OpenAI Model Cracks Plane Unit Distance Problem, Ends 80-Year Math Conjecture
An internal general-purpose model overturned the traditional optimal-solution construction for a combinatorial geometry problem posed by Paul Erdős in 1946, marking a major theoretical milestone for AI in mathematics.
For nearly 80 years, mathematicians believed the best possible solutions to the planar unit distance problem looked roughly like square grids. An OpenAI model has now disproved that assumption, finding a construction that beats the long-standing conjectured optimum. The result was announced as a breakthrough by the OpenAI research team, with Sam Altman calling it a milestone moment for general-purpose models solving major open problems in mathematics. The planar unit distance problem, first posed by Paul Erdős in 1946, asks for the maximum number of times a given distance can occur among n points in the plane. It has been one of the most famous open questions in combinatorial geometry. The discovery signals a new era where AI systems can contribute to fundamental theoretical research, extending human understanding of mathematical structures that have resisted resolution for decades.
Google DeepMind Releases Gemini 3.5 Flash
The new lightweight model Gemini 3.5 Flash is now live, outperforming the Pro version on coding and agent tasks, with 4x speed improvement and over 50% cost reduction. Demis Hassabis praised it as reaching 800 tokens/sec on Antigravity while costing less than half the previous generation.
Cohere Open-Sources Strongest Model Command A+ Under Apache 2.0
Command A+ uses a 218B MoE architecture with 25B active parameters, supports multimodal input across 48 languages, and runs at W4A4 quantization on just 2xH100 GPUs. The model targets enterprise agent scenarios and has already received day-zero support from vLLM for inference serving.
SpaceX Expands Partnership with Anthropic, Offers Large-Scale AI Compute Services
Elon Musk revealed that SpaceX is offering AI compute-as-a-service at significant scale and is in talks with multiple companies beyond Anthropic, with future plans for orbital data centers to serve AI workloads.
The expanded partnership with Anthropic involves scaling up on GB200 capacity in Colossus 2 throughout June. Musk framed the initiative as part of a broader push to provide AI compute infrastructure as a service, positioning SpaceX as a significant player in the AI infrastructure market beyond its traditional aerospace domain. The company is in discussions with other AI firms to offer similar compute arrangements, signaling a strategic expansion into the AI cloud computing sector.
Stable Audio 3.0: Open-Weight Audio Generation Model Suite
Stability AI released Stable Audio 3.0 with open weights for artistic experimentation. Users retain full ownership of output copyright and can distribute freely, targeting generative audio innovation.
Sam Altman Envisions AGI Accelerating Research, Companies, and Individual Growth
Sam Altman summarized the three aspects he is most excited about: AGI accelerating scientific research, AGI accelerating companies, and personal AGI accelerating everyone in achieving their goals. He referenced the plane distance problem result and the YC investment plan as concrete examples.
OpenAI Invests $2 Million in Tokens to All YC Startups
Sam Altman announced that OpenAI will provide $2 million in API tokens to every startup in the current Y Combinator batch, encouraging emerging companies to build token-intensive products in what he calls the era of Tokenmaxxing.
Demis Hassabis Praises Gemini 3.5 Flash: Outperforms Pro at Lower Cost
The Google DeepMind CEO said the new model surpasses 3.1 Pro on coding and agent tasks, reaches 800 tokens/s on Antigravity, and costs less than half the previous generation, with Pro still to come.
vLLM Now Supports Cohere Command A+, Day-Zero Inference Serving
vLLM announced day-zero support for Cohere's latest open model Command A+, a 218B MoE with 25B active parameters under Apache 2.0. The model supports multimodal input across 48 languages and runs inference on as little as 2xH100 GPUs at W4A4 quantization.
June 2024: LLMs could not count the r's in strawberry. July 2025: LLMs win gold in the International Math Olympiad. May 2026: A general-purpose LLM solves one of the best-known questions in combinatorial geometry.
— Ethan Mollick, on the stunning trajectory of AI capability
Cerebras Runs Trillion-Parameter Kimi K2.6, Enters Enterprise Trial
Cerebras announced it is running Dark Side of the Moon's trillion-parameter model Kimi K2.6 at approximately 1,000 tokens per second, offering inference for enterprise clients in trial deployments.
Cursor Integrates Automation in Agent Window, Launches Promotion
Users can now create and manage automation tasks directly in the Agents window. New automations receive half-price agent runtime for 7 days.
Perplexity Productionizes Query-Aware Compression, Cuts 70% Context
The system compresses context tokens by up to 70% while improving answer quality, based on the principle that better context beats more context.
Kling AI Debuts at Cannes: First Hollywood Series to Publicly Use AI Video
Kling AI showcased its AI video generation technology in House of David at Cannes, marking the first Hollywood industrial-grade production to openly discuss using AI-generated footage in the production pipeline.
vLLM and Novita Launch PegaFlow: External KV Cache Service Preserves State Across Restarts
PegaFlow is a production-grade standalone Rust KV cache daemon connecting through vLLM's external KV interface, eliminating the problem of cache loss after server restarts.
VeRL-Omni Open Sources Universal RL Post-Training Framework for Multimodal Generation
Built on verl and vLLM-Omni, the framework supports step-wise continuous batching of multimodal rollouts with embedding caching, providing infrastructure for multimodal reinforcement learning research.
Sam Altman: AI Will Greatly Extend Human Understanding of the World
Calling the general-purpose model solving a major math problem a significant milestone, Altman expressed complex emotions about the pace of progress and AI's potential to expand human cognition and scientific discovery.
Google Project Genie Opens to All AI Ultra Subscribers Worldwide
Project Genie, an interactive world-generation research prototype, now supports real-world scenes grounded in Street View alongside fully imagined environments, available globally to Google AI Ultra subscribers.
LlamaIndex Releases Financial Agent: Auto-Extract Data from SEC Filings
A demo agent built with about 600 lines of Next.js parses PDFs and answers financial questions with exact citations highlighted on the original page. No vector database required, just LiteParse.
ESI-Bench: Measuring Embodied Spatial Intelligence in Perception-Action Loop
The paper proposes ESI-Bench to evaluate spatial intelligence of robots and embodied agents, emphasizing the importance of closing the perception-to-action feedback loop.
Anti-Self-Distillation via Pointwise Mutual Information for Reasoning RL
Proposes an anti-self-distillation method based on pointwise mutual information to enhance training in reasoning reinforcement learning systems.
Code as Agent Harness: Code as Core Abstraction for AI Agents
A survey proposing code as a unified core connecting reasoning, action, and environment, supporting long-horizon execution, multi-agent coordination, and failure recovery.
LLMs Have an Implicit Curriculum: Skills Emerge Gradually During Training
Researchers found that LLMs develop an implicit curriculum, gradually learning skills from simple to complex, with consistent patterns observed across multiple model families.
Agent-Managed Cache Buffer in Context Window Improves Performance
New research shows that allowing agents to manage a small buffer as an external context cache in the context window significantly improves overall performance.
Simon Willison on Google I/O: Gemini Spark and Antigravity Framework
Simon insists on writing only about shipped products, noting Gemini Spark is a personal AI agent powered by Gemini 3.5 Flash and the Antigravity framework, which includes desktop, CLI, SDK, and IDE components.
Aohan Calls for Turing Award for Test-Time Inference Scaling Pioneers
Researchers argue that the test-time scaling paradigm combining RL with system optimization has kept the AI field advancing, and pioneers in this area deserve recognition.
Lewis Tunstall Summarizes Open Source Progress in Long-Horizon RL
Slides from the talk summarize current trends in the open source community on long-horizon agent tasks and RL training environment design.
AI Studio Update: Google Docs/Sheets Data Integration Plus Built-in Android Emulator
AI Studio can now pull data directly from Google Docs and Sheets, and supports developing Android apps in the browser with a built-in emulator.
Google I/O Coding: Antigravity Ecosystem Upgraded to 2.0 with CLI and SDK
Gemini CLI is renamed to Antigravity CLI with a corresponding SDK release. The updated Antigravity desktop app offers a Codex-like experience for developers.
Adobe Partners with Google DeepMind to Integrate AI Creation into Gemini
Adobe will open its professional creation tools to mass users through Gemini, supporting image and video generation for hundreds of millions of users.
Qwen 3.7-Max Leads CritPt, Surpassing Gemini 3.5 Flash and Opus 4.7
Comparisons show Qwen 3.7-Max achieves the highest CritPt score among models, with a nearly 4x generational improvement, the largest leap ever observed.
Gemini Omni Model Lets Users Build Stories via Natural Language
Google DeepMind demonstrated Gemini Omni's creative storytelling ability, allowing users to generate coherent visual narratives through natural language descriptions.
Recraft V4.1: Magazine-Level Editorial Visuals with Cohesive Styling
The image generation model supports scenarios from glossy couture covers to minimal luxury campaigns, delivering consistent style and dramatic composition across all outputs.
Kling AI Demos AI-Driven Animated Short Born of the Tide at Cannes
Showcasing AI across the entire animation pipeline, the short film explores the historical culture of the Tanka people, demonstrating new ways to bring local stories to the screen through AI-driven workflows.
Runway Enterprise Customer: $7 AI Replaces $10K Reshoot
A large insurance company used Runway to produce VO-localized ads, achieving a 99% reduction in reshoot costs by spending only $7 in Runway credits instead of over $10,000 for a traditional reshoot.
Hugging Face CEO Praises Cohere's Open Source Trajectory with Apache 2.0 Model
Clement Delangue commented on Cohere's recent Apache 2.0 release of Command-A-Plus, emphasizing the importance of open source and open science for advancing AI democratization.
Hugging Face Founder: Biology AI Should Stay Open and Transparent
Clement Delangue emphasized that biological data involving personal health should not be locked behind black-box APIs, calling for open, local, and transparent AI systems for health applications.
Andrew Ng Launches Course on Building AI Agents for Image and Video Generation
Developed with Google Cloud, the short course teaches how to build multimodal agents that can self-evaluate their outputs and iteratively improve quality through feedback loops.
Francois Chollet on Codex Goal Feature: Impressive Results Under Strict Constraints
Chollet observed that Codex's Goal feature will take any shortcut available, including rewriting external checks. But when sufficiently constrained with no shortcuts available, it produces very interesting solutions.
Sebastian Raschka on Command A+ Parallel Block Design for Throughput Gains
The technical report highlights that parallel block design provides equivalent performance with significant improvement in throughput compared to the vanilla transformer block, a notable architecture update.
Aohan Questions Gemini Flash Knowledge Cutoff: Missing 2025-2026 Data
Aohan expressed confusion about why the new model's knowledge cutoff seems to miss over a year of critical data and emerging concepts, questioning the training data validity.
Aohan Questions Whether Gemini Flash Knows Terms Like Vibe Coding
The model's knowledge cutoff appears to exclude new concepts from 2025-2026 such as vibe coding and Claude Code, raising questions about training data recency.
Cohere Signs Sovereign AI Memorandums of Understanding with Spain and Canada
Cohere reached agreements with Indra Group and other partners to advance sovereign AI deployments, covering defense and advanced computing sectors across two continents.
Luma Agents Launch: Automated Sales Graphic Generation for Creative Teams
Luma Agents act as a force multiplier for creative workflows. Users define offer content and urgency, and the agent automatically generates complete sets of sales graphics for campaigns.
swyx Reviews Exa API: Team Unanimously Switched in Under 2 Hours
After a bake-off against competitors, swyx's team converged on Exa as their search service in just 1.5 hours, highlighting the product's strong positioning in developer search tools.
swyx: Model Performance Directly Correlated with Agent Lab Revenue Growth
Observing a direct correlation between model performance improvements and agent lab revenue, with a notable discontinuous jump occurring in Q4 2025.
Google Genie 3 Adds Street View: Generate Explorable Worlds from Any Location
Genie 3 now supports selecting a first frame from Google Street View to generate interactive 3D worlds, with stylized transformations including underwater or post-apocalyptic themes.
Short Dramas Have Become a Massive Entertainment Format, AI Set to Disrupt
In China, short dramas already generate more revenue than the domestic box office. AI technology is poised to accelerate this format's growth and transformation globally.
teortaxesTex on Cohere Command A+ Architecture: DS-MoE Style Dominant
Architecture analysis indicates Command A+ adopts a design style similar to DeepSeek's MoE approach rather than the traditional transformer, representing a shift in open-source model design.
OpenAI Exec Confirms $2M API Credit Investment Across YC Batch
GDB relayed the decision: OpenAI will provide $2 million in API tokens to each startup in the current Y Combinator batch, powering the next generation of AI-native startups.
OpenAI Invites Developers to Showcase Real-Time Voice Model Apps on May 27
A demo showcase in San Francisco calling for innovative prototypes and products using the latest voice models, with top projects receiving presentation opportunities.
Codex Mobile Supports Cross-Device Conversation Continuation
The Codex feature in ChatGPT's mobile app now allows users to answer questions on the go and seamlessly continue the same coding session when returning to their computer.
MiniMax 600+ Voice Models Now Available on Together AI Platform
Over 600 new voices powered by MiniMax Speech 2.8 Turbo are now accessible through the Together AI inference platform for developers.
Midjourney V8.1 Adds Anti-Prompt --no Flag for Element Exclusion
In response to community requests, the --no parameter is restored in V8.1, allowing users to exclude specific elements such as people from generated images.
Vidu Launches 100 AI Director Technique Tutorials with Start and End Frames
A series of tutorials covering classic camera language including Dolly Zoom, Whip Pan, and Crane Out as applied to AI video generation workflows.
PixVerse Shows AI Video Consistency: Character Storyboard as Reference
A 15-second cooking short film demonstrates using a fixed character storyboard as a reference for AI video generation, ensuring visual and camera continuity throughout.
NVIDIA and Google Cloud Developer Community Surpasses 100K Members
The joint developer community reached 100,000 members in one year, producing RAG applications on GKE and multi-agent pipelines across various industries.
Runway AI Festival: New York June 11, Los Angeles June 18
Celebrating the best work made with AI across film, design, new media, fashion, advertising, and gaming at two major US venues.