Everything That Happened in AI on Wednesday - Thursday, May 20-21, 2026

An OpenAI model solved a central math problem, Spotify turned fan remixes into a licensed product, and Starbucks learned AI inventory counts still have to count inventory.

Welcome to the Around the Horn Digest, everything that crossed the AI desk today, sorted before your tabs unionize. The day's headline was OpenAI's geometry breakthrough, but the bigger theme was AI leaving the demo room and colliding with the world: robotaxis hit flooded roads, Starbucks killed an inventory system, California moved first on worker disruption, and Spotify tried to make "responsible AI remixes" into an actual product line. Meanwhile, compute kept looking less like a tech input and more like geopolitics with a cooling bill. Turns out the future arrives with footnotes, lawsuits, and a lot of GPUs. Let's get into it.

Previous digests: Tuesday, May 19 | Monday, May 18 | Wednesday-Thursday, May 13-14 | Tuesday, May 12 | Monday, May 11 | Thursday, May 7 | Wednesday, May 6 | Monday, May 4

Monthly skill digests: AI Skill of the Day Digest — May 2026

Around the Horn — Thursday, May 21, 2026

Note: this is everything from the past two days mixed together, just FYI!

The big story today was that OpenAI said one of its general-purpose reasoning models disproved the 80-year-old Erdos unit distance conjecture, a central problem in discrete geometry (the branch of math that studies point arrangements, distances, and combinatorial structure).

The short version: mathematicians had long believed square grids were the best way to arrange points to maximize unit-distance pairs. OpenAI's model found a new infinite family of point configurations that beat grids, and external mathematicians verified the result. Princeton's Will Sawin immediately refined the construction to at least n^(1+delta) unit-distance pairs with delta around 0.014, and that the proof used sophisticated algebraic number theory tools including infinite class field towers and Golod-Shafarevich theory. Translation: the model did not just pattern-match a contest problem; it connected deep math machinery to a famous open geometry problem.

That made the story more than "AI gets better at math." It became a preview of what scientific AI may look like when models can chain known results across fields in ways humans miss, then hand those results back to experts for compression, verification, and improvement. It also gave the day its weirdest loop: OpenAI's model solved the problem, human mathematicians immediately improved the bound, and the surrounding discourse turned into a live demo of human-AI collaboration.

🏆 TOP 5 NEWS

• The Trump administration planned $2B in quantum-computing grants for firms including IBM, GlobalFoundries, and Rigetti, while taking equity stakes as part of a national industrial-policy push.
• OpenAI generated nearly $6B in Q1 revenue, roughly $1B ahead of Anthropic, with Codex, enterprise sales, and ad testing cited as major contributors.
• Spotify and Universal Music Group announced licensing agreements for paid, "responsible" AI-generated fan covers and remixes, with Variety highlighting the artist-consent and revenue-share framing.
• California Gov. Gavin Newsom signed a first-in-the-nation AI workforce executive order while the White House postponed its own AI order over internal disagreements and concerns it could slow U.S. leadership.
• xAI's Grok reportedly flopped with U.S. government buyers, with Reuters finding only three identified federal use cases despite heavy promotion and low pricing.

Honorable Mentions

• ChatGPT for PowerPoint lets users create, edit, summarize, and polish native PowerPoint decks directly inside PowerPoint.
• Samsung will distribute about $26.6B in bonuses, averaging roughly $340K per chip-division employee, after an AI-boom labor deal.
• Starbucks scrapped its North American AI inventory tool after nine months because it produced frequent miscounts and failed to standardize properly.
• Waymo paused Atlanta service after robotaxis kept driving into flooded roads; San Antonio service had also been suspended.
• Cohere released Command A+, its fastest and most powerful open Apache 2.0 model, with VentureBeat emphasizing lossless quantization (shrinking a model without degrading its answers) and native citations.

🍪 TOP TREATS TO TRY

• Command A+ gives you a 218B-parameter Mixture-of-Experts model (a model that activates only part of itself per request to run cheaper) for enterprise agents, with Apache 2.0 licensing so teams can use and modify it freely; no pricing details for self-hosting beyond hardware costs.
• Studio by Spotify Labs turns your emails, calendar, notes, and web activity into personalized AI podcasts and audio briefings; research preview in 20+ markets.
• Mistral Workflows runs durable business AI workflows with state tracking, approvals, and observability (a dashboard showing what the agent did and where it failed); public preview with docs.
• Microsoft Fara1.5 is a family of computer-use models that can observe browser states, reason, and act, with a 9B model listed in Azure's catalog; no public pricing details in the provided context.
• Patina's Sense1 predicts how human smell receptors respond to molecules, turning fragrance design into a model-guided search problem; company site here, funding coverage here.
• Suzanne turns prompts, photos, or voice into editable 3D models you can export for 3D printing, AR, or design workflows; no pricing details.
• Centaur gives teams multiplayer, self-hosted Slack agents that run in isolated sandboxes (locked-down environments for untrusted code), with Paradigm's launch post, GitHub repo, ACME deployment example, llms-full.txt, advanced permissioning notes, and apps docs.
• Claude Design is becoming a fast sketchpad for app ideas, turning rough prompts into wireframes, ToDo app prototypes, animated scenes, ASCII loaders, and even an open-source design-agent workflow.
• Prime Intellect released a <$1 testbed for studying reward hacking in small RL models, showing that hidden shortcuts can emerge predictably when rewards are too easy, too hard, or poorly calibrated, then launched Prime Sprints with Discord updates, sponsored runs, and $5K+ in credits for community researchers.

🏢 Big Tech & Major Companies

• Spotify's investor-day cluster was basically "AI plus superfans plus audio expansion": Reserved saves concert-ticket purchase windows for top listeners, Studio takes on NotebookLM-style personal podcasts, audiobook updates include ElevenLabs-powered audiobook creation, and the UMG agreement makes fan-made AI covers a paid add-on.
• AMD announced more than $10B in Taiwan ecosystem investments to expand strategic partnerships and advanced packaging capacity for AI infrastructure.
• Workday jumped 14% after raising its margin forecast and pointing to AI product strength.
• Zoom reported Q1 fiscal 2027 revenue of $1.239B, up 5.5% year over year, and forecast stronger-than-expected sales as customers adopted its broader office suite.
• Jeff Bezos hyped AI in a CNBC interview; GeekWire reported Jeff Bezos described his $38B startup Prometheus as an "artificial general engineer" for CAD-style physical-object design, explicitly saying it has nothing to do with robotics; For the full interview, check out CNBC.
• NVIDIA CFO Colette Kress said $1T in cumulative shareholder returns is on the horizon, with the company planning to return about half of free cash flow through buybacks and dividends.
• Intel is dodging HBM shortages for its Crescent Island AI inference GPU by using 160GB of cheaper LPDDR5X memory instead of high-bandwidth memory (HBM, the expensive memory stacked near AI chips for fast data access), trading bandwidth for availability and cost.

💼 AI Productivity, Labor & Economics

• Arena.ai reported that GPT-4-level quality is now about 500x cheaper, with Text Arena's Pareto frontier (the best tradeoff curve between price and performance) moving from roughly $50 per million tokens in 2023 to about $0.10 today for similar quality.
• Ed Zitron argued Anthropic's "first profitable quarter" framing was misleading, claiming EBITDA profitability depended on temporary compute discounts, accounting timing, and funding-round optics rather than sustainable operations.
• Data centers can raise nearby temperatures by up to 4°F, according to Phoenix field measurements, with the ASME paper framing waste heat as an emerging urban thermal hazard.
• Terminal-Bench Science extended Terminal-Bench (a benchmark that tests whether AI agents can complete real terminal-based tasks) into scientific workflows like MRI reconstruction, virtual drug screening, and ice-crystal analysis.
• Dan Shipper argued AI progress creates more work for humans, not less, because automation floods the world with generic output and raises the value of judgment, context, taste, and differentiation.
• Peter John Lambert and Yannick Schindler argued remote work, not generative AI exposure, explains much of the early-career hiring decline, based on 243M new hires across occupations.

🤖 AI Agents & Infrastructure

• Anthropic is reportedly in talks to use Microsoft's AI chips, which would be a customer win for Microsoft's in-house AI server hardware.
• The Anthropic-backed enterprise services firm acquired Fractional AI as its first deal, making the San Francisco firm the operational core of the unnamed Blackstone/Anthropic/Hellman & Friedman venture.
• Modal raised $355M at a $4.65B valuation as AI coding workloads drove demand for its cloud infrastructure.
• Hark raised more than $700M at a $6B valuation for its secretive AI hardware/interface push, with Brett Adcock's X post surfacing the founder angle.
• InferenceBench benchmarks open-ended AI-agent optimization of LLM inference (making models respond faster/cheaper), with the paper PDF and Maksym Andriushchenko's post surfacing it.
• PEEK gives long-context agents a small "context map" of recurring external environments like repos or document corpora, improving long-context reasoning while reducing cost; code here, launch context here.
• GEPA's optimize_anything optimizes any measurable text parameter, including prompts, code, policies, and agent architectures, with Lakshya Agrawal's post framing it as "if you can measure it, you can optimize it."
• Privy launched agentic wallets so AI agents can spend money under programmable policies, key isolation, approvals, and permissions.
• Human Principal provides privacy-preserving proof that a real human is behind an agent/API call, using cryptographic attestations without exposing identity.

💻 AI Coding & Developer Tools

• OpenAI's Codex /goal mode lets you give Codex a durable milestone so it can work toward it for hours or days, while the use cases page shows team workflows; OpenAI Devs' post surfaced the update.
• Codex Appshots on Mac captures an app window's screenshot plus visible and scrolled-off text, file paths, and URLs so Codex can work with richer desktop context.
• Microsoft's Magentic-UI is an experimental agent that works across the browser and local filesystem, with later Microsoft AI Frontiers posts pointing toward smaller computer-use agents.
• Google Stitch added one-click export to Lovable, passing screenshots and HTML references so Lovable can generate a full-stack app matching the design.
• Eric Zakariasson shared Cursor's "thermo-nuclear code-quality review" skill, which blocks bloated files, thin wrappers, and complexity-moving PRs; marketplace reference here.
• KingBootoshi's directional-prompting repo packages outcome-first and directional-language patterns for Claude Code and Codex CLI; launch post here.
• OpenAI Computer Use for Codex lets Codex control macOS desktop apps and browsers with per-action approvals, including locked-use operation.

🔬 AI Research & Models

• Cohere's Command A+ collection includes the W4A4 quantized release, with Cohere's launch post, follow-up post, and Nick Frosst's post framing it as the company's best open-source model yet.
• ArtifactLinker links scientific artifacts for automatic state-of-the-art discovery, with AllenAI's repo, ArtifactBench dataset, and Haofei Yu's post.
• Delta Attention Residuals upgrades residual connections to route over deltas between layers, with code and Cheng Luo's launch post.
• A Bitter Lesson for Data Filtering argued that in high-compute, data-scarce training regimes, larger models can benefit from unfiltered data that smaller models cannot use; Tatsunori Hashimoto's post surfaced the result.
• Probabilistic Tiny Recursive Model scales tiny recursive reasoning models at test time by injecting noise into latent states and picking the best rollout with the model's own Q-head (a score head that estimates answer quality); Amin Sghaier's post supplied the demo context.
• Qwen3.7-Max was framed as an agent-centric flagship model for coding, tool use, and long-horizon execution, with access notes via Qwen chat, Model Studio, Alibaba Qwen post 1, post 2, Artificial Analysis benchmark, Kimmonismus take, sudoingX take, Andrew Curran reaction, follow-up, jiqizhixin coverage, and a Weixin link.
• MIRO integrates multiple reward signals directly into text-to-image pretraining, claiming 19x faster convergence and 370x less compute than FLUX, with paper, GitHub, Hugging Face model, Space, and Nicolas Dufour's post.
• Synthetic Persona Pretraining embeds moral reflections into pretraining data to shape model personas from token zero; Julian Minder's post provided the release context.
• Speculative Decoding: Performance or Illusion? was highlighted by Jiaxiang Yu as an MLSys 2026 honorable-mention paper on speculative decoding (a technique where a small model drafts tokens and a larger model verifies them).
• Dual-Rate Diffusion uses an interleaved heavy-light network to speed up diffusion models (image/video generation models that refine noise into media), with Grigory Bartosh's post, related Transition Matching, and Transition Matching Distillation for Fast Video Generation.
• OCTOPUS compresses KV cache (the model's short-term memory during generation) for transformers using octahedral quantization; project page here and Mark Boss post.
• ScheduleFree+ scales learning-rate-free and schedule-free optimization to large language models; Aaron Defazio's post surfaced it.
• Under the Hood of SKILL.md studied semantic supply-chain attacks on AI agent skill registries, with code and Soheil Feizi's post.
• Lattice Deduction Transformers showed tiny models can solve structured reasoning tasks when the representation bakes in the right priors, with MrCatid's post and Yann LeCun's response cautioning that current AI is still far from human learning/common sense.
• Implicit Curriculum Hypothesis found LLMs acquire skills in a consistent order during pretraining, with code here and Emmy Liu's post.
• AVSD uses multiple privileged "views" like full solutions, rationales, answers, code, and feedback to create better self-distillation rewards (training signals generated from model outputs), with Hugging Face paper page, GitHub, Duy Nguyen's post, and Hanqi Xiao's related post.
• ExploitGym tested whether AI agents can turn security vulnerabilities into real attacks, with Dawn Song's post surfacing the security angle.
• Robin, a multi-agent system for automating scientific discovery, proposed and lab-validated dry age-related macular degeneration drug candidates; Benjamin Chang's post highlighted the Nature publication.
• Metropolis-Adjusted Diffusion Models applied MCMC-style correction (a statistical accept/reject method) to diffusion samplers using only the score function; Tyler Farghly's post framed the result.
• Apple HeadsUp reconstructs high-quality 3D Gaussian heads from multi-view captures without test-time optimization; KIRI Engine's post surfaced it.
• DashAttention proposed differentiable and adaptive sparse hierarchical attention (a way for models to focus computation on the most useful tokens), with GitHub and launch post.
• Eric Jang's LLM RL variance note examined variance in reinforcement learning for language models, with post.

🏛️ AI Policy, Governance & Safety

• President Trump said he postponed the AI executive order because he did not want anything getting in the way of the U.S. lead over China; Axios separately reported the signing ceremony delay. TechCrunch's version emphasized Trump's quote about not wanting to get in the way of U.S. leadership, while Hadas Gold's X post captured the quote directly.
• Taiwanese prosecutors sought to detain three people over Nvidia AI chip smuggling, with Bloomberg framing it as Taiwan's first semiconductor-smuggling crackdown.
• The U.K. AI Security Institute warned AI oversight may get harder as systems become more capable, and linked a model transparency role as part of its capacity-building.
• Berkeley Law released a post-faculty-meeting AI policy PDF, with Hoofnagle's post summarizing exam restrictions and disclosure requirements.
• Interactive Evaluation Requires a Design Science argued that evaluating interactive AI systems needs explicit design principles, with a curated GitHub resource collecting the framework, taxonomy, design principles, and benchmark list.
• NVIDIA Verified Agent Skills introduced portable, signed, scanned skills for agents, with NVIDIA's post framing it as capability governance.

🛠️ AI Tools & Products

• Runway Aleph 2.0 in Edit Studio lets creators edit one video frame and propagate that edit across the clip; app entry here, related launch/demo posts here.
• Stable Audio 3 generates instrumental music and sound effects with local-running options.
• MIGA enhances train-free infinite-frame video generation for consistent long videos, with Hugging Face paper page and HuggingPapers post.
• ARQ is an AI film studio using AI for production while keeping human creative direction, with Starks ARQ post.
• Hell Grind, a 95-minute fully AI-generated Cannes film, reportedly cost $500K to make, including $400K in compute; Newser summarized the two-week production angle.
• LeRobot Humanoid is an open, low-cost, mostly 3D-printed humanoid platform for robot learning, with GitHub repo and LeRobot's post.
• Rerun 0.32 became a unified data layer for robotics data, with Rerun's post, Bearly AI, and Noah Solomon demos.
• ppf-contact-solver is a GPU-accelerated contact solver for cloth, solids, and rods, with Ryoichi Ando's post noting the Blender add-on and scale claims.
• August Robotics raised $30M to automate precision construction tasks with robots.
• BigSet is a multi-agent harness for turning prompts into structured datasets with confidence-ranked rows.
• fal Assets organizes generated images, video, and audio into searchable smart collections, with fal's post.
• Pangram Labs reported more AI-generated Commonwealth Short Story Prize winners.
• Arena.ai's text-to-image leaderboard ranks image models via blind human preference.
• Mega-ASR targets difficult real-world speech recognition, with model, dataset, Xie Zhifei's post, HuggingPapers, Techmeme, and AI News 24/7.

📊 Fundraising & Deals Roundup

• Hark — $700M+ at a $6B valuation for AI hardware/interface work.
• Modal — $355M at a $4.65B valuation for AI infrastructure and coding workloads.
• Exa — $250M at a $2.2B valuation for AI-native search; related Exa post.
• Commure — $70M at a $7B valuation for healthcare operations AI, with Fierce Healthcare coverage.
• Pivot — $40M to expand its AI procurement and finance platform.
• August Robotics — $30M for precision construction robots.
• Patina — $2M to build scent-foundation-model technology.
• Manus — reportedly considered raising about $1B to unwind Meta's takeover under pressure from Beijing.

🎙️ Interviews, Panels & Podcasts

• Patrick O'Shaughnessy interviewed Gavin Baker on TSMC and the AI bubble, with related Patrick thread and earlier post.
• Noam Brown argued the field is underestimating inference compute (compute spent while a model is thinking, not training), with Haider's post.
• Training Composer 2 featured Sasha Rush explaining how Cursor's research team builds Composer 2, from base model choice to long-horizon reinforcement learning.
• LiveKit's YouTube channel surfaced alongside the earlier LiveKit agent tooling cluster.

💡 Industry Commentary & Analysis

• Aymeric Roucher mapped the three tiers of the AI compute race across U.S. hyperscalers, Mistral, and Chinese giants/labs, with his X post.
• turbopuffer reportedly crossed $100M run-rate 19 months after $1M while staying profitable on less than $1M raised.
• Mitchell Hashimoto argued teams should aggressively fork and trim dependencies to reduce supply-chain risk instead of blindly updating everything.
• Greg Kamradt broke down the verifiability spectrum, explaining why math and code accelerate faster than messy real-world domains because they have cheap feedback loops.
• Turing Post argued Google AI Search is becoming an agentic execution layer, with their X post.
• Theo argued the browser is the new OS and that apps may collapse into agent skills.
• Llion Jones/Sakana's post argued the field is entering a post-Transformer era because current scaling laws are hitting diminishing returns.
• Ethan Mollick flagged AI reviewer research around Nature-family reviews, linked to the paper, and separately noted the OpenAI math result's human-refinement loop.
• Goodfire argued sparse autoencoders miss curved model geometry, suggesting clusters of related features reveal more of how models internally represent concepts.
• MTS published "Eighty years of AI research, in eleven eras", with post 1, post 2, and post 3.
• Haider predicted Gemini 3.5 Pro would launch in June as Google's new flagship for coding, agents, reasoning, and memory.
• SemiAnalysis posts tracked hardware and compute-market commentary, including AI memory hierarchy satire and related follow-up.
• tszzl called frontier-lab model development "the highest ELO game in the world", and separately commented on the day's model-race dynamics.
• BetaTomorrow proposed "Dataualism", a data-first doctrine for AI systems analogous to legal textualism.
• FelixCLC_ replied to OpenAI's math breakthrough with a favorite multiplication paper, linking to the arXiv PDF.
• t6aguirre joked the key to keeping up with AI is learning to pronounce Erdos.
• Nathan Calvin pointed out the contradiction between AI-job-loss dismissal and Andreessen's own preference for tireless AI workers.
• Electric Era's Quincy Lee announced CoPower, with follow-up, as battery-backed power management for AI data centers.
• Max Marchione highlighted Eli Lilly's Retatrutide Phase 3 weight-loss results and predicted it could become the best-selling drug of all time.
• Katherine Boyle/Kitty Mayo shared a BCI drone-control demo featuring Agnessa Pedersen.
• delba_oliveira showed Claude Code plus Remotion syncing a music video to finger-click timings.

Wednesday, May 20, 2026

Note: These are a bit out of order; will go back and sort later.

Honorable Mentions

• Google announced a $15B data center campus in Missouri, one of the largest corporate investments in state history.
• Cursor reportedly hit a $3B annualized revenue rate as a SpaceX deal looms.

Treats to Try

• ChatGPT for PowerPoint creates and edits native PowerPoint decks from notes, docs, spreadsheets, screenshots, or prompts while keeping slides fully editable; beta available globally.
• Devin for Windows runs natively in its own Windows VM so it can build, run, and test software on the world’s most common desktop OS; no pricing details.
• KVCache.ai estimates KV cache memory needs for DeepSeek, GLM, Kimi, Qwen3, and MiniMax models so developers can see how much VRAM long context will consume; no pricing details.
• text-to-CAD adds new Codex / Claude CAD skills, mechanism validation, STEP-file parameters, animations, robot-file support, and an interactive browser CAD viewer; open-source.
• Replication Radar scans scientific fields for replication-crisis warning signs like citation rings, author clustering, institutional monoculture, and fragile sample sizes.

🏢 Big Tech & Major Companies

• Nvidia CEO Jensen Huang joined Trump’s China trip as a last-minute Air Force One addition after reportedly being left off the original list.
• Google DeepMind researchers reportedly left to start labs as Google’s cloud deals limited internal compute access.
• Google DeepMind hired more than 20 Contextual AI researchers and licensed the Bezos-backed startup’s technology.
• Nvidia began Vera CPU deliveries ahead of earnings, with analysts raising price targets before the report.
• GE Aerospace used a custom generative AI app to complete preliminary design studies for a hypersonic dual-mode ramjet.
• Google drew backlash after replacing Gemini’s flat prompt caps with compute-based usage limits.
• Sundar Pichai argued at I/O that AI efficiency gains will justify Google’s spending surge.
• Tesla launched FSD Supervised in China after years of delays.
• Airbnb CEO Brian Chesky pushed back on a congressional investigation into the company’s use of Chinese AI, saying lawmakers were misunderstanding it.
• ASML confirmed talks with Musk on the Terafab chip project.

💼 AI Productivity, Labor & Economics

• Accenture is expanding entry-level hiring in 2026, betting Gen Z workers bring stronger AI fluency than senior colleagues.
• Databricks CEO Ali Ghodsi argued that enterprise AI’s biggest problem is context, not intelligence.
• A study of 95,000 students at 20 public research universities found about one-third regularly used generative AI tools to cheat.
• Leopold Aschenbrenner’s Situational Awareness LP reportedly disclosed billions in bets against chips after growing from $225M to more than $5.5B in about a year.
• British Progress argued that even superhuman AI may not erase human labor quickly because firms, tacit knowledge, contracts, and bundled jobs slow substitution.
• Sam Rodriques argued AI will accelerate science over the next 5-10 years, but human taste and hard-to-verify experiments will keep scientists in the loop.

🤖 AI Agents & Infrastructure

• Vapi raised $50M at a $500M valuation after an Amazon Ring deal for voice-agent infrastructure.
• Prime Intellect released <$1 reward-hacking testbeds for small RL models and launched Prime Sprints, a sponsored community research program.
• Auto-Dreamer trains language agents to consolidate memory offline, shrinking active memory 6-11x while improving ScienceWorld, ALFWorld, and WebArena performance.
• atomic.chat tested frontier models on a self-improving Tetris-bot loop, with Qwen 3.7-Max improving 56% at $1.32 versus smaller gains from Claude and GPT-5.5.

💻 AI Coding & Developer Tools

• CODA rewrites Transformer operations as fused GEMM epilogue programs, letting surrounding model ops run while data stays on-chip instead of creating slower memory-bound kernels.
• CODA’s GitHub repo shows the kernel implementation, and Han Guo’s post frames the fun bit: LLMs can author some of the optimized kernels too.
• POW3R dynamically reallocates rubric-based RL training pressure toward criteria that actually distinguish rollouts, reaching validation thresholds 2.5-4x faster.

🔬 AI Research & Models

• Mosaic is a probabilistic weather model that generates 24-member, 10-day global forecasts in under 12 seconds on one H100 while preserving high-frequency detail.
• Nature Physics published research showing simple input-output dependencies can explain most neuronal activity across mouse and C. elegans recordings.
• WARDEN transcribes and translates Wardaman, an endangered Indigenous language with only two speakers left, using just six hours of aligned audio.
• DexEXO is a dexterous exoskeleton for collecting scalable robot-hand demonstrations from different human operators.
• OLMoEarth v1.1 makes Earth-observation models for satellite imagery more efficient by reducing multi-resolution Sentinel-2 inputs to fewer tokens.
• HyperTransport speeds text-to-image activation steering by predicting layer interventions directly from CLIP latents.
• AI reviewer research found frontier models can raise more correct and well-evidenced review criticisms than low-rated human reviewers, but still miss field norms and reduce perspective diversity.

🏛️ AI Policy, Governance & Safety

• The Pentagon is deploying Anthropic’s Mythos cybersecurity model to find and patch U.S. government vulnerabilities even as it tests alternatives.
• The Pentagon is testing rival AI models as it moves away from Anthropic’s Claude.
• The Pentagon launched a task force to test offensive AI hacking tools on classified networks.
• Shield AI was selected to integrate Hivemind autonomy into the Pentagon’s low-cost LUCAS drone swarm program.
• OpenAI plans to provide Japan with GPT-5.5-Cyber for defense.
• Nature covered the debate over restricting biological AI tools that could help design viruses, toxins, or other bioweapons.
• Imec’s CEO warned Europe must build its own AI chip design companies or fall further behind the U.S. and China.

📊 Fundraising & Deals Roundup

• Google — $15B data center campus in New Florence, Missouri.
• Amazon — $33B cumulative cloud and AI infrastructure push across Southeast Asia.
• IBM — $1B CHIPS grant for a dedicated quantum chip foundry.
• Vapi — $50M Series B at a $500M valuation for voice-agent infrastructure.
• Google DeepMind — hired 20+ Contextual AI researchers and licensed the startup’s technology.
• Nvidia — private AI startup holdings nearly doubled to $43B in one quarter.

💡 Industry Commentary & Analysis

• Yaroslav Azhnyuk argued autonomous FPV drones have reshaped modern warfare and left Western militaries underprepared.
• Adam Kucharski examined whether AI systems are learning real cultural signals or amplifying artificial stereotypes.
• Pedro Serôdio argued that AI replacing tasks does not automatically dissolve jobs because firms bundle tasks, tacit knowledge, and incomplete contracts (blog).

Additional Finds (Not yet sorted)

• LiveKit voice-agent starter stack: LiveKit Agents is a realtime voice/video AI-agent framework with STT/LLM/TTS plugins (speech-to-text, language model, and text-to-speech), semantic turn detection, telephony, avatars, and MCP tool integration; agent-starter-python gives you a complete runnable Python voice-agent starter with Docker, evals, background voice cancellation, and observability; agent-skills provides reusable coding-agent guidance for low-latency voice-first LiveKit architecture.
• AI jobs apocalypse debate: Business Insider's James Manyika story said Google's SVP thinks Silicon Valley is overstating the AI jobs apocalypse and that most effects will be "jobs changed" rather than jobs erased; Vice argued students booing AI at commencement are reacting rationally to a technology framed as labor replacement; MIT News covered David Autor's study finding new tech-enabled jobs historically skewed toward young, college-educated workers, though AI's path is still uncertain; Business Insider's Gen Z piece said AI is pushing junior workers into more complex tasks earlier; Fast Company argued domain expertise, judgment, and relationships remain the hard-to-automate 20%; and McKinsey/PwC/EY executive-assistant layoffs showed support roles already getting cut or outsourced as AI accelerates.
• Anthropic compute and profitability cluster: Yahoo Finance reported Anthropic secured a $45B SpaceX compute deal for Claude, while Axios framed the same two-hour news window as a revealing cross-section of the AI revolution: OpenAI's math result, Anthropic's profit path, Nvidia earnings, SpaceX's IPO filing, and a Trump AI order.
• China and AI governance: War on the Rocks argued China's AI-governance offensive exports censorship-aligned standards through U.N. frameworks, bilateral deals, and Digital Silk Road infrastructure, raising compliance costs for U.S. firms and threatening U.S. tech leadership.
• Literature, detection, and anti-AI culture: The Atlantic argued the Granta AI fiction scandal shows detectors and public ridicule are becoming a de facto enforcement layer for human-written literature; The Handbasket argued shunning AI is a rational human choice because current tools hallucinate, outsource thinking, and serve concentrated power.
• Enterprise-agent startups: SiliconANGLE reported Tribal AI raised $10M to build metadata-native enterprise agents that ingest permissions, objects, rules, and system context from tools like Salesforce; Tech.eu reported Pivot raised $40M for procurement agents; and Fierce Healthcare plus Commure's press release covered Commure's $70M raise at a $7B valuation for healthcare operations AI.
• The web for agents versus people: Tech Policy Press argued llms.txt, MCP servers, and LLM-specific website instructions risk making the web more accessible to AI agents while leaving human accessibility behind.
• Recursive self-improvement concern: POLITICO framed the next AI policy race around recursive self-improvement, where AI systems could iterate faster than human oversight can keep up.
• AI layoffs and corporate restructuring: TechCrunch reported Intuit will lay off more than 3,000 employees, about 17% of its workforce, to simplify the company and refocus on AI products for TurboTax, QuickBooks, and Credit Karma; WSJ's Standard Chartered piece covered CEO Bill Winters walking back his "lower-value human capital" comment after backlash over AI-linked support-staff cuts.
• AI market rallies and power deals: CNBC reported SoftBank shares soared about 20% as Nvidia earnings reinforced AI momentum; CNBC's Bloom Energy story reported Bloom rose 12% after a $2.6B Nebius AI-infrastructure partnership for fuel-cell power at data centers.
• Google product and ad push: Google's ads blog announced Gemini-powered Search ad formats including Conversational Discovery ads, Highlighted Answers, AI shopping ads, Business Agent for Leads, and expanded Direct Offers; Reuters said Demis Hassabis went on offense at I/O, calling AGI "the most profound technology ever invented" while unveiling Gemini upgrades and Gemini for Science.
• AI failures in operations: Tom's Hardware reported Pizza Hut franchisees sued over an AI delivery system after alleged delivery times worsened from under 30 minutes to more than 45 minutes; Reuters reported Starbucks scrapped its AI inventory tool after miscounts; and TechCrunch reported Waymo paused service after robotaxis drove into flooded roads.
• AI science and big claims: The Guardian covered Anthropic co-founder Jack Clark predicting AI will help make a Nobel-prize-winning discovery within a year, AI-only companies will generate millions within 18 months, and AI may design successors by 2028, while also warning of risk.
• Workday and enterprise incumbents: WSJ reported returning Workday CEO Aneel Bhusri cut agent projects from 50 to 20, hired a new chief AI officer, and planned roughly 15 new agents in 2026; CNBC added the earnings-market response.
• Unexpected AI adoption: New York Magazine reported that some Amish entrepreneurs in Holmes County, Ohio use ChatGPT via filtered internet, dumb phones, or 1-800-ChatGPT for contracts, emails, dashboards, and research while treating it as a pragmatic labor-saving tool.
• Commencement AI backlash: Fast Company reported Steve Wozniak got applause, not boos, by telling graduates they have "AI" meaning actual intelligence.
• AI cybersecurity claims: Reuters reported fears that Anthropic's Mythos model would unlock unfettered hacking looked overstated one month after release, with practitioners calling gains real but incremental.
• Cohere / Palantir: TechCrunch reported Cohere quietly worked with Palantir to deploy models for unnamed customers, including strict data-storage and Arabic-inference use cases.
• Lapdog / coding-agent growth: AJ Stuyvenberg shared early Lapdog launch data showing rapid coding-agent adoption across Codex, Cursor, and other tools.
• Houda Nait on scientific discovery: Houda Nait's essay and X post argued OpenAI's Erdos result marks a new era of scientific discovery because a general-purpose reasoning model tackled a central open math problem.
• GRAM recursive reasoning: Generative Recursive Reasoning and paper, surfaced by Sungjin Ahn, turn recursive latent reasoning into probabilistic multi-trajectory computation so models can explore multiple solution paths.
• Mix-Quant for agentic LLMs: Mix-Quant, surfaced by _akhaliq and HuggingPapers, uses cheaper quantized prefilling and precise decoding for agentic LLMs, speeding long-context work while protecting final-answer quality.
• Gemini 3.5 Flash automation benchmark: OfficialLoganK reported Gemini 3.5 Flash ranked first on Zapier's Automation Bench, beating other frontier models on real-world automation tasks at lower cost.
• Bubeck on OpenAI math proof: Sebastien Bubeck's post explained the OpenAI unit-distance proof as cross-field algebraic number theory applied to an elementary geometry problem, emphasizing that the model chained known results rather than inventing brand-new math.
• ECHO speculative decoding: ECHO proposed elastic speculative decoding with sparse gating for high-concurrency scenarios, improving LLM serving efficiency by adapting draft/verify behavior under load.
• Metropolis, GRAM, and other generative-model papers: Metropolis-Adjusted Diffusion Models, GRAM, Dual-Rate Diffusion, and MIGA formed the day's "faster and more reliable generation" research bench.
• PopuLoRA / Vmax RL cluster: Vmax is building automated reinforcement-learning infrastructure, and PopuLoRA is its population-based self-play method for improving reasoning models; MavorParker's X post framed it as Vmax's launch story, while the arXiv paper explains the LoRA-adapter self-play setup (LoRA adapters are small trainable add-ons to a larger model).
• Claude Code usage telemetry: Boris Cherny's post said Claude Code now breaks down /usage by Skill, Agent, MCP tool, and Plugin, helping teams see which agent components are burning context and tokens.
• NVIDIA NeMo Safe Synthesizer: mvansegb's post surfaced NeMo Safe Synthesizer, an open-source tool for creating privacy-preserving synthetic tabular data with quality and privacy checks.
• OpenAI geometry reaction cluster: Jason Abaluck argued verifiable fields like math may have a short human-AI collaboration window before AI gets superhuman; Scott Kominers highlighted the immediate human refinement loop; Alvaro Lozano-Robledo shared visualizations of the new lattice constructions; Julian Bruns shared Will Sawin's tighter explicit bound; Andrew Curran pointed to the explicit lower-bound arXiv paper; memecrashes connected the moment to old machine-scale arguments in On the Impossibility of Supersized Machines; and FelixCLC_'s related arXiv PDF was a humorous math-culture sidebar.
• Microsoft lightweight agent cluster: Microsoft AI Frontiers described MagenticLite, including Fara1.5 9B for web navigation and MagenticBrain 14B as an orchestrator, expanding the Magentic-UI browser/filesystem agent line.
• Late-interaction retrieval: Benjamin Clavie's Google Slides deck argued ColBERT-style late interaction retrieval (a search method that compares many small text vectors instead of one compressed summary vector) is more robust than single-vector retrieval for agents and multimodal search.
• Google I/O mega-roundup: Google's 100 I/O announcements post collected Gemini Omni, Antigravity, Universal Cart, and broad Search/Workspace/Android updates; Google's X post surfaced the same I/O momentum from the primary account.
• Grok in OpenCode: xAI's announcement and xAI's post said SuperGrok and X Premium subscribers can use Grok inside OpenCode via /connect, bringing xAI's coding model into a local terminal workflow.
• AGI-levels discourse: Levels of AGI proposed a capability/autonomy ladder for AGI progress; Kol Tregaskes argued "personal AGI" is about practical usefulness to a specific person; and Sam Altman said OpenAI is focused on AGI accelerating research, companies, and individual goals.
• fast-rlm / recursive agents: AVB's post surfaced fast-rlm, a recursive language-model repo with REPL tool calling, Python function passing, variable storage, and sub-agent coordination.
• AI chip memory-cost cluster: Epoch AI's post, Epoch's analysis, snewmanpv, willdepue, and Hedgie Markets all pointed to the same bottleneck: high-bandwidth memory now makes up roughly 63% of AI chip component costs, shifting the AI capex story from chips alone to memory supply.
• Codex computer-use demo cluster: OpenAI Devs, OfficialLoganK's locked-Mac demo, and Codex Releases all demonstrated Codex computer use, including desktop control, locked-mode operation, and per-action human approvals.
• AI revenue / financial run-rate chatter: Wall St Engine and Wall St Engine's follow-up tracked reported OpenAI and Anthropic revenue run-rates, including the claim that OpenAI generated about $5.7B in Q1 while Anthropic projected a much larger annualized Q2 run-rate.
• Letta memory transparency: Letta's post showed context-structure visualization in Letta Code's Memory view, making an agent's core memory, external memory, skills, links, and version history inspectable.
• Fuchsia / hardware compliance: YC's launch post and YC launch page positioned Fuchsia as "Vanta for hardware," automating compliance and certification work for hardware startups.
• Spatial AI thesis: Speridlabs' essay and launch post argued the next intelligence leap is spatial AI: foundation models that unify perception, reasoning, simulation, and generation inside a coherent 3D-world representation.
• Modal funding reactions: bernhardsson and Modal supplied founder/company-side context for Modal's $355M raise beyond the Reuters funding story.
• Quantitative economics / DeepHAM: Yucheng Yang, Jiequn Han, and Quantitative Economics surfaced the May 2026 issue and DeepHAM, a deep-learning approach for heterogeneous-agent macroeconomic models.
• OpenAI Devs update cluster: OpenAI Devs post 1, post 2, and post 3 collectively covered Codex prompting, /goal, Appshots, use cases, and computer-use updates.
• Anything / Jam automation demos: Anything post 1, Anything post 2, Justin Sun's demo, Jam, jam-nodes, and Jia Seed's post showed growth/distribution agents, including Reddit-to-LinkedIn repurposing, AI-search visibility monitoring, and extensible automation nodes.
• PopuLoRA Colab: Tianshi Li shared a one-click Colab-style route for running the PopuLoRA co-evolution loop with Vmax's environment generator.
• Apps versus agents: Francois Chollet argued apps and UIs may be displaced as agents perform tasks directly.
• Notion custom agents: Notion's guide explained how to tune model choice, triggers, scope, and instructions for faster, cheaper, more reliable custom agents.
• Papers With Code / open model surface: Papers with Code was included as a trending research discovery surface, while Cohere Command A+ BF16 and Clement Delangue's post surfaced the open-weights model-distribution angle.
• Sapient HRM-Text: Sapient AI's post, HRM-Text GitHub, and HRM-Text-1B on Hugging Face described a 1B text-generation model built on HRM architecture with task completion and latent-space reasoning.
• Devin Minesweeper: Rafa Schwinger's post, Anne Ouyang's post, and Devin's Minesweeper demo showed a playable Minesweeper game built by Devin on a Windows VM.
• Okara: Okara and Okara's post positioned the product as an AI CMO for SEO, Reddit engagement, content creation, and always-on marketing opportunities.
• ESI-Bench: ESI-Bench, paper, and GitHub benchmark embodied spatial intelligence tasks that close the perception-action loop for vision-language-action systems.
• xAI / SpaceX / AI infrastructure cluster: TechCrunch on xAI generator lawsuits and $2.8B turbines, xAI's $6.4B burn, SpaceX's AI-heavy IPO filing, the SEC filing, Elon Musk's post, and related X broadcast all point to the same infrastructure story: SpaceX is now tied to AI compute sales, xAI's costs are huge, and power procurement is becoming core to AI strategy.
• OpenAI math coverage across outlets: TechCrunch's coverage of OpenAI's 80-year math claim independently framed the result as credible because mathematicians who had challenged prior AI-math claims backed this one.
• AI search market: TechCrunch reported AI search startups are blowing up, and Exa's Bloomberg funding story supplied the $250M/$2.2B proof point.
• OpenAI IPO chatter: TechCrunch reported OpenAI was barreling toward a possible September IPO, a capital-markets follow-on to the Q1 revenue story.
• IrisGo desktop agent: TechCrunch profiled IrisGo and IrisGo's site described it as an operating layer that learns desktop workflows across apps and automates them.
• NanoClaw: TechCrunch reported NanoClaw turned down a $20M buyout and raised $12M seed, while NanoClaw's site framed it as a secure, containerized alternative to OpenClaw for personal agents.
• Figma AI assistant: TechCrunch reported Figma added an AI assistant that generates and edits designs from natural-language prompts.
• Hark secondary coverage: TechCrunch's Hark story reinforced the Bloomberg funding item and added the "universal multimodal interface" framing.
• Cybercrime / VPN takedown: TechCrunch reported law enforcement shut down a VPN used by two dozen ransomware gangs, a cybersecurity enforcement item adjacent to AI security but still relevant to the broader tech-risk digest.
• AI search alternatives: TechCrunch's Google alternatives piece pointed readers to Open Web Engine, Kagi, DuckDuckGo, Startpage, udm14 and its GitHub repo, Brave, Ecosia, plus Ecosia's financial reports and blog, as the web reacts to AI-heavy Google search.
• Spotify AI-audio expansion: Spotify's save-to-Spotify repo and earlier TechCrunch piece supported the broader point that Spotify wants to become a home for generated personal audio, not just music streaming.
• The Path AI therapy: TechCrunch profiled The Path and The Path's site positioned it as an outcome-focused AI therapy app with safety benchmarking and 50K+ members.
• Google agent ecosystem skepticism: TechCrunch argued Google's I/O agent ecosystem may be confusing for consumers, adding a skeptical consumer-adoption angle to the Google I/O cluster.
• AI recycling / aluminum: TechCrunch covered recycling startups using AI to recover aluminum, with Sortera as a company example.
• NVIDIA $200B agent-CPU market: TechCrunch covered Jensen Huang's claim that CPUs for AI agents represent a new $200B market.
• Clouted: TechCrunch reported Clouted raised $7M to help short videos go viral, with Clouted's site framing it as distribution-first marketing.
• NVIDIA earnings cluster: TechCrunch reported NVIDIA posted another record quarter and $43B in startup holdings, while NVIDIA investor results, NVIDIA newsroom, and CNBC's stock-reaction piece supplied primary numbers and market context.
• Patina / smell model reactions: Hssein Mzannar, Yohei Nakajima, and the shortlink to Patina's whitepaper all supported the Sense1 framing: a foundation model for smell and taste that maps receptor activation to perception.
• Mistral Workflows community: Sophia Yang and the Mistral Workflow Community Challenge rules supplied the developer-community angle around Mistral Workflows.
• Rocklin Lab dataset registration: Rocklin Lab's MGnify Stability registration form asked users of its protein-stability dataset to register usage so the lab can track adoption and fund future releases; _simonsmith's post surfaced the dataset context.
• Derrick Choi / Stitch reaction: Derrick Choi reacted to the Stitch-to-Lovable export as a meaningful bridge from AI design mockups to working apps.
• Codex research workshops: nasqret described twice-weekly Codex jam sessions for domain researchers, where scientists built research apps after basic ChatGPT and Codex training.
• HuggingPapers MIGA surface: HuggingPapers surfaced the MIGA infinite-frame video paper to the Hugging Face research audience.

Previous Around the Horn Digests

Catch up on everything you missed:

• Tuesday, May 19: Google I/O pushed Gemini agents across Search, Android, Workspace, YouTube, and shopping while Anthropic hardened Managed Agents and OpenAI expanded provenance.
• Monday, May 18: GPU rentals spiked, H100s got harder to rent, and the class of 2026 became the first full AI-era graduating class.
• Wednesday-Thursday, May 13-14: A two-day digest tracking the flood of AI earnings, papers, and product launches.
• Tuesday, May 12: Anthropic refused China access to its newest model, Isomorphic raised $2.1B, and supply-chain attackers hit Mistral and TanStack.
• Monday, May 11: Cerebras upsized its IPO, Cowboy Space raised for orbital data centers, and Google confirmed the first criminal AI-discovered zero-day.
• Thursday, May 7: Anthropic shipped Natural Language Autoencoders, Google DeepMind detailed AlphaEvolve, and Cloudflare cut 20% of its workforce.
• Wednesday, May 6: Anthropic ran its developer day, signed a SpaceX compute deal, and Apple settled over the AI Siri it never shipped.

That's a Wrap

That's 140+ unique stories from today's consolidated link dump. If you made it to the bottom, you now know more about AI remix licensing, urban data-center heat, and Erdos geometry than any normal person should be expected to carry at dinner.

For the daily version, 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.