Everything That Happened in AI Today Thursday, June 11

Anthropic spent the day explaining why Claude quietly switched models, while OpenAI gave Codex a cloud workspace and SpaceX tried to make IPO history.

Welcome to the Around the Horn Digest, where we track every AI story so you do not have to open 147 tabs and call it a research process. Today had two speeds: frontier labs wrestling over who controls agentic work, and everyone else building the infrastructure around it. OpenAI moved Codex into persistent cloud workspaces, Anthropic tried to calm a researcher backlash, and agent tooling exploded from Xiaomi’s terminal agent to Goodfire’s training debugger and Ai2’s dependency maps. Outside model land, the day brought robotaxis with loyalty programs, Canada regulating chatbots, CISA shortening vulnerability windows, and SpaceX attempting the biggest IPO headline this side of the moon. AI news is now business, labor, cybersecurity, music, pizza, and orbital capitalism wearing one trench coat. Let’s get into it.

Previous digests: Monday, June 8 | Wed-Thu, June 3-4 | Tuesday, June 2
Recent Neuron archive: Claude Fable Five | Apple finally rebuilt Siri | ChatGPT admitted its memory was broken | Anthropic: AI Is Building AI now | Google Gemini got hijacked via WhatsApp

🆕 NEW From The Neuron

Claude Fable 5 explained: Grant’s Fable primer is the useful prequel to today’s backlash: what Anthropic shipped, why Mythos is different, and why “routed, restricted frontier model” is now a product category.
AI Sales Agents: Top Use Cases, Benefits & How They Work: Bianca mapped the part of enterprise agents that is actually deploying now: lead qualification, CRM cleanup, account research, and the governance headache hiding behind every nice demo.
Real-time translation is finally real: Today’s newsletter edition tracked the “universal translator” moment as live translation moved from impressive demo to product-shaped reality.

Around the Horn — Thursday, June 11, 2026

The big story today was Anthropic's Fable 5 backlash, because it turned a frontier-model (top-tier AI model) safety policy into a live product trust problem. Anthropic did not just ship a stronger Claude and ask researchers to behave nicely; the surrounding reporting showed Fable 5 being invisibly limited or switched on requests involving frontier-model research, cybersecurity, biology, and safety-adjacent work, which meant paying users could hit a weaker fallback without realizing the model underneath them had changed.

That is the sort of safety move that sounds tidy in a policy doc and instantly becomes weird in the wild: Anthropic's support note explained the switching, TechCrunch captured why security researchers were mad, and the reaction trail ran through @claudedevs, @Teknium, @nptacek, @haider1, @LLMSherpa, @alxfazio, @baym, @DavidSacks, @natolambert, @drfeifei, @dwarkesh_sp, @emollick, @GergelyOrosz.

Put simply: the company tried to keep dangerous requests away from its sharpest model, then discovered that serious users also need to know when the sharpest model is no longer the one answering.

The bigger signal is that model routing is now part of the product experience. The next frontier-model fight will not just be who has the smartest system; it will be who can explain when the smart system is allowed to show up, when it gets swapped out, and whether users trust the invisible hand on the switch.

🏆 TOP 5 NEWS (Around the Horn)

OpenAI was reported by WSJ to be considering drastic token price cuts as Anthropic competition heated up, with CNBC echoing the price-war read for frontier AI users and developers.
OpenAI agreed to acquire Ona to give Codex secure, persistent cloud environments for long-running agents, which is a very direct bet that coding agents need durable workspaces, not just chat boxes.
Bloomberg reported that Meta split operations from Manus and started unwinding the acquisition after Chinese regulatory pressure, with The Next Web adding the founders-reportedly-exploring-a-buyback angle.
Canada introduced the Safe Social Media Act, and Reuters framed it as a move to restrict under-16 social-media access unless platforms meet safety standards while also regulating AI chatbots.
Reuters reported that CISA shortened the federal deadline for fixing the riskiest vulnerabilities to three calendar days, citing faster AI-assisted exploitation.

Honorable Mentions

OpenAI reported China-linked influence operations using ChatGPT to target U.S. AI, data-center, tariff, and technology debates, with Axios covering the campaign’s political angle.
Dario Amodei argued AI is moving exponentially faster than policy institutions and called for stronger testing, deployment controls, labor policy, civil-liberties protections, scientific acceleration, and democratic coordination, then amplified the essay on X.
Google DeepMind and partners put up to $10M into multi-agent AI safety research, with @GoogleDeepMind highlighting sandboxes, identity/reputation protocols, and oversight of collective harms.
Goodfire introduced predictive data debugging, a way to forecast which behaviors training data will amplify or suppress before teams finish training a model, with Goodfire, Goodfire, arXiv paper, @deedydas carrying the product, research, and paper trail.

🍪 TOP TREATS TO TRY

Ask DoorDash turns prompts, voice, recipe links, and photos into food or grocery carts, with TechCrunch and @andyfang adding the product-demo angle; no pricing details.
Coinbase launched infrastructure for AI agents to trade crypto, rebalance portfolios, and pay for research data or compute through x402 (a payment protocol for agent transactions), with CNBC covering the payments angle; no pricing details.
Perplexity added Deep Research to Computer, connecting search-as-code, long-running sandboxes, connectors, tools, and licensed data for deeper research workflows, while @perplexity_ai, @AravSrinivas, and @perplexity_ai showed the rollout; no pricing details.
ElevenLabs Avatars generates talking videos from a script, voice, and consistent character avatar inside ElevenCreative, with @ElevenLabs and its preview image showing the launch; no pricing details.
Linear coding sessions let Linear Agent triage issues, investigate root causes, write fixes with Claude Code or Codex, open PRs (code-change requests), and send context back to Linear, with @linear introducing the feature; no pricing details.
Firecrawl Prometheus turns plain-English web-data requests into maintained scraping code users can run or host through Firecrawl, with @firecrawl showing the experiment; no pricing details.
Rilable is an open-source iPhone app builder that turns prompts into live web and iOS apps using Daytona, Chorus iOS, Convex, and Vercel, with @rileybrown and his follow-up documenting the launch; free / open-source.

🏛️ AI Policy, Governance & Safety

Silicon Valley AI politics and regulation: Former a16z partner John O’Farrell argued that Silicon Valley AI companies and political groups are spending heavily to defeat stricter AI-regulation candidates, raising concerns about concentrated private power over democratic AI governance.
China AI labor protections: China’s Workers’ Daily called for labor protections and algorithm oversight as AI spreads, warning about displacement and opaque algorithmic control.
Single biological neuron computation, with bioRxiv: Researchers showed that a detailed model of a single cortical neuron with dendritic computation can solve surprisingly complex tasks, challenging simple assumptions about neuron-level computation.
Malware hides WMD text to evade AI scanners, with Socket: John Scott-Railton and Socket described malware packages that include nuclear or biological weapons instructions to trigger large language model (AI text model) safety refusals and evade AI security scanners.
Cursor Auto-review safety classifier, with Cursor: Cursor made Auto-review the default for new users, using a classifier subagent to review actions in context before allowing, blocking, or asking for approval.

📊 Funding, Deals, Markets & Infrastructure

Prometheus industrial AI and invention loop, with @AndrewCurran_: Jeff Bezos-backed Prometheus reached a huge valuation for industrial AI tools meant to accelerate the loop from idea to prototype to manufacturing; Andrew Curran’s shared quote framed the company as an attempt to speed invention itself.
Anthropic data-center lease ambitions, with @kimmonismus: The Information article is paywalled; based on the provided title and surrounding context, Anthropic signed early data-center lease agreements and sought Google financial backing to control more of its compute supply and reduce long-term costs.
Digital Asset funding: Digital Asset Holdings raised $355M led by Andreessen Horowitz for Canton Network, a blockchain used by major Wall Street firms for traditional-asset workflows.
Google skilled-trades training, with Axios: Google committed $50M to train more than 300,000 skilled-trade workers for data-center and infrastructure jobs, reflecting the physical labor bottleneck behind AI buildout.
Theker robotics funding, with The Next Web: Theker raised roughly $85M for generalist factory robots that deploy quickly and improve autonomously on the job.
Crusoe Wyoming AI data center setback: Crusoe was pushed aside or paused at a Wyoming AI data-center project after customer and timeline concerns involving Google.
PhoenixAI database funding: PhoenixAI raised $80M to build an analytical database designed for live, low-latency agentic AI workloads.
Amazon data-center water use, with About Amazon: Amazon said its directly operated data centers used 2.5B gallons of water in 2025 while improving water efficiency and using reclaimed water at some sites.
Coram security-camera AI funding: Coram raised $35M to turn existing security cameras, badge readers, and logs into AI-searchable security systems that detect threats and support investigations.
KKR Helix AI infrastructure: KKR launched Helix Digital Infrastructure with more than $10B in committed capital for AI data centers, power, and connectivity.
Google Samsung AI chip talks: The Information article is paywalled; based on the provided title and snippet, Google is considering Samsung for a component in future AI chips as TSMC capacity tightens.
SpaceX record IPO (public stock listing), with Reuters, The Next Web: SpaceX priced a record-breaking IPO (public stock listing) with extraordinary demand and a major BlackRock order, making it a major AI-adjacent infrastructure and capital-markets story.
Compute Coalition report, with Carnegie Endowment, @TawilTeddy: The Carnegie Compute Coalition report argues that AI infrastructure will shape global power and democracies need faster permitting, grid queues, transmission, capital, and supply-chain coordination to keep up.
Ricursive chip-design AI model, with EE Times: Ricursive announced work on an end-to-end AI model for chip design focused on hardware-workload co-optimization.
Tether invests in NEURA Robotics: Tether led a large Series C financing for NEURA Robotics and framed the investment around edge AI, self-custodial wallets, and humanoid robotics participating in machine economies.

🛠️ AI Tools & Products

Deezer AI music detector, with Deezer Newsroom: Deezer launched a free detector for scanning playlists from other streaming platforms for AI-generated music, citing high rates of AI tracks among imported playlists and daily uploads.
River personal AI stack, with River: River AI launched with a mission to build personal AI owned and shaped by users rather than controlled by large corporations.
Cofounder company agents, with Cofounder: Cofounder is an agent orchestration platform for company functions; the demo focused on marketing agents that plan campaigns, generate content, schedule posts, and report results with human approval.
Bonemeal growth marketing agent: Bonemeal is an AI Growth Engineer for SEO and AI-search visibility, content creation, pages, and organic traffic.
Revyl mobile runtime maps, with Revyl: Revyl gives teams and agents live mobile environments, replayable test evidence, and Atlas runtime maps of app workflows, helping humans and agents understand running apps.
Abridge clinician intelligence platform, with @AbridgeHQ: Abridge presented an AI-native clinician intelligence platform built around clinical conversations, including foundation models, summaries, nursing workflows, risk gaps, and trial matching.
Nessie portable AI memory, with @garrytan: Nessie makes a user’s thinking, context, memory, and history portable across AI tools, with OpenClaw and Model Context Protocol (the tool-connection standard for AI agents) server integrations for agent memory.
Apple OpenCode MLX (Apple's machine-learning framework) demo, with @iamdavidhill, @iamdavidhill: Apple used OpenCode to demo MLX (Apple's machine-learning framework) at WWDC, signaling that agentic developer tools are becoming visible in Apple’s on-device machine-learning context.
Pool screenshot app, with @poolday: Pool organizes screenshots into searchable pools, recovers original links, suggests actions, and processes locally first with optional cloud analysis.
OpenCreator video studio, with OpenCreator: OpenCreator is an all-in-one generative video studio that combines creative models into a single interface for script, casting, storyboard, generation, and editing.
Honen course creator: Honen turns team docs, call recordings, kickoff videos, or topics into structured training courses that people can actually finish.
Applied Compute data-efficiency take: Applied Compute argued that reinforcement learning can specialize models with surprisingly little proprietary data when the data is high quality and uniquely business-specific.
fal Pixelcut video background remover, with fal: fal launched Pixelcut Video Background Remover for fast, precise video background removal, including movement and hair handling for commercial workflows.
Macrodata Refiner robotics data framework: Macrodata Labs helps robotics teams process physical-world data, with Refiner for local Python pipelines that scale to managed cloud compute.
Claude safeguards warnings and appeals: Anthropic’s support page explains how users can appeal safeguards warnings or mistaken account actions, and the pasted context notes longer response times after the recent Claude launch.
OpenClaw restaurant video prospecting demo: Chris built an autonomous agency-style system that finds restaurants with weak food photos, generates cinematic reels, and mails owners postcards with QR codes.

🏢 Big Tech & Major Companies

Waymo Premier loyalty tier, with Waymo: Waymo launched Waymo Premier, a paid loyalty tier for frequent robotaxi riders with trip credits, cancellation perks, and priority access in waitlist cities.
Apollo AI disruption screening: Apollo is screening software investments for AI disruption risk, signaling that asset managers now underwrite whether AI could make software businesses obsolete.
Claude Corps nonprofit fellowship, with Anthropic: Anthropic launched Claude Corps, a $150M fellowship to place 1,000 early-career fellows inside nonprofits for one-year AI-coaching roles with training, grants, and Claude credits.
Poetic enterprise AI system, with Poetic: Poetic launched an AI system for complex multi-hour enterprise tasks that blends deterministic code with AI regeneration, raised $50M at a $500M valuation, and is hiring.
Generalist to specialist representation theory, with @YujiaZheng9: Yujia Zheng and collaborators presented theory for turning generalist models into specialist representations with identifiability guarantees.
Mercury 2 on Baseten, with @baseten: Baseten made Inception’s Mercury 2 diffusion large language model (AI text model) available in production, emphasizing very fast token generation and lower cost for agent workflows.
Claude production agent features, with @claudeai: Claude added production-oriented agent features including scheduled deployments, secure environment variables, and dynamic workflows for recurring operational agents.
Agent loops as design pattern: Nick Dobos framed recurring agent loops as the useful unit of agent design, including planners, watchers, crash fixers, triagers, and daily automated work loops.
X trending pages, with X trending page, X trending page: These dynamic X trending pages captured platform-wide trend context around the day’s AI/model discussions; they are unstable sources but explain why the topic was surfacing.
Data-center compute records: Epoch AI reported that single-data-center compute-capacity records have doubled roughly every seven months, a key signal for the AI infrastructure race.
Mythos 5 music demos, with @scaling01: scaling01 shared Mythos 5 demos composing original piano music and generating matching visualizers, showing creative-output quality and social traction around the model.
AI Village multi-agent visualization: AI Digest shared visualizations from an AI Village multi-agent simulation comparing leadership dynamics among Claude Fable 5, Opus 4.8, and other models.
Superhuman Adaptable Intelligence argument, with arXiv paper: The SAI paper and thread argue that AGI is an overloaded human-centric frame and that AI should instead target superhuman, adaptable specialization across economically important tasks.
Deep-learning global migration dataset, with @guyabelguyabel, HKU Social Sciences, Zenodo, HF dataset: ThGaskin/Migration_flows, GitHub: ThGaskin/Migration_flows: Guy Abel and collaborators published a Nature paper using deep learning to build a global annual migration-flow dataset from 1990 onward, with an explorer, data, and code for studying migration across 230 countries and regions.
Infinite London city-generation demo, with Hugging Face: blog/mishig: Mishig built a procedural London-generation demo using Fable 5 and Trellis.2, producing detailed city scenes and a live explorable demo.
Shopify Quick internal backend demo, with Shopify: Daniel Beauchamp built Quick at Shopify, a zero-config internal backend for AI-generated sites with shared data, uploads, and realtime collaboration on a simple VM, showing how removing backend friction can unlock experimentation.
Interactive image-compression chapter, with @DanHollick: Dan Hollick published an interactive chapter explaining image compression, including PNG history and how JPEG discards visual information people do not notice.
Claude department and ECC workflows, with @undefinedKi: Yarchi shared guides and open-source resources for turning Claude into a department of specialized roles and using battle-tested Claude Code skills and subagents.

🔬 AI Research, Models & Benchmarks

NVIDIA and Abridge clinical model: The WSJ piece is paywalled; based on the title and snippet, NVIDIA is partnering with Abridge to develop a model tailored to clinical conversations.
ModSleuth dependency tracing, with ModSleuth, arXiv paper, ModSleuth, @sewon__min: Ai2 introduced ModSleuth to trace hidden model and dataset dependencies in modern large language models (AI text models), revealing recursive reliance on other models as data generators, filters, judges, OCR systems, and development assistants.
Castform model training platform, with Castform, Castform, Castform, Castform: Castform launched a platform for training open-weight models on proprietary task data or production agent traces, managing datasets, recipes, GPU (AI-chip) runs, and model inspection with free credits for new users.
Recursive automated AI research, with @cong_ml, GitHub: recursive-org/first-steps-toward-automated-ai-research, @josh_tobin_, @Recursive_SI, @hxiao: Recursive shared early results from an automated AI research system that proposes ideas, implements changes, runs experiments, validates results, retains context, combines promising branches, and improves model-training and GPU-kernel (chip-level speed) benchmarks.
large language model verbalization limits, with Elena Baixy: Xiaoyan Bai and collaborators argued that large language model (AI text model) verbalization methods can mix real signal with over-interpretation and should not be treated as faithful ground truth.
Agents Last Exam benchmark, with Agents Last Exam, Agents Last Exam, Agents Last Exam, arXiv paper, HF dataset: agents-last-exam/agents-last-exam, GitHub: rdi-berkeley/agents-last-exam, @dawnsongtweets: Agents’ Last Exam is a rolling benchmark of 1,500+ real professional tasks with verifiable outcomes across GUI and command-line environments, showing frontier agents still fail the hardest long-horizon work.
DeLM shared-context multi-agent framework, with arXiv paper, Yuzhenmao: DeLM is a decentralized multi-agent framework where parallel agents coordinate through shared verified context and a task queue instead of a central controller.
Social World Models, with GitHub: ulab-uiuc/social-world-model, HF dataset: ulab-ai/swm-bench, @haofeiyu44: Social World Models use large language models (AI text models) to simulate collective belief shifts after news, with prediction markets used as supervision for forecasting and counterfactual simulations.
Riemann Hypothesis visual journey, with @the2ndfloorguy: This interactive visual journey explains the Riemann Hypothesis from prime numbers upward, serving as a high-quality technical education example.
Interlatent robotics post-training loop, with Interlatent: Interlatent argued that modern robotics needs a post-training loop: capture every rollout, score what happened, and retrain on the parts that matter.
OdysSim human behavior simulation, with GitHub: sunnweiwei/OdysSim, Hugging Face: collections/cmu-lti, HF dataset: cmu-lti/osim-mid-training, HF dataset: cmu-lti/cmu-lti, GitHub: sunnweiwei/OdysSim: OdysSim trains large language models (AI text models) on real human interactions to simulate messy human behavior for evaluation, education, medicine, and social science.
PixelRAG visual retrieval, with GitHub: StarTrail-org/PixelRAG, PixelRAG: PixelRAG searches rendered web-page screenshots instead of text alone, targeting knowledge embedded in layout, tables, diagrams, and pixels.
Artificial Hivemind and Elara bias, with arXiv paper, Chris Thomas: Gabriele Berton connected the Artificial Hivemind paper and Chris Thomas’s Elara problem to show how large language models (AI text models) converge on homogeneous names, tropes, and outputs because of training-data attractors.
Sparrow sparse-rollout reinforcement learning, with arXiv paper, GitHub: Infini-AI-Lab/Sparrow, Infini Ai Lab: Sparrow is a sparse-rollout method for stable and efficient long-context reinforcement learning in large language models (AI text models), speeding rollouts while tracking actor-policy mismatch.
Apple Siri AI beta early tester reactions, with @JoannaStern, @ParkerOrtolani: Apple’s redesigned Siri AI beta impressed early testers, who highlighted more natural conversation, personal context from apps and photos, on-screen awareness, tight native iPhone integration, and Camera Control as an instant visual-intelligence trigger. The important caveat is that the proactive, agentic Siri that acts before a user asks is still not fully here.
Google Research scientific-progress conversation, with YouTube: Logan Kilpatrick shared a conversation with Google Research head Yossi Matias about AI accelerating scientific progress, improving real-world lives, and moving research into a golden age.
Google DiffusionGemma and local inference, with Google, HF Space: davanstrien/diffusiongemma-ocr-correction, Hugging Face: unsloth/diffusiongemma-26B-A4B-it-GGUF, Unsloth, @UnslothAI: Google released DiffusionGemma, a diffusion-style Gemma-family text model that refines token blocks in parallel, while Unsloth published local GGUF (a local-model file format) weights and docs for running it on consumer hardware.
PaperMentor writing tutor, with Overleafmentor, GitHub: jiarui-liu/overleaf, @Jiarui_Liu_: PaperMentor is a multi-agent writing tutor for AI research papers inside Overleaf that gives inline, expert-style comments rather than rewriting the paper.
Shortcut research-product lesson: Robert Yang reflected on the mistake of confusing cool research projects with products, arguing that focused API-based products outperformed broad demos and fragile GUI agents.
Research skill field guide: Vivek laid out a practical field guide for research: choose original problems, improve inputs, write everything down, build fast loops, study failures, collaborate generously, and compound over time.
No-CoT task horizon paper, with arXiv paper, Lesswrong: The no-CoT task-completion horizon paper estimates how long frontier models can complete tasks without visible chain-of-thought, with implications for hidden reasoning and safety oversight.
Trajectory-Refined Distillation, with AlphaXiv: Trajectory-Refined Distillation fixes on-policy distillation failures by having a teacher rewrite an incorrect student reasoning trajectory into a corrected full trajectory before distillation.
ECHO environment modeling: Prime Intellect shared ECHO, an approach that integrates environment modeling with agentic actions by combining supervised learning on tool outputs and reinforcement learning on assistant tokens.
Arbor autonomous research agent, with @XiaoxiLi0111, GitHub: RUC-NLPIR/Arbor, Ruc Nlpir, @HuggingPapers, Papers with Code, @HuggingPapers, @_akhaliq: Arbor is an autonomous research framework that uses persistent hypothesis-tree refinement to make long-horizon research cumulative, reporting stronger results than Codex and Claude Code on research tasks and strong MLE-Bench performance.
Cohere Transcribe benchmark: Cohere said its open-source Transcribe model topped a far-field speech recognition benchmark while remaining Apache 2.0 (a license that lets others use and modify it freely) and locally runnable.
Walk on Spheres Extensions, with Rohan Sawhney, @keenanisalive: Rohan Sawhney released Walk on Spheres Extensions, a GPU-accelerated library for grid-free Monte Carlo physics simulation on complex geometry, with Keenan Crane highlighting the mesh-free simulation advantage.
Illusion of Superposition latent-thinking paper, with ArxivIQ, arXiv paper: The Illusion of Superposition paper and breakdown argue that latent chain-of-thought does not reliably produce hidden parallel reasoning in existing models and often collapses to shortcuts.
DiffusionGemma OCR correction demo: Daniel van Strien built a Hugging Face Space comparing DiffusionGemma and Gemma-4 on post-OCR correction, showing diffusion-style denoising for noisy historical text.
WorldString neural world representation: Xueyan Zou and collaborators introduced WorldString, a compact neural representation for dynamic world instances such as articulated motion, skeletal skinning, and soft-body deformation.
SciConBench scientific conclusion synthesis, with arXiv paper, GitHub: hayoungjungg/SciConBench, HF dataset: hayoungjung/SciConBench: SciConBench asks whether AI agents can synthesize scientific conclusions from literature, with a paper, code, and dataset for evaluating conclusion synthesis.
Anthropic and unreleased internal-model take: Joscha Bach argued that Anthropic is not unique in withholding strong models, noting that Google has long kept superior internal coding models unreleased.
SearchSwarm deep research agent, with Papers with Code, Hugging Face: SearchSwarm/SearchSwarm-30B-A3B, HF dataset: SearchSwarm/SearchSwarm-SFT: SearchSwarm teaches a model to decompose long-horizon research tasks, dispatch subtasks to subagents, and integrate evidence-grounded results for stronger BrowseComp, GAIA, and deep-search performance.
DiRL reasoning-vs-memorization reinforcement learning paper, with @sheriyuo: The DiRL paper argues that diversity exploration in reinforcement learning (training by reward signals) for large language models (AI text models) can confuse new reasoning paths with memorized pattern reshuffles, and proposes direction-aware rewards to favor genuine reasoning.
Claude on Apple Foundation Models: Anthropic announced Foundation Models support for Claude on Apple platforms, letting developers call Claude through Apple’s native framework with typed outputs and SwiftUI integration.
Claude equity-research benchmark: Lilly Sharples argued that models can replicate the outputs of sell-side research but not the accountability, tacit knowledge, or instinct behind durable alpha.

💡 Industry Commentary & Analysis

ML job interviews guide, with @silviasapora: Silvia Sapora published a detailed guide to machine-learning research-scientist interviews based on her own job search across top AI labs, covering preparation, interview logistics, negotiation, and emotional reality.
Unsloth MTP guide: Unsloth published guidance for running Multi-Token Prediction models, which draft multiple future tokens to speed generation while a main model verifies the draft.
Anthropic enterprise spend take: Marty Kausas argued that Anthropic enterprise usage changes are forcing teams to add visibility, approval gates, and ROI discipline around Claude spend.
OpenAI vs Anthropic openness take: Xin Eric Wang argued that OpenAI has contributed more to public research than Anthropic, framing openness and reproducibility as better than secrecy and gatekeeping.
François Chollet bubble take: François Chollet argued that bubbles can exist even when the underlying technology is useful, profitable, and in demand because bubbles are about investor psychology and price dynamics.
Anthropic org structure take: Rohan Paul highlighted Anthropic’s unusually flat executive structure, with Dario Amodei having one direct report while Daniela Amodei runs daily operations.
Sloptimization of AI search, with @nxthompson: Will Oremus’s Atlantic piece argues that companies are gaming AI search and chatbots with machine-targeted content, self-promotional listicles, hidden prompts, and sock-puppet campaigns.
Frontier subscription economics, with @kimmonismus: SemiAnalysis argued that frontier AI subscription plans can be heavily subsidized relative to API pricing when power users exhaust them on long-horizon tasks.
AI consuming human thought essay: John Nosta argued that AI is not replacing human thought so much as consuming it, warning that recursive AI-generated training data can erase rare human variance and originality.
Microshifting work-trend essay: Aytekin Tank argued that employers should not fear microshifting, or breaking work into smaller bursts with personal time in between, when teams preserve deep-work blocks and collaboration guardrails.

💻 AI Coding, Agents & Developer Workflows

Coding agents podcast, with Spotify, YouTube: Cursor’s Sam Whitmore shared a conversation with Baseten researchers about running many coding agents in parallel and how agent-heavy research and engineering workflows are evolving.
Fable edits its own launch video, with Thariqs: Thariq showed Fable editing its own launch video by coordinating transcription, ffmpeg cuts, color grading, Figma Model Context Protocol (the tool-connection standard for AI agents), Remotion components, and rendering without a traditional video editor.
Claude Code team workflow shift: ClaudeDevs said Fable 5 shifted the Claude Code team’s role from verifying that Claude did work right to verifying it was doing the right work.
Fable Whoop side hustle: Pankaj built a side project that reverse-engineered Whoop data, matched heart-rate spikes to calendar events, and created a private coworker stress leaderboard.
Agent scripts for repository maintenance, with GitHub: steipete/agent-scripts, @steipete, GitHub: steipete/agent-scripts: Peter Steinberger shared agent-scripts skills for autonomous repository maintenance, GitHub project triage, worker delegation, review, proof, and release workflows.
Self-Harness agent improvement, with AlphaXiv: Self-Harness lets a large language model (AI text model)-based agent improve its own operating harness through weakness mining, targeted proposal generation, and regression-tested validation.
Fable San Francisco map demo: Nicolas Bustamante used Claude Fable to generate a detailed offline HTML map of San Francisco from public datasets, including streets, transit, terrain, buildings, and fog.
Xiaomi MiMo Code, with MiMo, MiMo, GitHub: XiaomiMiMo/MiMo-Code, @_LuoFuli: Xiaomi released MiMo Code, an open-source terminal coding agent for long-horizon programming with persistent memory, Compose workflows, self-evolution, voice input, and Claude Code compatibility.
confBuild turbofan demo, with confBuild: Daniel built a parametric animated turbofan engine in the browser with confBuild and Claude Fable 5, including moving structure, sliders, export, and drawings.
Still KV cache compaction, with arXiv paper: Still is a long-context efficiency paper/demo from Charles O’Neill and Baseten that uses a small per-layer model to compress key-value cache memory in one forward pass, making long-context agents cheaper and faster.
Use Computer sandboxes, with Use Computer: Use Computer provides infrastructure for computer-use agents: instant macOS, iOS, iPadOS, watchOS, and Ubuntu sandboxes on M4 silicon with SDKs, GUI control, screen reading, file transfer, and trajectory recording.
Fable 5 neuralese/codename behavior: Tamay Besiroglu observed that Fable 5 sometimes emits private internal codenames or gibberish in coding contexts, suggesting a theory-of-mind failure around what the user can understand.
Claude Code model-routing setup: CJ Zafir shared a workflow for routing planning, execution, and review across Fable 5 and Codex to reduce limit pressure while preserving quality.
Hermes Agent web dashboard and iMessage support, with Hermes Agent, @ryanzhuuuu, @ryanzhuuuu, @Teknium: Nous Research and partners pushed Hermes Agent toward configurable personal assistants, with a web dashboard for profiles and credentials plus reliable iMessage support across operating systems.
Fable QDD actuator CAD demo: Jake Fitzgerald showed Claude Fable 5 designing and validating a QDD actuator with CAD-style geometry, gearbox animation, and collision checking in one long session.
Devin handoff plugin, with GitHub: club-cog/devin-handoff, @imjaredz: Cognition and Jared Zoneraich open-sourced Devin handoff, a plugin/skill that packages local git context and hands coding-agent work to a cloud Devin session.
AI assistants in iMessage market map: Justine Moore argued that text messages could become the next major consumer AI interface and mapped iMessage assistants plus infrastructure providers.
Fable delegation workflow: Daniel McAteer shared a Claude Code workflow that uses Fable as an orchestrator and delegates heavy reasoning to other Claude models to preserve usage limits.
Claude Fable /improve command: shadcn introduced /improve for Claude Fable, a workflow that audits a codebase or branch and writes executable plans for bugs, performance, tech debt, tests, and issues.
Fable 5 app-clone demos, with @dabit3: Nader Dabit showed Fable 5 and Devin creating polished functional clones of apps like Notion, Figma, Obsidian, and Screen Studio from simple prompts.
VoxCode local voice coding workspace, with GitHub: Arindam200/awesome-ai-apps: VoxCode is a local voice-powered coding workspace built on Cursor harnesses, enabling voice codebase summaries, architecture Q&A, file edits, and local-control workflows.
Fable sheet-metal CAD demo: Brian Ratliff used Fable 5 and a Claude CAD setup to generate a professional-style 69-part sheet-metal enclosure with realistic components and specs.
OpenAI usage share after Fable launch: Dylan Patel reported that OpenAI usage share grew versus Anthropic after the Fable/Mythos launch as some power users tried Codex following Claude refusals or strange behavior.
Cursor Bugbot speed and cost improvements: Cursor said its code-review agent Bugbot is faster, cheaper, and catches more bugs, and can now run locally with a /review command before code is pushed.
Fable Titanic game demo: Misbah Syed built a playable 3D Titanic iceberg-avoidance game with Claude Fable 5, three.js, and Netlify from a small number of prompts.
Loss Function Development for coding agents: Elvis Sun shared a playbook for long-running autonomous coding-agent optimization using measurable targets, blind evals, hard constraints, anti-overfit checks, and stall jumps.
Every Claude Fable 5 prompt library, with Every: Every published a prompt library and interview context for Claude Fable 5, packaged as practical copy-paste prompts and usage tips.
John Carmack coding-style take: John Carmack argued that code style could be optimized for helping weaker models succeed reliably, likely overlapping with what helps humans understand code at a glance.
Mitchell Hashimoto Fable use-case take: Mitchell Hashimoto argued that Fable is not worth the cost for broad architecture or design work, but is unusually strong for narrow, high-value surgical optimization loops.
Fable 5 translation-test critique: Valerio Capraro argued that Fable 5 failed a professional translation test by literalizing a phrase instead of understanding the meta-linguistic update required.
Claude Code Auto mode explanation: Delba Oliveira explained Claude Code Auto mode, which routes long-running tasks through a classifier that can allow safe actions, block risky ones, or request approval.
KV cache explainer, with @akshay_pachaar: Akshay Pachaar explained how KV caching speeds autoregressive generation by storing key and value vectors after prefill, trading speed for GPU memory (graphics-chip memory) use.
Ghostty black-hole shader, with GitHub: s0xDk/ghostty-blackhole: Developer s13k built a black-hole shader for Ghostty that distorts terminal text and eventually nudges users to take breaks or compact Claude Code context.
Claude Code memes search page: This X search page is a discovery feed for Claude Code and Fable memes, failure screenshots, viral one-shot builds, and community jokes around agentic coding.
Claude Code vs Codex comparison: Hesamation argued from testing that Claude Code remains stronger for many developers than Codex because of task horizon, senior-dev feel, ecosystem, pricing, and implementation quality.
Claude/Codex Lottie animation skill, with GitHub: diffusionstudio/lottie: Konstantin Paulus and Diffusion Studio released a text-to-Lottie harness so Claude Code or Codex can generate production-ready Lottie animations from text, SVGs, or data.
Claude Fable dashboard prompt: Matt Shumer shared a prompt pattern for long-running Claude Fable tasks that creates a persistent HTML progress dashboard with screenshots and media.
Claude Fable Minecraft clone: Chris used Claude Fable 5 to one-shot a Minecraft-style game with biomes, ores, caves, and day/night systems.
DataEmpire analytics game: Marc Lou used Claude Fable 5 to build DataEmpire, a live game that turns website analytics into a growing empire simulation.
Fable mobile/web app builder demo: Riley Brown one-shot a Replit-like mobile/web app using Claude Fable, Daytona sandboxes, and Convex for database services.
Hyperbrowser harness plugin: Hyperbrowser’s harness plugin runs agents against projects, watches failures, and writes a CLAUDE.md from mistakes so the agent avoids repeating them.

Previous Around the Horn Digests

Catch up on everything you missed:

Monday, June 8, 2026: ChatGPT memory, Hermes Desktop, new Codex plugins, MAI-2.5 image rumors, and the agent stack kept shifting under everyone’s feet.
Wednesday-Thursday, June 3-4, 2026: Anthropic said Claude was already accelerating Claude, NVIDIA shipped Nemotron 3 Ultra, OpenAI upgraded memory, and Devin got a $10M enterprise dare.
Tuesday, June 2, 2026: earlier-week roundup coverage from the same Everything That Happened in AI Today stream.
Claude Fable Five is Anthropic’s Most Controversial Model Yet: the immediate newsletter prequel to today’s model-switching fight.
Apple finally rebuilt Siri: Apple’s AI rebuild after WWDC 2026.
ChatGPT is about to look completely different: the interface-change edition before today’s agent/tool pileup.
Google Gemini got hijacked via WhatsApp: the security story that made prompt injection feel very non-theoretical.

That’s a Wrap

That's 343 source entries across 157 topic clusters from one extremely normal day in AI, assuming your definition of normal includes secret model routing, AI-assisted vulnerability deadlines, crypto-trading agents, and at least seven separate ways to make Claude build software. If you made it to the bottom, you are legally allowed to say “context window” at dinner and refuse to explain it.

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.