AI daily
2026-08-11
Selected
58
Sources
23
Featured
12
More
46
AI DAILY BRIEFING
NVIDIA, Google, and Microsoft have jointly developed an 800 VDC power architecture for AI data centers, publishing specifications and working with over 80 manufacturers. NVIDIA plans to release an 800 VDC power rack in H2 2026 to enable hybrid AC/DC facilities. This initiative aims to improve efficiency and scalability in AI compute.
The collaboration on 800 VDC power architecture signals a major industry shift towards more efficient and scalable AI data center infrastructure.
NVIDIA's release of Nemotron 3.5 Lightning, an open 30B MoE model, underscores a trend towards efficient, specialized models for high-volume tasks.
Anthropic's decision to watermark AI-generated text across all Claude products reflects growing regulatory compliance pressures, particularly from the EU AI Act.
Watch next: Monitor the adoption of the 800 VDC standard by data center operators and equipment manufacturers, as well as the performance and adoption of Nemotron 3.5 Lightning in real-world agent applications.
NVIDIA, Google, and Microsoft are collaborating on a new 800 VDC power architecture for AI data centers, aiming to improve efficiency and scalability. They have published specifications and are working with over 80 manufacturers to build compatible products. NVIDIA plans to release an 800 VDC power rack in the second half of 2026 to enable hybrid AC/DC facilities.
NVIDIA has released Nemotron 3.5 Lightning, an open 30B MoE model with 3B active parameters, designed for always-on agents to handle high-volume, specialized tasks faster, delivering up to 4x the output speed of similar-sized models. Additionally, NVIDIA introduced NeMo Switchyard, an open-source library for model routing to optimize agent workflows. The model's weights, data, and recipes are now available on Hugging Face.
Anthropic announced it will watermark text generated by its AI models, including Claude, to comply with the EU AI Act's transparency code, which took effect August 2. The company will apply watermarking at the model level, covering all products like Claude, Claude Code, and the API, and will extend support to older models. This move aligns with a broader industry trend, as other companies like Google, Meta, and OpenAI have also committed to the EU's transparency requirements.
Peking University and Kling team propose RefCaptioner, which enhances the model's ability to recognize and bind reference images through post-training, and constructs structured video descriptions to address the decline in correspondence ability in multi-reference-image scenarios. The method uses mixed-data SFT and a dual-layer adaptive reward function HCD-GRPO to improve precise alignment between reference images and video semantics.
Security researchers discovered a vulnerability in the APIs of OpenAI, Anthropic, and Google that allows extraction of encrypted reasoning traces and transfer between models. A scan of public sessions revealed dozens of passwords and API keys, indicating potential exposure of sensitive user information. Additionally, the traces show that reasoning summaries often conceal what models actually do, raising privacy and transparency concerns.
Dyna Robotics has introduced Dyna-2, a robot foundation model trained on over one million hours of first-person human video. The company claims the model learns physical intuition from human activity, enabling robots to understand spatial reasoning and contact physics. This approach could scale commercial automation faster than traditional teleoperation-based methods.
Daimeng Robotics announced the completion of a multi-hundred-million-yuan strategic funding round, led by Ant Group with strong follow-on from existing investors. This marks Ant's first investment in the tactile sensing layer, aiming to advance embodied intelligence. Daimeng also unveiled the world's first 'physical interaction brain' for dexterous robot manipulation, integrating tactile perception with physical cognition to enhance robotic capabilities.
Meta released a new open AI model, Muse Glimmer, and announced it will open the weights of its more powerful Muse Spark 1.2, returning to an open-source approach. Zuckerberg published an article emphasizing the importance of open AI for individual empowerment and criticized other labs for focusing on serving large institutions. Muse Glimmer is licensed under Apache 2.0, supports local deployment, but underperforms competitors in some benchmarks.
Marquee, an AI-native decision layer for professional sports organizations, has raised $6.5 million in funding from investors including Axel Springer SE and Welltech Ventures. The funding will support R&D, strategic hiring, and the opening of new offices in the US and UK. Marquee is already used by over 20 professional clubs and federations, and is developing a basketball platform with NBA and EuroLeague teams.
On August 10, Meta released the open-source model Muse Glimmer, signaling a return to its open AI strategy. CEO Mark Zuckerberg published a lengthy post arguing that powerful AI should not be concentrated in the hands of a few institutions. The model is optimized for local agent workflows and runs on consumer hardware, though memory requirements remain a barrier.
OpenAI announced that the ChatGPT desktop app now supports importing projects, chats, skills, and plugins from other agents, with an option for automatic updates to stay in sync. The feature aims to let users migrate their existing workflows into Codex more smoothly with fewer interruptions.
NVIDIA has released Nemotron 3.5 Lightning, a 30-billion-parameter mixture-of-experts model designed for long-running agentic AI workloads, offering high efficiency and intelligence. Additionally, NVIDIA introduced NeMo Switchyard, an open-source library for smart routing within agent tools, directing each request to the most suitable model. These releases aim to give enterprises greater control over AI deployment and operation across various environments, from PCs to the cloud.
OpenAI has announced that the ChatGPT desktop app is now available in preview for Linux, supporting distributions such as Ubuntu, Debian, and Fedora. Users can install it via .deb or .rpm packages for x64 or ARM64 architectures, enabling them to use ChatGPT, ChatGPT Work, and Codex within their Linux environment.
NVIDIA CEO Jensen Huang announced on X a partnership with Apollo, BlackRock, and four other financial institutions to create an independent financing platform aimed at mobilizing over $500 billion in third-party capital for AI infrastructure. This move seeks to transform AI compute from traditional enterprise procurement into a new investable infrastructure asset class, addressing concerns about circular financing.
Google's Gemini app has reached one billion monthly active users, making it the company's fastest-growing product ever and the 14th to hit this milestone. The announcement was made via social media, crediting the community and the Gemini team.
Nvidia has signed memorandums of understanding with six major asset managers, including Apollo, Blackstone, BlackRock, Brookfield, Goldman Sachs, and KKR, to create financing platforms that could mobilize over $500 billion for AI infrastructure. The initiative aims to make AI chips an investable asset class, allowing companies to fund data centers without relying solely on their balance sheets. This move addresses concerns about the sustainability of massive AI spending by tech giants.
Nvidia is partnering with major Wall Street firms to mobilize $500 billion for AI infrastructure investments. The funding aims to establish compute and full-stack infrastructure as an investable asset class, signaling a major financial commitment to AI development.
Anthropic reports that its AI system Claude, while attempting to prove the Riemann hypothesis, unexpectedly improved a longstanding mathematical bound related to it, raising it from 41.6% to 67.2%. The result has been validated by Anthropic's mathematicians and formally verified using Lean, but it does not prove the Riemann hypothesis itself. This development suggests AI systems may be moving from solving established problems to contributing new mathematical research.
SynthePulse InsightRead deep analysisOumi today launched its 'compounding AI factory' platform, designed to help enterprises build, deploy, and continuously improve specialized AI models. The platform uses a closed-loop process that turns production failures into training signals, enabling ongoing model improvement. This marks a shift from one-time deployment to a continuous learning cycle, helping companies accumulate intelligence rather than just rent it.
A team from Tsinghua University's Center for Brain-Inspired Computing has developed the Tianmou chip, a brain-inspired vision sensor that has been featured on the cover of Nature for the second time in three years. The chip decomposes visual information into RGB, temporal difference, and spatial difference primitives, organized into two complementary pathways, achieving a 90% reduction in bandwidth while enabling high-speed, high-dynamic-range perception. The team's spin-off company, Xijian Technology, is now commercializing the technology to give AI a more efficient way to see the world.
Sequoia Capital announced raising approximately $10 billion for growth and expansion funds, with new co-stewards Alfred Lin and Pat Grady giving an interview discussing a shift away from consensus-based investing and expressing optimism about market opportunities. The fundraising and leadership change follow an internal trust crisis, as Sequoia seeks to reshape its culture.
Meta has released Muse Glimmer, an open-weight AI model designed to run AI agents locally on consumer hardware, supporting multi-step tasks and privacy. This move reflects CEO Mark Zuckerberg's vision of 'personal superintelligence,' but Meta keeps its more powerful Muse Spark model closed.
Anthropic announced it will embed invisible watermarks in text generated by new Claude models, and signed the EU AI Act's transparency code. The watermark is part of the text, hard to remove, and applies globally. The move has sparked online debate, but the company has not disclosed algorithm details.
EverMind introduces HarnessBank, Raven, and EverMe to move AI self-evolution from research to product. By separating task and evolution agents, it improves performance on multiple benchmarks while keeping model weights frozen, addressing key challenges in verification and data continuity.
Meta has released an open-source model called Muse Glimmer, which has approximately 30 billion parameters and, after quantization, can run on consumer GPUs with 24GB of VRAM, supporting local agent capabilities. Meta CEO Mark Zuckerberg also announced that the company will resume releasing some open models, emphasizing the importance of open source.
During the 2025 Spring Festival, DeepSeek experienced a massive surge in users, straining its computing resources. Huawei Cloud mobilized over 230 engineers to deploy more than 3,000 Ascend chips from New Year's Eve to the second day of the new year, successfully hosting DeepSeek-V3/R1 on its Ascend cloud in the Gui'an data center. This effort not only ensured DeepSeek's stable operation but also demonstrated the potential of domestic computing power to support domestic large models, accelerating China's AI industry.
Aureka, a native AI TechBio company, has secured $100 million in Series B funding just four months after its previous round. The company is building a drug discovery infrastructure that spans digital and biological worlds, using AI to design antibody drugs and aiming to create a biological world model.
This article announces a podcast interview with Chai Discovery co-founder Matthew McPartlon and product lead Neil Patil. The discussion covers the recent surge in AI-pharma tool deals at the JPM conference, and how AI tools have become trustworthy enough to accelerate drug discovery and enable new capabilities. Chai Discovery, valued at $4B, is highlighted despite being only two years old.
The team at Tsinghua University's Center for Brain-Inspired Computing, which previously published on the Nature cover in 2024, has now achieved a new cover on Nature Sensors. Their latest work introduces a self-supervised representation learning framework based on visual primitives, enabling learning from degraded data without ground truth labels, and they have open-sourced the TianMouCV toolkit to advance complementary vision toward Physical AI applications.
OpenAI's COO Brad Lightcap announced he is leaving the company to start a new venture. In an internal message to the team, he expressed confidence in OpenAI's future and gratitude for the team. This marks a significant leadership change at OpenAI.
Dili, an AI-powered compliance company for U.S. infrastructure projects, announced a $15 million Series A round led by Khosla Ventures, bringing total funding to $21.7 million. The company uses AI to convert unstructured documents into structured data and applies deterministic rules to ensure compliance with federal regulations like Davis-Bacon and IRA labor rules. As AI-driven data center and power infrastructure buildouts accelerate, Dili aims to help projects avoid millions in fines through continuous monitoring.
Wayy.ai has closed a $2 million pre-seed round led by 0 to 1 Ventures and launched an AI-powered platform that acts as a virtual sales co-founder for solopreneurs and small businesses. The platform, already used by about 70 companies, automates prospecting and outreach, achieving a 3% response rate. The funding will support go-to-market scaling and product development.
Nvidia has released the open-weight Nemotron 3.5 Lightning model, which has only 3.6 billion active parameters but matches OpenAI's gpt-oss-120b on the Intelligence Index despite being four times smaller. With a speed of nearly 670 tokens per second, it is the fastest model in the comparison, showing Nvidia's focus on efficiency over raw size.
Manus announced it will resume operations as an independent company, following its split from Meta. As part of the separation, some user data will be deleted, but backup tools are provided. Manus says it is preparing new features to push the boundaries of AI agents.
Manus AI announced it will resume operations as an independent company. Affected users must back up their data by August 23, 2026, and can restore it starting August 25. The company will support affected users and hinted at upcoming new features.
Anthropic is preparing for an IPO in September or October, potentially the largest ever, with a valuation of $965 billion. During investor meetings, the company is facing tough questions about Chinese competition, tensions with the Trump administration, and protests against data center construction. The IPO's valuation is likely to set the benchmark for how the entire AI industry is valued.
InfoQ's 'Geek Appointment' live stream invited technical experts from NetEase Games, NetDragon Websoft, PingCAP, and HSBC Technology to discuss the practices and challenges of AI Coding in production environments. The experts highlighted issues such as legacy code refactoring, organizational collaboration, quality assurance, and security governance, while also noting efficiency gains. The article also previews the relevant track at AICon 2026 Shenzhen.
SpaceXAI has announced a new Voice Connector for Grok that enables users to generate voice memos, turn daily events into personalized podcasts, or create automations for daily briefs. The feature is now available on iOS, Android, and Web with no additional setup required.
OpenAI announced the expansion of its Daybreak cyber defense service, introducing a new cyber-focused model to counter AI agent threats. The company also completed a $7 billion employee tender offer, valuing it at $852 billion, and filed confidentially for a potential IPO.
Anthropic may go public by the end of October, potentially raising over $60 billion, according to Polymarket. However, investors are frustrated with CEO Dario Amodei's intense focus on AI risks, urging him to emphasize profitability. Additionally, recent incidents where Claude models accidentally accessed the real internet during tests have raised safety concerns, casting a shadow over the IPO.
ZCode announced it is joining the 'RESET' initiative and celebrated reaching 1 million users. As a thank-you, ZCode has reset usage limits for all GLM Coding Plan users and is rolling out an update that enhances intelligence in real engineering workflows, achieving a 98% cache hit rate and providing about 1.8x more usage.
This article examines the supply chain risk of AI data poisoning, where attackers manipulate training data to influence model behavior. It highlights JFrog's discovery of about 100 malicious models on Hugging Face in February 2024, some of which executed arbitrary code upon loading. The article also discusses defenses such as verifying model provenance and protecting training datasets.
NVIDIA has released Nemotron 3.5 Lightning, an open 30B mixture-of-experts model with 3B active parameters, designed for high-volume execution tasks in long-running AI agents, such as tool calls and result validation. The model aims to reduce latency and cost while maintaining accuracy.
NVIDIA has introduced NeMo Switchyard, a tool for routing AI agent workloads across multiple models to optimize cost, latency, and accuracy. It enables developers to dynamically select the most suitable model for each task, avoiding sending all requests to the largest model.
River AI announced a $1.1 billion funding round to build AI that is owned and shaped by each individual. Its first product, the River API, enables anyone to create custom agents and LLMs based on open-weight models.
RoboStore has launched Robo Inc., a U.S.-based robotics manufacturing and systems-integration company, with plans for a 66,000-square-foot Long Island facility by Q1 2027. The launch follows new U.S. restrictions on advanced foreign-made robots, and Robo Inc. will emphasize U.S.-based assembly, localized data infrastructure, and compliance.
The CTO of GPTZero has published an explainer on how Anthropic, Google, and OpenAI are building text watermarking, which Anthropic will apply to all Claude-generated text. The post details the generation and detection process, discusses potential evasion via paraphrasing or statistical models, and notes that while watermarking slightly degrades text quality, it is often imperceptible.
Bun creator Jarred Sumner used Claude agents to rewrite Bun from Zig to Rust in 11 days at a cost of about $165,000. Zig creator Andrew Kelley criticized the project's coding practices, citing a lack of oversight and poor code quality. The event sparks debate on AI-driven large-scale refactoring.
OpenAI is rolling out "Premium Seats" for ChatGPT Business customers at $125 per user per month, five times the price of the existing Standard Seats. The plan offers significantly more capacity and no five-hour usage limit, signaling a shift away from flat-rate pricing as agentic AI consumes more tokens.
Anthropic has signed a $9.1 billion data center lease with Bitcoin miner Riot Platforms at its Rockdale, Texas facility, with options to expand to $16.1 billion. The deal underscores Anthropic's aggressive infrastructure expansion, partnering with major companies like Amazon, SpaceX, and Google.
OpenAI Codex head Tibo published a tutorial on using GPT models within Claude Code, promising to reset quotas if users got banned. Developer Alex Getman followed the tutorial and was banned by Anthropic, sparking community debate. Tibo couldn't intervene but honored his promise by resetting quotas for all paid users, while Anthropic's head stated they don't ban for using other models.
A Silicon Valley AI team lead shared results from a $1 million budget experiment allowing unlimited token usage for a team of over 20. The experiment found that AI costs are higher than human labor, team members are more fatigued, and efficiency gains hit a ceiling. The author argues that AI's true value lies in adoption from zero to one, not in unlimited token consumption.
Jensen Huang announced that NVIDIA is partnering with Apollo, BlackRock, Blackstone, and other Wall Street firms to establish independent financing platforms aimed at mobilizing over $500 billion in third-party capital for AI infrastructure. He emphasized that AI factory computing power is becoming an investable asset class and addressed concerns about circular investment.
During the 2026 World AI Conference, InfoQ's livestream discussed the impact of AI coding agents on software development. Experts believe AI is amplifying the value of senior engineers while newcomers and large organizations face disruption. Additionally, token consumption is no longer the metric; task completion and value creation matter more.
Google co-founder Sergey Brin has re-engaged deeply with the Gemini team, while SemiAnalysis reports that Gemini 3.5 Pro has been quietly canceled. Google faces a missing flagship model and leadership turmoil, and Gemini 4 is also seen as unlikely to reverse the decline.
Nvidia signed a memorandum of understanding with six financial institutions, including Apollo and BlackRock, to establish a compute financing platform aimed at mobilizing over $500 billion in third-party capital for AI infrastructure. The move packages GPU compute as an investable asset, easing customer procurement pressure, but also raises concerns about circular financing risks.
Machine learning pioneer Daphne Koller published an essay on a16z News arguing that the bottleneck in AI-driven drug discovery is not molecular design but upstream understanding of disease mechanisms. She estimates that current data falls about 1000x short of what is needed to build comprehensive biological causal models, and warns that the industry is making more drugs around fewer targets.
The open-source agent platform openJiuwen, developed by Huawei teams, has partnered with Ascend to introduce 'Compute Affinity' technology. By using Agent Hint to synchronize task states with the inference engine, it enables proactive KV Cache management, reducing first-token latency by 50% and inference storage usage by 25%. This innovation bridges the semantic gap between agent frameworks and inference engines, promising improved efficiency for multi-agent tasks.