Daily AI Briefing
The 38 stories that mattered in AI, curated and summarized from dozens of sources by AIBriefs.
Qwen3.7-Plus supports text, video, and image inputs at $0.40/$1.60 per million tokens — 60% cheaper than text-only Qwen3.7-Max. The proprietary model unifies vision and language for agent tasks.
OpenAI confidentially filed its S-1 with the SEC on June 8, following Anthropic's similar filing on June 1. Both AI labs are preparing for public offerings, with Anthropic selecting Morgan Stanley and Goldman Sachs to lead its IPO.
DeepSeek permanently applied its 75% discount on V4 Pro API, pricing at $0.435/1M input and $0.87/1M output tokens. Xiaomi's MiMo V2.5 Pro matched these prices, slashing its rates by up to 99%.
Suno closed a $400 million Series D round, valuing the company at $5.4 billion. The round more than doubles its valuation from a $250 million raise seven months ago, and Suno teased its first licensed model.
Hugging Face's Reachy Mini robot now runs fully locally for near real-time conversations without cloud dependency. Version 1.8.0 adds MCP (Model Context Protocol) support for integrating external tools.
OpenAI launched a new memory system called Dreaming that better retains user preferences across conversations, rolling out to Plus and Pro users in the US. The system is described as a significantly more capable and compute-efficient architecture built on top of dreaming.
The bill, SB 315, requires frontier AI labs like OpenAI and Anthropic to have safety practices verified by third-party auditors. Illinois Governor JB Pritzker says he will sign it, making it the nation's leading AI safety law.
Professors at UC Berkeley report a surge in failing grades in computer science courses, attributing it to increased reliance on AI and declining math skills among students. A New York Times article examines how California's university system's embrace of AI is leading to unintended consequences.
Mistral AI has signed partnerships with Airbus and BMW to expand into physical AI, as reported by Bloomberg. The partnerships mark Mistral's move beyond language models into physical world applications.
Roche will acquire PathAI to combine its diagnostics strength with PathAI's AI-powered pathology platform. PathAI's image management system and AI analysis capabilities aim to enhance laboratory efficiency and accelerate companion diagnostics.
Robinhood allows users to create a separate account for an AI agent, pre-load it with funds, and let the agent trade stocks automatically. The company pitches it as a way to experiment with AI-driven trading while maintaining control over risk.
Ali Behrouz, a Cornell grad student and Google researcher, discusses his Nested Learning paper, which aims to enable models to adapt to new context while preserving core knowledge. Jeff Dean praised it as a potential paradigm shift.
In a randomized controlled trial, LGBTQ+ youth at risk of self-harm who used Purrble, a socially assistive tactile robot, showed improvements in emotion regulation. The robot was integrated into safety planning.
ViCuR uses visual cues as a recoverable privilege signal to improve multimodal on-policy distillation. The method trains a student on its own policy trajectories under teacher supervision, enhancing reasoning performance.
The book, a follow-up to the New York Times bestseller Co-Intelligence, explores living and working with AIs that are sometimes smarter than humans. Pre-orders are available at co-existence.ai.
Paper introduces CHASE, a framework using reinforcement learning for adversarial red-blue teaming to generate prompt-rewriting attacks like persona modulation. Experiments show it improves safety alignment against such bypass attacks on frontier models.
A LessWrong analysis by Linch Zhang found that paragraphs of Pope Leo XIV's encyclical 'Magnifica Humanitas' were 40-100% likely AI-written. The encyclical itself uses AI as a lens to discuss concentrated power and eroding democracy, according to TechCrunch.
OpenAI researchers Alexander Wei, Hongxun Wu, and Lijie Chen discuss how their model found a counterexample to the Erdős unit distance conjecture, a problem unsolved for 80 years. Mathematician Timothy Gowers called it a 'major open problem' solved by AI.
At Snowflake Summit 26 in San Francisco, the focus shifted from LLM wonder to the agentic enterprise. The conference emphasized deploying autonomous AI agents for business workflows.
Created by Artificial Analysis and IBM Research, ITBench-AA tests SRE tasks. Frontier models like GPT-4o and Claude 4.5 scored under 50%, highlighting gaps in agentic enterprise AI.
Cloudflare's Durable Objects offer addressable, persistent, stateful compute with 15ms London latency, making them suitable for AI agents. The Agents SDK builds on this foundation.
Automated reasoning strategy design reduced token usage by 69.5% while maintaining performance. The method eliminates handcrafted test-time scaling strategies, relying on program search to discover optimal reasoning patterns.
Attackers are using AI to automate EDR evasion testing. Python scripts test malware against Sophos, CrowdStrike, and Windows Defender agents.
BitCPM-CANN achieves 1.58-bit ternary quantization with 6x less VRAM than BF16 while retaining most capability. The model family includes sizes from 2B to 8B and is fully open-source on HuggingFace.
AI-written submissions won three of five categories in the Commonwealth Short Story Prize, including one selection in Granta magazine. The discovery raises concerns about AI detection in literary competitions.
Sesame's iOS app is now available in Preview, featuring four personal AI voice agents. The agents offer state-of-the-art real-time voice interaction, web search, reminders, and memory. The startup was founded by the co-founders of Oculus.
LangChain warns that AI model providers lock users in at the harness layer, restricting flexibility. The blog advocates for neutral, open-source frameworks to maintain model neutrality and prevent vendor dependency.
Hermes Desktop app from Nous Research is now available on macOS, Windows, and Linux, offering a GUI for the Hermes Agent. It has 140K GitHub stars in 3 months and is the most used agent on OpenRouter.
The symposium on June 8, 2026, featured sessions on academic research, industry deployment, and a keynote panel. Speakers included Stanford leaders and industry experts discussing responsible AI for psychiatric care.
Microsoft and Mayo Clinic announced a partnership to develop a frontier AI model for healthcare, owned by Mayo Clinic. The model will combine Mayo's clinical expertise with Microsoft's technologies.