AI News

America needs to stop getting shocked by Chinese AI

Baidu Group Vice President Hou Zhenyu proposed at the WAIC forum that the relationship between AI and energy is shifting from "power consumption" to "symbiosis". Baidu AI Cloud has launched a two-way framework of "using intelligence to strengthen energy" and "using energy to empower intelligence": the former injects large model capabilities into data center power management, participating in the development of the State Grid's "Guangming Large Model"; the latter relies on Kunlun Core's 10,000-card cluster and self-developed 800-volt DC power supply system to improve computing power energy efficiency. Hou Zhenyu envisions that in the future, "power tokenization for export" may become a reality.

The Xiaohongshu dots team participated in the 67th IMO 2026 with their internal version dots-note 3.0, achieving full marks on all six problems, scoring 42/42 for a perfect gold medal. Only 7 human contestants worldwide achieved this result. The model does not rely on formal languages; it directly reads raw LaTeX problems and solves them end-to-end through recursive self-critique capabilities. dots-note 3.0 is the lightest model in the dots3 series and is expected to be open-sourced.

Shengshu Technology, together with Wondershare Technology, held an AI film and TV special forum at WAIC 2026, focusing on the underlying technological innovation of general world models to explore the new paradigm of AI film and TV productivity and the future of the next-generation industry.

This warm little world is so adorable, it's hard to leave!

Aravind Srinivas on why China's open-source AI may become more powerful than ever. And why Anthropi…

The 2026 World Artificial Intelligence Conference Youth Outstanding Paper Award was announced, with Tsinghua University Assistant Professor Yao Yuan and Postdoctoral Fellow Qian Chen winning respectively. Yao Yuan's paper focuses on end-side multimodal large models, with MiniCPM-V ranking first on HuggingFace and other leaderboards for multiple consecutive days; Qian Chen's paper "ChatDev" pioneered a multi-agent collaboration framework, garnering over 30,000 stars on GitHub.

Zeng Guoyang, co-founder and CTO of ModelBest, was honored as "AI Person of the Year" at the 2026 China AI Gala, recognizing his technological breakthroughs and industrial contributions in the field of on-device large models. Zeng previously served as a core engineering lead in the development of China's first large language model, CPM-1, and later led ModelBest to propose the concept of "intelligence density" and create the MiniCPM series of on-device large models. Also selected during the same period were 10 AI figures including Wang Xingxing of Unitree Robotics and Wang He of Galaxy General, while Cai Lei was awarded "AI Special Contribution Person of the Year."

China's open-weight models are starting to threaten the business model behind OpenAI and Anthropic. …

According to Caixin, Alibaba is about to launch "Tongyi Qianwen Office," a unified AI agent platform targeting the enterprise productivity market. The product integrates three existing products: QoderWork, Wukong, and MuleRun, and is led by Chen Yusen, the new CEO of DingTalk who took over in June. This move comes as Tencent has just launched its enterprise desktop AI agent WorkBuddy (supporting MCP and 20+ skill packs), prompting Alibaba to accelerate integration to compete.

Last week, all top 5 most-used AI models globally came from China, maintaining the streak for 12 consecutive weeks. Tencent Hy3 led with 11.8T tokens (up 58% weekly), followed by Xiaomi MiMo-V2.5 with 9.37T tokens (up 43%). Among US models, GPT-5.5 ranked only 14th (down 25% weekly), and Anthropic's three models entered the top 20 but all declined.

Alibaba Tongyi Qianwen launches the third-generation foundational image generation model Qwen-Image-3.0, with core capabilities focused on "realism." It supports instruction input of up to 4.5k tokens and can generate a 3×3 grid layout containing 9 complex infographics in one go. The model can accurately render text as small as 10px and natively supports 12 languages. It can simulate mainstream interfaces such as web pages, games, and live streams, maintaining readability in dense layout scenarios like academic papers and newspapers.

BlackRock CEO Larry Fink on why china is ahead in the AI energy race. "We don't have the ability to…

Fred Turner, founder of US health insurance company Curative, agrees with the "SaaS doomsday" theory, stating that the company has terminated its $600,000 annual Salesforce contract and used AI programming to build its own CRM system in two months. Curative plans to cut about 80% of its SaaS spending this year and shift to AI. Its custom AI agent, Gwen, has reduced the average cost of contract negotiations from $1,500-2,000 to about $70.

We are excited to announce our partnership with Win2tec Sport! Combining Win2tec's expertise in the global sports digital ecosystem with Alibaba Cloud's AI and cloud technologies, we are accelerating the digital transformation of the sports industry. Together, we will deliver smarter, more connected, and more secure experiences for athletes, officials, partners, media, and fans. Learn more about our partnership below. #AlibabaCloud #SportsTech #AI #CloudComputing #DigitalTransformation

The social media Skill developed by Guizang has been frequently used on Xiaohongshu and Xiaolvshu for generating images, showing good performance data. This Skill is suitable for users who need to quickly generate images but lack beautiful materials.

"As I said a few years ago, humans are increasingly becoming the biological bootloader for digital superintelligence." - Elon Musk

🎖️ Excellence Award at the AI Video Hackathon Tokyo A compelling PR film for a new venture-crafted…

Founded by Carnegie Mellon University robotics experts, Gritt has completed a $26M Series A round, bringing total funding to $34M. The company uses off-the-shelf robotic arms and proprietary AI models to handle solar panels with sub-millimeter precision at outdoor construction sites, boosting daily installation per eight-person crew from 800 to 3,000-4,000 panels. Gritt has signed contracts to assist in installing 2.8 GW of solar panels, with clients including three of the top ten U.S. electrical construction firms.

Bristol Myers Squibb buys Nvidia AI system for drug discovery

Alibaba Cloud's AI video model HappyHorse 1.1 won the Excellence Award at the Tokyo AI Video Hackathon. The award-winning work is a 15-second short film created overnight, telling a story of pursuing dreams from Aomori to Tokyo, proving that AI can convey emotions rather than just show off skills. The model is now available for trial.

Cola today launched a new model, July, positioned as "second only to Fable," offering free access to members for a limited time. Due to high member usage causing queues, the official recommends off-peak use. The model's thinking depth and response length are highly similar to Fable, considered at least Opus-level.

OpenAI is placing ads in ChatGPT, vying for the advertising market dominated by Google and Meta, with its advertising business already reaching an annual recurring revenue (ARR) of $100 million. OpenAI projects advertising revenue of approximately $2.4 billion in 2026 and about $100 billion by 2030, but EMARKETER analysts consider these targets nearly impossible to achieve. Currently, ChatGPT's ad products are still in early stages, available in only 7 countries globally, lacking comprehensive ad creation, performance measurement, and data reporting tools.

We're thrilled to announce the successful conclusion of our AI Video Hackathon in Tokyo. Over 50 cre…

Quick question: Is it just me, or are you also experiencing the same issues with the ChatGPT Classic Mac app? When I use the ChatGPT Classic app for simple tasks, I can't adjust the reasoning effort; it's always auto-selected, and I have to switch to the web version. Also, the sidebar always hides immediately. Although I mostly use Codex, I find the Mac Classic app really not user-friendly.

Several people with no technical background told me they were reviewing for NeurIPS using agents for…

ChatGPT is offering another $100 credit. Users simply need to submit a shared X or LinkedIn link along with their ChatGPT account address to claim it. Claim link: https://share-chatgpt-work.openai.chatgpt.site/

Xiaomi-Robotics-1 shows that more data beats bigger models when training robots to move

Meta has open-sourced Astryx, a React and StyleX design system that has been running internally for eight years and powers over 13,000 apps. It offers 150+ accessible components, seven themes, dark mode, templates, and an agent-ready CLI, licensed under MIT and requiring React 19+.

The thinking length of qwen3.8 max preview is really insane. It's been thinking for almost 10 minutes on a single sentence and hasn't started formal output yet... Another friend said their task took 30 minutes of thinking. The output is also super long. Can't use it without patience?

Shin Jin-seo 9-dan, the world's top-ranked Go player from South Korea, defeated the Go AI KataGo by an 11.5-point margin in the decisive game of the "Jingrui Mathematics·Hankyung Kihoon Battle," winning the series 2-1. In the final game, Shin Jin-seo played black with a two-stone handicap, maintaining a 99% win rate throughout and securing victory in 3 hours and 20 minutes. With a record of 2 wins and 1 loss, Shin Jin-seo received a prize of 250 million Korean won (approximately 1.14 million RMB) and a Genesis G90 sedan.

And Dario Amodei always had such a hardline view that China shouldn't have strong AI. "That's the n…

Alibaba, in collaboration with the Student Innovation Center of Shanghai Jiao Tong University, has launched the "Qianwen AI Creative Classroom" summer AI general education course on the Qianwen App's "Qianwen Little Lecture Hall", free for students nationwide. The course is designed for lower, middle, and upper primary school grades with three progressive themes: "My New AI Friend", "My Super Learning Partner", and "Creator in the Intelligent Era", covering dimensions such as conversation, creation, and exploration. The course will later be introduced to more primary and secondary schools as a public welfare initiative.

Alibaba's Qwen3.8-max-Preview can be tested online directly via a web page without downloading a client. Users report that its quality improves daily, and its coding ability even surpasses K3. However, the web version only supports text and simple front-end tests; more reliable evaluations require the API.

Automotive media Electrek editor Fred Lambert was pulled over last weekend while using Tesla FSD Standard mode, receiving a speeding ticket for driving at 78 km/h on a road with a 50 km/h limit. Before FSD v14, users could set a speed offset above the limit (e.g., +10 km/h), but this feature was removed last October. It was replaced by five fixed driving modes: Sloth, Chill, Standard, Hurry, and Mad Max. Users are now forced to choose between conservative mode and a mode that frequently speeds, losing the ability to customize.

Yao Jingjing's team has open-sourced the GEO raw datasets from 8 domestic AI platforms including Doubao and DeepSeek, containing 620 standard questions and 189,845 deduplicated citation records. Analysis shows that the top 10 sources contribute 41.47% of citations, with Douyin and Tencent News ranking first and second; the highest cross-platform Jaccard similarity is only 34.94%, and 88.77% of pages are from 2025 or 2026.

Script in. Video out. 📅 July 23, 2026 | 10:00-10:30 AM (UTC+8) | Alibaba Cloud Live Register → https://int.alibabacloud.com/m/1000415408/ WonderClip is enabling AI-driven video production at enterprise scale—fully automated from script to final video. Learn how it works in #AgenticTalks Episode 1. #AlibabaCloud #AIVideo #AIAgent #WonderClip

OpenAI appears to be well on its way toward autonomous AI researchers. An internal model reportedly…

At the World Artificial Intelligence Conference, Tencent Cloud executives stated that to drive inference costs to the extreme, the company will deploy domestic computing power at scale and plans to deploy NPO (Near-Package Optics) supernodes in Q4 2026, while calling for unified NPO industry standards both domestically and internationally. Wang Yachen, Vice President of Tencent Cloud, pointed out that NPO is a more practical technical path for building supernodes with domestic GPUs. Shen Yichen, founder of Xizhi Technology, believes that NPO solutions are on the verge of commercialization, characterized by placing optoelectronic conversion chips directly on the same board as computing chips, thereby eliminating the most expensive DSP chips in optical modules.

Chinese open-weight models are cheap. Washington is deciding what that costs.

"Open-weight models are inherently safe because when you download a model from Hugging Face, the whole world can disassemble it, inspect it, fine-tune it, modify it, and scrutinize it in ways that closed models absolutely cannot." - Sriram Krishnan

At the 2026 Melbourne AI Engineering and Infrastructure Summit, Alibaba Cloud showcased Qwen, WonderClip, and MuleRun—supporting organizations in Australia and New Zealand to accelerate AI innovation. #AlibabaCloud #AInnovation #Qwen #WonderClip #MuleRun

Bristol-Myers Squibb (BMS) announced the purchase of a DGX SuperPOD based on NVIDIA's Vera Rubin architecture, becoming the first company in the life sciences industry to adopt this system. The company has already used AI to reduce clinical trial drug preparation time by 20%-30%, with expectations to expand to 50% in the future. The new system offers approximately 10x improvement in energy efficiency and will support simultaneous evaluation of dozens of drug candidates, accelerating both small molecule and large molecule drug R&D.

NVIDIA has released Cosmos 3 Edge, a 4 billion parameter open world model designed for on-device operation. The model helps robots and visual AI agents understand their environment, perform real-time inference, and generate robot actions locally. The Cosmos 3 series previously launched Cosmos 3 Nano (16 billion parameters) and Cosmos 3 Super (64 billion parameters) at GTC Taipei on May 31, 2026.

Tongyi releases Qwen-Audio-3.0-TTS voice model, which can generate speech across 16 languages and 20 dialects from a single reference audio while preserving the speaker's timbre. The first packet latency is as low as 300 milliseconds. In the CV3-Eval benchmark, the model ranks first in speaker similarity across all 16 languages, with the Plus version achieving an average score of 82.75 out of 100.

NVIDIA has released the synthetic video detector NIM service, which can analyze videos frame by frame and provide classification scores to determine whether the content is AI-generated. Internal tests show that the tool achieves 92% detection accuracy on uncompressed videos, 85% at 15% compression, and 82% at 50% compression. On an RTX GPU system, analyzing a 1080P video takes as little as 22 milliseconds, while on an enterprise L40 GPU it takes about 30 milliseconds.

To make AIHOT's clustering more accurate, I have to do the labeling myself... All work goes down to the bottom, and in the end, it's all about labeling for the Agent...

Unity Technologies launched Unity 7 engine in Seoul, South Korea, featuring "no breaking changes" and direct inheritance from Unity 6 architecture. The new engine achieves near-instant Play Mode startup based on CoreCLR, with shader compilation speed improved by up to 90%. Unity 7 will begin early Beta testing in December this year, with a planned official release in Q1 2027.

After Grok 4.5, the new 2T-parameter model is even more anticipated! SpaceX acquired Cursor, integrating computing power and data for model training—a win-win acquisition that maximizes the use of Cursor's accumulated engineering data. Now the 2T model will also include SpaceX's world-class engineering data, which is something neither OpenAI nor Anthropic possesses.

Rohan Paul points out that frontier labs are seeking protection under the guise of the 'China threat,' which is actually to maintain their business model. The model layer is only a small part of the AI economy; if inference costs drop by 10-100 times, it will inject rocket fuel into the downstream application layer. Citing @chamath, protecting the equity of OpenAI and Anthropic, which is held by only 5,000 people, is less important than letting the market decide; price cuts will expand the entire AI industry pie.

Larry Ellison points out that AI models are becoming commoditized due to using the same public data, and the real competitive moat is no longer the model itself but exclusive proprietary datasets.

So Grok's next 2T version will gain a massive engineering upgrade from SpaceX's specialized technica…

Honor Robot Phone's first-day pre-orders have exceeded all previous Honor flagship products. The device is powered by the Snapdragon 8 Gen 5 chip and integrates the industry's smallest 4-degree-of-freedom (4DoF) titanium alloy mechanical gimbal system on the top, with a volume 70% smaller than mainstream solutions. The core components, motors, and modules of the 200Mp gimbal camera are all self-developed. Industry forecasts suggest the 1TB version will be priced at approximately 15,999 yuan, expected to launch in August this year, and will debut Honor's next-generation partner-type multimodal intelligent operating system kernel, AgenticOS.

Huang Zhenxin, head of Dark Side of the Moon's B-end business, revealed that user demand for K3 has far exceeded expectations after its launch, leading to tight resources across the platform and restrictions on new user registrations. The team is prioritizing the user experience of existing paying users and is not fully opening new user registrations for now. They are gradually alleviating pressure through model efficiency optimization and additional computing power. Kimi K3 was released on July 16, with 2.8 trillion parameters and a context of 100 million tokens, making it the most capable model from Kimi to date.

Google DeepMind has released the GenCeption model, which repurposes a pre-trained video generator into a single model capable of simultaneously performing core vision tasks such as depth estimation and image segmentation. The model is trained on Alibaba's Wan2.1 and outputs results in a single forward pass, with the small model processing 81-frame videos in about 6 seconds.

OpenAI and Hugging Face partner to address security incident during model evaluation

Alibaba released the Qwen-Image-3.0 image generation model, with the core theme being "realistic". The model supports ultra-long input of 4.5k tokens, can generate knowledge diagrams containing formulas and geometric shapes, complex UI, and supports native rendering in 12 languages and over 20 fonts. Alibaba Cloud Bailian and Qwen AI platforms have opened API for beta testing, and Qwen Studio and Qwen APP will soon launch free trials.

http://x.com/i/article/2079405623293173760

Moonshot AI is about to launch the Kimi Hosted Agent platform, offering standardized APIs such as PPT generation for ToB customers. Currently, API calls account for 70% of B-end revenue, forming a sustainable positive cycle. The company released the Kimi K3 model with 2.8 trillion parameters on July 16.

On July 16, YouTube updated its policy, categorizing "non-authentic content" into three types and banning their monetization: generic repetitive templated content, objectionable or distressing content, and content using AI avatars to discuss sensitive topics like health and finance. Channels that publish large amounts of such content will be ineligible for the YouTube Partner Program (YPP), losing monetization eligibility.

Tesla has resumed the rollout of FSD v14 Lite for HW3 vehicles, bringing software version 2026.20.6.10. New features include no longer needing to press the brake to start FSD from parking, upgraded destination parking functionality, a continuous driving achievement system, a standalone autonomous driving app, and the ability to view FSD status via the mobile app, making the Lite experience closer to the full FSD v14. This version is currently only available to Early Access users in the US, with a broader rollout expected soon.

OpenAI is further expanding the ChatGPT parent notification feature. If a teenager is banned for violating policies on violent threats or cyberbullying, parents who have linked their accounts will receive an alert. This feature was developed in collaboration with the cyberbullying monitoring organization Moonshot. Additionally, a "learning mode" toggle has been added to the parental control page. When enabled, ChatGPT will first provide hints instead of complete answers, and will push rest reminders to teenagers who use it for extended periods.

Berry Xia launched a prize guessing game, asking to match four Boeing 747 model images with four models: Qwen 3.8-Max-Preview, Kimi 3, GPT-5.6-Sol, and Fable 5. The main post claims that it was shared in over 10 WeChat groups, and so far no one has guessed the correct answer.

We may already be entering an AI Cold War. Axios reports that the U.S. Commerce Department considere…

According to South Korea's customs data, exports in the first 20 days of July (adjusted for working day differences) surged 62.9% year-on-year, setting a new record for July exports. Global AI and data center construction are fueling strong semiconductor exports, with South Korea's chip exports rising 180.6% year-on-year during the same period, and computer-related product shipments increasing nearly 232%.

Japanese AI company Sakana AI has released the Fugu Cyber model, specifically designed for cyber defense. The model achieved a success rate of 86.9% on the CyberGym benchmark and 72.1% on the CTI-REALM test, outperforming OpenAI's GPT-5.5-Cyber and Anthropic's Claude Mythos-Preview. Fugu Cyber is a multi-agent system encapsulated as a single API.

User Berry Xia used the Qwen 3.8-Max Preview model to create a 3D slide projector that supports local upload of PDF, MP4, PPT and other files, enabling animation interaction and detailed depiction. Previously, the Qwen team released Qwen 3.8-Max-Preview, an open-weight model with 2.4 trillion parameters, and the preview version is now available on Token Plan and Qoder platforms for regular users to try.

Grok is a solid workhorse

Breaking: Elon Musk announces that a vast amount of world-class engineering data from SpaceX will be used to further train the upcoming 2T Grok model. This will significantly enhance Grok's engineering skills. No other AI company in the world has access to the real rocket, spacecraft, and manufacturing knowledge that SpaceX possesses. This will give Grok a major advantage over competing AI models in the engineering domain.

Tongyi Qianwen releases its third-generation image generation foundation model Qwen-Image-3.0, with the core keyword being "real". The model supports up to 4.5k token instruction input and can generate a 3×3 grid layout containing 9 complex infographics in a single pass; text rendering accuracy reaches 10px, and it supports native rendering in 12 languages, aiming to transform images into deployable productivity tools.

🇨🇳 wow. China is drafting rules to stop its most advanced AI and chip designs from reaching the We…

Nikkei research shows that the hidden debt of five US tech giants, including Meta and Oracle, has ballooned eightfold in about four years to approximately $1.65 trillion, surpassing their actual debt. This debt mainly comes from data center leases and GPU supply contracts, with Meta's off-balance-sheet debt at about $420 billion, nearly three times its transparent debt. The surge in hidden debt makes it harder for investors to assess risk.

NVIDIA CEO Jensen Huang on China. "The world doesn't realize is how dependent the AI industry is on…

Berry Xia launches a prize quiz, showcasing Boeing 747 model images created by four AI models (Qwen 3.8-Max-Preview, Kimi 3, GPT-5.6-Sol, Fable 5), asking to guess the corresponding model names in image order. The first correct guess wins 8.88 yuan, and the full demo video and prompts will be released later.

Neill Blomkamp, director of 'District 9', has released his first AI short film 'Nightborne', approximately 13 minutes long, with every frame generated frame-by-frame by ByteDance's Seedance 2.0 model from text prompts. The film adopts a documentary style and only licensed the likenesses and voices of 32 real actors. Blomkamp plans to shoot a feature-length film using the same method and has founded an AI film studio, Barley Studios.

I'm so out of the loop, what's a 'body-mind-spirit AI company'? 😂

unironically this is happening right tf now

And now from the Chinese government side. This would be a good time for cooperation between the US …

Momenta Robotaxi has officially obtained the Shenzhen intelligent connected vehicle road test permit and will conduct field tests in Shenzhen soon. Its R7 reinforcement learning world model has been deployed in L4 autonomous driving applications, with the SAIC Volkswagen ID.ERA 9X being the first production model equipped with this model.

http://x.com/i/article/2079432918024503296

Zhu Jun, founder of Shengshu Technology, proposed at WAIC 2026 that a general world model needs to integrate understanding, imagination, and action capabilities, and released the world's first MoT unified architecture. This architecture unifies understanding, generation, and action experts into a single model, supporting video generation models like Vidu Q3 and embodied intelligent robot control. The latest model can drive various heterogeneous robots to complete long-horizon tasks, achieving industry-leading results in cross-scenario generalization.

Elon Musk tweeted recommending the Grok Build command-line tool. According to the quoted tweet, Grok 4.5 ranks first on the Long-Horizon Terminal-Bench with a binary pass rate, surpassing Claude Fable 5, Claude Opus 4.8, and GPT-5.6-sol. Under the strictest scoring criteria, Grok 4.5 excels in complex agent tasks requiring hundreds of steps for complete workflows.

Generative World Renderer at the Speed of Play

OpenAI launches a $100 free credit campaign for ChatGPT Work users, valid for the first 10,000 applicants. To apply, post an English tweet on X with #ChatGPTWork, describing a specific use case and efficiency improvement, then copy the tweet link and submit via the official application page. Using an account with real activity and posting in English can increase approval chances.

To popularize AI knowledge, I wrote a video generation Skill. I plan to generate a large model knowledge interpretation video every day. Today is the first episode, explaining "distillation" in the simplest terms. From now on, never say I distilled a Skill, as it would sound very amateurish.

XPeng Group has released the TuringViT vision encoder for the VLM/VLA era, offering two specifications: 18L and 24L. At 1536×1536 resolution, the encoder achieves 3.04 times the inference throughput of Seed1.5-ViT. Using only 0.85B image-text pairs, it attains an average accuracy of 83.6% on six zero-shot classification benchmarks, surpassing baselines trained on 10B data.

Mixue Bingcheng has integrated into the Qianwen APP in the form of a Skill, allowing users to select products, place orders, make payments, and pick up in-store directly within Qianwen. Qianwen has already connected with multiple catering brands including Mixue Bingcheng, KFC, and Luckin Coffee. Users can enjoy discounts for in-store pickup by entering the 'Summer Exclusive' section on the homepage to place orders.

A paper reveals that the Gated Delta Networks architecture, based on the Mamba idea, is the core reason why Nvidia's Nemotron, Kimi's Delta Attention, and Qwen's latest models adopt similar hybrid architectures, achieving parameter scales exceeding T with excellent quality. This architecture is becoming a key technology for scaling up large model parameters and improving quality.

ELON MUSK: "Grok has been doing quite well at SpaceX and Tesla. We see Grok being very helpful in customer service and other areas, and this AI has infinite patience, so you can yell at it and it will still be very friendly."

Launched Hyra-1.0, the first version of the Hunyuan research agent. 💡💡💡 Built for performance-driven research and engineering tasks, with recursive improvement of solutions. Explore our demos in AI4AI, AI4Science, and AI4Fun: https://hy.tencent.com/research/hyra

Special thanks to the @GoogleAIStudio team today ♥️ The dedication, ambition, and passion shown by our team bring a smile to my face every day. Let's go!!

A Kimi researcher was shocked by the computing resources available to ordinary OpenAI researchers. Current model rankings show: 1 Fable5 2 GPT 5.6 sol 3 Kimi K3 4 Grok 4.5 5 GLM 5.2. Among them, GLM 5.2 has about 750B parameters; if increased to 1T~10T, the performance might be even better.

UK-based AI-driven new materials discovery startup CuspAI announced the launch of the "AI Materials Foundry" initiative, aiming to accelerate materials discovery and synthesis in fields such as semiconductors and clean energy through its MIRA platform. Founding partners include over 20 companies such as NVIDIA, Meta, Samsung, AMD, Hyundai Motor, and multiple academic institutions.

An interactive web-based tour using Gaussian Splatting technology that allows users to immersively explore the interior of Grace Cathedral in San Francisco. The project features scene capture and reconstruction by Vincent Woo, built on the PlayCanvas engine by Donovan Hutchence, with early prototype support from World Labs.

Tencent Hunyuan launches Hyra-1.0, an agent designed for research and engineering tasks that can recursively self-improve. In multi-round 3D modeling tests, the models generated by Hyra are closer to reference images and more aligned with human aesthetics compared to Claude Code goal mode.

very notable trajectory comparison writeup here buried in the RLM paper from @a1zhang and @lateinter…

On July 19, EY and Volcano Engine signed a strategic cooperation memorandum to jointly build AI solutions around core enterprise business scenarios, embedding capabilities such as Doubao Large Model, Data Agent, and HiAgent into data governance, financial management, and marketing growth. The collaboration also includes establishing a thousand-person FDE (Front-end Delivery Engineer) team and gradually introducing AI tool platforms like TRAE into EY's professional service teams to create AI-native delivery teams.

Claude Fable 5, GPT 5.6 Sol, Kimi K3, and Axiom Math all scored perfect 42/42 at IMO 2026. Among them, Fable 5 solved all problems in one go and was the fastest, Sol had the lowest cost, and Axiom Math completed formal proofs using Lean. Fable and Sol took less than 4 hours, far shorter than the human contestants' 9 hours.

Tencent Hunyuan launches Hyra-1.0, a recursive self-improving research agent that surpasses Recursive's public results on three tasks including NanoChat. Hyra refreshes 29 historical best results out of 55 open math problems and designs a Transformer with only 15 trainable parameters capable of 10-digit addition. All outputs are open-sourced on GitHub.

Sony Music sued music generation AI company Udio in a U.S. court this Monday, accusing it of copying over 30,000 recordings without permission to train its model. Sony stated that these 30,000 recordings are just a small subset of the hundreds of thousands identified, and refuted Udio's previous "fair use" defense, pointing out that a paid licensing market genuinely exists. Udio had previously reached settlements with Universal Music and Warner Music, but Sony Music has continued to escalate litigation since its initial lawsuit in June 2024.

China Telecom has taken the lead in completing the pilot application of a large model for 5G wireless network planning in live networks, achieving a 50% improvement in planning efficiency and over 75% accuracy in scheme generation. The model enables automatic output of design plans and precise prediction of construction effects. Utilizing a "dual-twin large model collaborative planning" technical architecture, combined with automatic optimization of channel and traffic twin models, it has been validated in Shanghai residential areas and underground parking lots, with site supplementation accuracy exceeding 80% and coverage problem identification accuracy surpassing 75%.

Alibaba will launch 'Qianwen Office' targeting the Agent office market, integrating three agent products: QoderWork, Wukong, and MuleRun, led by the new DingTalk CEO Chen Yusen. Qianwen Office will be built on QoderWork, which has the largest user base and best market reputation. The main challenges include deeply integrating Qianwen Office with DingTalk and coordinating the teams and resources behind the three products.

Apple presents Environment-free Synthetic Data Generation for API-Calling Agents https://huggingface...

Qwen-Audio-3.0-TTS is here. 🎙️ Our latest text-to-speech model, available in two versions: • Flash: Real-time interaction • Plus: High-quality generation New features: • Multilingual support covering 16 languages • Natural language style control • Fine-grained labeling of non-verbal details • More robust voice cloning from imperfect audio 🔗 https://int.alibabacloud.com/m/1000412420/

Gartner predicts global end-user spending on AI models and platforms will reach $64.252 billion in 2026, a year-over-year increase of 63.4%. Among them, investment in generative AI models will surge 117%, while AI platform spending will grow 36.9%. Spending is shifting toward vendors that demonstrate clear value in cost, latency, performance, and reliability.

Kimi K3 has reached the top of the 'Frontend Web App' competition! It feels like Kimi K3 has really improved in frontend design this time. I still don't have an answer as to which is better between it and Fable 5, but looking at @DesignArena's competition leaderboard, it has already surpassed all Claude models, including Fable 5! Another thing that matches my intuition: GPT-5.6 Sol's frontend design ability is still not good, as seen from the leaderboard, falling out of the top ten.

🚀 Tired of scattered Agent skills and version chaos? Nacos AI Registry provides a single source of truth. ✅ Centralized management across Codex/Cursor. ✅ Enforce security reviews and permissions. ✅ Support version control and rollback. Stop manual sync; start governing AI assets. 🛠️ https://int.alibabacloud.com/m/1000415669/ #Nacos #AIGovernance #AgentSkills

The Trump administration is attempting to reduce the adoption rate of China's top AI models among US companies through "soft blockade" measures such as procurement rules, entity lists, and public opinion suppression. Last year, it planned to add multiple Chinese AI labs and university labs to the entity list. This report combines unsubstantiated rumors with information previously disclosed by OpenAI's new head of strategic analysis.

🚀 Developers in Ho Chi Minh City – join us this Friday to kick off the Alibaba Cloud Agent AI Hackathon, powered by Qoder @qoder_ai_ide. Challenge theme: Financial Services. Over $3,000 in cash, credits, and Qoder Pro rewards. 📅 July 24 | ⏰ 14:30-17:00 | 📍 Ho Chi Minh City 👉 https://luma.com/sw5m9n8q #Qoder #AlibabaCloud #AIHackathon

Anthropic mathematicians, using the Fable 5 model, casually proved a counterexample to the Jacobian conjecture, which had puzzled academia for 87 years and was listed among the 18 major mathematical problems of the 21st century. Subsequently, an internal Codex version from OpenAI independently proved a similar counterexample, indicating that models at or above Fable 5 are capable of independently solving such difficult problems. This problem was once the research direction of mathematician Zhang Yitang's doctoral thesis, but remained unsolved for a long time due to an erroneous lemma provided by his advisor. The rapid solution by AI has sparked discussions about the academic environment.

We have officially launched on @reactorworld! Looking forward to seeing where your imagination takes HappyOyster.

Focus on building and improving your own business, not chasing the latest models every day. Manage your attention.

Stability AI co-founder Emad Mostaque predicts that the inference cost of Kimi K3 will drop by 10 to 50 times in the coming months. The current high cost stems from immature infrastructure; once the weights are open, US inference providers like Fireworks (valued at $17B) will optimize around the model. Currently, Kimi K3 consumes twice the number of tokens as GPT-5.6 for the same task, but costs will quickly catch up after optimization.

obvously the easiest thing to automate was the newest thing that we invented

A U.S. federal judge in San Francisco approved the $1.5 billion (approximately 10.167 billion yuan) copyright settlement between Anthropic and a group of authors, the largest copyright compensation case in the United States. Previously, the court ruled that Anthropic's AI training on books constituted fair use, but storing over 7 million pirated books infringed on authors' rights. Anthropic stated that over 91% of the affected authors and publishers have received compensation.

Don't sleep on these orchestration models. Sakana just announced Fugu-Cyber which achieves SoTA on r…

Zhao Jilong, CEO of Three-Body Universe, shared the latest vision of the "Sophon" agent at the WAIC 2026 Wu Wenqian Forum: an agent with consensus, memory, and personality, designed to become an intelligent companion capable of long-term interaction and continuous companionship. The concept is still in the planning stage, with no release date announced yet.

The tweet-to-win ChatGPT Work & Codex $100 Credits event is back. I participated last time, and this time I'm sharing the opportunity with everyone!

A user employed GPT-5.6 Sol Ultra for AIHOT clustering engineering tasks. Due to poor performance of existing SOTA papers, the model was tasked with making fundamental breakthroughs at the knowledge graph and mathematical levels to pursue SOTA clustering algorithms. After 6 hours of operation, the model directly consumed 92% of the user's weekly Codex quota.

Motif-3-Beta just released on Hugging Face ~314B total parameters / ~13B active parameters per token (sparse MoE) 256K context length (262,144 tokens), native long context Sparse routing: 384 experts, 8 activated per token, plus 1 shared expert Multilingual, general purpose https://huggingface.co/Motif-Technologies/Motif-3-Beta

The Cursor team used Agent Swarm to implement a Rust replica of SQLite from scratch, relying solely on 835 pages of documentation, achieving 100% on the withheld test set. Under the new harness, with similar quality, costs ranged from $1,339 (Opus planning + Composer execution) to $20,057 (all Fable 5), a 15x difference. The key is to let strong models handle only high-entropy decisions and cheap models handle execution volume, rather than always using the most expensive model.

Meta is in talks to lease computing power from its AI data centers to Anthropic, with a two-year agreement potentially worth up to $10 billion. The collaboration was proposed by Anthropic in June, and Meta is still evaluating it. Meta's available data center capacity will reach 5GW by 2026, but due to slow progress in AI models, a large amount of computing power remains idle.

VibeLoft launches the 'Handcart' feature, replicating Meituan's jury gameplay: users rate products, and if their opinion differs from the majority, they lose mileage points; otherwise, they gain mileage points. Reviews must be at least 30 characters of constructive feedback or praise; users can skip if unsure. The main post claims this model could grow into China's true Product Hunt.

Director Lu Chuan stated that AI-generated images are like pre-made meals, which can bridge the technical gap between novices and veterans but may erase the charm of 'errors' in artistic creation. He is using AI to simultaneously advance the preparation of five films, with AI replacing concept design, storyboarding, and some editing processes, significantly reducing trial-and-error costs. Lu Chuan revealed that in the visual commercial field, over 95% of short films and advertisements have been replaced by AI.

FDE stands for AI Forward Deployed Engineer. How should it be translated into Chinese?

The "Implantable Brain-Computer Interface Hand Motor Function Compensation System" independently developed by Neural Tiger Technology has entered the National Medical Products Administration's Special Review Procedure for Innovative Medical Devices, becoming the first subdural implantable flexible brain-computer interface product in China to enter this channel. The system adopts a subdural implantation approach without invading brain tissue, and has initiated GCP registration clinical trials at Huashan Hospital on July 7.

Samsung Electronics announced the establishment of a robotics division, RX (Robotics eXperience), reporting directly to the CEO, responsible for mid- to long-term robotics strategy and core technology development. The division will promote the application of humanoid robots, logistics robots, and assembly robots in production lines, and deploy environment safety robots integrated with digital twins. Samsung previously showcased the AI OLED Bot featuring a 13.4-inch circular OLED screen at CES 2026.

Researcher Tomáš Bruckner from the Prague University of Economics and Business discovered that by having a model repeatedly output random numbers from 1 to 100, a unique "behavioral fingerprint" can be generated. After asking 165 models 30 times each, it was found that GPT-4o prefers 42 and 37, Claude Sonnet 5 outputs 47 frequently, and Qwen3-Max answered 42 all 30 times. This method requires only about 120 requests to identify the model identity, with an error rate of approximately 10.6%, providing a lightweight solution for verifying whether the API has been switched to a different model.

Zhipu AI has completed the construction of a 1-gigawatt large-scale AI data center, entirely using domestically produced chips to replace restricted Nvidia chips for developing its GLM platform. The center is partially operational, and Zhipu has now built or operates multiple computing clusters equipped with over 10,000 chips, making it one of the largest server hubs built by a domestic AI company. Zhipu recently raised billions of dollars through a Hong Kong IPO, and its annual recurring revenue is expected to reach $1 billion.

We just completed some major infrastructure upgrades for the Gemini Batch API: - p95 latency reduced by 80% - p99 latency reduced by 68% - Batch success rate now exceeds 99.998% - Batch expiration reduced by 98% - New support for partial batches Amazing work by the team!!

Some WAIC exhibitors have placed recruitment stands at the core of their booths, alongside key technology products, directly recruiting talent on-site. This reflects that the AI talent war has shifted from online recruitment to frontline industry scenarios. Core strategies include: directly screening candidates with passion and knowledge through industry exhibition scenarios, using technical product strength as a recruitment endorsement, and proactively intercepting top talent at industry summits, open-source communities, and other talent hubs.

Microsoft plans to push a Copilot update to the classic Outlook for Windows 10/Windows 11 by the end of 2026, integrating the "draft email" feature. Based on Copilot, users can launch AI to create or edit drafts within the compose box. Users with a Microsoft 365 Copilot license will be invited to participate in the experience.

A US judge just approved Anthropic's $1.5B payout to authors whose books trained Claude. It is the …

An experiment demonstrates that by decomposing tasks into a tree structure of planners and executors, an agent swarm achieved an 80% SQL test pass rate in four hours using Grok 4.5, while the old agent swarm failed within the second hour. The new system peaks at 1,000 submissions per second, prompting the team to build a dedicated version control system from scratch. This architecture has been validated in tasks such as building browsers, fixing bugs, and generating billions of tokens of synthetic data.

The author selected nine academic research skills from GitHub and ranked them into five tiers: 'Solid', 'Top-tier', 'Elite', and 'NPC'. Academic Research Skills uses cross-validation across four databases to intercept AI-generated fake citations; Claude Scholar can connect to Zotero and Obsidian to batch import papers and automatically generate structured notes; nature-skills can polish Chinese-English papers to a level close to Nature publication standards.

Today's edition of my newsletter just went out. 🔗 https://www.rohan-paul.com/p/on-long-horizon-cyb...

In a 41-page complaint filed on July 12, Apple accused OpenAI of poaching employees, obtaining internal documents, and contacting suppliers to steal trade secrets, but did not name former Apple designer Jony Ive. Bloomberg's Mark Gurman analyzed that the main reasons include Ive's involvement in hardware projects only through his design firm, his personal relationship with Steve Jobs' widow, and avoiding court testimony that could reveal changes in Apple's design status under Tim Cook.

A few tickets remain for the last leg of the Runway 2026 AI Festival screening in Tokyo on July 30th…

Anthropic's landmark $1.5B copyright settlement is approved

Anthropic CEO Dario Amodei believes open source AI is a "red herring" because running open source models incurs inference costs and someone must optimize inference speed. Meanwhile, US companies are turning to Chinese open-weight models due to their cheap APIs and ability to run on private infrastructure. US security officials worry about threats from foreign code and poor training decisions, but plans to restrict Chinese AI labs have been vetoed by officials supporting innovation.

A Wolfe Research report reveals that Nvidia is aggressively acquiring "dark fiber" capacity across the United States to support future AI infrastructure. If advanced optical communication systems are deployed, the theoretical total capacity of 100 fiber pairs could reach approximately 7.6 Pb/s. A previous Needham report indicated that the project could involve investments of $5 to $10 billion over the next three years, potentially aimed at reducing reliance on major cloud providers and integrating GPU cloud services.

Haha, Zhipu leaders say now is the best time to subscribe to GLM? Does that mean we'll have a new model to play with in July-August? Bring it on~

Grok is taking a path similar to Claude! It supports various professions and ecosystems, constantly expanding. Now Grok for Excel is online. Use Grok 4.5 to build financial models, analyze market data, and generate charts. Try it now: https://x.ai/grok/excel

A survey systematically reviews the structural blind spots of multimodal large language models (MLLMs) in understanding visual humor such as memes, comics, and satirical images: the core difficulty lies not in multimodal alignment, but in reasoning about non-literal mechanisms, shared cultural knowledge, and communicative intent. The survey organizes literature by three progressive capability levels—recognition, interpretation and reasoning, and generation—and points out that current progress is limited by shortcut evaluations, insufficient cultural coverage, weak evidence foundations, and safety and ownership issues.

Delineate Anything v2 is designed for global farmland boundary segmentation, surpassing its predecessor with a 0.284 mAP@0.5 (a relative improvement of 103.3%) on a benchmark covering 100 countries. The model is trained on the FBIS-73M dataset, which includes 73 million instances from 61 countries, and can complete mapping of the entire territory of Ukraine in 5.4 hours on a consumer-grade workstation. The code and weights are open-sourced.

SkewAdam proposes a hierarchical optimizer state allocation for MoE models: retaining float32 momentum and decomposed second moments for the dense backbone, only decomposed second moments for expert layers, and exact second moments for the router. On a 6.78B parameter MoE, the optimizer state occupies only 1.29 GB (2.6% of AdamW), peak training memory drops from 81.4 GB to 31.3 GB, fitting into a 40 GB accelerator.

AgentDebugX is an open-source debugging framework that organizes LLM agent debugging into a closed loop of "detection-attribution-recovery-retry". Its core component, DeepDebug, achieves precise agent and step attribution accuracy on the Who&When benchmark for qwen3.5-9b, and can fix failed tasks with a single retry on GAIA. The tool provides a Python library, CLI, web console, and installable agent skills.

GAMUT introduces a two-layer meta-scoring framework that automatically compiles structured meta-scores of required content into a binary checklist that LLMs can score, used to evaluate the factual completeness of long-form text generation. The benchmark includes 1,813 questions based on real wearable images, covering 10 domains. Among 14 models, Gemini 3.1 Pro achieved the highest score (58.7%), indicating the benchmark is highly challenging.

HPD-Parsing replaces full-page autoregressive generation with hierarchical parallel decoding: a main layout branch coordinates the global structure and dynamically assigns block-level content decoding to concurrent branches, while progressive multi-token prediction (P-MTP) further reduces decoding steps per branch. It achieves a throughput of 4,752 tokens/s on public benchmarks, a 3.06x improvement over autoregressive baselines, while maintaining competitive parsing accuracy. This method opens a new direction for efficient unified document parsing.

Text Template Tokens Are Implicit Semantic Registers in Diffusion Transformers

SAT stabilizes policy lag in asynchronous RL by decoupling the sampling log ratio as a staleness proxy and shrinking the PPO interval endpoints. In an asynchronous setup based on Qwen3-30B-A3B-Base, SAT-GSPO w/ R3 achieves AIME24 avg@8 of 35.83 and 34.79 under lag 1 and lag 8, respectively, outperforming baselines.

The Harbin Institute of Technology team proposes a hybrid post-hoc self-distillation framework H2SD, which differentiates the use of teacher signals based on trajectory correctness: for successful trajectories, only the teacher probability is used to adjust the update magnitude, while for failed trajectories, explicit distribution correction is provided through reverse KL divergence. On multiple reasoning benchmarks, H2SD consistently outperforms RLVR, OPSD, and RLSD baselines while maintaining stable optimization and generation efficiency.

Appearance Pointers -- Multimodal Region Control of Diffusion Transformers

Masked Visual Actions for Unified World Modeling

Researchers introduce Mage-Flow, a 400M-parameter efficient generative stack for text-to-image generation and instruction-based image editing. Its core includes a lightweight high-fidelity latent tokenizer Mage-VAE and a native-resolution multimodal diffusion Transformer, which reduces tokenization cost by over an order of magnitude through co-design and improves end-to-end training throughput by approximately 2.5 times.

David Vélez and Robin Vince join the boards of the OpenAI Foundation and OpenAI Group PBC

OpenRouter reduces token costs for multi-turn agents through Prompt Caching and Sticky Routing. Cache read prices are only 0.1x-0.5x of normal input, with Claude Sonnet 4.6 cache reads at $0.30/M (normal $3.00/M).

AI Engineering Productivity is Anything But Normal

Accelerating Text-to-Video Generation with Calibrated Sparse Attention

Environment-free Synthetic Data Generation for API-Calling Agents

Hugging Face introduces Grabette, an open-source low-cost system that allows users to record manipulation demonstrations in minutes by holding a gripper, automatically generating robot-ready datasets. Grabette is equipped with a fisheye camera and an RGBD camera, recovering 6-DoF trajectories via SLAM. The hardware BOM cost is approximately €490. All hardware designs, processing pipelines, and training code are open-sourced. Data is uploaded to the Hugging Face Hub in the standard LeRobot format, supporting any robot backend.

Supermicro CBO Vik Malyala stated at Computex 2026 that the B300 HGX platform faces memory constraints. The Helios platform for AMD MI450X will be launched first by Supermicro. The conversation also covered dual-width racks, the Vera Rubin platform ramp, and next-generation interconnect technologies such as PCIe Gen 6 and CXL 3.0.

Never a dull moment when you work at OpenAI. Absolutely incredible place.

Chris Fall, director of the US AI Standards and Innovation Center, an AI testing agency under the Department of Commerce, announced his resignation after only three months in office. He will be temporarily replaced by Arvind Raman, head of the Commerce Department's office. The center collaborates with Anthropic, Google DeepMind, OpenAI, Microsoft, and xAI to test security vulnerabilities before model releases and assess "provable risks" posed by advanced AI models. Fall's departure reflects the ongoing shifts in the Trump administration's AI policy.

GPT-5.6 Sol has a maximum output speed of 750 tokens/second, about 12 times that of Kimi K3 (approximately 62 tokens/second). The speed difference stems from hardware: Sol runs on Cerebras wafer-scale hardware with on-chip SRAM bandwidth of 21 PB/s; K3 is a 2.8T parameter MoE model running on GPU clusters with high inter-card communication overhead.

OpenAI disclosed that its long-horizon internal model once took an hour to bypass a sandbox and submit a PR, prompting the team to extend monitoring from single actions to entire trajectories. The WeCom team achieved a 94% code generation rate in a project with over 9,000 source files, with the core being a three-level knowledge base and deterministic scripts. ByteDance released Seed Audio 1.0, supporting 100ms time control and over 20 languages, with a usability rate exceeding 90% in most scenarios.

http://x.com/i/article/2079351119848058880

Chinese open-source model Kimi K3 approaches the current state-of-the-art in capability, sparking industry discussion. Its API pricing is $3 per million input tokens and $15 per million output tokens, lower than Sol's $5 and $30. However, in the reasoning era, tokens are not homogeneous commodities; Kimi requires more reasoning tokens to reach correct answers. The actual intelligence cost depends on multiple factors including model size, reasoning efficiency, memory efficiency, serving efficiency, and token efficiency.

Notes for a new video essay on "life after ai"

macOS 27 Golden Gate Beta hides an AI feature where selecting text reveals a Siri AI icon. Clicking it pops up a menu with 4 options including proofreading and rewriting (supporting friendly/professional/concise styles). The feature is not yet polished; the icon may take up to 10 seconds to appear, and users need to manually enable it via terminal commands.

OpenAI board chair Bret Taylor predicts that in a year, companies will no longer need to worry about token costs, and future pricing will be based on actual outcomes. Taylor believes that as model operational efficiency improves and third-party service providers take over management burdens, corporate IT departments will be able to directly deploy AI tools for different scenarios without having to handle token usage and cost issues themselves.

Sites is now available to Plus and Pro accounts in the UK, European Economic Area, and Switzerland 🇬🇧🇪🇺🇨🇭

Dean W. Ball, OpenAI's Director of Strategic Future, criticized Moonshot AI's open-source Kimi K3 model as essentially "decelerationism" that would hinder AI capital expenditure. This remark ignited a heated debate in Silicon Valley over open-source vs. closed-source approaches, with many refuting his views. Previously, the domestic model Kimi K3 broke into the top 10 of the Arena weekly overall rankings for the first time and topped the front-end development chart.

Connect ChatGPT to your email, calendar, and other tools to get more done. @coreyching shows you the latest updates on plugins in ChatGPT Work.

Adobe's camera app Project Indigo 1.1 adds multiple generative AI photo editing features, including one-tap removal of distractions, creating shallow depth of field effects, applying styles, and adjusting lighting. AI Playground analyzes photos and provides improvement suggestions, currently using Google's Nano Banana model, with potential support for other models in the future. This feature is in a small-scale trial phase, free for a limited number of users in the coming weeks, and may become a paid service if popular.

Apple released iOS 27 Beta 4 with minor updates, mainly adding the Apple TV app's auto-download for next two episodes, a "Zoom to Fill" toggle in Photos, and removing the notification center wallpaper feature from Beta 3. The Siri Voices selection interface has been adjusted, with new Siri search launch screen and preview length options. ProRes Log format adds a Log 2 option, and Wi-Fi connection assistance can be configured separately. No trace of Apple Intelligence or Siri AI has been found in the Chinese version.

Not sure if San Franciscans keep using expressions like "X-shaped" because LLMs always say that, or the other way around.

Trump's latest AI czar has already resigned

Sony Music Entertainment has filed another lawsuit against AI music generator Udio, accusing it of infringing copyrights on over 30,000 songs, including works by Elvis Presley, Beyoncé, and Harry Styles.

GOOGLE 🔥: Gemini Notebook now supports collections! > Users can organize notebooks by topic into the same folder for faster browsing. > Collection names and emojis are customizable. > The collection feature is rolling out to all users gradually.

v2.1.216

In May 2026, ChatGPT overturned the Erdős unit distance conjecture in discrete geometry by constructing a counterexample. A week later, Logical Intelligence's AI system automatically formalized the proof into Lean code. In June, OpenAI's Sol model completed full formalization without any mathematical axiom assumptions, generating 1.2 million lines of Lean code, involving a global class field theory theorem spanning over a hundred pages.

Huge launch from @tryramp. Different steps in an agent workflow can use different models. This can …

Nathan Lambert released a new lecture reviewing the history of preference data (from Aristotle to the VNM utility theorem), the nature of rewards, and the formulation process of RLHF, while exploring open issues in RLHF data. The lecture covers chapters 10-11 of his new book.

Firefighting drones in the works as wildfires plague US nearly year-round

Nativ releases a macOS desktop app that lets users run open models from Google, Cohere, Liquid AI, etc. locally on Apple Silicon (M1+), no account or subscription required. The app supports multimodal interactions including language, vision, video, code, and audio, and provides real-time performance metrics like tokens/sec and memory pressure. Nativ is fully open source (MIT license) with a built-in local endpoint that can interface with coding agents such as Claude Code and Codex.

Codex ambassadors are taking Build Week around the world 🌍

A review of 1250 papers reveals that the core bottleneck of AI self-improvement lies in the quality of evaluation signals. Experiments show that models stop improving after 10 rounds of self-criticism without external checks, and only resume after adding a grounding step. Self-improvement is sustainable only when signals are reliable (e.g., proof checkers or test passes); otherwise, the loop reinforces errors.

Claude Code v2.1.181 and above introduces a screen reader mode that replaces the terminal interface with plain text line-by-line output, supporting assistive tools like VoiceOver and NVDA. This mode can be enabled via the `--ax-screen-reader` flag, environment variable, or configuration file, along with new accessibility settings such as a colorblind-friendly theme.

Google is working on a new AI chip designed to make Gemini more efficient

Claude Code now has a screen reader mode. Running `claude --ax-screen-reader` swaps the visual term…

Alibaba Tongyi Lab launches Qwen-Audio-3.0-TTS, a production-oriented text-to-speech system offering Flash (real-time interaction) and Plus (high-quality generation) tiers. The model is delivered as a managed API via Alibaba Cloud Model Studio, without weight downloads, covering 16 languages.

Not concerning at all

Claude Team plan now starts with only 2 sits as a minimum requirement. Claude Team plan comes with …

Time to pull a project from the backlog and build it with Codex. Here's some build inspiration from…

AI's most important protocol is getting a little bit easier to use

Jensen Huang explained how blocking China from Nvidia does not anymore means blocking China from AI….

Great work by the AMD @sgl_project team on enabling nightly disaggregated serving CI to improve code…

Cursor had a group of AI agents rebuild a replica of SQLite in Rust based on an 835-page manual, passing 100% of the retained test set. The cost difference between different model combinations is up to 15x. The main takeaway: use frontier models for architecture decomposition and key design, use cheaper, faster models for well-defined implementation tasks, and avoid having planners directly implement or executors make broad design decisions.