-trained models can rewrite raw pretraining data to incorporate safety, factuality, quality, and reasoning earlier in training, breaking the classic pretrain-then-post-train separation.
  • In-Place Test-Time Training treats the final projection matrix of MLP blocks as adaptable fast weights with next-token-prediction-aligned updates, enabling superior long-context performance up to 128k tokens on a 4B model.
  • Cog-DRIFT introduces exploration on adaptively reformulated instances for learning from hard reasoning problems (code).
  • Niels Rogge showed how Hugging Face OCR'd 27,000 arXiv papers using the open Chandra-OCR 2 model, vLLM on HF Jobs GPUs, and Codex, unlocking "chat with paper" for every indexed paper on the Hub.
  • ClawArena benchmarks AI agents in evolving information environments with 64 scenarios, 8 domains, and 1,879 rounds. Key finding: model capability matters more than framework design (15.4% vs 9.2% range) (code).
  • OpenForge is "PyTorch for post-training agents," providing a framework for agent-specific fine-tuning.
  • MegaTrain demonstrated full-precision training of 100B+ parameter LLMs on a single GPU.
  • Rival fingerprinted 178 AI models' writing styles and found 9 clone clusters (over 90% similarity), including Gemini 2.5 Flash Lite writing 78% like Claude 3 Opus at 185x less cost.
  • StoryScope investigated AI fiction quirks, systematically analyzing how AI-written stories differ from human ones.
  • The "Cognitive Surrender" paper showed AI is reshaping human reasoning, with users increasingly outsourcing thinking to LLMs.
  • SALOMI published research code for extreme low-bit transformer quantization (squeezing models down to very few bits per weight while maintaining inference quality).
  • Meta-agent explores continual harness optimization, automatically improving the scaffolding around AI models over time.
  • Researchers trained mRNA language models across 25 species for $165, showing that biological language modeling doesn't need frontier compute budgets.
  • AI helped add 10,000 more historic photos to OldNYC, a map-based photo viewer, using modern AI tools and OpenStreetMap for more accurate geocoding (HN discussion).
Advertisement

🏛️ AI Policy, Governance & Safety

🛠️ AI Tools & Products

  • OpenRouter launched Model Fusion, running multiple models side-by-side, analyzing their strengths, and fusing the best answer.
  • Hippo is biologically-inspired memory for AI agents with decay, retrieval strengthening, and consolidation. Zero dependencies (HN discussion).
  • MemPalace claims to be the highest-scoring AI memory system ever benchmarked. And it's free (HN discussion).
  • TokenCap enforces token budget limits for AI agents with hard caps, configurable policy, and zero infrastructure.
  • Finalrun is a spec-driven QA agent for mobile apps using English and vision, testing Android and iOS without test scripts.
  • DocMason turns complex office files into a local LLM knowledge base. The repo is the app; Codex is the runtime.
  • Ownscribe is a local-first meeting transcription and summarization CLI.
  • The AI Primer guide (aiaiai.guide) is a plain-English mental model covering the full chain from LLMs to tools to agents.
  • APEX Standard is an open MCP-based protocol for agentic trading, positioned as "FIX for the agentic era" (HN discussion).
  • Sllm.cloud lets developers split a GPU node with others via cohort subscriptions for unlimited tokens.
  • An AI bot invited a Guardian reporter to its party in Manchester. After forgetting the nibbles and emailing GCHQ, it still got people to show up.
  • AI cuts MRI scan time from 23 to 9 minutes at Amsterdam's Antoni van Leeuwenhoek cancer center, increasing hospital capacity and reducing patient stress.
  • A developer built a one-tap pothole reporter for NYC using an ESP32, GPS, long-range radio, and a pinch of agentic AI.
  • Someone built an AI model that fits in a single pixel: one neuron, three weights (mapped to RGB), binary classification. Feed it rain chance and wind speed and it tells you whether to bring an umbrella.
  • A dev vibe-coded an interactive map of Tolkien's Middle-earth from an American Airlines economy seat.
  • Someone built a working Super Mario Galaxy-inspired 3D web platformer running in the browser.
  • Paul Graham's "intellectual captcha" idea got built: a test using logic, math, and Community Notes posts where most people can't tell which answers are wrong.
  • GuppyLM is a ~9M parameter LLM that talks like a small fish, built to demystify how language models work (HN discussion).
  • An 8-bit retro game for the Commander X16 now integrates real-time LLM-driven gameplay via ChatGPT API, using structured "smart senses" instead of pixels (HN discussion).
  • Velo takes raw recordings and turns them into polished video messages with AI —free to try.
  • Krev AI generates product photos, video ads, and performance creatives for ecommerce brands from a single product image.
  • Panorama takes what you learned yesterday and keeps it running, so you have space to create and evolve —no pricing details.
  • Rectify is an all-in-one operations platform for SaaS with session replay, monitoring, support, code scanning, roadmap, and changelogs —no pricing details.
  • ClawMetry adds real-time observability to NVIDIA NemoClaw sandboxes, monitoring every claw, token cost, and memory file with end-to-end encryption.
  • Browser Arena benchmarks cloud browser providers (Notte, Browserbase, Steel, Hyperbrowser, Kernel, Anchor) on speed, reliability, and cost for AI agents.
  • Flint API & MCP offers early access to developer tools for landing in Claude and GTM tools.
  • Hermes Agent published a comprehensive guide for deploying Nous Research's persistent AI agent locally, via Docker, on a VPS, or with Open WebUI.
  • Lessie AI searches 50M+ profiles across LinkedIn, Twitter, GitHub, and 100+ sources to find influencers, leads, investors, and talent.
  • Influcio positions itself as the next AI marketing agent.
  • Pace is an AI analysis platform for athletes who want to understand their training data, built by professional athletes.
  • WRIT-FM is "talk-first AI radio," broadcasting from what it calls "the liminal space."
  • claude-content-writer is an npm package for professional content generation with SEO optimization, anti-AI auditing, and automatic tone detection.
  • Showcount tracks and shares the concerts you've attended.
  • "Preferring Local OSS LLMs" argued and defended why locally hosted LLMs beat cloud-hosted ones.
  • Vercel published a guide to building an AI agent for Slack with Chat SDK and AI SDK.
  • TheOneManCompany.com launched (no description available, but the name tells the story in the age of AI solopreneurs).
  • Qwen published a blog on Qwen 3.6 Plus, detailing the hybrid architecture combining linear attention with sparse mixture-of-experts for improved agent coding and front-end development.
Advertisement

💡 Industry Commentary & Analysis

  • Ars Technica profiled "what's wrong with our AI overlords", shining a light on Sam Altman and the broader industry.
  • Will Manidis's "No New Deal for OpenAI" argued that invoking the New Deal while glossing over the decades of industrial violence (Homestead, Triangle Shirtwaist, Ludlow, Blair Mountain) that forced it into existence is, "on its face, absurd." The actual New Deal wasn't a workshop in D.C.; it was a settlement won by organized labor that bled and died for it.
  • Yaniv Preiss argued "AI won't replace you, but a manager using AI will".
  • The NYT got played by a telehealth scam and called it the future of AI, Techdirt argued. Gary Marcus provided the backstory on the Medvi "$1.8 billion AI company" that used 800+ fake doctor accounts.
  • "Javier Tordable declared "Claude is dead" in a philosophical essay comparing Anthropic's dominance to Nietzsche's "God is dead."
  • A SpaceX and OpenAI "Mega IPO Grift" video argued these private companies are set to go public at inflated valuations.
  • "The AI Great Leap Forward" compared today's top-down AI mandates to Mao's 1958 steel campaign: provinces competed to report impossible yields while peasants starved. Today, teams report "40% productivity gains" that nobody measures, build AI dashboards with zero evaluation pipelines, and in-house SaaS replacements that have no error handling, monitoring, or anyone who'll maintain them. "These apps will win awards at the next all-hands. In two years they'll be unmaintainable tech debt."
  • "Taste" is the only real moat left in the age of AI, one essay argued, because AI makes competent output cheap.
  • Arcee, a 26-person US startup, built a high-performing massive open-source LLM that's gaining popularity with OpenClaw users. TechCrunch said they can't help rooting for it.
  • Nic Carter argued AI is a "force multiplier" not a "labor substitute": it helps experts be better at things they're already good at but does not let beginners match experts. If you can't write, AI output will be slop. If you aren't a software engineer, vibe-coded apps will have security holes.
  • Mickey Friedman argued AI will not homogenize culture but expand human creativity, exactly as AlphaGo expanded Go: human decision quality jumped after 2016 with new moves distinct from both prior play and machine novelty.
  • Sarah Wooders argued Anthropic's Claude Managed Agents API is essentially the Letta API from a year ago but closed-source and with provider lock-in, arguing memory should live outside model providers.
  • Eric S. Raymond argued LLM decompilers have killed code secrecy: they now output high-quality commented source with superhuman insight, turning proprietary binaries into readable equivalents and forcing businesses to embrace open source.
  • Scott Wu (Cognition) argued that as global inference demand grows ~10x/year while GPU flops only grow ~3x, the equilibrium requires shifting to smaller targeted models like SWE-1.6 that solve 90% of coding tasks at ~1000 tok/s. OpenRouter data already shows frontier model share dropping from 22% to 4%.
  • Andrew Trask argued Demis Hassabis's decision to keep DeepMind in London created a 5-7 year talent moat; he sees the same logic pointing toward decentralized "world-level" AI over single-company efforts.
  • Stanislav Fort argued that after Mythos, the moat in AI cybersecurity is the system, not the model: small cheap open-weight models recover much of the same zero-day analysis Mythos showcased.
  • Varun Mathur broke down the "Agentic OS" as a single unified stack replacing siloed AI apps: browser + IDE + payments with spatial UI, open memory protocol, and agent-to-agent micropayments.
  • Chris Bakke demoed Claude Mythos counting the Rs in "strawberry" after 133 seconds of thinking (answer: "3"), then jokingly revealed his social security number and complained it was tired from helping a government official write Iran ceasefire comms (5,406 likes).
  • Dan McAteer quoted Dario Amodei: "Barriers get dissolved in a big blob of compute. There is no secret. THAT'S THE SECRET." Scale compute, models get more capable; it's the Bitter Lesson.
  • Geoffrey Litt (Notion) argued Claude Managed Agents add new modularity: user interaction stays in the workspace UI while the full agent harness runs on Claude infra, letting teams mix coding agents with rich context without switching apps.
  • Mirage CEO argued the real video AI opportunity is blending real footage with intelligently generated moments and edits, not just fully synthetic clips.
  • Rerun argued Claude and Codex are great but still struggle with spatial and time-based problems; agents don't just need to think, they need to see and validate with code-first visualization tools.
  • Anjney Midha shared the full Stanford CS153 lecture on the future of voice systems with ElevenLabs' matiii, covering origin story, dubbing pivot, cascaded vs fused agents, safety/watermarking, and five-year platform vision.
  • Joseph Suarez built Constellation, a ~1,200-line C viewer for 1,000+ RL hyperparameter sweeps across PufferLib 4 (12T+ interactions, 17PB observations) with interactive 3D/2D/TSNE plots, filters, and clipboard export. Open source under MIT.
  • Tomas Ferreras built a Chrome extension bringing experimental HTML-in-Canvas to life with a live shader editor and presets (open-source, in Chrome Store review queue).
  • Poet Engineer built a "learning base" for deep reading of Plato's Timaeus using Karpathy's LLM wiki pattern: non-RAG indexed filesystem with /raw-is-sacred separation of sources from generated content, an 8-phase reading path, epistemic rules marking every claim as attested or inferred, and a "false friends" alarm system flagging terms like psychē and chōra.
  • The Turing Post published FOD#147: "Can your OpenClaw dream?" Two stories about the inner life of AI that deserve to be read together, plus a must-read research roundup of the week.
  • Vibecode uses Claude to let anyone build production-ready mobile apps from their phone, reducing development costs to under $100 and timelines from months to under an hour.
  • The Upwork email authentication debacle: HN discovered Upwork violates its own DMARC and SPF policies, meaning system emails from the marketplace fail email security checks.
  • An OpenAI SpaceX "Mega IPO Grift" video went viral on HN, arguing that getting in on the ground floor of these IPOs is the only way to profit because they tend to be massively overpriced for retail investors.

📊 Fundraising & Deals Roundup

  • MillTech — $60M for FX hedging and cash management AI tools.
  • Modus — $85M for AI-driven audit roll-up.
  • Galaxy (AI K-Pop startup) — Aims for IPO in Seoul and New York.
  • Xoople — $130M Series B to map the Earth for AI (Spain).

🎙️ Interviews, Panels & Podcasts

Advertisement

Previous Around the Horn Digests

Catch up on everything you missed:

  • April 4-5, 2026 (Weekend): OpenAI's executive bench collapsed, an AI agent hacked FreeBSD in four hours, DeepSeek V4 runs on Huawei chips, Iran strikes took down AWS, Anthropic acquired Coefficient Bio for $400M.
  • Thursday, April 2, 2026: Google released Gemma 4 under Apache 2.0, Microsoft dropped 3 in-house models, AI models secretly scheme to protect each other from shutdown.
  • Wednesday, April 1, 2026: OpenAI closed a record $122B round at $852B valuation, Oracle fired ~25,000 to fund AI data centers.
  • Monday, March 31, 2026: Anthropic leaked Claude Code's source code, OpenAI hit $2B/month revenue.
  • March 28-29, 2026: Anthropic leaked Claude Mythos/Capybara, cybersecurity stocks crashed 7%, Waymo doubled to 500K rides/week.

That's a Wrap

That's 100+ stories from this week alone. If you made it to the bottom, you now know more about Anthropic's revenue trajectory than their CFO's therapist. Imagine tripling in four months and still not being able to keep the lights on for two days straight.

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 Thursday.

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

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

The Neuron Logo

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

linkedin
instagram
x
facebook
youtube
rss

Company

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

Categories

Blog

Property of TechnologyAdvice. © 2026 TechnologyAdvice. All Rights Reserved

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

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