Everything That Happened in AI Today Wed-Thur, June 3-4

Anthropic said Claude now writes most of Anthropic's code, NVIDIA dropped a giant open agent model, and the rest of the day felt like everyone sprinting toward the same finish line.

Welcome to the Around the Horn Digest, your one-page attempt to sound dangerously informed after an AI news day that refused to pick a lane. The official headline was Anthropic saying Claude is already accelerating Claude. Around it, NVIDIA pushed open models toward long-running agents, Google tried to put a strong multimodal model on your laptop, OpenAI gave ChatGPT a fresher memory system, and Cognition basically dared enterprises to measure whether Devin is worth the money. At some point, the industry stopped shipping features and started shipping proofs of acceleration. Let's get into it.

Around the Horn: Thursday, June 4, 2026

The lead story was Anthropic's RSI data, because it turned the abstract “AI will help build better AI” debate into operating-company data. Anthropic said Claude now authors more than 80% of merged code in its own codebase, typical engineers are shipping about 8x more code than prior baselines, open-ended coding success rose from roughly 26% to 76% in six months, and recursive self-improvement test speedups jumped from about 3x to 52x.

The important part is the mechanism, not the sci-fi label. Recursive self-improvement means an AI system helps build the next, more capable AI system. Anthropic is saying that loop is already visible in code, research planning, and training-code optimization, while outside observers like Ethan Mollick, Yuchen Jin, Andrew Curran, and Jasmine Wang focused on the organizational and coordination problem: even if a lab wants to slow down, any meaningful pause would need verification across the whole frontier model race (the competition to build the most capable AI systems).

🏆 TOP 5 NEWS (Around the Horn)

NVIDIA released Nemotron 3 Ultra, a 550B-total / 55B-active open model (it has 550B parameters total, but uses about 55B per request to run cheaper) for long-running agents, with a 1M-token context window, downloadable weights, tuning recipes, Unsloth GGUF builds, NVFP4 weights, and discussion on cheaper agent fleets from Joshim5.
OpenAI rolled out a new ChatGPT memory system that keeps useful context fresh across chats and gives users a summary they can review and steer.
Google released Gemma 4 12B, a laptop-friendly multimodal model (it handles text, images, and audio) that can run with 16GB VRAM (graphics memory) or newer-Mac unified memory.
xAI released Grok Build 0.1, a fast coding model for agentic software work, with a 256K-token context window, $1 input / $2 output per million-token pricing, Vercel access, xAI docs, and console links for API keys and the chat playground.
1X launched World Model Lab, arguing humanoid robots need world models (systems that predict how the physical world changes) rather than more fine-tuning alone (extra training on a narrower dataset).

Honorable Mentions

Cognition backed Devin with up to $10M in credits if the software-engineering agent delivers less measured value than an enterprise customer pays for.
Meta began building AI data centers in tents to speed infrastructure buildout.
Pinterest committed $4B to AWS through 2031 for AI-driven visual search, recommendations, multimodal models (models that handle images and text), and conversational discovery.
Perplexity and the U.S. Small Business Administration launched a $25M Main Street AI Accelerator for eligible small businesses, with Perplexity Computer credits and a related Perplexity announcement.
Coralogix raised $200M to build monitoring and troubleshooting infrastructure for production AI agents.
Major AI labs backed stronger screening for synthetic DNA/RNA providers (companies that manufacture custom genetic material), while IFP and FAI argued the DNA supply chain is the highest-leverage chokepoint for AI-enabled biothreats.
US Rep. John Moolenaar introduced the GUARD Act to ban Chinese-made robot dogs and humanoids, citing espionage risks and possible PLA-accessible backdoors.

🍪 TOP TREATS TO TRY

Raindrop monitors production AI agents, catches silent failures, alerts teams in Slack, and tries to fix issues automatically; no pricing details.
Locally lets you access your largest desktop local models from an iPhone or iPad through an end to end encrypted LM Link connection (LM Studio teaser, launch details); no pricing details.
Stemdeck separates a song into vocals, drums, bass, piano, and guitar stems (individual audio tracks) from a YouTube link or MP3, with local processing, BPM/key detection, and per-stem exports (How To AI demo); free / open-source.
Higgsfield MCP turns Claude into a company-building workflow for brand identity, app screens, motion videos, founder posts, ads, and viral-potential testing (Higgsfield demo); no pricing details.
Tasklet for Teams turns personal agent workflows into shared company workspaces for tools, knowledge, agents, model access, billing, and usage limits (workspaces, members, knowledge); no pricing details.
OpenProse turns agent workflows into reusable .prose.md programs (plain-English instructions that agents can run, review, version, and audit); no pricing details.
Spiral 4.0 learns your writing style from examples and gives teams, agents, CLI (command-line), API, and MCP (tool-connection) access to that voice system (Every update); starts at $15/mo.

🏢 Big Tech & Major Companies

U.K. regulators required Google to offer publishers an opt-out tool for generative AI search features, with U.K. testing first and a global rollout planned.
Meta released a Facebook creator assistant that answers questions about when to post, what commenters are saying, and how content is performing; no pricing details.
Lovable signed an expanded multiyear Google Cloud deal to increase usage 5x and gain broader access to Anthropic Claude.
two voice-AI founders left Goldman Sachs and Meta to build for Africa and the Middle East, with their stack now handling more than 17,000 calls per day.
Amazon began showing AI-generated product images in search results, using visual search and generative images to match shoppers' queries to products.
Microsoft released Adaptive Spec-driven Scoring, a framework that creates AI behavior tests and regression evaluations from text descriptions; free / open-source.
Meta's WhatsApp Business agent lets businesses automate customer conversations globally, with pricing based on token usage.
Microsoft Scout brings OpenClaw-style personal assistant capabilities into Microsoft 365; no pricing details.
U.S. Defense Department denied Anthropic's request to reverse its national-security blacklisting as a supply-chain risk, leaving Anthropic's federal legal challenge ongoing.
The Financial Times reported that the NSA is using Anthropic's Mythos model for cyber operations, while outside discussion focused on anonymous sourcing and what government deployment means for frontier model oversight.
Meta The Information reported that Meta is considering charging up to $200/month for its planned consumer Hatch AI agent, with tiered usage limits.
Reddit SEO report reported that companies are spamming subreddits with AI optimized posts to manipulate ChatGPT and Google AI search results.
The Financial Times examined whether Anthropic can preserve its safety-first founding principles while scaling into a major commercial AI company.
Financial Times reported that AI legal-tech challengers like Harvey, Legora, and Anthropic are pressuring Thomson Reuters and Relx/LexisNexis, while incumbents respond with product upgrades and trusted legal content.
Google introduced Search profiles that let creators, publishers, and brands claim a dedicated space to highlight their work and online presence in Search.
Nicky observed that Google started citing and ranking Vellum AI at #1 across AI Overviews, AI Mode, and Gemini after recent search updates.
Codex plugins OpenAI previously published role-specific Codex plugin templates on GitHub as example builds for customizing coding agents with defined roles and behaviors.
Logan Kilpatrick shared that the team is building what they believe will be the world’s best vibe coding app for Android and iOS.
Microsoft launched MAI-Image-2.5, which debuted at No. 2 on the Arena leaderboard for image editing.
Prince Canuma and others shared practical demos and usage of MLX-VLM for local VLM inference and fine-tuning on Apple Silicon.
Kwindla highlighted NVIDIA's Nemotron 3.5 ASR as a low-latency, open multilingual speech-to-text model that voice-agent builders can self-host more cheaply.
Zhengyao Jiang said the Aiden autonomous research agent beat all 1,016 human participants in OpenAI's Parameter Golf challenge by filing high-impact pull requests and setting multiple records over 22 days.
Georgi Gerganov highlighted recent advances in multi-GPU and tensor parallel support in llama.cpp, where maintainers and NVIDIA engineers improved ggml multi-GPU performance for significant gains on RTX systems and hardware-agnostic tensor parallelism groundwork (quoted NVIDIA RTX Spark post on on-device AI agents; likes ~394, reposts ~49).
Palantir showed the USDA using its Ontology system for national food-supply security, arguing governed data layers make AI more auditable than disconnected databases.
Notion highlighted a Custom Agents build that recreated The Office in under 48 hours, using long-term context, character reflection loops, and proximity reasoning.
OpenAI Developers upgraded the Build iOS Apps plugin in Codex so it can view and test apps in an in-app browser, open SwiftUI previews, and hot-reload edits.
AWS launched Startup Advisor, a builder assistant that gives founders architecture guidance inside their IDE, suggests AWS services, warns against overbuilding, and creates migration plans.
ElevenLabs gives you lifelike speech, voice agents, APIs, and SDKs with 5,000+ voices across 70+ languages; free to try.
Amazon built a next-generation Proteus warehouse robot that workers can speak to for routing and prioritization, then framed it as worker-support automation inside a €10B European fulfillment investment.
Meta kept delaying developer access to its latest AI model, with CNET noting developers are still waiting for the API and tools needed to build on it.

💼 AI Productivity, Labor & Economics

Cyera is seeking a $12B valuation at roughly 80x ARR while raising a $300M round led by Evolution Equity Partners despite operating losses.
Phoenix data centers The Wall Street Journal reported that Phoenix is becoming a test case for AI power costs, with Arizona's largest utility proposing a 45% rate hike for data centers and 14.5% for households.
Applied Materials plans to expand its Southeast Asia workforce by 25% this year to capture growing chip-equipment demand.
Artificial Lawyer argues that token costs plus enterprise AI licenses could become a major law-firm budgeting issue by 2029 as agent workflows and power-user adoption expand.
Foxconn and Intel announced a strategic collaboration to develop AI infrastructure, including server racks, interconnects, cooling, energy-efficient designs, and custom chips for data centers and edge deployments.
Lisan al Gaib said another API appeared to offer claude-oceanus-v1-p and suggested Mythos / Oceanus pricing may land around $16 per million input tokens and $80 per million output tokens.
Gautam Kamath and collaborators won the 2026 Gödel Prize for robust high dimensional statistics work that made it possible to learn from corrupted data without accuracy falling apart as dimensions grow.
Carina Hong closed Axiom Math's $200M Series A at a $1.6B valuation and argued Lean-style formal verification (machine-checkable math proofs) is the bottleneck for compounding mathematical AI.
QuiverAI highlighted Synth GEPA, a prompt-optimization method that uses agents as proposers and searches for better interventions across tasks.
Synth GEPA scales compute for prompt optimization and coding agents by using agents as proposers to generate interventions, tracking a Pareto frontier across objectives (performance, cost, time), and running tasks in modular containers that expose a simple HTTP contract (open-source on GitHub).
- Synth GEPA uses agents as prompt-improvement proposers, then searches across performance, cost, and time to find stronger coding-agent interventions in modular task containers.
Amit Tandon argued model routing is becoming core AI architecture as companies like Factory, Harvey, and Rekursor choose models dynamically by cost, quality, advisor type, or inspectable skill.
Dwarkesh Patel shared Sasha Rush explaining targeted on policy self-distillation, a training trick that places hint tokens at an agent's exact failure point so the model learns to avoid that mistake later. (2,106 likes, 137 reposts)
Mastra launched Agent Builder, a governed agent-building platform where developers provide tools and workflows while non-developers assemble, test, and publish agents through a browser UI or natural language (docs, announcement, Alex Booker).
Flo Crivello moved 100% of Lindy traffic from Anthropic to DeepSeek v4, saying Lindy saved millions while improving many core workflows after building internal tooling that makes model swaps easier. (1,919 likes, 111 reposts)
Anthropic explained its Claude analytics setup: governed datasets, a semantic layer (plain-English definitions for data), reference docs, offline evals, adversarial review, provenance tracking, and correction loops helped push self-service accuracy above 95%.
Esan Durrani introduced Honen, a reskilling platform that turns company documents into interactive courses, and said the company is partnering with NVIDIA to deliver AI literacy training to 250,000 learners.
The Financial Times reported Americans show the lowest support among 15 large economies for AI data-center expansion, while Heatmap found U.S. opposition to nearby data centers has surged since last fall.
TSMC warned AI-fueled chip demand will outstrip supply for years, while CEO C.C. Wei told shareholders the company will keep prices stable; Taipei Times framed the same pressure as a near-term AI supply bottleneck.
The Financial Times reported France’s €110B AI buildout is testing Macron’s tech ambitions as approvals and local opposition threaten to slow data-center projects.

🤖 AI Agents & Infrastructure

MIT released a new study covered by Ars Technica that found Waymo robotaxis drive empty 44% of the time, roughly matching ride-hailing deadheading and undercutting claims that robotaxis reduce traffic.
Airspeed uses agents to process customer conversations, emails, support tickets, and CRM data into follow-ups, risk flags, CRM updates, and deal-moving actions; it raised $20M and reports 200 customers across 20 countries.
Julien C. flagged the same agent-tooling wave as Hyperbrowser /web and Mastra Agent Builder: developers are giving agents structured maps of websites, tools, and workflows instead of asking them to guess from raw pages.
Sarah Wang shared Exa CEO William Bryk's argument that agents need search built for exhaustive, fast, filterable semantic and keyword retrieval, rather than human-style top-10 search results. (15 likes)
Hugging Face updated its hf CLI as an agent-friendly way to work with the Hub from scripts, terminals, and automated workflows.
Lassie is building AI office operations for medical practices and says it is already used by 700+ practices across the U.S.
Jongwon Park discussed systems that convert real repositories into reinforcement-learning environments for software agents, stressing continuous QA to prevent reward hacking.
Daniel Kundel showed Codex goal mode, where you set a verifiable outcome and the agent keeps working until it reaches the goal, using numeric exit criteria, progress checks, and final review. (202 likes, 13 reposts)
n2parko built a Slack-powered /canvas inside Cursor's publishable canvas feature, turning agent output into a shareable internal dashboard or tool. (15 likes)
SemiAnalysis reported that agentic traffic has surpassed human traffic for HTML webpages worldwide according to Cloudflare Radar data (likes: 1026, reposts: 114).
Cameron Wolfe shared that sliding-window context compaction performed best (surprisingly) versus summary-based or append-only strategies when training smaller agents with GRPO on GSM8K, and called for stronger evals of compaction techniques especially for long horizon coding (links to Agent-R1 paper; likes: 44).
Haozhe Zhao released Crafter, a multi-agent system for turning text, PDFs, sketches, masks, and key elements into editable scientific figures, posters, and infographics (GitHub).
Remoroo automates robot-training data collection by running unattended real-world episodes, varying poses and lighting, logging outcomes, and producing verified training-ready datasets.
Ben Burtenshaw shared a runnable single-GPU script for improving agents from their own traces (logs of what an agent tried), using Gemma 4, evaluator feedback, and a LoRA adapter (a small model add-on). (31 likes, 10 reposts)
Clement Delangue argued agent traces should default to private Hugging Face storage so teams can preserve run history, debug failures, and post-train models (teach later models from prior runs). (82 likes, 15 reposts)
Astro open-sourced Flue, a Claude Code framework that uses virtual sandboxes instead of real containers to run headless agent fleets (agents running without a visible app window) more cheaply. (19 likes, 3 reposts)
Lindy Drope described a “Vercel for internal agents” where employees deploy Claude Code agents while IT controls hosting, permissions, execution logs, and per-agent token spend (AI usage costs). (325 likes, 9 reposts)
ElevenLabs Developers built a Hermes agent that can answer phone calls through ElevenLabs Conversational AI using OpenClaw tools, memory, and skills over an OpenAI-compatible API; the follow-up added turn-taking, TTS/STT (text-to-speech and speech-to-text), and Twilio phone support.
datapointai highlighted the agent-evaluation gap that Agent Arena is trying to solve: measuring live task success, steerability, error recovery, and tool hallucination instead of relying only on static benchmarks.
OSU-NLP Group introduced ACuRL, an autonomous curriculum-learning method that lets computer-use agents improve through exploration and generated tasks, with the GitHub repo available for developers.
Alibaba is turning Qwen into a digital concierge for everyday tasks, from ordering fried chicken to booking flights, while opening Qwen to third-party agents as Tencent readies a WeChat rival.
mvanhorn highlighted the Suno agent-native CLI, which gives Claude Code, Cursor, Gemini CLI, and Copilot an offline-first library, local SQLite database, skill, and MCP server for controlling Suno.
ParseBench launched a document-parsing benchmark for agents with 2,000 human-verified pages, 169K deterministic rules, and five capability dimensions, while the paper, LlamaIndex, and Jerry Liu framed it as a test of whether parsers preserve tables, charts, formatting, and visual grounding.
AutoLab benchmarked 36 long-horizon research tasks and found frontier agents succeed most when they persist, benchmark, and fold in feedback, with DAIR.AI spotlighting the same lesson.
NVIDIA announced physical-AI agent skills powered by Cosmos 3 for data generation, simulation, policy training, and evaluation, while Omar Sar connected it to the broader model-to-workflow shift alongside Nous Research Nemotron/Hermes access.
Matt Gunnin said OpenClaw’s hindsight skill beats Supermemory, Mem, and other memory tools for Hermes agents, then described a shared cortex where Ella and Lyra collaborate for hours and only ping him for high-stakes decisions.
Hello Robot released the fourth-generation Stretch home-assistance robot, continuing its bet that general-purpose robots can work inside real homes.

💻 AI Coding & Developer Tools

Hedgie argued UC Berkeley's CS 10 failure rate jumped to 35.3% as students used LLMs for assignments, skipped learning, and then failed exams, creating a pipeline problem for future engineers.
Firecrawl said it has fetched more than 8B pages, reached 1.25M developers, and passed 2.5M weekly downloads two years after launch.
OpenAI moved Codex toward token-based billing with pooled credits across plans, with Ed Zitron noting the shift makes Codex usage look more like API pricing after introductory periods.
Anthropic open-sourced a defending-code reference harness with skills for threat modeling, scanning, triage, patching, and autonomous code-security scans, while _mattata called it a clean template that still needs more pieces for full vulnerability work.
Justin Lebar used AI to fuzz GPU and CPU compilers (stress-test them with random inputs), spending about $10K and finding dozens of AMD and x86 LLVM miscompilation bugs; SemiAnalysis published the full writeup.
Vishakh Padmakumar introduced Offloading Score, a counterfactual measure of how much cognitive work developers hand to AI, with the GitHub repo, paper, and project page showing 43% higher reliance under time pressure.
Matt Gunnin said OpenClaw’s autoreview skill replaced his branch / PR / CI / CodeRabbit flow, while eglyman surfaced the same local-review loop through Greptile CLI and ClaudeDevs tied it to reusable Claude Code-style skills.

🔬 AI Research & Models

SemiAnalysis argues that space datacenters become economically plausible only under severe terrestrial power and chip constraints, with its model showing space compute about 4x more expensive today and base-case parity closer to 2040.
European Commission The European Commission asked households to reduce peak-hour electricity use as AI data centers strain grids, with Ireland's data centers already consuming 22% of national power.
Liquid AI announced ShieldFlow, an on-device privacy layer for AI apps that keeps sensitive processing closer to the user's device, available through an early-access request form.
alphaXiv highlighted Trust Region On-Policy Distillation, a training method that stabilizes long reasoning by focusing on reliable teacher reasoning paths and handling outliers separately.
Depth-Attention gm8xx8 shows that Depth-Attention adds cross-layer reuse inside Transformers by letting queries attend to earlier-layer keys at the same position and mixing values into the normal V-cache slot, improving Qwen3-style 1.5B-3B models by up to +2.3 accuracy points with <0.01% extra compute and no added parameters or persistent state (likes ~49).
Epoch AI Anson Ho et al. (Epoch AI) released “A Rosetta Stone for AI Benchmarks”; a statistical framework that places model capabilities and benchmark difficulties on a single comparable scale, enabling long-run trend measurement, forecasting, algorithmic efficiency estimation, and detection of accelerations without strong assumptions about capability growth.
floatingpoint launched a dedicated effort to build the data layer for pixels and vision, moving from document AI into infrastructure for understanding model behavior on images and video. (63 likes, 11 reposts)
Asymmetric VLM post-training found that standard post-training improves reasoning more than perception in vision-language models (models that read images and text), then proposed loss reweighting and perception-aware rewards to close the gap.
Merve shared a Gemma 4 + Hermes / Claw fix: put text after images because the model expects a specific order for mixed image-and-text messages. Merve apparently switched from Qwen3.6 35B 8-bit quantization to Gemma 12B BF16 for local coding with Hermes, trading model size for a higher-precision local setup.
MLX-VLM helps Mac users run and fine-tune vision-language models with MLX, Apple's machine-learning framework.
Microsoft AI pointed developers to UserLM-8B, an MIT-licensed model that simulates realistic multi-turn users so teams can test assistants before putting them in front of people.
NVIDIA released ArtiFixer code and weights, a SIGGRAPH 2026 system that repairs 3D reconstruction artifacts and extends novel views, with Haithem Turki noting it beats prior methods while preserving generative ability.
Google Research showed smartphone front-camera videos can estimate heart rate and resting heart rate after face unlock, with under 10% average percent error and under 5 bpm error across skin tones on 350K+ clips.
Flourish is using wet-lab neuroscience and brain imaging to hunt for the brain’s core learning algorithm, and reportedly raised $500M at a $2.5B valuation with backing from Jeff Bezos, GV, Lux Capital, and Catalio.
Q Labs introduced q0, a hyper-epoch pretraining approach that trains diverse model populations and distills them into a more data-efficient model, with industriaalist highlighting the snapshot-and-distillation angle.
MAI-Thinking-1 argues AI progress needs a hill-climbing machine that treats model development as a system-level optimization loop, with alphaXiv highlighting the paper’s compounding-improvement thesis.
Brian Christian and co-authors proved the optimal policy for Feynman’s restaurant problem from handwritten notes, and Christian said a 2,520-person study found humans use near-optimal shortcuts.
Science published a petavoxel fragment of human cerebral cortex reconstructed at 4 nm resolution from 1.05 mm³ of temporal cortex, with Hao Yin highlighting the nanoscale H01 dataset.
Alex Nam, Aimee Li, and Natasha Jaques released MIPO, while Natasha Jaques explained that its contrastive preference pairs improve personalization and reasoning without extra data or verifiers.

🏛️ AI Policy, Governance & Safety

Javier Milei argued that AI should develop without premature regulation as Argentina tries to position itself as a friendly home for AI advancement.
synthwavedd reported Anthropic paused Mythos red-teaming (adversarial safety testing) after a red teamer allegedly resold Oceanus checkpoint access through a Chinese API proxy. (1,781 likes, 78 reposts)
Andrew Curran tied OpenAI policy work and Anthropic recursive self-improvement data to the same verification problem: a frontier model slowdown only works if labs can prove everyone else is slowing too (earlier context).
Séb Krier joked that AI is moving too fast for society to adapt safely, saying the ideal lab speed is 56 kbps because institutions can only adapt at dial-up pace. He added that summer breaks already make conformity assessments hard to run, so the tech may have to wait anyway.
OpenAI published a policy agenda covering frontier model safety, recursive self-improvement monitoring, youth protections, synthetic CSAM safeguards, provenance, election/deepfake protections, and workforce transition, then shared it on X.
Sam Altman said he does not plan to donate to the 2026 U.S. midterms, unlike some other Silicon Valley billionaires.
Canada launched AI for All, a five-year national strategy for trust, opportunity, sovereignty, safety legislation, AI literacy, jobs, healthcare missions, and public compute; the official strategy page lays out the plan, while Canadian AI observers praised adoption goals but asked for clearer timelines and governance details.
AI super PACs have already spent nearly $24M and promised more than $100M more to influence the 2026 midterms and AI regulation.

🛠️ AI Tools & Products

Founders Fund posted MAFIA EP 001, a social-deduction video where tech figures play Mafia and try to identify the hidden traitors.
Jerrod Lew launched Reve 2.0, which generates 4K images automatically split into editable layers for reprompting and layout control.
NotebookLM added Source Attribution so users can see the exact prompts and sources behind each generated artifact, then iterate directly from that evidence.
Boson AI released Higgs Audio v3, a text-to-speech model supporting 100 languages with controllable emotion, speaking style, prosody (voice rhythm), and sound effects, available through API, Workspace, and open weights. (134 likes, 23 reposts)
Poke brings proactive agents for planning, calendar, health, smart home, and photo editing into messaging apps, with TechCrunch reporting it became the first AI agent approved for Apple Messages for Business and Poke touting speed and reliability upgrades.
ElevenLabs introduced Flows Agent in ElevenCreative, which builds creative workflows from natural-language requests by choosing models, creating nodes, wiring connections, and adding approval gates for expensive steps.
Google Magenta released Magenta RealTime 2, an open music model for real-time generation and editing with local Apple Silicon support, hosted options, weights, and an Omar Sanseviero launch thread.
Paper added SVG editing with pen tools, path editing, fill/stroke controls, and drawing new shapes from scratch, with the demo on YouTube.
Unsora / Seedance 2.0 added path control for video generation, where you draw a route on an image and the model follows that camera path while removing the guide line.
Motion generates motion graphics and animated videos for product launches, explainers, brand docs, and logo animations from prompts or templates.

📊 Fundraising & Deals Roundup

Waymo struck a deal with B2U Storage Solutions to repurpose retired robotaxi battery packs as grid storage.
SpaceX secured a 35-year, 100% property-tax exemption for its planned $55B Texas Terafab chip factory, despite resident backlash over transparency, infrastructure strain, and environmental impact.
Supabase raised $500M at a $10.5B valuation as vibe-coding demand lifted its open-source Postgres platform into agent infrastructure, with the company calling the round a push to accelerate agentic infrastructure.
Generalist AI raised $400M at a $2B valuation to build general intelligence for robotics and real-world physical tasks, with SiliconANGLE noting Radical Ventures, Nvidia’s NVentures, and Bezos Expeditions in the backer mix.
Honeycomb raised $40M while using automated underwriting to price landlord insurance without inspections, and Calcalist reported it ended 2025 at $275M in gross written premium.

💡 Industry Commentary & Analysis

Kevin O'Leary said he will shrink his planned Stratos AI data center campus in Utah by roughly half after broad political backlash.
The Atlantic argues that AI is not conscious; treating current systems as conscious leads to absurd and morally confused conclusions.
Hiten Shah argued that “skills”, structured, reusable packages of judgment, process, examples, and edge cases, will be the real AI advantage for companies, more than data access or connectors.
Naithan Jones demonstrated a Hermes + Obsidian setup that acts like an external brain for ADHD, capturing stray thoughts and turning them into reminders, newsletters, and priority checks.
Claude featured Lovable CEO Anton Osika arguing trust, built through craft, care, and obsession, is the most underrated moat in AI. (1,748 likes, 115 reposts)
AdExchanger noted that AI SEO is pushing brands toward Reddit astroturfing as traditional search traffic shifts toward AI-generated answers.
Logan Bolton noted that GPT-5.5-Pro appeared to read Teen Vogue in its chain of thought while solving a hard math problem, a viral example of how strange long reasoning traces can look.
Alberto Romero argued AI companies are rushing toward IPOs because burn rates, infrastructure costs, weak profitability, and enterprise distrust make private financing harder to sustain.
Ethan Mollick argued humans are moving from co-intelligence to co-existence with autonomous AI systems, which means deciding when to delegate, when to refuse help, and when to use AI as critic or gatekeeper.
Every.to explored why humans may still keep valuable roles even if AI can do almost every task, with the human edge shifting toward taste, accountability, relationships, and judgment.
How To AI broke down Apple's GSM-Symbolic study, which found that leading LLMs (large language models) can drop up to 65% on grade-school math word problems when one irrelevant sentence is added, because they often pattern-match numbers instead of ignoring distractions (paper).
Satya Nadella framed Microsoft’s AI strategy around frontier models, OpenAI, trusted infrastructure, capex, and agent platforms, while the No Priors x Latent Space crossover dug into Build 2026 and Microsoft’s developer ecosystem.
Cloudflare said VoidZero is joining the company while keeping Vite open-source and vendor-agnostic, and VoidZero said Vite, Vitest, Rolldown, Oxc, and Vite+ will stay MIT-licensed under Evan You and the existing team.
Logan Thorneloe recommended CMU’s ML training infrastructure guide as a clear overview of hardware, memory, and experimentation, while Jay Palat explained why GPUs won matrix-heavy workloads and how teams should validate systems with MLPerf before production decisions.

🎙️ Podcasts & Deep Dives

OpenAI Podcast broke down how a reasoning model found a counterexample to an 80-year-old Erdős conjecture, with related coverage on YouTube and Apple; Slater Stich also interviewed Noam Brown on AI for math, the Erdős unit-distance problem, and the future of mathematical research.

Previous Around the Horn Digests

Catch up on our recent roundups:

Tuesday, June 2, 2026: OpenAI pushed Codex into knowledge work, the White House narrowed AI oversight, Microsoft added Windows agent security, and Axiom verified economics in Lean.
Monday, June 1, 2026: NVIDIA turned the PC into an agent computer, Anthropic filed confidentially for an IPO, MiniMax released M3, and Bernie Sanders proposed public ownership of AI labs.
Weekend, May 29-31, 2026: Kog pushed real-time inference toward 3,000 tokens per second, OpenAI launched Rosalind Biodefense, and Microsoft worked on a Copilot super app.
Thursday, May 28, 2026: Claude Opus 4.8 arrived, Anthropic raised a $65B Series H, IBM committed $10B to quantum, and Amazon killed an AI usage leaderboard.
Wednesday, May 27, 2026: Robinhood gave agents brokerage access, AxiomProver moved verified math into papers, OpenAI and Thrive built tax agents, and Google launched AI Threat Defense.
Tuesday, May 26, 2026: China curbed private-sector AI talent travel, Qualcomm struck a ByteDance chip deal, OpenRouter raised $113M, and xAI finished Grok V9-Medium.
Thursday, May 21, 2026: OpenAI said a reasoning model disproved the 80-year Erdős unit distance conjecture, Spotify and UMG licensed AI fan remixes, and Waymo paused service.
Tuesday, May 19, 2026: Google I/O pushed Gemini agents across Search, Android, Workspace, YouTube, and shopping while Anthropic hardened Managed Agents.
Monday, May 18, 2026: Microsoft open-sourced ECHO, Odyssey launched real-time AI simulators, and OpenAI added bank connections to ChatGPT.
Wednesday-Thursday, May 13-14, 2026: Nvidia H200 sales cleared but stalled, Americans opposed AI data centers, and Meta planned layoffs.
Tuesday, May 12, 2026: Anthropic refused China access to its newest model, Isomorphic raised $2.1B, and Google pushed Gemini deeper into Android.
Monday, May 11, 2026: Cerebras upsized its $4.8B IPO, Cowboy Space raised $275M for orbital data centers, and Google confirmed an AI-found zero-day.
Weekend, May 9-10, 2026: The Trump administration drafted an AI security order, Apple and Intel reached a chip-making agreement, and Cerebras' IPO heated up.

That's a Wrap

That is 200+ stories from Thursday alone. If you made it to the bottom, you now know more about Anthropic's recursive self-improvement metrics, NVIDIA's open-model push, and local-model iPhone dreams than at least one person currently forwarding a screenshot in Slack. Please use this power responsibly.

For the daily version (bite-sized, five-minute reads), make sure you are subscribed to The Neuron. We send six issues a week, and yes, we read all of this so you do not have to.

See you tomorrow.

P.S: Know someone who would find this useful? Forward this to them and tell them to subscribe here.