Everything That Happened in AI Today Tuesday, May 26

Welcome to the Around the Horn Digest, everything that crossed our desk today, sorted. Tuesday was a chokepoint day: people, chips, compute, courts, data centers, and the right to build the next layer of AI in public. China moved to keep private-sector AI talent close. Qualcomm landed a ByteDance infrastructure win. OpenRouter raised like the model-routing layer is becoming default plumbing. Meanwhile, US law enforcement started naming “anti-tech extremism,” and the Pope’s AI warning gave the backlash a moral vocabulary. Nothing says “AI is eating the world” like zoning fights, new travel restrictions, and a surprise Vatican subplot. Let’s get into it.

Previous digests: Monday, May 18 | Wed-Thurs, May 13-14 | Tuesday, May 12 | Monday, May 11 | Weekend, May 9-10

BIG AI NEWS

China expanded overseas travel curbs to top AI talent at private firms such as Alibaba and DeepSeek, requiring Beijing approval for international trips to protect strategic AI know-how (Tom's Hardware).
Qualcomm struck a deal to supply millions of AI chips to ByteDance for data centers, a major step in its push beyond smartphone processors (Reuters, Sherwood).
OpenRouter raised a $113M Series B led by CapitalG, more than doubling its valuation to $1.3B as usage grew 5x in six months to 25T tokens a week (AFP, post).
Elon Musk said xAI finished training Grok V9-Medium, a 1.5T-parameter model trained with substantial Cursor data, with fine-tuning underway and public release expected in 2-3 weeks. X: 68K likes, 8.5K reposts.
Anthropic reportedly rented Colossus 1, a 220K+ GPU SpaceX/xAI cluster, to boost compute for Claude model scaling and future coding capabilities.
Modal raised $355M at a $4.65B valuation to keep building its production cloud for inference, agent runtimes, reinforcement learning, and sandboxed environments.
SenseTime open-sourced the full training codebase for SenseNova-U1, its 8B dense and A3B MoE unified multimodal series for text-to-image, image editing, interleaved generation, and text/vision understanding. X: 63 likes, 22 reposts.
EdgeRunnerAI released EDGE-OPD, an on-policy distillation method that uses guided rollouts, privileged context, and evidence masks to internalize rare or private behaviors from a model’s own generations without directly training on hidden data (post). X: 16 likes, 5 reposts.
US law enforcement warned that AI job-loss fears and data-center backlash could turn into “anti-tech violent extremism” (Gizmodo).
ByteDance reportedly offered special stock grants tied to its AI business unit to keep top AI employees from getting poached.
Mistral AI expanded its Harvey AI partnership to bring its models to more than 1,500 legal customers across 60+ countries.
Harvey reported that frontier legal agents still completed less than 10% of its Legal Agent Benchmark tasks end-to-end, with different models leading different practice areas and Opus 4.7 costing about $50.90 per task at roughly 22 minutes of latency (Gabe Pereyra, Ben Scharfstein, Vtrivedy10, James da Costa). X: 102 likes, 13 reposts on the main thread.
ValsAI released an independent evaluation showing Qwen 3.7 Max leading or tying the frontier on its internal coding, math, and agentic benchmarks while remaining open-weights. X: 89 likes, 14 reposts.
Google DeepMind solved or identified prior solutions for 9 open Erdős problems using autonomous LLM-Lean agents with formal verification before human review (source shortlink). X: 2,711 likes, 406 reposts.
Stanford HAI introduced Item Response Scaling Laws, a measurement-science approach that can reduce the compute needed to forecast large-model scaling by more than 99% while keeping or improving predictive accuracy (Obvious Ventures, scaling01).
Pony AI raised its 2026 robotaxi fleet goal to 3,500 vehicles after Q1 robotaxi revenue nearly quintupled (Pulse 2.0).
Spotify struck a Universal deal for controlled AI-generated covers and remixes, with chief Alex Norström saying Spotify wants to be the legal, controlled version of a trend already happening elsewhere (PC Gamer).
AI debt collectors are becoming a growth market, with companies deploying automated agents for persistent calls, texts, and emails on unpaid bills.
AI-assisted pro se lawsuits are flooding US court dockets with more numerous and complex filings from self-represented litigants (ABA Journal).
Scientists used an IBM quantum computer to train a pretrained language model, improving answers the base model missed.
The Verge argued that AI warfare is already here, with semi-autonomous systems and compressed kill chains crossing many of the red lines people associate with lethal autonomous weapons.
Sundar Pichai discussed Google’s AI-first restructuring, Search convergence, YouTube changes, and the open web’s future in an interview with The Verge.

TOP TREATS TO TRY

OLMoTrace gives researchers training-trace transparency for open language models, so they can inspect provenance instead of treating “open data” as the finish line.
Webwright gives web agents a terminal, local workspace, and reusable Playwright scripts, turning browser tasks into durable programs instead of one-off clicks (post). X: 2,257 likes.
agent-browser gives agents a Rust CLI for browser automation with compact accessibility-tree outputs and ref-based element selection.
FlashLib gives agents fast GPU-native classical ML operators like KMeans, KNN, PCA, HDBSCAN, TruncatedSVD, exact t-SNE, and MultinomialNB, with tolerance-driven routing and cost prediction so retrieval, clustering, compression, and verification can run inside online agent loops (GitHub, post). X: 88 likes, 28 reposts.
Claude Code’s security-guidance plugin scans file edits, end-of-turn diffs, commits, and pushes for injection, unsafe deserialization, unsafe DOM APIs, and custom org rules, then has Claude fix issues in-session (post) — free to try for Claude Code users. X: 7,888 likes, 688 reposts.
noahlofq demoed Lightcone, a fast computer-use agent for boring browser and desktop tasks, with a public app at lightcone.ai. X: 291 likes, 8 reposts.
LiveAvatar released plugins for LiveKit, Pipecat, Agora, and VisionAgents that turn conversational AI engines into real-time virtual avatars with lip-sync, expressions, and full-body animation.
RefusalBench lets you compare frontier models on biological-research refusal behavior, with sortable results, confidence intervals, safety tiers, and a paper arguing raw refusal rates misrank models (post). X: 10 likes, 3 reposts.
Pointer hit 83.6% on OSWorld-Verified with Claude Opus 4.7, beating the human baseline and showing how a feasibility gate, planner, and executor can improve computer-use agents (post). X: 137 likes, 32 reposts.
Shard compresses Llama-3.1-8B’s KV cache by about 10x at 8K context and 11.2x at 32K without measurable hits on NIAH or LongBench (post). X: 1,244 likes, 105 reposts.
CUA-Gym provides 32,122 verifiable training tuples across 110 environments for computer-use agent reinforcement learning (paper, dataset, GitHub, hub, post). X: 208 likes, 41 reposts.
AVTR-1 gives developers an open-weights real-time avatar model that listens while you talk, reacts with facial expressions, and can be self-hosted with model weights, inference code, and a streaming demo (GitHub, model weights, technical page)—free for personal, research, and commercial use under $10M ARR.
Willow Scribe turns your rough spoken thought into a finished message in your own voice, then rewrites highlighted text inside apps like email, Slack, iMessage, and Google Doc.
Parrot Speech-to-text API transcribes messy real-world voice-agent calls with low latency, cleaner Hindi-English code-switching, and Hindi-aware normalization for workflows where one bad word can break the next action — free options.
Brew turns a plain-English campaign idea into on-brand emails, lifecycle flows, production-ready HTML, audience logic, and automations that work across Gmail, Outlook, and Apple Mail — free to get started.
Simi by Lamina Labs turns a prompt or document into a whiteboard-style explainer video in seconds, including course lessons, student walkthroughs, customer training, and startup explainers (post). X: 1,513 likes, 71 reposts — free to try.
PrismML’s Bonsai Image 4B lets you generate high-quality images locally on phones or laptops with 1-bit (0.93GB) and ternary (1.21GB) diffusion models that are 6-8x smaller and up to 5.6x faster, with an iOS app, WebGPU demo, HF collection, GitHub, whitepaper, and post. X: 1,020 likes, 157 reposts
FloYo’s VOID Video Inpainting + SAM3 Text Masking workflow lets you remove an object from video by typing what to erase, having SAM3 create the mask, and using VOID to fill the missing frames coherently.
Moda MCP connects Claude, Codex, Cursor, and other agents to a tasteful, editable design canvas for slides, posters, ads, and social content (Anvisha, trq212).
MOSS-TTS-v1.5 gives you a ModelScope text-to-speech model from the OpenMOSS team for generating speech from text (post).
MiniCPM5-1B gives you a compact OpenBMB model release on ModelScope for lightweight local or hosted inference experiments (post).
AIventure is an open-source GitHub repo for building AI-powered adventure game prototypes (Logan Kilpatrick, Antigravity, Rihard Jarc, xenova) — free to try.
Google AI, Google Gemma, and Taiming Lu shared Gemini/Gemma workflow notes and prompting examples for multimodal video analysis, frame-by-frame review, and multi-turn agentic editing.

RESEARCH, MODELS, AND AGENTS

PiD replaces traditional VAE decoders with a high-resolution pixel-space diffusion decoder for sharper text, textures, and fine detail across FLUX.1/2, SD3, Z-Image, DINOv2, SigLIP, and RAE latents (post). X: 241 likes, 46 reposts.
HumanEgo used just 30 minutes of human egocentric video to teach robots zero-shot behavior through Interaction-Centric Tokens, reaching 90%+ success across embodiments without robot data (context). X: 202 likes, 25 reposts.
Josef Chen launched Epicure, a multilingual food-ingredient embedding model trained on 4.1M recipes across seven languages, compressing 1,790 ingredients into ~2MB embeddings that reveal emergent geometry in ingredient space. X: 2,183 likes, 189 reposts.
MiniMax teased M3’s GQA-based dynamic block-sparse attention, showing 9.7x faster prefill and 15.6x faster decode at 1M tokens versus M2. X: 2,136 likes, 202 reposts.
Elie Bakouch compared MiniMax’s MSA architecture with DeepSeek sparse attention, highlighting how MiniMax selects real KV blocks through a two-pass structure. X: 330 likes, 21 reposts.
@kimmonismus broke down MiniMax M3’s Index Branch plus Sparse Branch design, framing it as a major long-context speedup. X: 622 likes, 33 reposts.
Ahmad Osman highlighted MiniMax M3’s sparse-attention jump after the technical teaser. X: 243 likes, 12 reposts.
SMART turns frozen single-vector embedding models into stronger multi-vector retrievers using hidden states, lightweight adapters, or LoRA finetuning (post). X: 70 likes, 18 reposts.
AMUSE combines Muon orthogonalization with stable gradient evaluation for anytime training, showing gains across Llama pretraining and vision tasks (post). X: 261 likes, 34 reposts.
PEIRA introduced a non-contrastive self-supervised learning method that uses inter-view regressor alignment to learn predictive encoders without collapse (post). X: 118 likes, 19 reposts.
zxlzr released InnoEval, an ICML 2026 method for evaluating research ideas with knowledge-grounded, multi-perspective review boards, plus SciAtlas, a scientific knowledge graph with 43M+ papers, 157M entities, and 3B triples packaged as agent skills for literature review, idea evaluation, idea generation, trend synthesis, and researcher profiles (GitHub). X: 40 likes, 13 reposts.
SkillOpt trains reusable natural-language skills for frozen agents, letting them self-improve executive strategies without parameter updates (HF paper page, arXiv, post, Omar Sar).
LEO teaches goal-conditioned agents to learn policies and values for all goals at once, outperforming Hindsight Experience Replay on Craftax (GitHub, paper, post).
Nano World Models open-sourced a compact repo for world-model research built around diffusion-forcing, long-horizon rollouts, planning, and pretrained checkpoints.
EOPD proposed entropy-aware on-policy distillation for language models and released code for the method (GitHub, post, Papers with Code, Niels Rogge).
KPop introduced adaptive masking regions to reduce the training-inference mismatch in reinforcement learning (post).
mwxely464 released ParaVT, an agentic video reinforcement-learning method that tames the “Tool Prior Paradox” by having one agent choose temporal video crops while weight-sharing sub-agents process them in parallel before a final answer, achieving SOTA on long-video benchmarks with an open ParaVT-8B model (project page, GitHub, paper page). X: 13 likes, 9 reposts.
Renaissance Philanthropy launched Formal Frontier, an XTX Markets-funded program supporting responsible, scalable, open-source AI-driven autoformalization of mathematics in the Lean theorem prover’s Mathlib library. X: 32 likes, 6 reposts.
EvalAwareBench decomposed evaluation awareness into environmental trigger factors and model-side recognition/behavior change, with code and a dataset showing models rarely recognize non-safety evals but become more sensitive as cues stack (ChanglingXavier). X: 47 likes, 13 reposts.
CHI-Bench released a healthcare-agent benchmark with 101 long-horizon care-management, prior-authorization, and utilization-management tasks in high-fidelity clinical workflows (post).
MathCode gives you a terminal coding assistant with a built-in math formalization engine that turns plain-language math problems into Lean 4 theorems and attempts formal proofs (post).
PyTorch highlighted TLX Block Attention, a Blackwell-focused block-sparse attention kernel that delivers ~1.85x forward, ~2.5x backward, ~3.5x fused rotary-backward speedups, and a +30.6% MFU gain on production self-attention layers (Meta kernel library, post).
vLLM merged a Rust frontend as an opt-in alternative to the Python API server, enabled with VLLM_USE_RUST_FRONTEND=1, to cut CPU overhead on preprocess-heavy workloads (GitHub PR). X: 456 likes, 51 reposts.
DeepSWE launched a contamination-resistant long-horizon software-engineering benchmark with 113 original tasks across 91 repos and 5 languages, where GPT-5.5 scored 70% and models spread out far more than on saturated public leaderboards (Serena Ge, source shortlink, reaction, reaction shortlink, CodexReleases, Chubby, Peter Steinberger).
Awesome-Inference-Time-Trustworthiness organized inference-time methods for trustworthy LLMs across external controls, internal manipulations, and system-level orchestration, with a companion project page and Yuyang Bai thread. X: 142 likes, 31 reposts.
From Simulation to Enaction argued that post-trained language models can recognize and react to their own generations, showing on-policy outputs have 3-4x lower entropy than off-policy continuations and route implicit self-recognition differently from explicit verbal recognition (Asvin G., Clare Birch).
OdysSim and Ditto-8B trained human-behavior simulation with verbal feedback as a first-class RL signal, improving role-play, learner/user simulation, social skills, and theory-of-mind tasks by 36% over base and beating GPT-5.4 on 6 of 10 SOUL tasks (GitHub, model, post, Xuhui Zhou).
LLMs as Noisy Channels reframed model capacity and scaling laws through Shannon-style signal-to-noise limits, arguing that more parameters or data can amplify noise and cause U-shaped degradation without enough preserved learning signal (post, Arena).
LeJEPA proved it linearly recovers true world latents up to rotation only when the latent distribution is Gaussian, with Lean 4-verified theorems and empirical recovery across high-dimensional and pixel-control settings (David Klindt, Dean Ball, Interconnects, Nathan Lambert, Seunghyun Seo, Gabri Berton).
From Entropy to Epiplexity proposed epiplexity as a bounded-intelligence information measure for systems that are limited by computation, not just Shannon uncertainty (Andrew Gordon Wilson, Hedgie Markets, Thomas Wolf).
From Model Scaling to System Scaling argued that agentic AI progress increasingly depends on scaling the harness around models, not just scaling the model itself.
Language Models Need Sleep explored whether LLMs benefit from sleep-like consolidation phases that reorganize learned knowledge (alphaXiv discussion).
The Age of Curiosity Meets the Age of AI benchmarked child safety in large language models, focusing on how systems answer curiosity-driven questions from younger users.

VIDEO, ROBOTS, AND 3D

Runway shared Project Luxo and “The Rogue,” a 10-minute AI film made by one person in under a month, as early evidence that AI video is crossing the uncanny valley (post, Cristóbal Valenzuela). X: 234 likes, 36 reposts.
Researchers demonstrated AthenaZero, a bimanual robot that learns to juggle barehanded with multi-fingered hands and onboard vision, mastering patterns like cascade in under 10 minutes of real hardware training by composing reusable skills.
Vision-Language Binding showed how FLUX.2 routes reference-image information through text tokens during in-context image generation, with code and a paper released (GitHub, arXiv, post). X: 172 likes, 31 reposts.
Soap2Soap open-sourced a multi-agent system for long cinematic video remaking by breaking scenes into shots, rewriting scripts, and regenerating footage with consistent characters and style (GitHub, post). X: 38 likes, 9 reposts.
TriSplat reconstructs simulation-ready 3D scenes from sparse images or casual video in 0.57 seconds, producing oriented-triangle meshes that can go into Unity or Isaac Sim (post). X: 140 likes, 22 reposts.
Visual Jenga tests whether models understand object dependencies by removing objects one by one from an image while keeping the rest stable (W&B, GitHub, reproduction, post). X: 251 likes, 15 reposts.
MARCO released a CVPR 2026 Oral model for semantic correspondence that generalizes from sparse supervision to unseen keypoints and new categories.
GenRecon bridges generative priors for high-fidelity PBR mesh reconstruction of 3D indoor scenes from sparse images or casual videos (reaction).

SAFETY, POLICY, AND SOCIETY

Simon Willison pulled out key passages from Pope Leo XIV’s AI encyclical, highlighting its concerns about interpretability, cultural bias, environmental impact, compassion in automated decisions, and data as a public good (Axios).
@kimmonismus argued Anthropic’s collaboration with Pope Leo XIV is a major PR win because it positions Claude as the morally legitimate AI option for 1.4B Catholics. X: 1,972 likes, 182 reposts.
RefusalBench found that raw refusal rates can misrank frontier models on biological-research prompts, especially when models over-refuse benign or borderline questions (paper, follow-up).
AI warfare moved from abstract debate to operational reality as the Pentagon, Anthropic, and military AI systems clash over red lines and contract terms.
AI in debt collection showed how automation is moving into one of the most disliked parts of consumer finance.
AI-assisted lawsuits highlighted a legal-system bottleneck: easier filing can mean more court bandwidth spent on self-represented cases.
US law enforcement is tracking possible “anti-tech extremism” as AI data centers and job-loss fears become local political flashpoints.
Anthropic explained how it contains Claude across claude.ai, Claude Code, and Cowork using ephemeral containers, OS-level sandboxes, local VMs, model steering, and audited tools to limit blast radius as agents gain more autonomy (post).
Chris Olah argued in remarks on Pope Leo XIV’s AI encyclical that AI development needs external oversight and dialogue with independent critics because commercial, research, geopolitical, and personal incentives can overpower even sincere intentions.
Neetu Arnold argued that allowing AI on take-home math and statistics exams may now be the only defensible policy because detection is impossible, AI tutors already exist, and bans create unequal access, while proctored in-person exams remain the high-stakes standard (UC Student Success).

INTELLIGENT INSIGHTS

Dan Shipper argued that the AI paradox means more automation creates more work, more humans, and more oversight, with much of that work moving inside Codex or Claude Code.
Prof. Michael I. Jordan argued that intelligence is collective and social, emerging from economic systems rather than disembodied AGI, and that AI research should prioritize actionable explanations, incentives, and hybrid human-machine systems (homepage).
Demis Hassabis discussed DeepMind’s Co-Scientist for accelerating scientific discovery and disease research, plus progress toward the “Einstein Test” for measuring AI scientific creativity.
Tuhin Chakrabarty argued AI writes poorly because its training mix absorbed too much fanfic-style language, creating a slop loop that now pollutes grants, academic writing, and literary culture (Sewon Min).
John Carmack praised SemiAnalysis for systems-level datacenter and semiconductor analysis, pointing to EV-commoditized 800VDC designs and 10kV SiC MOSFETs that could let AI data centers connect directly to medium-voltage AC. X: 1,471 likes, 54 reposts.
Francois Chauba argued that current LLM stacks are trapped on the human training manifold, so test-time compute will not discover genuinely novel algorithms without AlphaZero-style first-principles search over the broader hypothesis space. X: 22 likes, 3 reposts.
Epoch AI estimated token demand is growing roughly 10x per year while global inference supply grows about 3.4x, making a compute crunch likely near or already here, especially for long-context agent workloads (post). X: 176 likes, 26 reposts.
James da Costa and Angela Strange argued compliance is AI’s biggest boring enterprise opportunity: a $40B+ labor market where agents can turn regulation into code, replace legacy GRC systems, and automate end-to-end workflows like SAR narratives (a16z post, thread). X: 221 likes, 35 reposts.
Sushant Sachdeva shared that GPT-5.5 Pro one-shot produced a tighter electrical-flow localization proof than the previous log²n bound, while Sebastien Bubeck said the right harness let Mythos and public GPT-5.5 reproduce an internal model’s one-shot unit-distance result. X: 328 likes, 42 reposts.
Joseph Jacks argued AlphaFold’s representational limits run deeper than quantum van der Waals forces because textbook biology includes the tubulin code and microtubule dynamics AlphaFold cannot directly represent.
DAIR.AI Academy posted an event on autonomous long-running coding agents, covering workflows like /goal in Claude Code, Hermes Agent, Codex, and Cursor (Omar Sar, Anvisha).
Tasklet CEO Andrew Lee discussed why three kinds of software survive and how horizontal agent platforms are competing around file systems, orchestration, and long-running workflows, while Claude Code engineers walked through workflows for orchestrating agents instead of babysitting them.
Aleksa Gordić published “Inside the Transformer: The Life of a Token,” a deep technical walkthrough of YaRN, hybrid attention, soft capping, QK normalization, FLOPs per token, and cluster sizing (post).
Daniel Miessler argued Policy + SOPs + AI are the core ingredients for secure, scalable systems because agents should execute inside explicit boundaries instead of improvising freely (post).
Vintage Data argued the economy is splitting into “The AI Decoupling,” where inference innovations compress old software margins while high-margin models absorb more application value (post).
Sundar Pichai framed Google’s AI era as a full company reorganization around Gemini, Search, YouTube, agents, and an open web that still needs traffic to flow.
The Verge argued that the debate over lethal autonomous weapons is catching up to systems already used in targeting, surveillance, and battlefield decision loops.

ADDITIONAL SOURCE NOTES

alphaXiv, Zhuofeng Li, Anjney Midha, Derek Thompson, and Katie Roof added supporting discussion around the day’s research, agent, startup, and AI-economy threads; none introduced a separate paper, product page, or funding source beyond the linked items above.
hardmaru, Sakana AI Labs, Greg Kamradt, and ARC Prize added related commentary around AI reasoning, theorem-proving, and puzzle-solving progress that sits alongside the DeepMind Erdős and math-agent items above.

Previous Around the Horn Digests

Catch up on everything you missed:

Monday, May 18, 2026: Microsoft open-sourced ECHO, Odyssey launched real-time AI simulators, and OpenAI added bank connections to ChatGPT.
Wednesday-Thursday, May 13-14, 2026: Nvidia H200 sales cleared but stalled, Americans opposed AI data centers, and Meta planned layoffs.
Tuesday, May 12, 2026: Anthropic refused China model access, Isomorphic raised $2.1B, and Google pushed Gemini deeper into Android.
Monday, May 11, 2026: Cerebras upsized its $4.8B IPO, Cowboy Space raised $275M for orbital data centers, and Google confirmed the first criminal AI-found zero-day.
Weekend, May 9-10, 2026: The Trump administration drafted an AI security order, Apple and Intel reached a preliminary chip-making agreement, French prosecutors escalated their Musk and X probe, and Cerebras’ IPO heated up.

That's a Wrap

That’s 200+ source links from one Tuesday alone. If you made it to the bottom, you now know more about sparse attention, avatar latency, local image models, and AI talent travel restrictions than most people currently arguing about all four online. Use this power only for Slack dominance.

For the daily version, bite-sized and actually readable, 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.