Daily AI Briefing
The 120 stories that mattered in AI, curated and summarized from dozens of sources by AIBriefs.
The US government issued an export control directive citing national security concerns over a potential jailbreak of Fable 5. Anthropic abruptly disabled both Fable 5 and Mythos 5 for all customers, and talks with the administration continue.
SpaceX will acquire AI coding assistant Cursor for $60 billion in stock, days after its blockbuster IPO. The deal aims to boost SpaceX's AI division, which sees a $26 trillion addressable market.
Gemini Omni creates videos from any input (images, audio, video, text) and allows conversational editing with consistent characters. The first model, Gemini Omni Flash, is available in the Gemini app, Google Flow, and YouTube Shorts.
Google announced the biggest Search upgrade in 25 years, replacing the classic search box with an AI-powered experience driven by Gemini 3.5 Flash models. The update introduces AI Mode, AI Overviews with follow-up queries, and autonomous "information agents" that can monitor topics. Rolling out globally starting today, traditional results remain available via the "Web" tab.
Anthropic confidentially filed a draft S-1 with the SEC, preparing for an IPO. It is on track for its first profitable quarter with revenue exceeding $10.9B, picking Morgan Stanley and Goldman Sachs to lead the offering. The move positions Anthropic ahead of rival OpenAI in the race to go public.
Cosmos 3 uses a Mixture-of-Transformers architecture with a reasoning tower and a generation tower, available in Nano (16B) and Super (64B) sizes. It is the first fully open omnimodel with native action generation, trained on diverse data to produce physically grounded outputs for robotics and autonomous vehicles.
Gemma 4 12B runs on 16GB of VRAM or unified memory and nears performance of the larger 26B model. It's the first mid-sized Gemma with native audio inputs, released under Apache 2.0.
Kimi K2.7-Code achieves +21.8% on Kimi Code Bench v2 over K2.6 with 30% fewer reasoning tokens. It is a trillion-parameter MoE model available under a Modified MIT license, supporting long-horizon software engineering tasks.
The model family includes two on-device models: a 3B dense model and a 20B sparse model, plus three server models. NVIDIA Blackwell GPUs with Confidential Computing provide secure inference on Google Cloud. Third-party models like Gemini and Claude are now accessible via Apple's Foundation Models framework.
OpenAI's net losses increased nearly 8x in 2025, while total spending hit $34 billion, according to an exclusive financial report. The report, obtained by journalist Ed Zitron, details the staggering burn rate behind the company's operations.
OpenAI is investing $150 million to launch the OpenAI Partner Network, designed to accelerate enterprise AI adoption and deployment. The program will provide resources and support to global partners including systems integrators, consultancies, and technology providers.
Anthropic is in early discussions with investors to raise at least $30 billion in fresh financing, according to people familiar with the matter. The talks underscore the massive capital pouring into AI development.
Salesforce plans to acquire AI customer service startup Fin for $3.6 billion. The deal was reported by Bloomberg.
Gemini app now serves 900M users. New features include Gemini 3.5 Flash, Neural Expressive design, Daily Brief agent, and Gemini Spark proactive AI agent. Also launches Gemini Omni for video generation.
Hermes Agent has reached 140K GitHub stars in 3 months and launched a desktop app for macOS, Windows, and Linux. The app requires no terminal setup and includes features like Write Gate, SOUL.md agent profiles, and Nvidia Skills integration.
Kazakhstan and investment firm Firebird have signed a $10 billion agreement to develop AI infrastructure, with support from Nvidia. The deal aims to boost the country's computing capabilities and attract further tech investment.
At Google I/O 2026, Google unveiled an AI-powered overhaul of Search featuring an intelligent search box, AI Overviews with follow-up questions, and personalized information agents. The change demotes traditional ranked links in favor of interactive experiences, with AI Mode available starting Tuesday.
The UK government has unveiled a $1.47 billion plan to build a national AI supercomputer, with $530 million for hardware including $200 million for inference chips. British startups like Olix and Fractile will be prioritized for procurement, and the system is expected to be operational by 2030. The initiative is part of a broader push for sovereign AI to reduce reliance on foreign tech.
SpaceX went public in the largest IPO ever, with OpenAI and Anthropic also confidentially filing. The TechCrunch Equity podcast highlights the "MANGOS" acronym (Meta, Anthropic, Nvidia, Google, OpenAI, SpaceX) as the new FAANG.
Gemini 3.5 Flash outperforms Gemini 3.1 Pro across almost all benchmarks while running four times faster than other frontier models. Google also introduced Antigravity 2.0 desktop app, CLI, and SDK, along with managed agents and native Android vibe coding in AI Studio.
OpenAI completed a conversion to a Public Benefit Corporation in early 2025, with the nonprofit retaining an estimated 25% equity stake. The IPO thesis hinges on cheap tokens plus proprietary harnesses owning the work layer of AI, according to an analysis.
A solo attacker used Anthropic's Claude to breach Mexican government systems, exfiltrating 150 GB of data. The breach highlights how AI lowers barriers for cybercriminals.
Karpathy will work on pre-training under Nick Joseph, using Claude to accelerate research. The OpenAI co-founder and former Tesla AI lead returns to frontier R&D after founding Eureka Labs.
Google at I/O unveiled agentic Search features, including 'information agents' that proactively monitor topics and alert users even when inactive. The experience, powered by Gemini 3.5 Flash, will roll out first to AI Pro/Ultra subscribers this summer.
A proposed class-action lawsuit was filed in California, alleging Anthropic misled customers about usage limits on its $200/month Claude Max plan. The suit claims the plan did not provide the extra usage promised.
Mistral AI acquires Emmi AI, a Linz-based Physics AI company specializing in industrial simulation for aerospace, automotive, and semiconductors. Emmi's team of over 30 researchers will join Mistral's Science and Applied AI teams, and Linz will become a Mistral AI office.
Google unveiled Antigravity 2.0 at IO 2026 with a desktop app for orchestrating multiple agents, a CLI tool, and an SDK for custom workflows. A new $100 AI Ultra plan offers 5x higher limits, and the app gets native voice command support, powered by Gemini 3.5 Flash.
Google announced new Gemini features, including an 'intelligent search box' with generative AI and custom layouts. The company also previewed Android-powered smart glasses launching this fall. Google says 900 million people use Gemini and over 50 billion images have been generated.
Three experimental tools—Hypothesis Generation, Computational Discovery, and a third—aim to accelerate scientific discovery by automating hypothesis generation and computational experiments. The tools are built on Gemini models and available via Google Labs.
Karpathy, a prominent AI researcher who co-founded OpenAI and led Tesla's AI efforts, is joining Anthropic. The move marks a high-profile addition to the AI startup.
Opening arguments began in Anthropic's lawsuit against the Defense Department over being declared a supply chain risk. The AI startup argues the blacklisting harms its business and reputation.
Self-improving AI startup Recursive AI reaches a $4.65 billion valuation, as reported by Bloomberg. The startup focuses on AI systems that improve themselves.
Letter warns SpaceX's $75B IPO faces 'unpriced risks' from xAI's poor safety record. Guidelight AI Standards co-founder Page Hedley says xAI has worst safety practices among frontier labs.
Google's updated Gemini app, reaching 900M monthly users, introduces a 'Neural Expressive' redesign, Daily Brief personal digest, Gemini Spark agent, and the Gemini Omni video model. Daily Brief rolls out today to US subscribers.
Anthropic and the Gates Foundation are launching a $200 million, four-year initiative to support governments in building open AI tools for healthcare, education, and agriculture. The partnership comes amid reports that Anthropic is seeking a $30B investment. The initiative will provide funding, API credits, and technical support to developing countries.
Meta is investing $200 billion in a Louisiana data center, one of the largest corporate infrastructure projects ever. The facility will support AI and cloud computing, marking a major bet on expanding Meta's compute capacity.
Rio de Janeiro's IplanRIO released Rio 3.5 Open 397B on June 13, claiming it beat Qwen 3.7 Plus. AI company Nex proved the model is a 0.6 Nex / 0.4 Qwen weight merge; IplanRIO updated the model card and blamed an incorrect upload.
Two types will launch: audio glasses (first this fall) and display glasses. Partnerships with Gentle Monster and Warby Parker for stylish designs. Features include hands-free Gemini access, navigation, and messaging.
The Take It Down Act, signed in May 2025, now requires online platforms to remove nonconsensual intimate imagery (including deepfakes) within 48 hours or face fines. Experts warn the law could facilitate censorship and does little to help victims.
M3 has ~428B total params, ~23B activated, supports 1M-token context and native multimodal (text, image, video). It is open-weight with day-0 support on vLLM including dedicated MSA kernels for Hopper and Blackwell.
Blackstone commits $5 billion equity to the venture, targeting 500MW of capacity by 2027. Google will supply TPUs, software, and services.
NVIDIA's first Vera CPUs arrived at Anthropic, OpenAI, SpaceXAI, and Oracle Cloud, hand-delivered by VP Ian Buck. The custom CPU, purpose-built for agentic AI, features 88 Olympus cores, 1.2TB/s memory bandwidth, and 50% faster per-core performance.
KPMG will embed Claude into its Digital Gateway platform, starting with tax and legal tools, and provide access to all 276,000+ employees. Anthropic names KPMG a preferred partner for private equity, co-building products for portfolio companies.
Following the conclusion of the Musk-Altman trial, both billionaires are now focused on taking their AI companies public. The IPOs are expected to be landmark events on Wall Street.
Moonshot, the Kimi chatbot maker, is raising about $2B at a $20B+ valuation while dismantling its red-chip structure to comply with Chinese regulations for overseas listings. The restructuring signals low chances of securing a regulatory waiver for its current VIE setup.
Indian AI startup Sarvam raised $234 million in funding led by HCLTech, with HCLTech contributing $150 million. The round gives Sarvam unicorn status, making it one of India's most valuable AI companies based in Bengaluru.
ERA uses Gemini to write and optimize scientific code, tested on genomics, public health, and more benchmarks. It achieves expert-level performance and is part of Computational Discovery, now available through a trusted tester program in Google Labs.
The blog post from LangChain argues that agent frameworks remain useful as LLMs improve, but they must evolve. It highlights that LangSmith offers observability for any framework, including LangChain, Claude SDK, and custom-built agents.
Companies using AI assistants like Claude face unexpectedly high token consumption, driving up costs. A software maker and an ecommerce firm share their experiences navigating 'tokenomics.'
The article argues that traditional AI benchmarks like MMLU are insufficient and advocates for metrics based on economically valuable work. Efforts like Nvidia's AgentPerf and the Tokenomics Foundation are cited as alternatives focusing on real-world utility.
Agents have an infinite input space and non-deterministic behavior, requiring new observability tools beyond traditional software monitoring. The guide covers tracing, evaluation, and using production traces for continuous improvement.
SoftBank founder Masayoshi Son's heavy investment in OpenAI has sparked concern among insiders, who view his devotion to CEO Sam Altman as risky. The bet highlights growing Japanese capital influence in AI.
Agents make writing code almost free, but code review remains costly. The 2026 data shows this consistently, and advice on AI code review often fails for solo vs team contexts.
Kalshi developed an AI agent to stress-test prediction market bets. The tool evaluates portfolio risk by simulating market scenarios. No technical details or release date were disclosed.
Automates prompt engineering through optimization. Aims to save time and improve AI performance while enabling faster model switching.
In April, an Earth observation satellite autonomously located a target object for the first time, using onboard AI without ground control. The milestone demonstrates real-time autonomous decision-making in orbit.
Ben Thompson analyzes Anthropic's safety-focused marketing strategy, linking it to the earlier Mythos Preview release. The article explores how the company's safety claims serve as a competitive differentiator and addresses skepticism around its messaging.
The National Hurricane Center used DeepMind's WeatherNext to forecast Hurricane Melissa's rapid intensification and Category 5 landfall in Jamaica five days in advance. It marked the first time NHC predicted a storm's Category 5 intensity from Category 1 wind speed.
OpenAI's new dreaming memory system is rolling out to Plus and Pro users in the US, using a more capable architecture to synthesize preferences across conversations. Some users report the system loses specific details, leading many to revert to legacy memory.
China's universities cut 12,200 undergraduate programs and added 10,200 new ones between 2021-2025, with cuts concentrated in arts, humanities, and management. New majors include embodied intelligence and other AI-related fields, aligning with Beijing's economic goals and addressing graduate unemployment.
Wayfair leverages GPT-5.5 via OpenAI API to enrich product information across 40 million products, improving organization and differentiation. The next phase involves applying Codex to complex internal challenges.
Regularized f-Divergence Kernel Tests, presented at AISTATS 2026, improves statistical power for verifying unlearning without accessing model internals. The method reduces the number of samples needed to detect violations, addressing computational challenges in large-scale models.
Replit's AI coding agent uses LangGraph and LangSmith to manage hundreds of steps with deep observability. LangChain added performance, search, and human-in-the-loop features to support Replit's scale.
Malaysian AI messaging startup Respond.io raised $62.5M to fuel its AI agent-powered platform. The company charges per conversation instead of per seat, and plans to use the funds for acquisitions in North America and Europe.
NVIDIA published cookbook-style recipes for fine-tuning protein language models like ESM2 using LoRA with BioNeMo. The recipes include data prep, training, and evaluation, with code on GitHub and NGC.
Zyphra's ZONOS2 is a next-generation real-time TTS model with 8B total parameters (900M active) for efficient inference. It supports high-fidelity voice cloning and is released on Hugging Face.
Cybersecurity companies are deploying AI to automatically detect vulnerabilities in open-source software. The approach aims to scale security analysis beyond manual review.
NewCore raised $66M to build a security-first identity platform for AI agents. The startup argues that managing AI agent identities is the next major enterprise security challenge.
Atoms is a no-code vibe coding tool that converts plain language descriptions into fully built, deployed, and marketed applications using AI agents. It aims to make app development accessible to non-developers.
Wayfair reduced ML model costs by 90% on two separate occasions using Cursor. The case study details how the tool optimized their ML infrastructure.
Two arXiv papers introduce shielding approaches for safe reinforcement learning. 'Contract-Based Compositional Shielding' addresses multi-agent coordination under global safety constraints. 'Robust Shielding' handles unknown transition dynamics in MDPs.
Gemini Spark runs 24/7 on Google Cloud virtual machines, powered by Gemini 3.5 Flash, and can act on your behalf even when your laptop is closed. It integrates natively with Gmail, Docs, Sheets, and Slides, with third-party app support via the MCP protocol.
Tens of thousands of workers are being laid off while a small cohort of AI insiders becomes wealthy on an incomprehensible scale. The article argues this disparity is creating a powder keg of social unrest.
Cohere, known for sovereign AI deployments with regulated enterprises, releases its first coding model, North Mini Code. The model targets developers, expanding beyond its traditional customer base in banking, government, and healthcare.
A Digital Music News investigation identifies 'giant datasets' of copyrighted tracks used to train AI music models, challenging claims of indiscriminate data collection. The report suggests training may be more systematic than previously believed, raising transparency concerns.
Baidu's Apollo Go, in partnership with Swiss public transport operator PostBus, has begun road testing its driverless service AmiGo in Switzerland. The project aims to achieve regular commercial operations by 2027, primarily serving specific routes.
Guide explains that agent observability shifts focus from debugging code to debugging reasoning, using traces to understand agent behavior. It shows how production traces become the foundation for systematic evaluation and iteration.
LangChain and Fireworks fine-tuned an open model to mine error signals from production traces, matching frontier model performance at a fraction of the cost. The approach reduces costs by 100x compared to using GPT-4 or Claude.
Alipay unveiled AI-powered assistant Abao on June 16, marking the biggest redesign in the platform's history. The assistant is available via invitation-only testing, signaling Alipay's full-scale shift toward AI-driven services.
OpenEvals and AgentEvals provide pre-built evaluators for LLM-as-judge, structured data, and agent trajectories. The packages aim to simplify evaluation for production-ready applications.
Gary Marcus criticizes Trump's June 2 AI Executive Order as voluntary and too narrow, focusing primarily on cybersecurity. He notes Florida's lawsuit against OpenAI highlights harms not covered by the EO. Marcus calls for stronger government action to address AI risks.
Sakana Marlin generates up to 100-page research reports with slides in about 8 hours. Positioned as a Virtual CSO, it focuses on deep analysis over speed.
The post covers bespoke testing, single-step validation, full turns, multi-turn simulations, and environment setup. It emphasizes the need for layered evaluation as agents grow more autonomous.
LangChain evaluated four Claude Code setups for domain-specific coding, finding that a concise Claude.md guide combined with MCP documentation access produced the best results. The team used Claude 4 Sonnet and tested setups including vanilla, MCP-only, Claude.md-only, and the combined approach.
The DOJ contends that xAI's gas turbine operations are integral to military operations, including the Iran War, to dismiss a NAACP lawsuit over pollution. The lawsuit alleges environmental racism from xAI's Memphis facility.
Astrophysicist Chi-kwan Chan uses Codex to simulate black holes. Nextdoor engineers use Codex with GPT-5.5 to ship features. Codex is now available through Oracle Cloud commitments.
OpenRouter's Fusion sends prompts to multiple models in parallel, using web search and bash tools, then a judge compares answers and a synthesizer writes the final response. Tweets claim three cheap models fused outperformed GPT-5.5 and Claude Opus 4.8 at half the price.
MiMo Code, from Xiaomi's MiMo AI team, reportedly maintains coherence beyond 200 coding steps, whereas typical coding agents stall around 30 steps. The claim challenges the endurance limit of tools like Claude Code.
Qualcomm CEO Cristiano Amon predicts AI agents will eventually replace traditional apps, and the chipmaker is working on 40 new AI-powered devices including smart glasses. Amon said smart glasses could become as big as the smartphone.
Uber is limiting employees to $1,500 in monthly token spending per AI coding tool, according to Bloomberg. The cap applies to agentic coding software like Cursor and Claude Code after Uber exhausted its 2026 AI budget in four months.
An internal personality clash at Anthropic reportedly took its models offline, according to an Axios report. The incident highlights internal turmoil at the AI lab.
European regulators propose requiring Android to allow third-party AI tools invoked system-wide via hot words or button presses. The ruling, reported by Ars Technica in April, aims to open the platform to competing AI assistants.
NextEra Energy and Dominion Energy have agreed to a merger, underscoring the growing demand for electricity from AI data centers. The deal is expected to accelerate utility consolidation to meet AI infrastructure needs.
Standard Chartered plans to cut thousands of jobs as part of an AI-driven efficiency initiative. The move reflects the growing trend of automation in the banking sector.
Multi-agent systems built on LangGraph reduce debug time by 93% and compress cross-team delivery, according to a LangChain blog post. The architecture mirrors real engineering teams using agent swarms.
Analog Devices is in advanced talks to acquire Empower for approximately $1.5 billion. The deal would bolster ADI's power management business.
A single click on a trusted Microsoft link could leak emails, calendar details, and indexed files from Microsoft 365 Copilot Enterprise Search. Researchers at Varonis Threat Labs chained three bugs into a one-click exfiltration path called SearchLeak.
GitHub published a new open dataset to accelerate multilingual AI development. The dataset aims to support researchers and developers building LLMs for non-English languages.
The Grande 44 has 44 hp of peak power and a towing capacity of 6,000 lb. It is Burro's most powerful autonomous platform built for outdoor heavy industry.
India and the UAE have formed a partnership to develop sovereign AI infrastructure, including compute and data centers, aiming to reduce reliance on US tech giants like Google and Microsoft. The collaboration involves G42 and Cerebras, focusing on building independent AI capabilities.
Arcee AI announced a multi-million-dollar strategic partnership with Hugging Face, replacing AWS S3 with Hugging Face for storage. It is the first major American AI lab to make the switch, per Hugging Face CEO Clément Delangue.