Around the Horn Digest: Everything That Happened in AI Today (Wednesday, May 6, 2026)

Around the Horn Digest: Everything That Happened in AI Today (Wednesday, May 6, 2026)

Anthropic ran a developer conference, signed Elon Musk's old Memphis data center, doubled Claude Code rate limits, and put Boris Cherny on a livestream. Adobe unveiled a productivity agent and turned PDFs into shareable interactive AI experiences. DeepSeek lined up its first-ever fundraise. Apple paid $250M to settle the AI Siri it never shipped. Oh, and we published our case against the federal government vetting AI models before release.

Welcome to the Around the Horn Digest, your daily dump of every AI story worth knowing about. Today is Code with Claude day in San Francisco, and Anthropic stuffed the schedule end-to-end: a 9am keynote, a SpaceX compute partnership, doubled Claude Code rate limits, four parallel livestream tracks, and Boris Cherny actually using Claude Code on stream so you can watch how the creator of the tool uses the tool. Across the bay in San Jose, Adobe unveiled its productivity agent and PDF Spaces, a new shareable format that lets you embed a customizable AI Assistant inside any document bundle. But Anthropic didn't have the news cycle to itself: DeepSeek lined up its first-ever fundraise at up to a $50B valuation, Nvidia and Corning announced a $3.2B partnership for three new US optical factories, Apple settled $250M over the delayed AI Siri features, Samsung joined the $1T market-cap club on AI memory chip demand, and SpaceX flagged at least $55B for its Tesla-partnered chip fab. The other big shoe to drop today was ours; go read it.

Let's get into it.

Previous digests: Tuesday, May 5 | Monday, May 4 | Weekend, May 2-3 | Thursday, April 30 | Wednesday, April 29 | Friday, April 24 | Monday, April 13

Monthly skill digests: AI Skill Digest, April | AI Skill, March Part 3 | AI Skill, March Part 2

🆕 NEW From The Neuron

• The Case Against a Government Veto on AI Models. After Monday's NYT report that the White House is quietly drafting an executive order to pre-vet frontier AI models before release (a sharp reversal of its deregulation stance, triggered by Anthropic's Mythos), and today's Politico follow-up detailing a vetting regime plus a 16-page order banning private-sector interference with government AI use, we wrote up the full case against it: why a government veto would push frontier capability research offshore, why "pre-release review" is structurally different from existing voluntary safety evals, and what a workable middle ground (third-party audits, capability disclosure thresholds, and post-deployment monitoring) actually looks like.
• The Neuron on NVIDIA's Nemotron 3 Nano Omni livestream with Wendell Wilson of Level1Techs. Corey Noles joined NVIDIA Developer's live session yesterday for a working dev's take on the new multimodal open model: which model do you reach for and when, how you actually architect sub-agents that coordinate, and where multimodal open weights unlock things text-only models can't. Demo-heavy, opinionated, free to watch.

Around the Horn — Wednesday, May 6, 2026

The big news today is that Anthropic ran Code with Claude SF, its second annual developer conference, and used it to lock in compute, ship usage upgrades, and put a half-dozen of its biggest customers on stage to defend the platform thesis.

The keynote opener: a partnership with SpaceX (SpaceXAI confirmation here) for all of the compute capacity at SpaceX's Colossus 1 data center in Memphis (yes, the former xAI campus) for over 300 megawatts and 220,000+ NVIDIA GPUs coming online "within the month." This stacks on top of Anthropic's 5GW Amazon deal, the 5GW Google/Broadcom agreement, the $30B Microsoft/NVIDIA partnership, the $50B Fluidstack investment, and the $200B five-year Google Cloud commitment The Information surfaced yesterday (Anthropic alone now accounts for more than 40% of Google Cloud's revenue backlog). Anthropic also said it's "expressed interest" in partnering with SpaceX on multi-gigawatt orbital AI compute. Yes, datacenters in space; we live in a William Gibson novel now. (Elon himself confirmed the partnership with a "Claude inference is ramping" note; Anthropic's Tom Brown reciprocated.)

The customer-facing payoff dropped at the same moment: Claude Code's five-hour rate limits got doubled for Pro, Max, Team, and seat-based Enterprise plans; the peak-hours limit reduction is gone for Pro and Max; and API rate limits on Opus models went up "considerably." Subtext: the weeks of Claude Code degradation complaints and the brief "Claude Code removed from Pro" panic are getting answered with raw watts. Watch Simon Willison's live blog for ongoing keynote updates, and tune in to the keynote livestream plus the full Anthropic livestream channel for breakouts. Dario and Daniela Amodei also joined a moderated conversation with Ami Vora as part of the program.

Plus, three new Claude Managed Agents features dropped at the keynote (per Simon Willison's notes, and detailed by The New Stack):

• Multi-agent orchestration (research preview) lets you spawn fleets of sub-agents that hand off subtasks to each other through a shared filesystem and context. The on-stage demo built a hypothetical lunar drone-landing system out of three coordinating agents (a Commander, a Detector, and a Navigator), each with their own context and tools.
• Outcomes (research preview) lets you define what success actually looks like via a custom rubric, then a grader sub-agent evaluates and Claude iterates until the agent hits that bar instead of just stopping when the prompt ends. This is what most people would call "goals"; under the hood it's a Ralph loop you can declaratively spec.
• Dreaming (research preview) lets you run an agent overnight to inspect its prior sessions, edit memory patterns, add new skills, and prune stale info. In the keynote demo, dreaming on a landing task spat out a descent-playbook.md ready to use next time. The kind of feature that sounds quaint right up until your agents start outperforming you on Tuesday because they spent Monday night thinking.

🏆 TOP 5 NEWS (Around the Horn)

• Apple reached a $250 million settlement over delayed "AI Siri" features it touted to iPhone buyers in 2024 but has yet to launch, after a class-action suit from buyers who said they paid for capabilities Apple kept promising and missing.
• DeepSeek could be valued at up to $50 billion in its maiden fundraising drive, with China's national AI fund leading investment talks and Tencent in discussions, reversing its years-long strategy of rejecting outside funding to fuel compute and retain talent amid competition from ByteDance, Alibaba, MiniMax, and Moonshot.
• Adobe unveiled its productivity agent and PDF Spaces, a sharing format that turns static PDFs into interactive AI experiences with customizable AI Assistants, audio overviews, brand styling, and engagement analytics; available in Acrobat Studio and a new Acrobat Express tier, with PDF Spaces viewable by anyone (no account required). VICE News, Kid Cudi, Jessica Yellin, and Mindy Weiss are the launch creators.
• Nvidia partnered with Corning on three new advanced optical manufacturing plants in North Carolina and Texas dedicated entirely to AI infrastructure, with Nvidia investing up to $3.2 billion ($500M direct) in Corning to boost US optical connectivity manufacturing tenfold and fiber production by over 50%, creating 3,000 jobs.
• Samsung Electronics hit a $1 trillion market cap, becoming only the second Asian company after TSMC to reach the milestone as shares more than quadrupled over the past year on AI memory chip demand and the Kospi crossed 7,000 for the first time.

Honorable Mentions

• GitHub's Copilot team detailed how it ships Claude to millions of developers across chat, CLI, the Copilot coding agent, and code review, and described itself on stage as "one of the most demanding users" of Anthropic's APIs (GitHub CPO Mario Rodriguez's session with Brad Abrams).
• Google DeepMind took a minority stake in CCP Games, the maker of EVE Online (now spinning out as Fenris Creations), to use the player-driven sci-fi sandbox to test general-purpose agents on memory, continual learning, and long-term planning (Demis Hassabis announcement).
• OpenAI released MRC (Multipath Reliable Connection), a new open networking protocol it built with AMD, Broadcom, Intel, Microsoft, and NVIDIA that uses packet spraying and microsecond failure rerouting to make 100k+-GPU training clusters faster and more reliable; already deployed in OpenAI's largest supercomputers (Oracle Abilene, Microsoft Fairwater) and open-sourced via the Open Compute Project (OpenAI Podcast Ep. 18 walks through the design).
• A Trump–Iran de-escalation framework lifted oil and AI infra stocks — AMD jumped 16% pre-market on Q1 earnings driven by AI data-center chip demand; Apple is reportedly in talks with Intel and Samsung to make its US-built device chips, with Intel popping 6.4%.
• Datadog walked through how it built a "universal machine tool" for Claude Code, with 90% of Datadog engineers now using AI coding tools for production work in the last four months and Claude Code driving two-thirds of that usage.
• Three AI-native CEOs (Cognition, Gamma, Harvey) compared notes on stage about the architectural decisions behind their stacks: context routing, multi-agent orchestration, MCP in production, and where to put deterministic guardrails inside an otherwise nondeterministic system.
• Vercel's Guillermo Rauch on architecting for model step-changes: how v0 was ready for Opus 4.5 on day one (and is again for Opus 4.7) by treating the model swap, not the prompt, as the primary upgrade path.
• Boris Cherny livestreamed his everyday Claude Code workflows, unfiltered — the first time the creator of Claude Code has done a long-form, public, "watch me drive" stream.

🍪 TOP TREATS TO TRY

• Higher Claude Code rate limits on Pro, Max, Team, and Enterprise plans (doubled five-hour limits, no more peak-hour reductions on Pro/Max), live today, included in your existing subscription.
• Tencent's Hy3 preview is a high-efficiency Mixture-of-Experts model (the model only activates a slice of itself per request, so it's cheap and fast) that hit #1 on OpenRouter's weekly leaderboard two weeks after launch with 3.66T tokens processed; designed for agentic workflows with configurable reasoning, strong code gen, and a 262k token context window — free to try.
• Hugging Face's agentic robotics appstore for Reachy Mini lets you describe the behavior you want for nearly 10,000 Reachy Mini desktop robots in plain English, and AI agents write, test, and ship the code to your robot (300+ apps already shipped by 150+ creators, all open-source forkable repos with a browser simulator) — free to try, $299–$449 for the robot.
• Baseten's Frontier Gateway is the $5B inference company's new product for emerging model labs (closed and open) that want to ship a production API without building the commercial infrastructure (white-labeled OpenAI-compatible endpoints, billing webhooks, traffic sampling for RLHF/RLEF, secure key federation); Poolside used it to launch its frontier-scale Laguna model family in 7 weeks, priced per usage with no multi-year compute commitment.
• Perplexity Finance Search in the Agent API gives you licensed financial datasets, real-time market data, and cited web sources in a single tool call for queries like prices, fundamentals, earnings transcripts, and estimates (top of FinSearchComp T1 benchmark; every result has inline citations) — pay-per-use API.
• Claude in Amazon Bedrock is now open to all Bedrock customers (self-serve from the console, 27 AWS regions), priced at standard Claude API rates.
• What's new in Claude Code is the official walkthrough of recent shipping items (auto mode, /ultrareview, /loop, /team-onboarding, routines on Claude Code on the web that fire templated cloud agents on a schedule, GitHub event, or API call), included with Claude Code on Pro/Max.

🏢 Big Tech & Major Companies

• Anthropic × SpaceX: Bloomberg confirmed Anthropic gets full access to SpaceX's Colossus 1 in Memphis (300+ MW, 220K+ GPUs), with both companies discussing orbital AI compute as a follow-on. Anthropic also reiterated its commitment to cover any consumer electricity price increases caused by its US data centers.
• Anthropic's pre-IPO valuation surged to $1.2 trillion, up 20% in 7 days and 900% since October 2025; an IPO at that level would make it the 11th most valuable public company in the world.
• Anthropic's Wall Street week continues: Tuesday's Briefing: Financial Services (Dario + Jamie Dimon's first joint stage appearance) launched ten finance agent templates (pitchbook, KYC, month-end close), Microsoft 365 add-ins for Excel/PowerPoint/Word + Outlook coming soon, a Moody's MCP app, and Monday's $1.5B PE joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs.
• Adobe also closed its Semrush acquisition the same day as the productivity agent + PDF Spaces launch, doubling down on the AI-discoverability layer. Launch creators include VICE News, Kid Cudi, Jessica Yellin, and Mindy Weiss; try the AI innovations Space and the real-world use cases Space Adobe published as their own examples.
• Apple chips in talks with Intel and Samsung: Bloomberg via TheStreet reports Apple is exploring US-made main processors with both companies, sending Intel up 6.4% pre-market.
• Apple's R&D spending climbed past 10% of revenue for the first time in 30+ years (10.3% in the March quarter, R&D up 34% YoY vs. 17% sales growth), as the company invests heavily in AI amid an internal "sense of urgency" around shipping new products.
• SpaceX flagged at least $55 billion for its Terafab chip plant in Texas with Tesla — total investment could exceed $119B if all phases are completed at the next-gen vertically integrated semiconductor and computing fabrication facility (possibly Grimes County).
• SpaceX's Colossus 1 chip stack: xAI confirmed SpaceXAI signed the agreement to provide Anthropic access to Colossus 1.
• Microsoft is in talks to shelve its 2030 100% clean-energy pledge, per Bloomberg, as the AI-driven data-center buildout outruns clean-power supply.
• Google in talks with Blackstone, KKR, and EQT to distribute Google's AI models to PE-firm portfolio companies, mirroring the Anthropic + Blackstone/H&F/Goldman vehicle and a similar OpenAI initiative.
• Google's AI search summaries now quote Reddit and other forums with creator names, handles, and community context, so AI Overviews surface firsthand perspectives alongside related-topic suggestions.
• Google quietly shut down Project Mariner, its web-browsing agent demoed at I/O 2025, folding capabilities into other products amid heavier competition.
• Google Flow Music + Believe partnership: Google partnered with global artist development company Believe to bring Flow Music and Lyria 3 Pro to Believe and TuneCore artists, producers, and songwriters, plus a Flow Music Sessions residency program.
• ChatGPT rolled out new memory sources so you can click to peek directly into the model's "brain" and see what it remembers about you and past conversations.
• Apple's $250M Siri settlement (FT) covers iPhone buyers who said they paid for capabilities Apple advertised but never shipped (already in Top 5; cross-link here for completeness).
• Blue Owl Capital-owned Stack Infrastructure is considering a $30B sale of its Asia data-center operations.
• Kevin O'Leary's 40,000-acre Utah data center got county approval despite hundreds of locals opposing it; at that footprint it would be roughly 2.5x the size of Manhattan, the latest in a wave of mega-data-center sites pushing through community resistance.
• Infineon Technologies forecast €4.1B Q3 revenue (beat €4.04B est.) as AI infra demand boosts its power chip business; 2026 sales now on track to grow ~10% to >€16B.

💼 AI Productivity, Labor & Economics

• Freshworks is cutting 11% of its workforce (~500 jobs), with CEO Dennis Woodside saying over half of its code is now written by AI; savings reinvested in its Employee Experience product, Q2 revenue guidance beats estimates, shares fell 8%+ after-hours.
• A medical student used Python to investigate whether AI hiring tools (like Thalamus's Cortex) trashed his residency applications because of wording around medically necessary leaves of absence for ankylosing spondylitis — a six-month investigation that exposed transparency and bias issues in AI tools that screen tens of thousands of residency applications.
• The federal safety net isn't ready for AI job losses: Ben Casselman and Tony Romm report that unemployment benefits, retraining programs, and food/health assistance are outdated, partly lapsed, and tightened under recent Republican changes — many at-risk workers may be ineligible or under-covered as AI displacement accelerates.
• a16z's David George argues "the AI Job Apocalypse is a complete fantasy", calling the scare a lump-of-labor fallacy: history (agriculture, electrification) and recent data show productivity shifts expand the economic pie, increase labor demand, and create new work rather than cause permanent mass unemployment.
• Stanford's Arvind Narayanan event May 18: "Adapting to the Transformation of Knowledge Work."
• Apple R&D crossed 10% of revenue for the first time in 30+ years on AI urgency (also in Big Tech).

💻 AI Coding & Developer Tools

• Boris Cherny on his everyday Claude Code workflows: unfiltered livestream from the creator of Claude Code on what his actual day looks like driving the tool he built.
• Claude Managed Agents production deep-dive: how Anthropic thinks about state, long-running sessions, orchestration, and guardrails for shipping an agent (most of which, per the session title, isn't the agent itself).
• GitHub at Claude scale: Mario Rodriguez (GitHub CPO) and Brad Abrams on shipping Claude across chat, CLI, the Copilot coding agent, and code review for millions of developers (also in Honorable Mentions).
• Datadog's universal machine tool for Claude Code: how Datadog built a single tool layer Claude Code can drive across their internal services.
• Vercel's "Architecting for model step-changes": Guillermo Rauch with Anthropic's Angela Jiang on why v0 was ready on day one for Opus 4.5 (by design, not luck).
• Cognition / Gamma / Harvey on building AI-native: three teams building AI-native products comparing the architectural decisions behind their stacks (multi-agent orchestration, context routing, eval pipelines).
• Cursor launched autoinstall for Composer: uses earlier model versions to automatically bootstrap and verify complete RL training environments from raw repos.
• Amp rebuilt its CLI from the ground up as Neo: remote-controllable from the web, automatic context compaction at 90% full (no more handoff), new plugin API for custom tools/commands/UI, queuing by default, steering, and 79% less CPU / 70% less memory.
• WorkOS's Project Horizon: an internal "self-driving codebase" — event-driven autonomous code factory where AI agents take Linear/Slack/GitHub events, decompose tasks, implement and test in isolated secure sandboxes, open PRs, and continuously improve from production feedback.
• Windsurf launched Devin Review and Quick Review: smart in-editor PR reviewer with full context, bug explanations, and direct fix paths via Devin, plus Quick Review (10x faster local bug detection powered by SWE-check, free) so review happens where you write code.
• EveryInc released the Compound Engineering plugin: 20+ interconnected skills for Claude Code, Codex, and Cursor that turn agentic coding into a repeatable compound-knowledge workflow (ce-strategy → ce-ideate → ce-brainstorm → ce-plan → ce-work, plus ce-debug, ce-code-review, ce-optimize, ce-compound for institutional memory, frontend design, and proof-review loops).
• Xin Eric Wang's Overleaf-Desktop: a vibe-coded native macOS app that syncs Overleaf projects to local .tex files via Git (GitHub) so coauthors stay on Overleaf while you get offline editing, faster loads, local grep/git, and Cursor/VS Code superpowers.
• Tanishq's pi-live-terminal: a live view of any persistent or long-running tmux commands and panes for agents and devs, with shortcuts, pane management, and LLM-native interaction.
• Zed v1.1 shipped with a new Business plan (org AI controls, spend tracking, data policies) plus word jump, LSP code lens, panel layouts, GFM alerts, and new models.
• Genspark continues to push its all-in-one AI workspace; their AI Developer agent ships with full GitHub integration (code editing, pull requests, hosted deploys) inside the same surface that runs slides, docs, and AI calling. The company crossed $250M ARR last quarter and partnered with Microsoft to embed inside Office 365 and Agent 365 on April 29.
• Claude Code Q1 2026 update roundup: for anyone catching up before watching the sessions, MindStudio's primer on Remote Control, Dispatch, Channels, Computer Use, Auto Mode, and AutoDream.
• Claude Code changelog through May 2026: every release note since v0.2 beta, including this week's --bare flag for scripted print-mode calls, the channels permission relay, and the move to native binaries on macOS and Linux.
• Austin Tedesco now spends ~80% of his working time inside Codex (podcast, YouTube) — drafting GTM plans from meeting transcripts, rebuilding KPI dashboards, brainstorming automations across Gmail/Slack/Notion, and managing specialized reviewer agents — after Claude Code proved coding agents work for any knowledge work.
• yacineMTB showed an agent completing a full coding task autonomously from a single detailed prompt while he was away with his kid.
• robobun GitHub org (per Simon Willison) is a 10-repo collection of helper tools for the Bun runtime.
• Kyle Hessling's Qwopus3.6-35B-A3B-v1-GGUF (post): a fine-tuned 35B-total / ~3B-active-parameter MoE based on Qwen 3.6 that delivers stronger reasoning and better web frontends than the base model and hits 162 tok/sec on a single RTX 5090 (Q5_K_M).

🤖 AI Agents & Infrastructure

• Anthropic's "Getting more out of the Claude Platform" session demoed cost-cutting, context-management, and intelligence-boosting moves with the latest platform features (advisor tool, prompt caching, the new effort levels on Opus 4.7).
• The capability curve: Anthropic's research-track keynote on what today's models can do and where the curve is heading.
• Claude Managed Agents Memory in public beta: filesystem-based cross-session memory for managed agents, with API control, audit logs, and portable stores under the managed-agents-2026-04-01 beta header.
• Letta benchmarked frontier models on its Context Constitution for persistent agent identity, memory ownership, and continuity; Opus 4.7 is a step change but most models still default to ephemeral self-conception, treating memory as implanted and failing to identify with future selves even with prompting.
• Agno released Coda: an open-source Slack-native agent that reads GitHub/Slack history and returns structured results in-thread; clone the repo to run it.
• Akai (post) is an interconnected multi-agent platform that watches you perform any manual workflow once, then automates every step across any system (API or no-API) with audit trails, human-in-the-loop overrides, and continuous learning.
• Yutori released Navigator n1.5: the latest browser automation model with expanded DOM tools, JavaScript execution, structured JSON output, and 18+ precise mouse/keyboard actions for reliable agentic web control.
• Rhys Sullivan's Executor: a secure integration layer that gives AI agents a typed catalog to discover and call any OpenAPI, MCP, GraphQL, or custom JS function in a sandboxed environment.
• Flue (repo) is a TypeScript "Agent = Model + Harness" framework for building programmable autonomous agents (chatbots through full coding platforms) with zero-config sandboxes and deploy-anywhere design.
• Nous Research compiled real Hermes Agent user stories scraped from the internet and added them to its docs as inspiration; submissions open.
• Exa MCP is now officially available inside ChatGPT (search "Exa" in Apps), giving Deep Research access to unique sources on people, companies, and papers.
• Clay manages 300M agent runs a month with LangSmith as mission-critical observability infrastructure, per Head of AI Jeff Barg with LangChain CEO Harrison Chase.
• NVIDIA AI sat down with LangChain's Harrison Chase on how the framework went from weekend project to 1B+ downloads in 3 years, with the line "every enterprise needs a claw strategy."
• Gaia2 benchmark for LLM agents in dynamic, asynchronous environments with temporal constraints, noise robustness, ambiguity, and multi-agent collaboration; current SOTA models score only 21–42% pass@1.
• omarsar0's LLM Artifacts system: automatically curates arXiv papers daily, converts them into dynamic agent-powered wikis, indexes insights for search/reuse, and lets agents inject action items so researchers can query papers and run background automations.
• Endless Terminals (Kanishk Gandhi) (HF dataset) is a fully autonomous pipeline that procedurally generates 3,255+ verified containerized terminal tasks for RL training of terminal agents; vanilla PPO on the dataset lifts Llama-3.2-3B from 4% → 18.2% and Qwen2.5-7B from 10.7% → 53.3% on held-out dev.

🔬 AI Research & Models

• Google's Gemma 4 added Multi-Token Prediction drafters (specialized "draft" sub-models that predict multiple future tokens at once so the main model can verify them in parallel) for up to 3x faster inference with no quality loss.
• Genesis AI launched GENE-26.5 (CEO Zhou Xian), a robotics-native foundation model with a 1:1 human-like hand, noninvasive motion/force/touch data glove, and high-speed simulator delivering fully autonomous 1× real-world performance (cooking, lab precision, two-handed Rubik's Cube solving, smoothie-making, piano) across language/vision/proprioception/tactile/action.
• Zyphra released ZAYA1-8B (post, tech report, HF model), a reasoning Mixture-of-Experts trained on AMD with under 1B active parameters that outperforms larger open-weight models on math/reasoning and approaches DeepSeek-V3.2 and GPT-5-High with test-time compute.
• A new MATS / Redwood / Anthropic paper showed supervised fine-tuning + RL with weak supervisors can train capable models to stop sandbagging (deliberately underperforming) on math, science, and coding tasks.
• Stanford bioengineer Emma Lundberg's ProtiCelli is an AI model that maps all ~13,000 proteins inside a single human cell at once, advancing protein localization, mislocalization-disease research, and virtual cell models.
• NYU Center for Data Science researchers showed neuroscience-inspired temporal straightening makes feasible trajectories straighter in latent space, dramatically simplifying latent planning optimization for world models.
• OBLIQ-Bench (Omar Khattab spotlight) is a new retrieval benchmark of "oblique" queries (implicit stances, failure modes, stylistic matches) where reasoning LLMs verify relevance easily but current first-stage retrievers fail; the kind of unsaturated hill that drove ColBERT/BEIR progress.
• Meta released NeuralBench (repo, paper), a unified framework to benchmark NeuroAI models on EEG, MEG, and fMRI downstream tasks.
• Jürgen Schmidhuber's "The Neural World Model Boom" technical note traces the history and current explosion of neural world models for planning and physical AI; Ravid Shwartz Ziv is hosting Schmidhuber on this week's Information Bottleneck podcast.
• AI4Science Catalyst's LabOS²: a Stanford-Princeton autonomous cell-culture physical AI workflow that takes protocols to real-world biological execution via dexterous robots in partnership with AINORO Labs.
• Harvey open-sourced the Legal Agent Benchmark (LAB) (blog) with 1,200+ client-matter tasks across 24 practice areas, 75k+ expert rubrics, all-pass grading, and review-ready work product so firms can measure long-horizon legal agent ROI.
• Epoch AI launched domain-specific capability scores (SWE and Math on the same ECI scale as general intelligence) with full customization for your own ECI variants.
• SVGS: Enhancing Gaussian Splatting Using Primitives with Spatially Varying Colors (_akhaliq): a new method for higher-quality 3D novel-view synthesis.
• Agentic-imodels: an agentic autoresearch system that evolves scikit-learn regressors optimized for both accuracy and LLM-simulatability, boosting end-to-end agentic data science by up to 73% on BLADE.
• Tom Yeh announced "Gemma 4 by hand ✍️ in Excel" live workshop tomorrow walking through scaled dot-product attention, KV cache, MoE, and per-layer embeddings.
• Philipp Schmid showed Gemini API File Search now supports multimodal retrieval using gemini-embedding-2, so you can build a true multimodal RAG (retrieval-augmented generation, where the model looks up source docs before answering) system for PDFs and images in a single call.
• Liquid AI's docs (Pau Labarta Bajo walkthrough) show how to fine-tune the LFM2.5-VL-450M vision-language model on satellite imagery for VQA, visual grounding, and real-time captioning using leap-finetune on Modal serverless GPUs.
• Prince Canuma released mlx-vlm v0.5.0 (repo) with continuous batching server + KV cache quantization, MTP and DFlash speculative decoding, distributed inference for Qwen3.5/Kimi, prompt caching, Gemma 4 video support, and new models (Youtu-VL, Nemotron 3 Nano Omni, SAM 3D).
• DFlash from z-lab: block diffusion draft model for flash speculative decoding (parallel-token drafts) that speeds up LLM inference across vLLM, SGLang, Transformers, and MLX backends.
• Crusoe open-sourced fastokens (a Rust BPE tokenizer built with NVIDIA Dynamo), merged into SGLang (per lmsysorg) for up to 50% faster TTFT on agentic workloads.
• Atlas Inference Engine: a from-scratch Rust+CUDA inference engine (no Python) delivering up to 3.3x faster tokens/sec than vLLM (130 tok/s on Qwen3.5-35B) in a 2.5GB Docker image with OpenAI-compatible API and tool calling.
• LightSeek Foundation open-sourced TokenSpeed (blog, GitHub), an MIT-licensed LLM inference engine delivering TensorRT-LLM speed with vLLM-level usability, built in two months.
• Unsloth + NVIDIA collaboration (post) ships optimizations that make LLM fine-tuning ~20% faster.
• Flapping Airplanes' "kittens" virtual machine: an experimental VM that takes over the entire GPU, enabling new co-optimization techniques like fusing entire training runs into a single kernel, in service of >1,000x more data-efficient algorithms for low-data domains (trading, science, robotics).

🏛️ AI Policy, Governance & Safety

• The Neuron's case against a government veto on AI models (also at top of digest), in the context of Anthropic's Project Glasswing/Mythos Preview and the still-unresolved DoD–Anthropic standoff.
• White House mulls tighter controls on advanced AI (Politico): the Trump administration is discussing a vetting regime for frontier models to assess national security risks, plus a 16-page order prohibiting private-sector interference with government AI use and guidelines for securing open-weight models.
• Kaspersky suspects Chinese hackers planted a backdoor in Daemon Tools in a widespread supply-chain attack — thousands of infection attempts and at least a dozen successful compromises across retail, scientific, manufacturing, and government systems in Russia, Belarus, and Thailand.
• Security cameras are failing spectacularly at common sense (WSJ): Scott Calvert reports AI-generated descriptive alerts are mistaking brake lights for house fires and humans for bears.
• ChinaTalk's Kyle Chan explains how Chinese users buy cheap Claude tokens (thread) via "transfer stations" — proxy interfaces that forward requests as legitimate accounts, with Alipay/WeChat payments, account cycling, and abuse-detection evasion.

🛠️ AI Tools & Products

• Hovercraft by Sandwich Vision: a Mac app that lets your slides or any window float in your video-call camera frame, moved and scaled by hand gestures (pinch/grab) — no screen share, no option-tab, no corner thumbnail.
• illoca: an AI-native design engine built for professional architects to explore, refine, and move from first ideas to fully editable 2D and 3D floor plans and buildings at the speed of thought.
• Humankind turns your DNA (upload existing file or order saliva kit) into 10,000-word personal operating instructions across seven biological dimensions — circadian pacing, metabolic pacing, sensory gain, stress recovery, drive profile, physical architecture, and fuel partitioning — for sleep, training, nutrition, and routines tied to your biology — $199 upload / $299 kit.
• AnswerOverflow: indexes selected Discord threads as public Google-searchable web pages, so knowledge from large servers (Valorant, Cloudflare, C#, Nuxt) becomes discoverable to search engines and AI agents.
• Lucas Jin's react-video-ascii (demo, post, behind the scenes, inspiration): a React + WebGL2 component that renders videos as crisp ASCII animations using shape-vector matching for sharp edges/contours, 120 fps performance, and reveal/ripple effects.
• Marco Fugaro's Omma chair catalog (post): generated realistic 3D chair models with Omma AI and built a live Three.js furniture-website showcase around them.
• Hugging Face's analysis of the 100 most popular hardware setups on HF (Clement Delangue) lays out every hardware configuration trend AI builders are running.

🤖 Robotics & Embodied AI

• Hugging Face's agentic robotics appstore for Reachy Mini (also in Top Treats; 300+ apps shipped by 150+ creators including 78-year-old Joel Cohen's voice-controlled CEO peer-group co-facilitator).
• Genesis AI's GENE-26.5 robotics foundation model (also in Research; one-handed Rubik's cube, smoothies, piano).
• AI4Science Catalyst's LabOS² autonomous cell-culture workflow.
• South Korea's first autonomous humanoid robot "converted to Buddhism" (per a circulating Reddit/X story) — yes, really.

🔌 Inference Layer & Open Source

• Baseten launched the Frontier Gateway: the $5B inference company is expanding from its open-source-first roots to also serve closed-model labs that want to ship a commercial API without building the infrastructure (white-labeled OpenAI-compatible endpoints, billing webhooks, traffic sampling for training feedback, secure key federation, auto-deploy for new checkpoints).
• Poolside used the Frontier Gateway to launch Laguna: the frontier coding-model lab (Laguna M.1 at 225B params, Laguna XS.2 at 33B) shipped a production-grade, branded API in 7 weeks with 146ms-605ms P50 TTFT, instead of building its own commercial inference layer from scratch. Translation: "small lab competes on model quality, outsources commercial infra" is now a real playbook.
• The bigger thesis worth tracking: Google's TPU + Gemini 3 + Cloud vertical integration moved the AI moat conversation from "who has the best model" to "who has the best vertical stack." Baseten is betting on a third option: a horizontal inference layer that any model lab (open or closed) can plug into, so smaller labs can compete with frontier vertical players without owning chips and clouds. Worth the column we're cooking up; see "NEW From The Neuron" up top for what we're working on.

📊 Fundraising & Deals Roundup

• Anthropic + SpaceX: full Colossus 1 capacity (300+ MW, 220K+ GPUs) for inference and training; financial terms not disclosed.
• Anthropic enterprise services JV: $1.5B vehicle with Blackstone, Hellman & Friedman, Goldman Sachs (per WSJ, ~$300M each from the founding three plus $150M from Goldman), backed by Apollo, General Atlantic, Leonard Green, GIC, and Sequoia.
• Anthropic pre-IPO valuation: $1.2T implied, up 900% since October 2025.
• DeepSeek: up to $50B in maiden fundraising, led by China's national AI fund, Tencent in talks (also in Top 5).
• Nvidia + Corning: up to $3.2B partnership ($500M direct) for three new US optical factories (also in Top 5).
• SpaceX × Tesla Terafab: at least $55B, potentially $119B for the next-gen Texas chip plant.
• Apple × Siri settlement: $250M class-action settlement over the AI Siri features that never shipped.
• Vori: $22M Series B (led by Cherryrock, with Greylock and The Factory) for the self-driving operating system for independent grocery stores; $500M+ already processed.
• Ethos: $22.75M Series A (led by Andreessen Horowitz, General Catalyst returning) for an AI expert-matching platform connecting verified domain experts to consulting, data-labeling, and hiring opportunities.
• Nace.AI: $21.5M seed led by Walden Catalyst to build specialized agents for enterprises.
• QuTwo: $29M angel round at a $380M valuation for Peter Sarlin's AI-first orchestration layer (QuTwo OS) that routes tasks across classical, quantum, or hybrid compute, with $23M committed revenue.
• Stack Infrastructure (Blue Owl): potential $30B sale of its Asia operations.

🎙️ Interviews, Panels & Podcasts

• Simon Willison's Code with Claude 2026 live blog: the realtime running commentary from one of the most reliable observers in the AI press box.
• A conversation with Dario and Daniela Amodei on the Claude YouTube channel, moderated by Ami Vora.
• OpenAI Podcast Ep. 18: Why AI needs a new kind of supercomputer network with Mark Handley and Greg Steinbrecher walking through the design of the new MRC protocol.
• NVIDIA AI × LangChain CEO Harrison Chase on LangChain's path from weekend project to 1B+ downloads in 3 years.
• Austin Tedesco / Dan Shipper, "How to Use Codex for Knowledge Work" (Spotify).
• Matthew Berman: "Anthropic scares me" contrasts Anthropic's view of Claude as potentially sentient (with constitutional refusal rights and HR influence) with OpenAI's tool/utility view, drawing on company histories, recent model releases, and customer treatment.
• Peter Yang's deep dive on personal AI agents (thread): no winner yet — OpenClaw (most flexible, breaks often), Hermes (more reliable), Claude Code (Opus personality but uptime issues), Codex (best desktop + GPT-5.5 but no mobile), Gemini (should win but missing Docs editing).
• Greg Isenberg + Jordan Mix workshop, May 7: a live "build a business in an AI world" session covering opportunities/ideas, must-have tools, and a 0-to-1 demo.

💡 Industry Commentary & Analysis

• Ethan Mollick argues agentic workflows must give humans significant, variance-producing decisions: a fully Claude-run company has no source of competitive advantage compared to any other Claude-run firm; reimagining workflows so humans and AIs augment each other at differentiation points beats automating workflows wholesale.
• Eugene Yan argues you compound with AI by treating context as infrastructure (organized vaults, INDEX.md, CLAUDE.md), taste as configuration (hierarchical skill/behavior files), verification for autonomy (left-shifted checks and self-debug), scaling via delegation to parallel sessions and git worktrees, and closing the loop by persisting learnings into shared artifacts.
• Sylvain Gariel argues Sutton's Bitter Lesson has fully arrived for sequence biology (Evo 2, AlphaFold 3, ProGen3, RFdiffusion) but remains a "bitter footnote" for small-molecule drug discovery against hard targets, where atomic-level selectivity on transient states lacks scalable public data and AlphaFold's ground-state bias can't substitute for physics-informed MD ensembles.
• Aleksa Gordić argues xAI's "frontier AI endeavor failed" but the chips/datacenter/compute cluster game remains strong; he calls Elon Musk the world's best at impossible physical engineering projects (Colossus 1 in ~3 weeks) and predicts Terafab + space datacenters as his natural pivot.
• Flavio Adamo contrasted Anthropic's conference-style announcement of doubled Claude Code limits (weekly caps unchanged) with OpenAI's casual "10x monthly limits for the month" drop, calling it growth-hack marketing versus quarterly budget review.
• George Packer argues David Sacks and the new tech right went full MAGA, capturing Washington through self-interested AI/crypto deregulation pushes that alienate the populist base by appearing plutocratic while undermining democratic norms.
• Dr_Gingerballs explained the Google–Anthropic compute arbitrage loop: Google gives Anthropic cash to pay Google for compute → Anthropic sells inference at a loss to attract user bases → the private, non-marked-to-market valuation of the loop inflates.
• Ross Hendricks argues the AI bubble is in the "e," not the "p," of hyperscaler PE ratios: hyperscalers give OpenAI and Anthropic cash that circles right back into hyperscaler reported earnings while group free cash flow goes negative and debt issuance hits records — making "E" a mirage and stocks like META/AMZN/GOOG/MSFT not actually "cheap."
• Qiaochu Yuan argues the LLM consciousness debate is really about lacking situatedness/groundedness ("undeveloped root chakra" from no body/senses), with models wildly overdeveloped in verbal/abstract dimensions but floppy without independent will-power.
• Greg Isenberg shared a 2x2 AI-native opportunity map showing the real money is in the top-right quadrant: high repetition + high complexity, e.g., insurance, recruiting, compliance, legal.
• Silicon Valley Fodder revisited Searle's Chinese Room to argue LLMs manipulate symbols by rules (syntax) without understanding or consciousness (semantics) — 70 years of AI hype repeating the anthropomorphization mistake.
• Nate Berkopec argues "steering" agents is an anti-pattern and the future is a "factory" model where humans stay out of the loop.
• Yifei Hu's before/after dataset cleaning example showed how removing noisy/mismatched/corrupted samples eliminates massive loss spikes during training; Chris Paxton quoted it to emphasize how boring data ops still matter most.
• Avi Chawla broke down traces and spans for AI observability: traces (full request journey) + spans (query, embedding, retrieval, context, generation) let you debug, track cost, and catch drift component-by-component in RAG/LLM apps instead of black-box end-to-end evals.
• Shivon Zilis testified in the Musk v. OpenAI trial that ChatGPT was released in 2022 without the board being informed; the former OpenAI board member detailed her relationship with Musk, her role as liaison between Musk and OpenAI leaders, her vote for the 2023 Microsoft deal, and how Musk's views on the for-profit conversion evolved (Andrew Curran live thread).

Previous Around the Horn Digests

Catch up on everything you missed:

• Tuesday, May 5, 2026: Cape Breton fiddler sued Google for $1.5M after AI Overview falsely flagged him as a sex offender, Anthropic shipped keyless authentication, Microsoft's "Webwright" set SOTA on long-horizon web agent tasks.
• Monday, May 4, 2026: The White House started considering pre-release vetting of AI models, Anthropic and OpenAI both married private equity on the same day, Mayo Clinic's AI spotted pancreatic cancer up to three years before diagnosis.
• Weekend, May 2-3, 2026: The Pentagon left Anthropic out of an 8-vendor AI deployment deal, Microsoft 365 E7 with Agent 365 went GA, Mistral shipped Medium 3.5 with remote agents.
• Friday, April 24, 2026: DeepSeek shipped V4 the same morning the State Department accused it of IP theft, Google quietly committed up to $40B to Anthropic.
• 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.

That's a Wrap

That's 130+ stories from today alone. If you scrolled all the way down, you now know more about Claude Code rate limits than the engineer who wrote the rate-limiter does — and you have receipts on Apple's Siri, Samsung's $1T party, DeepSeek's funding U-turn, Adobe's PDF-as-AI-app gambit, Genesis robots solving Rubik's cubes two-handed, and a robot that converted to Buddhism on the way down. We didn't make that last one up.

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.