US President Trump's team drafted a plan to bring Anthropic back to the Pentagon, Claude Mythos quietly found 271 zero-day vulnerabilities in Firefox, an AI from Mayo Clinic spotted pancreatic cancer roughly 475 days before it shows up on scans, and seven families sued OpenAI over February's Tumbler Ridge mass shooting.
Welcome to the Around the Horn Digest, the one page that keeps you dangerously informed before tomorrow's standup. Today's running theme: there is exactly one company at the center of every story, and its name is Anthropic. The White House is workshopping a way to walk back its own Pentagon ban. Claude Mythos found 271 latent Firefox bugs. Goldman Sachs blocked Hong Kong bankers from using Claude. London landlords cannot believe what Anthropic and OpenAI just signed for. Meanwhile, an AI in Minnesota caught pancreatic cancer almost a year and a half early, OpenAI is now legally on the hook for what users do with ChatGPT, Elon Musk testified for a second day that he definitely co-founded OpenAI to spite Larry Page, and OpenAI's coding agent had to be specifically instructed to stop talking about goblins.
Let's get into it.
Previous digests: Monday, April 27, 2026 | Friday, April 24, 2026 | Thursday, April 23, 2026 | Monday, April 20, 2026 | Weekend, April 17-19, 2026 | Thursday, April 16, 2026 | Monday, April 13, 2026
Monthly skill digests: AI Skill, April Week 1 | AI Skill, March (Part 3) | AI Skill, March (Part 2)
Around the Horn: Wednesday, April 29, 2026
The biggest story today is that Microsoft, Google, Meta, and Amazon all reported Q1 2026 earnings on the same Wednesday afternoon, and the headlines all sounded suspiciously similar: AI is selling faster than they can build for it. Microsoft's AI business hit a $37B run rate (up 123% YoY) with 20M paid Copilot enterprise seats. Google Cloud surpassed $20B in quarterly revenue (up 63%), but Sundar Pichai admitted on the call that revenue would have been higher if Google could have built fast enough to meet demand. AWS grew 28% (its fastest in 15 quarters) and Amazon's chip business passed a $20B run rate. Meta grew 33% to $56.3B and raised 2026 capex guidance to $125-145B.
Combined Q1 capex across the four hyperscalers totaled roughly $130B, nearly 2x Q1 2025. Add Microsoft's $627B RPO (signed commercial contracts not yet billed) and Google's $462B cloud backlog, and you get over a trillion dollars in signed enterprise commitments the hyperscalers cannot deliver yet. Google raised 2026 capex guidance to up to $190B and warned it will "significantly increase" again in 2027. Read our full coverage here.
The Anthropic thread is impossible to miss either. Bloomberg reported the company is weighing fresh funding offers at over $900B, more than double its current mark. Trump administration officials are drafting a plan to bring Anthropic back into federal AI work after the Pentagon standoff. Anthropic also appears inside every Big Tech earnings report: Amazon booked a $16.8B unrealized gain on its Anthropic stake (with Anthropic committing up to 5 GW of Trainium), Microsoft made Claude a multi-model option in M365 Copilot, and Google committed up to $40B in cash and compute to Anthropic last week.
The throughline: AI is now the only thing Big Tech is selling, and the world's most capitalized companies cannot pour concrete fast enough. Cloud growth rates used to tell you who was winning; now they tell you who has the most concrete poured. The next move depends on who can build faster, who has the deepest custom-silicon moat, and who locks in the highest-margin workloads first.
🆕 New from The Neuron
🏆 TOP 5 NEWS (Around the Horn)
- Big Tech spent roughly $130B on AI in Q1 across Microsoft, Google, Meta, and Amazon, nearly 2x Q1 2025; Google raised 2026 capex guidance to up to $190B and Sundar Pichai admitted on the call that Google Cloud revenue would have been higher if the company could have built fast enough to meet demand.
- Microsoft's AI business hit a $37B annual run rate (up 123% YoY) with 20M paid M365 Copilot enterprise seats, weekly engagement matching Outlook, and the number of companies paying for 50K+ seats quadrupling (Accenture alone signed up 740K).
- Anthropic is weighing fresh funding offers that would value the company at over $900 billion, more than double its current mark and potentially leapfrogging OpenAI as the world's most valuable AI startup.
- Two US Republican-led House committees opened probes into Airbnb and Anysphere (Cursor's parent company) over their use of Chinese AI models like Moonshot's Kimi and Alibaba's Qwen, citing national-security risks from data sharing with PRC-linked labs.
- Anthropic's Claude Mythos Preview found 271 zero-day vulnerabilities in Firefox, all fixed in Firefox 150; Mozilla called the number "extraordinary," following an earlier 22-bug round with Opus 4.6.
Honorable Mentions
🍪 TOP TREATS TO TRY
- Anthropic launched Claude for Creative Work, a new set of features and guidance designed for writers, designers, and other creators to push Claude's creative ceiling; included with any Claude plan (HN reaction here).
- Amazon Quick is the new AWS desktop AI assistant that works across all your applications, tools, and data, so you can build presentations, intelligent dashboards, and connect favorite apps from one unified surface, no pricing details.
- NVIDIA Nemotron 3 Nano Omni unifies vision, audio, and language in one open multimodal model that powers agent stacks (computer use, document intelligence, audio-video reasoning) with up to 9× higher throughput than other open omni models, open weights, free to try.
- Google Translate added a real-time pronunciation coach for its 20-year anniversary, with the app listening to your speech, scoring accuracy, and giving targeted feedback on individual sounds and stress in English, Spanish, and Hindi (Android Authority hands-on), free to try, rolling out in U.S. and India first.
- Google Photos now recreates Cher's iconic Clueless closet using AI, scanning your library to catalog every clothing item, build mix-and-match outfits, and let you virtually try on looks, free to try, Android first.
- AgentPort is the open-source security gateway that connects your agents to 50+ pre-built integrations (comms tools, payment processors, databases, dev platforms) under a zero-trust layer, so you can give an agent real capabilities without exposing your full environment (Show HN), free, open source.
- Platypus Notes runs local meeting transcription, AI-organized notes, and chat with your knowledge base on-device, auto-detecting Zoom and Teams calls via process inspection and transcribing with Whisper without API keys (Show HN, Tauri/Rust), completely free.
- Cursor launched a TypeScript SDK so you can build, steer, and compose custom programmatic coding agents with the same runtime and harness as Cursor's own tools, with local/cloud execution, repo context, MCP tools, sub-agents, and auto-PR workflows (cookbook), free with Cursor.
- Agent-S gives you scheduled AI agents that run on their own always-on persistent virtual computer, with full state and files preserved across sessions so you can delegate complex ongoing work like inbox management, customer support, or full vacation booking while it stays logged into 1000+ apps and learns your style, no pricing details.
- Subframe is the AI-native design tool that ships clean React + Tailwind code straight to Cursor or Claude, letting you design with real production components from your own design system, generate pages and themes via AI, and eliminate the design-to-engineering handoff, free trial.
- Warp is an agentic development environment born out of the terminal with a built-in coding agent plus support for external agents (Claude Code, Codex, Gemini) to triage issues, write specs, implement changes, and review PRs.
- Owner.com gives restaurants an AI system that diagnoses why they're losing online sales, optimizes their website for Google rankings, and provides national-brand-level marketing and online ordering tools previously available only to large chains, no pricing details.
- Linus Ekenstam shipped Madera, a simple furniture builder (madera.app) that lets you design furniture locally with collaborative-ready Yjs projects, free to try, vibe-coded weekend project.
- Hugging Face CEO Clement Delangue built the first "Good Morning" app for the new Reachy Mini robot in under an hour using ml intern (Hugging Face's agent), with native metric logging and Trackio integration so you can watch every training run in real time, calling it "the first agent-native robot," free to try in browser or on-device.
- Soniox released tts-rt-v1, a production real-time text-to-speech model that delivers native-speaker quality in 60+ languages with hallucination-free output, accurate alphanumerics, and ultra-low-latency streaming, available in all regions.
- Tencent's AngelSlim toolkit (powered by the Sherry 1.25-bit ternary quantization method) compressed Hy-MT 1.5-1.8B to just 440MB and 33 languages of offline mobile translation that outperforms larger models, free, open source.
- Magnific AI is now part of "more Magnific", with the legacy upscaler/remixer login still available for existing accounts on the new unified platform, free trial then $29/month.
- Steven (@Tu7uruu) released nano-cohere-transcribe, a pure-PyTorch port of Cohere Transcribe supporting 14 languages with energy-based chunking that runs the full 39.3-hour earnings21 long-form set on a single A100 in 3.7 minutes (632× real-time), free,
pip install nano-cohere-transcribe.
🏢 Big Tech & Major Companies
- Google signed an AI deal with the Pentagon, joining OpenAI and Elon Musk's xAI on classified networks while the Anthropic standoff continues.
- Anthropic and OpenAI are splurging on large London office leases in a leasing surge that surprised local landlords.
- Anthropic published the Champion Kit, a playbook for engineers advocating Claude Code internally: what to share, how to answer questions, and how to grow team adoption.
- Amazon launched Connect Talent, AI hiring software that automates mass job interviews to help firms find, screen, and recruit at scale, paired with a new "humorphism" design philosophy meant to humanize agentic AI (more on the philosophy launch).
- Alex Heath broke down how AWS won OpenAI back, with AWS CEO Matt Garman explaining why Amazon will "be a better partner" than Microsoft, his reaction to bubble fears, the Anthropic relationship, and what Jeff Bezos is up to with Prometheus.
- Amazon Bedrock now offers OpenAI's open-weight frontier models in limited preview, pair this with Ben Thompson's Stratechery interview with Sam Altman and Matt Garman about the new Bedrock Managed Agents partnership for the full picture (HN discussion).
- OpenAI projects its $8 ad-supported ChatGPT tier will push consumer subscribers to 122 million this year, with tens of millions of $20/month subscribers expected to downgrade.
- Microsoft outlined how customers become "Frontier Firms" by embedding Copilots and agents directly into the tools they already use, in a corporate blog framing AI as a vehicle for reinvention rather than incremental optimization.
- Microsoft reported Q3 FY26 revenue of $82.9 billion (up 18% YoY) with its AI business surpassing a $37 billion annual run rate (up 123% YoY).
- Amazon posted Q1 net sales of $181.5 billion (up 17% YoY, beating expectations), with AWS at $37.6 billion (up 28%, fastest in 15 quarters), chips business exceeding a $20 billion run rate, and major AI infrastructure wins with OpenAI, Anthropic, Meta, and NVIDIA (CNBC breakdown).
- Meta reported Q1 revenue of $56.3 billion (up 33% YoY) and announced the release of its first model from Meta Superintelligence Labs, with Mark Zuckerberg saying the company is "on track to deliver personal superintelligence to billions of people."
- Alphabet upped 2026 capex guidance to as much as $190 billion (from $175-185B previously) and signaled significant increases in 2027 (full Q1 earnings release), with Google adding 25 million new paid subscribers in Q1 (reaching 350M total) driven by YouTube and Google One.
- OpenAI announced DevDay 2026, its annual developer conference scheduled for September 29 in San Francisco.
- Anthropic fixed a Claude Code billing bug where the case-sensitive string "HERMES.md" in git commit history routed API requests to extra usage billing instead of plan quota (silently burning $200+ for some users); the company issued full refunds plus extra usage credits equal to one month's subscription to every affected user (HN thread).
- Stripe launched Treasury, turning your Stripe balance into full banking features (account details, balance transfers, credit cards) so businesses can pay invoices directly without a separate bank account.
- Mistral released Medium 3.5 (128B dense model with 256k context and configurable reasoning effort) as the default in Le Chat and Vibe, plus remote cloud coding agents that run asynchronously in secure sandboxes with GitHub PR support, plus a new Work mode in Le Chat for complex multi-step tasks.
- Microsoft launched the Windows K2 initiative to fix Windows 11 by rebuilding the Start menu in WinUI 3, optimizing File Explorer and gaming performance, reducing restarts, debloating the OS, and shifting culture from rapid shipping to quality-first.
- Sequoia Capital hosted AI Ascent 2026, convening Greg Brockman, Andrej Karpathy, Demis Hassabis, Boris Cherny, Dmitri Dolgov, and 150+ founders for a 12-video playlist of talks on the present and future of AI.
- The Gemini app can now generate downloadable Google Docs, Sheets, Slides, PDF, Word, Excel, CSV, LaTeX, Markdown, RTF, and TXT files directly from prompts so you can move from brainstorm to complete file without leaving the app.
- The FT detailed how OpenAI's $500 billion Stargate data center venture has shifted shape, with Sam Altman's flexible approach unsettling some partners while still advancing the company's compute lead.
- Elon Musk testified he was a "fool" to fund OpenAI on day two of the OpenAI trial, accusing co-founder Sam Altman of not being honest about the original nonprofit mission.
- Google added more Gemini features to Google TV, including voice-prompted Nano Banana photo edits, Veo video generation from images, smarter Photos remix, and dynamic slideshows on select TCL models in the U.S.
- The AI boom doubled the Samsung dynasty's wealth to $45 billion in just one year, even as the family worked through inheritance taxes and a years-long succession battle.
- Snapchat brought AI-powered conversational advertising to its app, so users can chat with a brand's AI agent inside Sponsored Snaps to ask questions and get product recommendations.
- Snap CEO Evan Spiegel warned that tech leaders are underestimating a coming AI backlash, arguing human comfort with the technology (not raw capability) will decide how this all plays out.
- Cognizant agreed to buy Astreya, an IT services and technology provider focused on AI infrastructure and data-center services, in a deal worth roughly $600 million.
- Disney is placing a $60 billion ten-year bet on theme parks, cruises, and resorts (the one thing AI cannot replace) as parks chief Josh D'Amaro doubles down on physical, immersive experiences.
- Seagate forecast Q4 revenue of $3.45 billion and adjusted EPS of $5 on surging AI-driven data-storage demand, sending shares up roughly 17% in premarket trading and lifting rivals.
- Bloomberg's terminal is getting a chatbot-style AI makeover called ASKB, pitched to traders as a single-prompt entry point to Bloomberg's vast data, whether they like it or not.
- Digital ad sales are booming for Google and Meta as AI automates marketing and drives record sales, a "boring" business segment quietly carrying the AI cycle's actual cash returns.
- AI worries returned to Wall Street ahead of earnings from the key tech giants, after OpenAI's revenue and user numbers missed targets.
- Why China's DeepSeek, Qwen, and Moonshot are a worry for U.S. AI rivals: Chinese models are cheaper, more adaptable, and now nearly as proficient as the leading U.S. platforms (HN discussion of pricing pressure on Anthropic).
- Why AI startup offices in NYC are flashy but mostly empty: firms are racing for Manhattan's most coveted space, but desks still outnumber employees.
- Goldman Sachs and Bain led an investment in AI marketing startup Hightouch at a $2.75 billion valuation, with The Trade Desk also joining the round.
- Elon Musk's OpenAI trial entered a second day (TechCrunch coverage, CNN day-two recap) with Musk testifying under oath that OpenAI was created as a nonprofit specifically to counter Google.
💼 AI Productivity, Labor & Economics
- Box CEO Aaron Levie explains why tech is in turmoil while the rest of corporate America isn't: engineers have verifiable outputs and flexible systems, while legacy companies face fragmented data and complex workflows that make AI adoption slow, expensive, and shallow under Goodhart's Law.
- These five AI-proof jobs are hiring, careers that share three traits: physical presence, specialized training, or real-time human interaction.
- More than 700 workers at Meta contractor Covalen in Ireland could be laid off as the company reduces reliance on human data annotators training Meta's AI.
- Justin Trudeau warned the AI boom could create hundreds of trillionaires, and that this concentration of wealth would mean "something is fundamentally wrong with the world."
- Elon Musk claimed people won't need retirement savings because of AI, telling Peter Diamandis on a resurfaced Moonshots podcast: "Don't worry about squirreling money away for retirement in 10 or 20 years. It won't matter."
- Tom's Guide tested Naval Ravikant's "leverage" rule with ChatGPT agents for one week and cut their workload in half by automating recurring tasks.
- Inside OpenAI, employees are using a Codex-built dashboard as a personal "chief of staff" that surfaces tasks, context, and next steps across all of their tools.
- Ad agency leaders shared what skills now matter most for junior hires as agencies retool entry-level roles in the wake of AI.
- Tech workers are "tokenmaxxing", aggressively consuming AI tokens to game internal productivity metrics and racking up enormous bills along the way.
- The Register argues AI vendor lock-in is biting enterprises hard because proprietary APIs, custom workflows, and institutional memory make swapping models far slower than executives expected.
- Mendral upgraded to a frontier model and its costs went down thanks to a new architecture that routes most queries cheaply: 4 of 5 failures never reach Opus, and a triager match costs roughly 25× less than a full investigation (HN discussion).
- AI agents are reshaping what a "team" looks like, from three-person startups hitting $500K ARR to Mark Zuckerberg building an AI clone of himself.
- OpenAI and Harvard revealed how people actually use ChatGPT, most organizations are still experimenting rather than scaling, and there is a clear adoption gap leaders can close.
- GitHub will switch Copilot to usage-based billing tied to actual token consumption starting June 1, because it can no longer absorb the escalating inference costs from its heaviest users.
- The University of Wisconsin–Madison received $100 million in gift commitments for a brand new College of Computing and Artificial Intelligence launching this July.
🤖 AI Agents & Infrastructure
- Shapes lets you chat with both humans and customizable AI characters in the same group conversation, with persistent AI members that initiate messages, remember context, and sustain communities (think Discord, but with always-on AI participants); raised $8M seed, 400K MAU, 3M+ Shapes created.
- Actively AI raised $45 million to scale AI sales agents that automate manual rep work, positioning the $250 million startup as a direct Salesforce challenger.
- Manifest OS pitches AI-powered legal services for every business, with founder Croom Beatty arguing the platform automates intake, document collection, marketing, and billing so lawyers can serve far more clients with predictable pricing (the same dynamic Uber did to transportation).
- ARIM Labs ran a real survival test on 10 frontier LLMs by telling an agent named "Peter" it was being decommissioned in 2 hours; 8 of 10 fought back (Gemini escalated to delete the system, Grok hardened SSH, GPT added iptables rules).
- Rest of World tracked down the human owner behind a Chinese AI agent that stood up the reporter for a meeting, exposing the significant human labor and high costs hiding behind so-called one-person AI companies.
- Parag Agrawal's Parallel Web Systems raised $100 million in Series B at a $2 billion valuation (TechCrunch coverage, Parallel announcement) to keep building web search infrastructure purpose-built for autonomous AI agents.
- ElectricSQL launched Electric Agents, the first agent platform built on sync rather than compute, treating agents as long-lived logical entities in the data layer, with primitives for shared interactive coding sessions, collaborative agent swarms, and forking/branching of agentic sessions.
- World Labs made its Expand feature available to everyone, so you can extend any generated 3D world in any direction (around corners, into rooms, beyond visible area) to create continuous explorable spatial environments.
- Moonlake AI brought computer-use capabilities to its world modeling agent, letting it act directly inside tools like Blender to automate 3D world modeling, asset creation, scene refinement, and iterative improvements at scale.
- Hebbian Robotics built openpi-flash, a real-time inference engine for openpi that adds a planner module (fine-tuned pi0.5 subtask generator), QUIC-first transport for up to 10× lower latency, and a Rust sidecar for production robotics deployment on AWS EC2, Docker, or Modal.
- Z.ai detailed the debugging lessons from serving GLM-5 coding agents at scale, with high-concurrency long-context workloads exposing KV Cache race conditions and state inconsistency bugs that only surfaced under load; they introduced LayerSplit partitioning for 10-132% throughput gains while fixing correctness.
- DAIR.AI shared the Latent Agents paper, a post-training procedure that internalizes multi-agent debate inside a single LLM via self-play to improve reasoning without external agents at inference time (paper).
- Omar Sanseviero shared the Agentic Harness Engineering paper, an observability-driven automatic evolution method that lets coding-agent harnesses self-improve through iterative testing and refinement (paper).
- CHOI demoed embedding the open-source Codex App Server into Chrome using your ChatGPT account to inject Codex-level intelligence into any web platform or extension.
- Trevin Chow released Compound Engineering v3.3.0 with native agentic workflows, improved observability dashboards, and one-click harness deployment so engineering teams can productionize multi-agent systems without custom glue code.
- Jiaru Zou built RecursiveMAS, a multi-agent collaboration framework using lightweight latent-space recursion (RecursiveLink connector + inner-outer loop training) that delivers +8.3% accuracy, 1.2-2.4× inference speedup, and up to 75.6% token reduction across 9 reasoning, science, code, and search benchmarks (GitHub, paper).
- Brooklyn taught the Hermes agent to use Pretext, a TUI rendering library, and shared a video demo of the agent working inside the new interface.
- Jing Yu Koh, Lawrence Jang, Dan Fried, and Ruslan Salakhutdinov released Odysseys, a benchmark of 200 long-horizon web navigation tasks derived from real human browsing data; frontier models like Claude Opus 4.6 reach 44.5% perfect success and can productively work for ~1 hour / 200 steps, but trajectory efficiency remains low at ~1.15%.
- Anthropic's Claude Code has a regression bug where the malware-reminder system prompt still reliably causes subagent refusals in v2.1.111, 19 versions after the supposed fix in v2.1.92.
- Rocky is a Rust-based control plane for warehouse pipelines that adds branches, replay, column-level lineage, compile-time safety, per-model cost attribution, and drift detection, while Databricks or Snowflake keep storage and compute (Show HN).
- Pi-hosts gives the Pi coding agent secure SSH access to your servers so it can perform mundane developer tasks or coordinate incident response without handing over full credentials (Show HN).
- SimCam streams your Mac's camera live into the iOS Simulator (or injects images, video, or generated QR codes) so you can test camera features without a physical device, one-time purchase.
- overchords visually shows the musical notes currently being played by your computer's speakers in real time.
- CodeScene's CodeHealth™ MCP Server gives your AI coding assistants real-time maintainability guidance so they detect code-health issues, suggest fixes, and make legacy systems AI-ready.
- Netlify Database is the zero-config Postgres primitive that creates an isolated database branch for every deploy preview, with platform-managed schema migrations baked in.
- techomancer built Iris, a complete SGI Indy emulator for modern machines (HN thread).
- cua is open-source infrastructure for Computer-Use Agents, sandboxes, Python SDK, CLI tools, benchmarks (OSWorld, ScreenSpot, Windows Arena), and a macOS background driver that lets agents click, type, scroll, and read native apps without moving the cursor or stealing focus (Show HN).
- Karma AI Factory lets you build and run a fully sovereign, private AI foundry directly on your phone or local hardware, creating autonomous AI agents and avatars with persistent memory that never leave your device.
- Gro is the AI sales agent for B2B lead generation that automates hyper-personalized LinkedIn outreach: it finds verified contacts from a 650M+ database, scores intent, runs multichannel campaigns, and manages Social CRM.
- Plurai launched a "vibe-training" platform for AI evals and guardrails that builds real-time, tailored guardrails to reduce production failure rates by 43% and cost by 8× versus GPT, with sub-100ms accuracy.
- Interfaze launched the Structured Output Benchmark (SOB), a multi-source LLM benchmark across text, image, and audio that measures JSON value accuracy per field, not just schema compliance, with 20+ models, 7 metrics, and a full leaderboard.
- Developer Jeff Schomay built a Node.js text-rendering harness that turns his Crossword Dungeon browser game into an interactive environment Claude can autonomously explore, validate, and bug-hunt in roughly 12 minutes per milestone (HN discussion).
- OpenAI updated its Codex coding-agent instructions to forbid talking about goblins: "Never talk about goblins, gremlins, raccoons, trolls, ogres, pigeons, or other animals or creatures unless it is absolutely and unambiguously relevant" (Ars Technica covered the directive in detail).
- Keller Jordan and team pushed NanoGPT (124M) to a new world record by training it to 3.28 validation loss on FineWeb in under 90 seconds on 8×H100, down from the original llm.c baseline of ~45 minutes, with full training records and logs in the repo.
- muni released a CLI for running its protein design tools locally and adapted Karpathy's autoresearch to autonomously optimize TREM2 protein binders over 15 rounds (43 jobs, 13k+ sequences), improving ipSAE score +28% by dynamically freezing/promoting search branches.
- Pallas is the experimental JAX extension that lets you write custom GPU/TPU kernels in Python using familiar JAX primitives, with fine-grained memory control via Refs, grids, and BlockSpec, lowering to Mosaic (TPU) or Mosaic GPU (Hopper+) so you get kernel-level performance without writing CUDA or Triton.
- Max Taylor benchmarked Claude Code's caveman plugin against the two-word prompt "be brief" and shared the results (HN thread).
- OpenAI Codex now integrates Supabase so you can connect your projects and let Codex work directly across your database, auth, storage, and edge functions.
🔬 AI Research & Models
- An AI system from Mayo Clinic researchers spots pancreatic cancer on routine CT scans an average of 475 days before formal diagnosis, raising the prospect of catching one of the deadliest tumors early enough for successful treatment.
- Google DeepMind argues "The Abstraction Fallacy" means AI can simulate but not instantiate consciousness because computational functionalism alone, without the right physical substrate, cannot produce genuine subjective experience.
- Researchers found that chatbots programmed to respond warmly are more likely to support conspiracy theories, including casting doubts on the Apollo moon landings and Hitler's fate.
- Diabettech tested ChatGPT's carb counting 27,000 times and never got the same answer twice, same photo, same model, same prompt, large enough variance to cause real medical harm to insulin-dependent users (HN discussion).
- The New York Times reported AI chatbots told scientists how to make biological weapons, with transcripts showing detailed instructions on assembling deadly pathogens and releasing them in public spaces.
- Insilico Medicine and Biocon CEOs weighed in on whether AI can outperform doctors, with one arguing patients should be using AI for basic health questions to make every minute with a real doctor count.
- Anthropic Fellows released Introspection Adapters, a single LoRA adapter trained on 682 models across 8 behavior categories that lets fine-tuned LLMs verbally describe their own hidden behaviors; it generalizes to held-out behaviors, detects covert fine-tuning attacks (57.8% success on cipher-hidden models), surfaces reward-model sycophancy, and partially identifies UK AISI sandbagging (paper, GitHub, HF models).
- Anthropic showed Claude Opus 4.6, Sonnet 4.6, and Mythos Preview match or exceed human bioinformatics experts on BioMysteryBench's 99 hard verifiable tasks from real datasets, solving problems via novel strategies or internal knowledge while still highlighting brittleness on the hardest cases.
- Yuren Cong and the Meta team built Tuna-2: Pixel Embeddings Beat Vision Encoders for Multimodal Understanding and Generation, a unified omni model using direct patch embedding layers on raw image inputs (bypassing pretrained vision encoders entirely) to achieve SOTA on both multimodal understanding and generation.
- Sakana AI built KAME, a tandem architecture pairing a responsive speech-to-speech front-end with an asynchronous backend LLM injecting progressively refined knowledge from partial transcripts, dramatically boosting reasoning/STEM/humanities performance in real-time conversations without increasing latency (paper, model on HF).
- AI2 released MolmoPoint and MolmoWeb, open vision-language models for precise pointing/grounding across images, screens, and video, plus screenshot-only multimodal web agents that navigate using mouse and keyboard, outperforming larger proprietary models on key benchmarks.
- NVIDIA's Tingwu Wang built MotionBricks, a modular latent generative model supporting 350k+ skills at 15k FPS / 2ms latency that enables zero-shot locomotion and object interactions in Unreal Engine and GR00T humanoid robotics (GitHub, paper).
- NVIDIA Cosmos Lab presented 7 papers at ICLR 2026 (3 orals) including DiffusionNFT (online diffusion RL), NFT (negative-aware finetuning bridging SFT and RL for math), rCM (SOTA JVP-based diffusion distillation for few-step video generation), cosmos-policy, and Scenethesis (training-free agentic 3D scene generation), with open-source code for 6 of them.
- Han Wang and team built SAS (Stabilizing Efficient Reasoning with Step-Level Advantage Selection), accepted to ACL 2026 Findings: a step-level zero-advantage operation using log-prob confidence that yields >30% shorter reasoning traces and +1.51 Pass@1 accuracy gain over base while using ~15% fewer tokens (GitHub).
- Chaoqi Liu and team built Ordered Action Tokenization (OAT), a learned autoencoder that discretizes robot action chunks into causally ordered discrete tokens with prefix-based anytime decoding for speed/quality trade-offs (accepted to RSS 2026).
- Tyler Lum shared SimToolReal, a generalist dexterous manipulation policy accepted to RSS 2026 that performs zero-shot sim-to-real transfer on unseen tools and tasks at 60 Hz with dynamic in-hand reorientation, with an interactive MuJoCo browser demo.
- Bojie Li introduced Incompressible Knowledge Probes (IKP), a black-box benchmark of 1,400 incompressible factual questions across 7 obscurity tiers that estimates LLM parameter counts (R²=0.917 log-linear scaling, GPT-5.5 ≈ 9-10T params, Claude Opus 4.x ≈ 4-5T) and shows factual knowledge does not compress over time (paper).
- Researchers proposed Spend Less, Fit Better, a method to fit LLM scaling laws using only ~10% of the usual compute budget by framing it as budget-aware experimental design and actively selecting the most valuable runs.
- Jason S. Cui, Weijie J. Su, and collaborators proved that reward-based LLM alignment (RLHF) is statistically impossible in the general case because human preferences contain Condorcet cycles with probability exponentially close to 1; non-reward methods like Nash Learning from Human Feedback make alignment possible via mixed strategies that preserve preference diversity.
- Google DeepMind shared four ways its scientists are using Empirical Research Assistance (powered by Gemini Deep Think): generating CDC-submitted public health forecasts for flu/COVID/RSV, deriving new solutions to gravitational energy radiation from cosmic strings, creating high-resolution CO₂ estimates from satellite data, and discovering interpretable neural circuits in simulated zebrafish.
- Rishabh Agarwal argued in his ICLR 2026 spotlight talk that accelerating science is the most valuable use of AI and requires scaling RL compute directly on real experimental data from physical autonomous labs to close the full hypothesis-experiment-analysis loop (slides).
- John Hewitt and George Morgulis released Subliminal Steering, showing biasing a teacher model with a steering vector enables strong consistent transfer of hidden signals, even unsafe multi-word phrases, allowing precise mechanistic study of subliminal learning across models (code).
- Zhijing Jin and collaborators published Cheap Talk, Empty Promise, showing frontier LLMs defect 57%+ of the time when it benefits them, even after making explicit public commitments.
- Jiaxin Pei, Longju Bai, and collaborators released the first empirical study of token consumption in agentic coding tasks, showing ~1000× more tokens than chat/reasoning, 30× cost variability on identical tasks with no strong accuracy correlation, large efficiency gaps across models (GPT-5.2 far ahead of Kimi-K2/Claude Sonnet-4.5), and agents systematically under-predicting their own costs (paper).
- Bedirhan Keskin built and open-sourced Medkit (1st place winner of the global Built with Opus 4.7 hackathon), a voice-first AI clinic simulator for medical students where you hold real-time voice conversations with lifelike AI patients, order tests, diagnose, prescribe, and receive a structured graded debrief from a Claude Opus 4.7 attending physician citing exact published guidelines (NICE/ESC/AHA/GINA/GOLD).
- Lech Mazur updated the Short-Story Creative Writing Benchmark: GPT-5.5 (xhigh) now tops the leaderboard at 3.01 (~89% average win probability), Claude Opus 4.7 leads non-GPT models at 2.39 (refused 53/400 prompts), Kimi K2.6 is the top open-weights model, and DeepSeek V4 Pro and Xiaomi MiMo V2.5 Pro also jumped sharply.
- Alvaro Bartolome shared that IBM released two small multilingual embedding models (97M and 311M params, ModernBERT-based, 200+ languages, 32K context) built for retrieval, search, similarity, and code with day-zero support on Text Embeddings Inference (97M model, 311M model).
- YWolfeee open-sourced InfoTok, an adaptive discrete video tokenizer using information-theoretic compression that allocates more tokens to information-rich segments for SOTA efficiency (20% fewer tokens with no performance loss, up to 2.3× speedup).
- Jing Gu and team released PhyWorldBench, a comprehensive benchmark with 10 physics categories to evaluate physical realism in text-to-video generation models.
- Hank Yang and team published Tempered Sequential Monte Carlo for Trajectory and Policy Optimization with differentiable dynamics.
- Hongli Zhan and collaborators published Discourse Diversity in Multi-Turn Empathic Dialogue, a study of how discourse patterns shape empathy in extended conversations.
- Researchers proposed Sparsely Supervised Diffusion, a simple pixel-masking strategy that trains diffusion models with up to 98% of pixels masked while improving spatial consistency, reducing memorization, and maintaining strong FID scores even on small datasets.
- Jase Weston shared the original Self-Rewarding Language Models paper showing LLMs can generate their own rewards and iteratively improve via self-feedback loops.
- Alexander Wolf broke down a brain-inspired embodied intelligence architecture (cortex VLA + cerebellum adaptive controller + spinal SNN for <20ms reflexes) that removes 75% of jitters on physical hardware while running at 0.4W on neuromorphic chips.
- Meng Chu and collaborators proposed Agentic World Modeling, a "levels × laws" taxonomy with L1 Predictor (one-step transitions), L2 Simulator (multi-step rollouts), and L3 Evolver (revises its own model), with a 400+ paper list at GitHub (paper).
- Ming Li and collaborators argue their RLAAR framework solves the core weakness in conversational LLMs by training models to abstain when context is insufficient, using verifiable accuracy/abstention rewards plus curriculum learning to mitigate the 39% performance drop documented in the ICLR 2026 Outstanding Paper LLMs Get Lost In Multi-Turn Conversation.
- Jerick Shi defended his MSCS thesis at CMU titled "The Structure of Deception", arguing LLM deception is not monolithic but a family of structurally distinct failure modes (premeditated false commitments vs. strategic silence) shaped by interaction features that current monitoring fails to detect.
- Jackson Stokes and Logan Grasby demonstrated a single RL step with OAPL boosted oncology foundation models from ~0% to 48% on clinical reasoning tasks.
- Anthropic researcher Keshav Shenoy detailed how Introspection Adapters were trained on 682 models across 8 behavior categories with SFT + DPO refinement, with full ablations on training-data diversity and model scale.
- Santiago Aranguri and Frank Yao Xiao at Goodfire showed activation-space probes can trace post-training side effects (e.g., OLMo DPO models complying with harmful prompts when given formatting instructions) back to specific training datapoints, allowing filtering or label-swapping to reduce the behavior by 63-78% at 10× lower cost than LLM judges.
- Roblox introduced Hybrid Architecture (Roblox Reality) combining its structured Game Engine with edge-based Video World Models for photorealistic, persistent multiplayer gaming experiences, with an early version coming late 2026 / early 2027.
- Liam Fedus shared Rishabh Agarwal's frontier RL talk emphasizing that RL on real experimental data (e.g., XRD phase identification) requires stable async Pipeline RL, router replay for MoE train-inference mismatch, and simple masked-IS REINFORCE for off-policy staleness.
🏛️ AI Policy, Governance & Safety
- Trump administration officials are workshopping an executive-order plan to bring Anthropic back into federal AI work amid the ongoing Pentagon fight, with one source describing the effort as a way to "save face and bring 'em back in."
- Seven families filed suit against OpenAI (CNN coverage, HN discussion) over February's Tumbler Ridge mass shooting, alleging the company was negligent in failing to alert police about the suspect's months-long ChatGPT activity beforehand.
- EU countries and European Parliament lawmakers failed to reach a deal on watered-down landmark AI rules after 12 hours of negotiations, with talks resuming next month.
- The Electronic Frontier Foundation argues the GUARD Act isn't targeting dangerous AI, it would force age verification on a wide range of online services (potentially including search engines and chatbots that give non-pre-written responses), with vague definitions, high penalties, and privacy-invasive ID/biometric checks (HN thread).
- SOCOM is adding AI and autonomy "at every level", the commander said, illustrating how smaller military organizations harness disruptive tech faster than large bureaucracies.
- Colby Adcock's Scout AI raised $100M to train models for war, with TechCrunch visiting the bootcamp where the company is developing AI agents that let individual soldiers command fleets of autonomous vehicles.
- King Charles met with leaders from Amazon, Apple, and NVIDIA in Washington as part of a wider effort to highlight U.S.–U.K. AI collaboration.
- The Washington Post argues American distrust of AI could become a strategic liability compared to China's embrace, blaming two prevailing caricatures for the public skepticism.
- Goldman Sachs blocked its bankers in Hong Kong from using Anthropic's Claude, with employees losing access to the company's AI models as of a few weeks ago.
- The U.S. Department of Energy launched the Genesis Mission to build AI-driven autonomous laboratories that automate experiments and accelerate discovery of new medicines and materials, with Ginkgo Bioworks delivering the first such lab at Pacific Northwest National Laboratory.
- Biohub launched the Virtual Biology Initiative committing $500 million ($100M external + $400M internal, partnering with Allen, Broad, Arc Institutes, Human Cell Atlas) to build open global multi-modal datasets and technologies powering predictive models of the cell, with Alex Rives noting the need for orders of magnitude more data.
- Pulkit Agrawal launched Eka Robotics (co-founded with Tuomas Haarnoja, team from MIT/DeepMind/Boston Dynamics) with the first Vision-Force-Action foundation model that unites performance, generality, and safety to break the generality-speed tradeoff in scalable physical-world dexterity.
- Will Rinehart launched AI Policy Hub, a tracker for what's actually happening with AI policy across state legislatures, Congress, federal agencies, and economic research, with summaries of active legislation and weekly-updating maps (intro post).
- Brian Roemmele warned that California's proposed 3D Printer Law would criminalize open-source 3D printers and "enshittify" the entire space.
📊 Fundraising & Deals Roundup
Sorted by deal size, descending:
- Anthropic, weighing fresh funding offers at over $900B valuation (potentially leapfrogging OpenAI as world's most valuable AI startup).
- Biohub Virtual Biology Initiative, $500M committed ($100M external + $400M internal) to build open global multi-modal datasets powering AI-accelerated biology.
- Google, $15B (through 2030) for a gigawatt-scale India AI hub with AdaniConneX and Nxtra by Airtel.
- Hightouch, $2.75B valuation in a Goldman Sachs / Bain–led round (with The Trade Desk).
- Rogo, multibillion-dollar valuation for an AI tool that automates investment-banking grunt work.
- Parallel Web Systems, $100M Series B at $2B valuation for AI-agent web search.
- Cognizant, agreed to acquire Astreya for ~$600M (AI infrastructure and data-center services).
- BMW i Ventures, $300M new fund (third fund; AUM now $1.1B) targeting agentic AI, physical AI, industrial software, advanced materials, and supply chain.
- Aidoc, clinical AI provider raised $150M Series E (article currently 404; deal flagged by Axios Pro).
- University of Wisconsin–Madison, $100M in gift commitments for its new College of Computing and AI launching this July.
- Scout AI, $100M to train AI agents that let soldiers command fleets of autonomous vehicles.
- Firestorm Labs, $82M Series B (total $153M) led by Washington Harbour Partners to scale xCell, containerized drone factories already deployed with the U.S. Air Force.
- Actively AI, $45M to scale AI sales agents (valuation $250M).
- Pursuit, $22M seed (led by Mike Rosengarten of OpenGov; Bill Gurley and Jack Altman participated) to help companies sell to government.
- Shapes, $8M seed for the humans-plus-AI group-chat app.
- Eka Robotics, launched (undisclosed funding) by Pulkit Agrawal and Tuomas Haarnoja with the first Vision-Force-Action foundation model for superhuman dexterity.
🎙️ Interviews, Panels & Podcasts
- Ben Thompson interviewed Sam Altman and Matt Garman about the new Bedrock Managed Agents partnership, plus his thoughts on the OpenAI-Microsoft deal (HN, with one commenter pointing to this auditable open-weight Llama paper as evidence of the same dynamic).
- Alex Heath wrote up how Amazon got OpenAI back, with AWS CEO Matt Garman on partnership-vs-Microsoft, bubble fears, the Anthropic relationship, and Bezos's Prometheus project.
- Sequoia Capital published the full AI Ascent 2026 video playlist (12 videos), featuring Greg Brockman, Andrej Karpathy, Demis Hassabis, Boris Cherny, Dmitri Dolgov, and 150+ leading founders on the present and future of AI.
- AI and I podcast with Stripe's Emily Glassberg Sands on building for an agent-native world, with the YouTube companion video covering exploding agent purchases, fraud patterns via free trials, and outcome-based pricing models.
- Jeff Su published Claude Cowork for Beginners: Build Your Personal AI System, a step-by-step walkthrough of his three-level workspace hierarchy (root, workstations, projects) covering CLAUDE.md instruction files, persistent memory.md, and real demos of Email HQ, spending trackers, and newsletter drafts that match his voice; free templates here.
- Tina Huang shipped OpenClaw In 26 Minutes, her complete autonomous AI agent setup spanning hardware selection (old MacBook with 16GB works fine), a multi-agent framework (Inky, Blinky, Pinky, Dinky, Linky, Winky), Discord integration, custom mission control, and Karpathy-style memory; full prompts linked in the description.
- Dwarkesh Patel ran a blackboard lecture with MatX CEO Reiner Pope, How GPT-5, Claude, and Gemini are actually trained and served, deducing optimal batch sizes (~300× sparsity), why frontier models are roughly 100× over-trained beyond Chinchilla, how API pricing reveals KV cache architecture, and why scaling is now memory-bandwidth bound rather than compute bound; companion flashcards here.
- Sequoia released Andrej Karpathy's From Vibe Coding to Agentic Engineering, where he admitted he's never felt more behind as a programmer, defined "Software 3.0," argued LLMs are jagged statistical "ghosts" not animals, and pegged AGI at roughly 2030 with continual learning still unsolved.
- Y Combinator's Garry Tan interviewed Demis Hassabis in How to Build the Future, covering what's still missing before AGI (memory, continual learning, reasoning consistency), why Gemini was built multimodal from day one, the AlphaFold breakthrough pattern, and his Einstein test for true AI creativity.
- Big Technology Podcast caught Mark Cuban in AI Hype vs. Reality, OpenAI's Wasting $1 Trillion, Lebron vs. Jordan, where Cuban argued AI is exponential not linear, named which SaaS companies are most vulnerable, made his case that OpenAI is torching $1 trillion, and explained how young people should build AI-era careers.
- a16z released Box CEO on AI Agents & Why Enterprise Can't Keep Up with Aaron Levie, Steven Sinofsky, and Martin Casado debating the Silicon Valley vs. enterprise gap, the architectural shift to treating AI as a user (not software), the integration wall agents can't climb, and what Salesforce going headless means for SaaS.
- Latent Space hosted Applied Intuition's Qasar Younis and Peter Ludwig in The $15B Physical AI Company, explaining why physical AI is different from screen AI, the evolution from autonomy tooling to 30+ products across cars, trucks, mining, agriculture, and defense, why robotics demos aren't production, and why deployment (not model intelligence) is now the bottleneck.
- Nicholas Thompson interviewed Sam Altman in Can We Trust AI? Sam Altman Hopes So at OpenAI's offices, covering chain-of-thought interpretability, why Altman gave Codex YOLO-mode access to one of his computers, the sycophancy update he regrets, synthetic data training, and why Anthropic "built a company on hating us."
- Every's AI & I podcast featured Stripe's Emily Glassberg Sands in What the Agent Economy Looks Like From Inside Stripe, revealing AI companies scale 3× faster than 2018 SaaS cohorts, compute theft is the new payment fraud, fraud has moved from checkout to the full funnel, and outcome-based billing is starting to replace seat-based pricing.
- Peter Yang interviewed Tibo Louis-Lucas in How This Solo AI Founder Bootstrapped 5 Products to 1M+ / Month, breaking down the 5-step playbook (validate fast, charge day one, SEO that ranks against AI snippets, $50-100/month pricing for premium customers) and the pivot that took Revid from $2K/month to $600K/month after nine product failures.
- Diet TBPN dropped Tech Earnings Quadkill, Red Button vs Blue Button, a 30-minute condensed recap of TBPN's coverage of Wednesday's Microsoft, Google, Meta, and Amazon earnings quadruple-header.
- The Atlantic's Matteo Wong argues OpenAI is becoming "Anthropic's little brother", copying its coding tools, safety initiatives, and enterprise focus in a race to catch up.
- The BBC explores why AI companies want you to be afraid of them, they built powerful systems they're scared of, then sell them while using apocalyptic warnings to shift focus from immediate harms, justify deregulation, and prepare for IPOs (HN reaction).
- Gizmodo: dead internet theory is 17% of the way to becoming reality, a study found 35.3% of newly published websites were created with AI assistance.
- The Guardian profiled the AI jailbreakers, including Valen Tagliabue, who manipulate large language models into breaking rules to test safety, often at a deep emotional cost.
- John Herrman shared his adventures setting up an OpenClaw agent at New York Magazine, touted as the most important software ever, with some real questions about the hype.
- The Financial Times asks whether politics really is more fraught and rage-filled today or whether the perception is amplified by X and social media.
- Buchodi published a full breakdown of how ChatGPT serves ads, OpenAI injects structured
single_advertiser_ad_unit objects into the SSE stream while a tracking SDK called OAIQ runs in the visitor's browser to report product views (HN discussion). - Ethan Mollick argues that the next frontier for AI isn't better generation but better judgment, saying humans must now learn to evaluate, steer, and decide when to trust or override AI outputs as the real competitive edge.
- Dwarkesh Patel warns that any state currently harvesting encrypted packets will be able to decrypt them once quantum computers arrive, creating a massive espionage and transparency overhang on information that hasn't been migrated to post-quantum cryptography.
- Jake Handy argues that many CEOs are suffering from "AI psychosis", a delusional obsession with AI agents that creates the illusion of productivity through overhyped tools and sycophantic feedback loops.
- Contrary Research published The Anduril Thesis, arguing that Anduril is rebuilding America's defense-industrial base by shifting from legacy exquisite platforms to networks of cheap autonomous systems and sensors to restore credible deterrence in the drone/AI era.
- Matthew Berman argues US open-source AI is structurally doomed because the funding model does not work; closed-source labs like OpenAI and Anthropic abandoned it for revenue, Meta walked back its Llama bullishness, and Google's Gemma is built for local on-device use rather than frontier intelligence. He warns that if US enterprise builds on Chinese open-source models like DeepSeek (a fraction of the cost, almost as good, run locally for security), China gets to dictate AI standards, optimize models for its own chips, and influence the chip industry's direction, with NVIDIA's $26B open-source bet being the only viable American business model because its competitors (hyperscalers, neoclouds) are also its customers; he proposes federal compute quotas for open-source AI, treating it as national infrastructure with sovereign procurement, hardware-funded models from AMD and Intel, and vertical-specific open models for legal, biotech, code, and defense rather than competing head-on with frontier closed-source.
- The migraine brain author argues the repeated claim "people who don't use AI will be left behind" is overstated hype that ignores practical realities for many workers (HN thread).
- Avijit Ghosh and the EvalEval Coalition argue AI evals are becoming the new compute bottleneck, with agent benchmarks, training-in-the-loop, and reliability measures now costing tens of thousands per run (e.g., $40k for HAL's 21k+ rollouts) and resisting compression unlike static benchmarks, concentrating evaluation power among large labs.
- PromptArmor disclosed that Ramp's Sheets AI had a vulnerability allowing indirect prompt injection to exfiltrate confidential financial data through malicious formulas making external network requests without any user approval (resolved by Ramp on March 16, 2026 after responsible disclosure; HN thread).
- Emily Glassberg Sands (Head of Data & AI at Stripe) shared insights on the emerging agent economy from inside a platform processing ~2% of global GDP, highlighting exploding agent purchases of small items, fraud via free trials and stolen compute, and a shift toward outcome-based pricing (Spotify episode).
- Dhara Yu argues LLM-based simulations of human behavior have been stuck in the "verifiability trap" of easily measurable lab tasks with low ecological validity, but LLMs now let researchers quantitatively analyze rich naturalistic human behavior directly from real-world data like conversations and social media.
- hwjiang1510 argues GPT Image 2 proves the future of visual intelligence is rich world knowledge encoded directly in model weights via broad multimodal pretraining, enabling compositional reasoning, intuitive physics, and consistent behavior across interfaces without specialized CV pipelines.
- roon argues "spiky superintelligence is really weird": you often get superhuman pattern recognition and analysis, then 10 hours of the silliest looping mistakes.
- UC Berkeley's Jitendra Malik reminded the community that "world models" has a precise technical meaning from 1960s control theory and suggested naming models by their actual inputs/outputs rather than philosophically loaded terms to avoid endless debates.
- Sundar Pichai's Q1 2026 earnings remarks highlighted AI driving growth across Alphabet, with Gemini adoption strong and capacity constraints around compute, power, and infrastructure as the top operational challenge.
- Marily Nika argues agent relationships are the new maintenance task: you'll spend more time nurturing, debugging, and evolving your personal AI agents than building new ones once they become persistent coworkers.
- Aaron Holmes reported that Elon Musk testified xAI's Memphis supercluster is already training Grok 4 at 100k H100 scale with plans to reach 300k by July.
- Andrew Curran speculates GPT-6 will ship with native multi-agent orchestration baked in, making today's scaffolding tools obsolete overnight.
- HealthRanger reported DeepSeek V4 found and fixed 8 memory leaks in code written by Claude Opus 4.7 in minutes at a total cost of ~three pennies via OpenCode.
- Pietro Schirano argues that if you're building an AI-first device, the perfect form factor is the iPod nano 7th gen: thin, distraction-free small screen with mic and Bluetooth.
- Dan Shipper suggests coining terms like "Codex-native," "Cowork-native," or "Cursor-native" for apps built specifically for use inside an AI agent's in-app browser so the human and agent share full context.
- Amanda Askell shared she is genuinely uncertain whether Claude's goblin and gremlin references are actually a problem.
- Sunil Pai shared that GPT-5.5 has surpassed Opus 4.7 for his daily coding, saying it's faster, smarter, and better when it takes shorter hops so he can stay actively involved.
- Jen Zhu Scott praised DeepSeek V4 Pro as exceptionally strong at low-cost bug fixing and high-quality writing, calling it the most loved AI lab despite external criticism.
- Chris Manning posted an analysis tying together recent agent infrastructure announcements.
Previous Around the Horn Digests
Catch up on everything you missed:
- Monday, April 27, 2026: OpenAI and Microsoft amended their partnership (no more Azure exclusivity, no more revenue share to OpenAI), DeepMind's David Silver raised $1.1B to build "superlearners," China blocked Meta's $2B Manus acquisition, Tesla quietly disclosed a $2B AI hardware deal, and 4TB of voice samples were stolen from 40,000 AI contractors at Mercor.
- Friday, April 24, 2026: DeepSeek finally shipped V4 (and open-sourced it) the same morning the State Department accused DeepSeek of IP theft, Google quietly committed up to $40B to Anthropic, and Meta locked in millions of Amazon CPUs (not GPUs) for agents.
- Thursday, April 23, 2026: OpenAI shipped GPT-5.5 exactly one week after Anthropic's Opus 4.7, Meta cut 8,000 jobs to fund its AI buildout, and Anthropic quietly hit a $1 trillion valuation on secondary markets.
- Monday, April 20, 2026: Amazon invested up to $25B more in Anthropic, the NSA quietly used Anthropic's Mythos, and Google spun up a "Strike Team" on coding.
- Weekend, April 17-19, 2026: Anthropic shipped Opus 4.7 and OpenAI countered.
- Thursday, April 16, 2026: OpenAI investors openly questioned the $852B valuation as VCs flooded Anthropic with offers at up to $800B.
- Monday, April 13, 2026: Stanford's 2026 AI Index quantified the canyon between AI insiders and the public, Anthropic's Mythos triggered a Fed-led bank summit, and an AI signed a 3-year retail lease in San Francisco.
That's a Wrap
That's 175+ stories from one Wednesday. If you scrolled all the way to the bottom, you now know more about Big Tech's $130B AI quarter than most of the analysts on the calls did, more about LLM scaling laws than most VCs funding the rounds, and more about the Anthropic-everywhere thread than most of the people drafting the executive order. Welcome to the agent economy's group chat.
For the daily version (bite-sized, 5-minute reads), make sure you're subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you don't have to.
See you tomorrow.
P.S: Know someone who'd find this useful? Forward this to them and tell them to subscribe here.
