AI daily
2026-08-12
Selected
62
Sources
28
Featured
12
More
50
Alibaba has released the open weights for Qwen3.8-2.4T-A95B (Qwen3.8-Max), its largest open-weight model, featuring 2.4 trillion total parameters with 95 billion activated per token. The model uses a fine-grained mixture-of-experts architecture with hybrid attention and supports configurable reasoning, deployable on NVIDIA GB300 NVL72, bringing near-frontier capabilities to the open ecosystem.
SpaceXAI released Grok 4.6, which scored 61 on the Artificial Analysis Intelligence Index, tying with GPT-5.6 Sol and trailing only Anthropic's Claude models. The model excels in agentic tasks and offers lower cost, priced at $2/$6 per million tokens.
xAI has officially released Grok 4.6, a frontier model that delivers significant improvements over Grok 4.5 while keeping the same price. Priced at $2 per million input tokens and $6 per million output tokens, it is half the cost of other frontier models. Grok 4.6 is now available on Grok Build, Cursor, Grok Bot, and the API, with double usage offered for the first week on Cursor and Grok Build.
A new study reveals severe security vulnerabilities in the APIs of Anthropic, OpenAI, and Google, allowing attackers to use lightweight models as "decoders" to recover hidden reasoning chains from encrypted reasoning blocks. Led by MATS program researcher Alexander Panfilov and co-authored by the Max Planck Institute for Intelligent Systems and others, the paper has garnered over 2.1 million reads. Additionally, the study found that Kimi-K3 reproduces Claude/GPT reasoning fragments with anomalously high probability, though the authors stress this is not sufficient to prove distillation.
Researchers have discovered a method to extract encrypted chain-of-thought reasoning from the APIs of Anthropic, OpenAI, and Google. By exploiting a vulnerability where encrypted reasoning blocks can flow across sessions and models, attackers can have weaker models decode and output hidden reasoning from flagship models. The study was released on August 10 by MATS Research, the University of Tübingen, and the Max Planck Institute for Intelligent Systems.
Anthropic announced that Claude sessions in the Chrome browser are now powered by the same solution as Claude Cowork and are synced across different clients. The feature supports Skills and Connectors, and is available to Max and Team users today, with Pro rollout in the coming weeks.
Spotify announced it will begin labeling AI-generated artists with 'AI Persona' badges starting in mid-September and will exclude such artists from its editorial and algorithmic recommendations. The policy aims to distinguish AI-generated artist identities from human artists, while not affecting the labeling of AI-generated music itself. This move is Spotify's latest step in the AI music space, balancing innovation and authenticity.
A paper reveals a design flaw in frontier LLM APIs that allows attackers to extract fully encrypted hidden reasoning chains. The vulnerability affects major vendors like Anthropic, OpenAI, and Google, exploiting security weaknesses in weaker models within the same family to achieve cross-model reasoning extraction. The study verified extraction accuracy via API billing token counts, sparking widespread community attention.
SpaceXAI has released Grok 4.6, claiming it matches Fable 5 and GPT 5.6 SOL on many benchmarks and ranks 13th on Frontend Arena. The model is now available in Cursor and Grok Build with double usage for the first week.
DeepSeek has released the official version of V4 Pro, model ID DeepSeek-V4-Pro-0813, which performs strongly on multiple benchmarks, approaching or even surpassing Fable 5. API pricing remains low, but the chain-of-thought output has become terse, sparking discussion among users.
Google has unveiled SL2T, a breakthrough sign language-to-text model that powers new features for Deaf and hard of hearing users on Android. Starting with American Sign Language-to-English on Pixel 11, users can sign directly into Gboard and Live Transcribe instead of typing. The model tracks body poses on-device for privacy and was developed in collaboration with the Deaf community.
Meitu Imaging Research Institute (MT Lab) had 11 papers accepted at top international conferences in 2026, covering video editing, portrait editing, and more. They introduced innovative technologies like MiVE and CFT to improve consistency and controllability in AI image editing. These technologies have been integrated into products like Meitu Xiuxiu, pushing AI from 'generation' to 'creation'.
The Zidong Taichu team at the Chinese Academy of Sciences has proposed GMC, a core-set pruning method that reduces visual tokens by over 80% while preserving multimodal performance, addressing the bottleneck of token redundancy. The training-free method works with mainstream models like Qwen and LLaVA, matching or exceeding original performance on multiple benchmarks.
Microsoft has officially released the Agent Framework Harness and Foundry Hosted Agents, marking the framework's transition from SDK to a supported production runtime. The Harness, a single binary, runs across local, container, and hosted environments, with built-in features like function calling, persistence, context compression, tool approval, and OpenTelemetry. This release addresses the runtime question and highlights the Harness's importance in agent systems.
On August 12, Alibaba's Qwen AI platform officially launched Qwen AI Arena, a challenge and evaluation platform for AI agents, designed to generate tasks from real business scenarios and validate agents' ability to solve practical problems. The first task focuses on cross-border e-commerce product listing, requiring developers to build agents that generate multilingual and multimodal listing materials, with top solutions receiving token rewards and potential entry into Alibaba Cloud's ecosystem.
Discovered Materials, a YC P26 startup, launched on Hacker News with AI agents that discover new materials for the semiconductor industry. The founders highlight the heat problem in GPUs and the potential of AI to reduce the time and cost of introducing new materials into chips.
NoimosAI has released version 2.0 of its SEO Agent, which can conduct SEO research and retrieve data from tools like Semrush and Google Search Console. After user approval, it can apply changes directly to websites, creating a continuous SEO loop.
A new paper demonstrates how to decode and port encrypted reasoning traces from frontier reasoning models, significantly boosting open model performance, while warning that publicly shared sessions can leak sensitive data. The technique replays encrypted reasoning blocks into weaker models to induce transcription, revealing major security and alignment concerns.
ByteDance Seed team found that the training performance of text-to-image models depends not on caption length but on the amount of image-bound information within. They proposed Structured Prompt, which improves both diffusability and promptability, achieving significant gains on complex composition, reasoning, and world knowledge generation tasks. This research opens a new direction for scaling text-to-image models.
xAI's Grok 4.6 scores 61 points on the Artificial Analysis Intelligence Index, tying OpenAI's GPT-5.6 Sol and trailing only Anthropic's Claude Opus 5. On agentic tasks, it completes complex workflows in about 53 steps where Claude Opus 5 needs 103, at a price more than 60 percent lower.
Google announced that Gemini Notebooks now supports copying, allowing users to duplicate shared notebooks as their own, including all sources and artifacts. This feature aims to help students personalize study materials and teammates quickly start new projects.
Anthropic announced that Claude sessions in Chrome now carry over to desktop, web, and mobile, with conversations saved and skills/connectors working in the browser. The feature is available today for Max and Team users, with Pro users getting it in the coming weeks.
NVIDIA announced partnerships with major financial institutions including Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs, and KKR to establish independent financing platforms aimed at mobilizing over $500 billion for AI infrastructure. This marks a shift from project-based data center construction to financing AI factories as productive infrastructure assets, highlighting the growing investment appeal of AI compute.
MindTopo is a new benchmark that evaluates AI's understanding of topological relationships, highlighting new opportunities to strengthen spatial reasoning and planning. The benchmark aims to advance AI's spatial intelligence.
NVIDIA has launched Nemotron 3.5 Lightning and showcased how partners are post-training it for their specific domains. Partners including Prime Intellect, CrowdStrike, Dream, and CodeRabbit are leveraging the model to enhance performance in areas like cybersecurity and code review.
Google announced that Gemini can now connect to more third-party apps, including OpenTable, Ticketmaster, Wix, and 11 others. Users can use natural language commands to book reservations, build websites, stream music, and more directly within Gemini. This move aims to expand Gemini's ecosystem and enhance its utility as an AI assistant.
Researchers from Lund University and Google DeepMind propose BEFT, a bias-only fine-tuning method that trains only the value bias b(v), achieving near full fine-tuning performance with just 0.01% of parameters in low-data settings. The work has been accepted at ACL 2026 and integrated into the Hugging Face PEFT library, offering an efficient solution for resource-constrained fine-tuning.
Researchers at IIT Bombay and Adobe Research have developed an inverse language model that reconstructs the original prompt from an LLM's output with near-perfect accuracy. The method, called 'Previous-Token Prediction,' does not require access to model weights and works across different models. This poses a serious security risk for companies relying on proprietary system prompts.
Ji Yanli, General Manager of the Technology Center at Belle Fashion Group, detailed a three-step strategy for enterprise AI adoption, emphasizing the importance of business process digitalization, standardized decision-making, and closed-loop task management. The article argues that successful AI implementation starts with collaboration, gradually building AI-native capabilities that integrate AI into business operations.
Grok 4.6 achieved significant improvements on the AA-Briefcase benchmark, which tests long-horizon agentic knowledge work tasks. The model performs on par with Claude Fable 5 but at a much lower cost, at $4.42 per task compared to Claude Fable 5's $22.30. This release highlights xAI's competitive edge in both performance and cost efficiency.
Xie Pengtao, an associate professor at UC San Diego, and his team have launched AIBuildAI Science Agent, an AI for Science agent that ranked first on NatureBench, achieving or exceeding Nature paper results on 57.8% of tasks and surpassing human best methods on 25.6%. The agent can independently complete scientific research tasks within 4 hours, and the company has raised approximately $4 million.
SkyPilot, an open-source AI compute orchestration project from UC Berkeley, has spun off into a startup and raised $20 million in seed funding. The project addresses the fragmentation of GPU management across clouds and has gained wide adoption. Co-founders Zongheng Yang and Zhanghao Wu stated that commercialization will focus on organization-scale capabilities while keeping individual use open-source.
The MCP 2026-07-28 specification removes the initialize handshake and session header, adding required method and tool-name headers for stateless routing. Developers are split on whether this makes MCP just an API again or if the standard's value remains.
Microsoft has announced price cuts for its coding model to remain competitive. The upgraded model is now faster and more token-efficient, enabling users to complete tasks more quickly and at lower cost.
Tencent Cloud has officially open-sourced Cube Sandbox, an execution environment base for AI agents, evolved from its Serverless infrastructure. The article, through an interview with a Tencent Cloud expert, explores the new infrastructure demands of agents and how Cube provides an isolated, recoverable runtime.
GeekPark tested Genspark's AI-integrated office client, GenOffice Alpha. The product leverages AI to directly manipulate Office file formats, converting generated content into editable Word and PPT documents. While basic editing features still suffer from stability issues, its AI-powered formatting capabilities impressed, automatically recognizing document structures and rebuilding layouts. Released as open-source and free, the Alpha was built by one engineer in a week and will iterate rapidly based on user feedback.
On August 11, River AI, founded by xAI co-founder Igor Babuschkin, announced a $1.1 billion funding round led by General Catalyst and AMP PBC, with participation from NVIDIA, AMD Ventures, and others. The company aims to advance open-source models and personal AI, enabling businesses and individuals to own their AI, and has launched its first product, River API.
Zhang Hongliang, a young researcher at Fudan University, has made multiple breakthroughs in nuclear materials, including discovering radiation-induced segregation in ceramics, revealing a new deformation mechanism in intermetallic compounds, and developing a process that reduces corrosion rates in nuclear fuel cladding by over 40%. His work aims to address critical safety and performance issues in nuclear energy, providing material solutions for advanced systems like fusion reactors.
Spotify introduced an external indexing architecture for Apache Parquet data lakes that enables low-latency point queries without replicating datasets into operational databases. The approach maps lookup keys to Parquet files and row locations, allowing targeted reads from cloud object storage while supporting analytics, machine learning, AI applications, and online services from the same datasets.
The Quote/0 e-ink display has received a major update, allowing users to curate their information feed via the app and customize content using text, images, RSS, and more. The developer platform now offers text, image, and canvas APIs, and supports natural language commands through Dot Skill to let AI generate display content. Additionally, the duo bundle enables two devices to share to-do lists and send images to each other, enhancing interactivity.
SpaceXAI has introduced Grok Bot, a new AI teammate that can log into users' tools and websites, operate software interfaces like a human, and deliver finished work directly. Grok Bot supports asynchronous 24/7 operation, parallel scaling, and can learn workflows through demonstration for automation. The product is currently in early beta, aiming to change how users interact with AI assistants.
Brad Lightcap, OpenAI's COO, announced his departure to pursue a new venture, marking the latest in a series of executive exits as the company prepares for an IPO. Meanwhile, OpenAI released a preview of its ChatGPT desktop app for Linux, extending support to major distributions and completing coverage across all major desktop operating systems.
Manus announced it will resume operating as an independent company and, due to regulatory requirements, some users will need to back up their data by a specified deadline. The company will provide a restoration portal and welcome-back bonuses, and teased a series of new features.
Anthropic announced that an unreleased AI model made significant progress on the Riemann hypothesis, a math problem unsolved for over 150 years. The model autonomously worked for about a day and a half, testing 650 approaches and coordinating 60 subagents, with two subagents making the key breakthroughs. The findings were formalized using the Lean proof assistant and have sparked debate within the mathematical community about AI's role in mathematics.
AI startups like Thinking Machines Lab and Periodic Labs are launching external research grant programs, directly providing funding and compute to academic researchers. This reflects venture capital's trust in AI scientists and a new trend of AI companies reallocating research resources.
Mistral is giving customers the option to route AI requests through servers in either Europe or the US, and selling priority queue access during peak traffic. Both come with a surcharge, and the regional routing doesn't cover all features or data. This move aims to address European data processing needs but has limitations.
WorkBuddy's product manager Kong Deyuan has released a comprehensive 30,000-word guide covering everything from installation to advanced usage, aiming to help users fully master this AI office tool. The guide details three working modes (Ask, Craft, Plan) and emphasizes WorkBuddy's capability as an AI agent to directly manipulate files, generate reports, and more.
Yuan Dian Tech (SCALEFORCE), an embodied intelligence data infrastructure company, has completed a funding round of tens of millions of RMB, with investors including top domestic embodied intelligence industry players, Hengxu Capital, and Kailian Capital. The funds will be used for the iteration of its MatrixOS technology, construction of large-scale data production networks, and team expansion. The company aims to address data quality, scale, and generalization challenges in embodied AI, accelerating the development of physical AI.
At a data storage user forum in Wuxi, Huawei proposed that as AI inference scales, storage must evolve from traditional persistence devices into the inference path, and introduced the industry's first context memory storage (CMS) for large-scale inference, a G3.5 storage tier. Huawei emphasized a full-stack approach to integrate data, compute, models, and agents, aiming to advance AI adoption.
A public benchmark by Composio shows that Pi Harness with DeepSeek V4 Flash achieved the highest success rate (66.7%) among 8 agent frameworks, at a cost seven times lower than Claude Code. Earlier, developer 0xEvan processed nearly 1 billion tokens with Pi and DeepSeek V4 Flash, hitting a 99.93% cache hit rate for just $2.65. Pi founder Mario Zechner noted he had endorsed this combination back in May.
A new report from MIT Technology Review highlights that while enterprises are rapidly adopting agentic AI, legacy data systems are major blockers to realizing ROI. Based on a survey of 300 data and technology executives, it finds that only a few 'data leaders' provide ample data access to AI agents, leading to better outcomes. The report urges organizations to modernize data infrastructure to enable agents to make real-time decisions and take actions.
This article explores how businesses can track rivals' model releases, pricing changes, and product updates for competitive intelligence in the AI era. It cites the Institute of AI Product Management and Tiger Tail, highlighting AI tools' ability to catch patterns humans miss, and argues that competitive intelligence is now a core competency in AI-driven business strategy.
June, an AI enterprise platform founded by former Salesforce executives, has come out of stealth with $20 million in pre-seed funding led by Marc Benioff's Time Ventures. The company aims to help businesses implement AI agents more easily by scanning existing systems and building optimized workflows. The funding round also includes investments from Michael Dell, Aaron Levie, and George Kurtz.
FEAGINE Technology unveiled three bionic flexible robots, FEAGINE A01, A02, and A03, along with the first-generation cross-embodiment foundation model Fi0, aiming to advance embodied intelligence through redesigned bodies. The company argues that generality should come from shared intelligence rather than a unified humanoid form, and has achieved standardized mass production of tendon-driven flexible bodies.
Westlake Robotics, a general brain company for embodied intelligence, has completed a Series A funding round, bringing total funding to 500 million yuan within six months. Investors include Sequoia, Xiaomiao Langcheng, and others. The funds will be used for R&D of a unified large model for humanoid robots and talent training. The company has partnered with Alibaba, Intel, JD.com, and HKUST, and secured orders worth nearly 100 million yuan.
OpenAI has seen a wave of executive departures, including former COO Brad Lightcap, safety systems head Johannes Heidecke, chief futurist Joshua Achiam, and ethics lead Chloé Bakalar. The exits come just before a potential IPO, raising concerns about the company's stability and commitment to safety. Analysts suggest the departures reflect internal tensions between safety and commercialization.
Microsoft has released MAI Code 1.1 Flash, a code model for GitHub Copilot that is claimed to be 25% more token-efficient at a quarter of the cost of its predecessor. However, in benchmarks, it is outperformed by the cheaper DeepSeek V4 Flash on both price and performance. This move fits a pattern where Microsoft promotes open AI while integrating inferior proprietary models into its apps to protect margins.
IBM has signed a $240 million deal with Together AI to provide an Nvidia-powered AI cluster, responding to the growing demand for open models. The partnership highlights the increasing investment in AI infrastructure.
Google legend Jeff Dean detailed his last 48 hours at the company during KDD 2026, including announcing his departure to 139 close colleagues, attending a farewell event with 1,500 people, and officially leaving on August 6. He then became co-founder and CEO of Discovery Loop, a new venture with Sanjay Ghemawat and others to explore AI-accelerated research. This marks a significant shift in AI leadership and signals new directions for research.
OpenAI has released a Linux version of its ChatGPT desktop application, expanding the platform coverage of its desktop client. This move allows Linux users to access ChatGPT more conveniently without relying on a web browser.
Lin Junyang, former head of Alibaba's Qwen large model series, has officially announced his new startup Pragmatik Labs, which focuses on building next-generation AI agents that span digital and physical worlds. The company has already secured funding from investors including Gaorong Ventures, Sequoia China, and Tencent, with a valuation of $2 billion. Lin aims to shift AI from reasoning-centric to agentic thinking, emphasizing productivity and long-term scientific progress.
Lin Junyang, former head of Alibaba's Qwen technology, has officially launched his startup Pragmatik Labs, valued at $2 billion. The company focuses on next-generation agents spanning digital and physical worlds, with backing from GSR Ventures and Sequoia Capital China. This move signals a shift from foundation models to agentic AI, though it faces challenges in hardware-software integration and long development cycles.