Live Feed
Max Planck Paper Proposes Multi-Stream LLMs to Break the Single-Thread Agent Bottleneck
Researchers at MPI Tubingen show that replacing sequential message formats with parallel computation streams lets language models think while reading, act while thinking, and monitor themselves while generating — fixing a structural limitation present in every deployed agent today.
Waymo Pauses Atlanta Service After Robotaxis Drive Into Floods — Two Cities Down, Same Failure Mode
Waymo has halted operations in Atlanta after multiple robotaxis drove into flooded streets. The pause follows a recall issued last week after a vehicle was swept into a creek in a separate market. Two commercial deployments suspended in a week over the same adverse-weather routing failure.
Cursor Composer 2.5 Takes #3 on AA Coding Agent Index at $0.07 Per Task — 60x Cheaper Than Codex
Artificial Analysis places Composer 2.5 third on its Coding Agent Index with a score of 62, behind only Claude Opus 4.7 (max) at 66 and GPT-5.5 (xhigh) in Codex at 65. Standard variant costs $0.07 per task versus $4.82 for Codex — making it the only agent above 60 on the cost-quality Pareto frontier.
Anthropic on Track for First Profitable Quarter as Q2 Revenue Heads to $10.9B
The Wall Street Journal reports Anthropic's Q2 revenue will more than double to $10.9 billion, the company's first profitable quarter since founding. The implied annual run-rate has crossed $43B — up sharply from the $30B figure disclosed in March.
Google Puts Ads Inside AI Mode: Conversational Discovery and Highlighted Answers Built on Gemini
Google introduced two new AI Mode ad formats at Google Marketing Live on May 20: Conversational Discovery ads that generate custom creative from nuanced prompts, and Highlighted Answers that insert ads directly into AI-generated recommendation lists. Both run on Gemini and carry sponsored labels alongside AI explainers.
TeamPCP Steals 3,800 GitHub Internal Repos via Poisoned VS Code Extension
The TeamPCP supply chain group compromised one GitHub employee's device through a malicious VS Code extension on May 18, 2026, exfiltrating roughly 3,800 internal repositories. GitHub confirmed no customer data was impacted and has rotated critical credentials. The data is being offered on Breached forum for $50,000.
OpenAI's Reasoning Model Disproves Erdős' 1946 Unit Distance Conjecture — First AI to Crack an Open Problem in a Major Math Field
A general-purpose OpenAI reasoning model disproved the Erdős planar unit distance conjecture, open since 1946, by finding that square grids are not optimal for unit distance problems. Fields Medalist Timothy Gowers called the proof ready for Annals of Mathematics. No specialized math model was used.
Cohere Releases Command A+: Open Weights, 86% Non-Hallucination, First Model in 14 Months
Cohere's Command A+ open-weights model lands at Artificial Analysis Intelligence Index 37, leading its tier on hallucination resistance while trailing on hard science and agentic coding. It is Cohere's first major model release since Command A.
Intuit Cuts 3,200 Jobs After a Record $8.6B Quarter — The AI Restructuring That Needs No Excuse
Intuit is eliminating 17% of its global workforce and closing its Los Angeles office, with CEO Sasan Goodarzi framing the move as an investment in AI, not a response to financial pressure. The quarter before the cuts was a record.
OpenAI Confidentially Files for IPO — Largest AI Startup Heads for Public Markets
OpenAI has submitted a confidential S-1 to the SEC, with a public listing potentially weeks away. Prediction markets now place OpenAI ahead of Anthropic in the race to go public first.
Anthropic Is Paying SpaceX $1.25B a Month — and Now Expanding to Colossus 2 With GB200
New financial terms reveal Anthropic's SpaceX deal runs to $15B a year through May 2029. The compute relationship is expanding: Colossus 2, previously reserved for Grok, is next in line, this time with GB200 chips.
SenseTime Open-Sources SenseNova U1: 38B Native Multimodal at 3B Active Parameters
SenseTime has open-sourced SenseNova U1, a native multimodal model that eliminates the visual encoder and VAE used in every major multimodal system. The MoE variant runs on 3B active parameters with 38B total, using a Mixture-of-Transformers backbone that fuses generation and understanding into one unified architecture.
Google Launches Gemini Omni at I/O 2026: World Model Targets Any Input, Any Output
DeepMind CEO Demis Hassabis unveiled Gemini Omni at I/O 2026 as Google's first model built to generate any output modality from any input. Gemini Omni Flash launches today in the Gemini app, Google Flow, and YouTube Shorts, starting with video and expanding to image and audio.
Anthropic Consulted 15+ Religious and Cross-Cultural Groups to Teach Claude Moral Stability Under Pressure
Anthropic held formal dialogues with philosophers, theologians, clergy, and ethicists from more than 15 religious and cultural traditions, treating AI character formation as an alignment problem rather than a product feature. A resulting self-reminder tool lets Claude pause mid-task to recall its own commitments before high-stakes actions, and reduced misaligned behaviour in internal tests.
Google Adds GEO Manipulation to Spam Policy — AI Overviews and AI Mode Now Formally Protected
Google updated its Search spam policy to explicitly cover attempts to game AI Overviews and AI Mode, naming a new category of abuse the industry calls GEO (Generative Engine Optimization). Violations carry the same penalties as traditional search spam: rank demotion or removal, detected by automated systems and human review.