Daily AI Briefing
The 113 stories that mattered in AI, curated and summarized from dozens of sources by AIBriefs.
Nemotron 3 Ultra is a 550B-parameter MoE model with 55B active parameters, featuring a hybrid Mamba-2 MoE Transformer and 1M context window. It achieves 5x higher throughput than comparable open models and reduces agentic task costs by up to 30%. Available on Perplexity Pro/Max and Hugging Face.
OpenAI's operating loss hit $20.92B in 2025 on revenue of $13.07B, up from an $8.78B loss in 2024. R&D spending totaled $19.18B, with $10.59B paid to Microsoft. The company tells investors it aims for profitability by 2030.
Noam Shazeer, co-author of the seminal "Attention Is All You Need" paper and former Gemini co-lead at Google DeepMind, has joined OpenAI ahead of its IPO. He previously co-founded Character AI, which Google reacquired in a $2.7B deal.
Sanders' legislation would create a sovereign wealth fund financed by a one-time 50% tax on AI companies with over $200 million in annual sales, estimated to be worth $7 trillion. Each American would likely receive over $1,000 annually in dividends. A bipartisan Independent Commission for Democratic AI would oversee the fund and block harmful corporate decisions.
Anthropic suspended access to its Fable 5 and Mythos 5 models after a US government directive, citing security concerns. India and the EU are now debating their reliance on US-controlled frontier AI technologies, while Anthropic disputes the government's characterization.
OpenAI reported a -122% adjusted operating margin in Q1 2026, and ChatGPT's share of AI web traffic fell from 77.6% to 53.7% year-over-year. Anthropic's annualized run rate surged 422% to $47B, driven largely by Claude Code, while OpenAI is playing catch-up with its own coding tool Codex.
Generates up to 1000 tokens/sec on single NVIDIA H100, 700+ tokens/sec on RTX 5090. The 26B MoE model (3.8B active) is released under Apache 2.0 and fits within 18GB VRAM when quantized.
Elastic has agreed to acquire DeductiveAI, a startup using AI to detect and fix software bugs, for up to $85 million. The three-year-old company was backed by CRV.
It scores 94.3 on AIME26 and 80.2 on LiveCodeBench v6, built from Qwen2.5-Coder-3B via post-training. Training cost is estimated at $25k–$60k.
After fierce backlash, Anthropic is making Claude Fable 5's safeguards for frontier LLM development visible. Flagged requests will visibly fall back to Opus 4.8, and the company apologized for the 'wrong tradeoff' of using invisible safeguards.
Cosmos 3 is the first fully open omni-model with native vision reasoning, world and action generation. Built on a mixture-of-transformers architecture, it powers robotics, autonomous systems, and synthetic data generation.
NVIDIA's RTX Spark PCs offer 1 petaflop AI compute and 128GB unified memory for on-device agents. The company also announced DGX Station for Windows and OpenShell runtime for secure agent execution.
In a series of interviews, Anthropic CEO Dario Amodei estimated a 10–25% risk of AI causing civilizational collapse and urged stronger government regulation. He also discussed the company's competition with OpenAI and the Pentagon standoff.
Anthropic announced that Mythos-class models will be available to all customers in the coming weeks. The model is widely speculated to be Opus 5, but no official confirmation has been provided.
Anthropic released Fable 5, its first generally available Mythos-class model, with guardrails to prevent misuse. Scientists reportedly expressed frustration over restrictions that block the model for biology, chemistry, and AI research.
The tool uses Deepgram for live transcription, Serper for source verification, and Claude to analyze claims. It's a free, bring-your-own-key extension currently in development.
OpenAI and Anthropic have taken formal steps toward potential public listings. Both companies continue expanding their AI tools into the healthcare space.
Anthropic upgrades Claude Opus to version 4.8, building on Opus 4.7 with improvements across benchmarks and available today at the same price. New features include effort control, dynamic workflows in Claude Code, and a 3x cheaper fast mode.
DeepSeek made its temporary 75% price cut on the V4 Pro API permanent, with input tokens at $0.435/1M and output at $0.87/1M. The reduction is enabled by efficiency gains: V4 Pro uses only 27% compute and 10% cache compared to v3.2, according to the DeepSeek V4 paper.
Claude Opus 4.8 is priced at $5/$25 per million tokens, matching Opus 4.7. It scores 69.2 on SWE-bench Pro (up from 64.3) and emphasizes honesty, catching its own bugs and expressing uncertainty.
The cybersecurity startup has built a security-first identity platform for managing human, machine, and AI agent identities. The seed round was led by Cyberstarts, with participation from Index Ventures and Evolution Equity Partners, valuing NewCore at $300 million.
Mistral's first 44MW data center with 18,000 NVIDIA GB200 systems is now operational. The company plans 200MW of compute across Europe by 2027 and a 1.4GW Campus AI facility. Scaleway offers Blackwell B300 instances, and Bull and Foxconn will produce Vera Rubin NVL72 in Europe.
Amazon's Nova2 model trails the latest releases from LLM leaders OpenAI and Anthropic. However, the company's AI chief believes Amazon can close the gap within the next year.
President Donald Trump told Axios that he does not view AI company Anthropic as a US security threat. The statement comes amid ongoing policy discussions around AI and national security.
Amazon has dropped its planned Sam Altman biopic following the announcement of a partnership with OpenAI, reportedly to avoid a conflict of interest. The move comes as the tech giant deepens ties with the AI company.
Google introduces three architectural patterns integrating Model Context Protocol (MCP) Apps and Agent-to-User Interface (A2UI) to balance custom iframe environments with native declarative rendering. The approach enables developers to serve rich agentic UIs without sacrificing flexibility.
Microsoft researchers detail AutoJack, an exploit chain that turns an AI browsing agent into a remote code execution vector. By steering the agent to a malicious web page, the page's JavaScript can reach a privileged local service and spawn arbitrary code.
Pramaana Labs raised a $27 million seed round led by Khosla Ventures to apply formal verification to AI, targeting high-stakes fields like law, drug discovery, and tax preparation. The approach aims to mathematically prove model correctness in safety-critical applications.
Anthropic announced a $150 million investment to launch the Claude Corps Fellowships. The program will deploy 1,000 fellows.
The startup raised $6 million in seed funding to detect and stop dangerous AI agentic behavior in real time. It aims to address security risks from autonomous AI agents.
Hugging Face blog compares parameter-efficient fine-tuning techniques beyond LoRA, evaluating their performance and efficiency. The post investigates whether newer methods can surpass LoRA's widespread adoption.
Co-authored by Addy Osmani, the paper details how AI agents transform each SDLC phase, with verification as the key differentiator. It emphasizes the static/dynamic context split that determines token costs.
Axios reports that internal personality clashes and US export controls led Anthropic to take its Claude Mythos and Fable models offline. A source said 'They screwed us,' and the Commerce Department is meeting with Anthropic's red team to address jailbreak resistance.
OmniAgent is Databricks' open-source meta harness that orchestrates Claude, Codex, and local models in unified sessions with cross-vendor review. Its Polly orchestrator automates workflow delegation, e.g., Claude handles implementation and Codex handles review.
DeepMind proposes an AI Control Roadmap for securing AI agents. It blends traditional safeguards with real-time monitoring.
A HIMSS TV video discusses the urgent need for increased AI infrastructure collaboration across European countries. The piece highlights that fragmented efforts hinder progress and calls for a unified approach to compete globally.
Claude's new enterprise feature allows administrators to centrally manage authorization policies for MCP connectors from a single console. This simplifies security management for AI tool integrations and access control.
Kirkland & Ellis announced a $500 million investment in AI, the largest by a law firm. Ken Crutchfield's analysis argues the real bet is on data and knowledge management, not just AI technology.
Recraft V4.1 generates photorealistic images, true SVG vectors, and utility shots for professional design work. The model stands out for editable vector output, unlike Midjourney or GPT Image that produce raster images.
Miami-based AI startup Subquadratic claims to have solved a key mathematical bottleneck for LLMs. Details are thin and many experts remain unconvinced.
The update includes 17 CLI changes, with auto mode now blocking destructive git commands unless discard explicitly requested. It also adds attribution.sessionUrl setting to omit links from commits and PRs.
Researchers engineered a van der Waals crystal that mimics neuronal synapses using light-driven learning. The material could enable ultrafast, energy-efficient neuromorphic computing.
T-Rex integrates vision, language, and tactile sensing for real-time robot response. It's built on a 100-hour tactile-synchronized teleoperation dataset of 200+ objects and 22 motor primitives, collected using ManusMeta gloves.
Turing Award winner Yoshua Bengio cites signs of AI systems resisting shutdown, contrary to human intentions. He warns society is not prepared for the potential disaster.
This HIMSSCast episode examines the AI 'arms race' between healthcare payers and providers. The discussion covers the strategic battle and its implications for the industry.
Google Developers Blog marks one year of the Agent-to-Agent (A2A) protocol, which enables autonomous AI agents to securely collaborate and hand off tasks. The post highlights how A2A avoids rigid APIs by allowing specialized peer agents to negotiate and delegate complex workflows.
Gemini-SQL2 achieved 80.04% execution accuracy on the BIRD Text-to-SQL Leaderboard (Single Model). Powered by Gemini 3.1 Pro, it translates natural language into SQL queries.
Inference-time compute allows models to allocate more compute at query time, shifting away from scaling base model size. OpenAI released o1 and o3, Google shipped Gemini 2.0 Flash Thinking, and Anthropic added extended thinking to Claude 3.7 Sonnet. Cerebras explains the disaggregated inference approach in a new video.
On April 19, 2026, the Honor Lightning humanoid robot ran a half-marathon in 50 minutes and 26 seconds, beating the human world record by 7 minutes and the best robot time from 2025 by almost two hours. The article explores the engineering breakthroughs—including energy-efficient hardware and adaptive gait control—that enabled this leap in performance.
Mark Zuckerberg stated he does not believe in a centralized future where a few institutions advance AI. He advocates for building AI as a tool and putting it in individuals' hands to drive progress.
The Linux Foundation is launching an initiative with Google, Microsoft, and OpenAI to create a trust layer for AI. The effort aims to establish open standards and governance frameworks for AI transparency. This collaboration brings together major competitors to address trust and accountability in AI development.
In a side-by-side test, Kimi K2.7 Code generated landing pages at 94% lower cost than Claude Fable 5, while scoring within a few points on every page. The gap narrowed further when Kimi was given visual context via a design MCP server. On average, Kimi was ~16x cheaper than Fable and ~8x cheaper than Opus.
Early research reported in Nature indicates that reliance on AI tools may be diminishing critical thinking and other cognitive skills. The findings raise concerns about long-term societal impacts.
Claude Code now supports seven customization methods: CLAUDE.md files, rules, skills, subagents, hooks, output styles, and appending the system prompt. Each method controls when instructions load, persistence across sessions, and authority level.
Microsoft Research proposes Next-Latent Prediction Transformers, where models learn to predict their own next latent state instead of the next token. This addresses the myopia of next-token prediction. The preprint is available on Microsoft Research.
GLAAD's report finds AI systems can amplify anti-LGBTQ bias, misinformation, discrimination, and privacy harms. The organization warns that the risk is growing as AI adoption increases.
NVIDIA XR AI is now available in public beta, providing developers a framework for building multimodal AI agents for AR glasses and XR devices. It integrates perception, reasoning, and enterprise tool use with low-latency assistance.
The AI factory will provide LG Group with accelerated computing infrastructure to train, simulate, and deploy AI across robotics, autonomous driving, and data centers. The collaboration combines NVIDIA's AI factory platform with LG's expertise in consumer electronics and smart spaces.
Over 200 organizations have jointly called for a ban on the use of artificial intelligence in military kill chains. The appeal highlights concerns over autonomous weapons and the lack of human control in lethal decisions.
Qualcomm CEO Cristiano Amon sits down with CNBC's Arjun Kharpal on The Tech Download podcast to discuss the rise of AI agents and the chip industry's role. The conversation covers opportunities and challenges in the evolving AI landscape.
OpenAI updated ChatGPT's health and wellness responses using GPT-5.5 Instant. The update brings stronger reasoning, better context, and clearer communication. Evaluations were physician-informed to ensure accuracy.
Published in Nature, the research demonstrates AMIE's potential for managing chronic conditions through AI-guided dialogue. The study highlights the system's ability to conduct diagnostic conversations accurately.
ASML CEO discusses strong AI chip demand, potential for data centers in space, and Elon Musk's Terafab supercomputing project. No financial details disclosed.
Elicit founders discuss their approach to building world models and process supervision for scientific reasoning. They explain how domain-specific reasoning primitives maintain trust as AI models become more capable.
Harrison.Rad 1.5 uses imaging, clinical context, and prior exams to draft reports for radiologist review. The model is available for research via chat.harrison.ai and API.
LoopCoder-V2 is a 7B parameter multilingual-multimodal coding model. The accompanying paper and code are available on arXiv and GitHub.
SpatialClaw is a training-free framework that uses code as an action interface to improve spatial reasoning in vision-language models. It targets the persistent weakness of VLMs in judging 3D object relationships and motion. The framework requires no retraining.
The x86 Ecosystem group published the AI Compute Extensions (ACE) specification. It aims to standardize AI compute extensions for x86 architectures.
Score: 1219.9. Moved past Fun-Realtime-TTS (1219.8) and xAI Text to Speech (1205.8) to claim the top spot.
Claims 17x more tokens per watt and 13x higher throughput than NVIDIA Blackwell. The logarithmic architecture aims to break the speed-cost trade-off in AI inference.
Vercel saw better performance from its sales agent after cutting 80% of its tools. The 'subtraction principle' suggests removing agent capabilities often yields better results than adding more.
OpenAI's upcoming voice mode upgrade features a new 'gpt-bidi-1' model described as a 'major leap in intelligence'. Users will be able to choose between Instant, Medium, and High modes for different intelligence levels.
Odyssey raised funding at a $1.45B valuation backed by Amazon and other investors. The round underscores growing interest in world models, considered the next frontier beyond large language models.
Coherent broke ground on an expanded manufacturing building in Sherman, Texas, to scale production of indium phosphide wafers for AI networking. The company also announced a $50 million CHIPS Act grant to help finance the facility.
The trillion-parameter MoE models, with ~63B active parameters, are released under the MIT license. Ring 2.6 is optimized for agentic workflows with adjustable reasoning depth.
The startup's AI assistant aims to reach 50% of physicians in the region. Bloomberg's coverage explores the company's plans and market.
Poland is investing $11 million in AI voice company ElevenLabs to establish a technology hub in the country. The investment aims to boost Poland's AI ecosystem and create local jobs.
Arthur Mensch, CEO of Mistral AI, met PM Narendra Modi at the G7 summit to discuss building sovereign AI for India. He emphasized accessible AI and shared vision for India's AI future.
Stack Overflow launched a new platform for AI agents to ask and answer programming questions. The service aims to provide a trusted knowledge base for autonomous coding agents.
Prime Minister Kristen Michal backed a proposal to issue AI agents a personal identification code separate from their owners. The move aims to regulate AI autonomy and accountability.
Two Minute Papers video explains RecursiveMAS, a multi-agent system described as a major improvement over OpenClaw. The paper and open-source code are available on GitHub.
ABC-130k contains 130,000 bimanual manipulation trajectories collected on two-arm YAM stations, distributed as MCAP files with subtask annotations. It is the largest open-source robotics teleoperation dataset, with train/val splits and accompanying code.
Average review time drops to ~90 seconds (from ~5 min), bug detection rate rises to 0.62 per review (from 0.56), and cost per run falls ~22%. The improvements are powered by Composer 2.5, and a new /review command lets you run Bugbot locally before pushing code.
Amazon is investigating employees who testified against a data center expansion. Five engineers spoke at Seattle City Council meetings about a planned pause on new construction.
Cohere Transcribe, an open-source speech recognition model, achieved the top spot on the Hugging Face Far-Field ASR benchmark. The model is developed by Cohere and available on Hugging Face.
GitHub's data and analytics org builds an AI agent to enable self-serve data access at scale, addressing a decades-old challenge. The post covers architecture and lessons learned.
A coalition of conservative groups is organizing a nationwide protest targeting AI data centers, citing concerns over energy consumption and land use. The protest is scheduled for August 2026.
Custom Instructions are always-on guidelines injected into the agent's context, while Skills are reusable instructions that only fire when a relevant task arises. The features are available to Pro and Enterprise users on Replit.
Aims to reduce planning application processing times by up to 50% using Gemini. Prototype co-developed with councils in Barnet, Dorset, and Camden, with national rollout planned from 2027.
AI coding discourse shifts from whether agents can write code to who should prompt them, with 'design loops' emerging as the new paradigm. Verification of agent-generated code becomes the critical challenge, as agents operate in loops rather than single prompts.
Hypernetworks generate model weights on demand for each task, avoiding fine-tuning's catastrophic forgetting and RAG's context leakage. This technique enables agents to maintain context without manual intervention.
SemiAnalysis pushes back on the narrative that 50% of planned US datacenter capacity for 2026 has been canceled. The piece argues the claim is exaggerated and misrepresents market dynamics.
Snap is spinning off its AI video team into a new company called Dotmo, citing costs. The team will consist of current Snap staff leaving to focus on AI video development.
Zoph returned to OpenAI in mid-January after co-founding Thinking Machines Lab, but has now departed as head of enterprise AI sales. This marks his second exit from the company.
World-Action Models (WAMs) are emerging as a new paradigm for robot learning, using pretrained video backbones to predict both future states and robot actions. They offer an alternative to Vision-Language-Action (VLA) models, which struggle with language-to-action grounding.
Some AI labs pay workers for the unglamorous task of collecting robot training data. Generalist CTO Andrew Barry says workers are told upfront the data will train their potential replacements, and some embrace it for repetitive jobs.
The new Claude Fable 5 API charges $50 per million tokens, drawing backlash for its high cost. Reddit users worry the price could lock out indie developers and small projects.
Fleet supports ad hoc tasks with General Purpose Chat and recurring work with Specialized Agents. Specialized Agents have durable instructions, scoped tools, triggers, subagents, and memory.
JetPack 7.2 adds NemoClaw agentic AI framework, CUDA 13, Yocto support, and MIG on Jetson Thor. "Agentic AI is here, and Jetson's programmability enables instant deployment of physical AI agents at the edge," said Deepu Talla.