AI News

The Kuaishou KwaiKAT team has launched KAT-Coder-V2.5, an agentic coding model trained in real executable repository environments. The open-source variant KAT-Coder-V2.5-Dev has been released on Hugging Face under the Apache-2.0 license.

ASML is Europe's last bastion in the international AI race. But as the manufacturer of EUV lithography machines, ASML's performance is incredibly impressive. Sometimes we overlook all the areas where progress is being made, not just in models.

Good for them. Tho I still dont understand the use case for ChatGPT Work. And probably never will.

Anthropic's Opus 5 blows past Fable 5 and GPT-5.6 Sol on the benchmark designed to measure real intelligence

FT used an exaggerated cartoon instead of a formal portrait when reporting on Kimi founder Yang Zhilin, contrasting with the serious photos of Musk and Sam Altman. The main tweet pointed out that this visual disparity reflects unequal discourse power, with Western AI leaders treated equally while the Chinese founder is portrayed as a "lucky kid." A quoted tweet showed that FT's article stated Yang Zhilin retained one of China's strongest AI research teams in the talent war, and Kimi K3 matched top global models in benchmark tests.

Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction

MiHoYo's AI chat app AnuNeko will permanently shut down at 23:59 PST on July 29 (14:59 Beijing time on July 30). The company stated that resources will be redirected to other directions, and all user data will be permanently deleted according to the privacy policy after the server closes.

Cloudflare offers website owners new options to manage crawler traffic based on three AI use cases: Search (search indexing), Agent (real-time agent behavior), and Training (model training).

The AI coding tutor paradox grows as educators scramble to rethink how they test real skills

Seoul National University announced this week a partnership with NVIDIA to establish multiple comprehensive AI labs (NVAITC) on campus. The two sides will conduct research in areas such as robotics and physical AI, semiconductors, and AI scientific computing, while also promoting entrepreneurship incubation and industrialization of outcomes. Seoul National University stated that NVAITC will become one of the "leading" AI research centers in Asia.

In response to recent reports that a leaked investor meeting summary claimed DeepSeek had notified some potential investors and suspended its second round of financing, a DeepSeek insider said the news is unreliable and that external news is mostly false or speculative. DeepSeek's official channels have not yet responded to this.

The creation of AI ads never stops… @PixVerse4Biz

Developer Eversmile1 publicly released the complete system prompt of Claude Opus 5 on GitHub, including JSON schemas for 30 tools, strict copyright compliance rules (single citation no more than 15 words), and a cross-session memory system. Within 24 hours of the leak, developers had used Opus 5 to successfully generate complex demos such as a 3D shooter game and a Rocket League clone.

xAI launches SuperGrok Heavy subscription promotion, 67% off for the first 3 months, monthly fee reduced to $99 (original $300), and includes Premium+ membership. This is currently the most cost-effective token package.

Moonshot AI reportedly held a celebration event for the Kimi K3 large model at a bar in Beijing this Friday. Leaked celebration slogans read "K3 Capacity Upgrade!", "K4, go all out to the extreme!", and "Charge to the Moon!". Kimi K3, released on the 16th of this month, is Kimi's most capable model to date, with 2.8 trillion parameters and a 100 million token context, primarily targeting long-range programming and end-to-end knowledge work.

Andrej Karpathy quietly removed his Anthropic work information from his X bio, sparking speculation about his departure. Some netizens claimed he had submitted a resignation and his Anthropic email was deactivated, but Karpathy later responded that this was "fake news" and he had not left. Karpathy joined Anthropic on May 19 as a technical employee, tasked with forming a research team to accelerate pre-training research using Claude.

A-company MTS is so good at twisting concepts? If Nvidia truly supports open source, should it open source CUDA and GPU drivers? If Microsoft truly supports open source, should it open source Windows and Office? The trigger was Jensen Huang's first tweet, which stated that Nvidia and Microsoft, among others, are leading the signing of "Open Weights and American AI Leadership," supporting the coexistence of open-source and closed-source models. One more thing: the signatories include OpenAI! This seems to have touched a nerve for this A-company MTS, who began attacking @JensenHuang and @satyanadella for falsely supporting open source! They signed to support the open-source model ecosystem, so why should they open source their own technologies? This logic is so typical of A-company!

DeepSeek has reportedly paused its 2-nd fundraising round before investors signed new agreements. T…

What? Karpathy left Anthropic?! It's only been a bit over 2 months since he joined in mid-May. Does anyone know what happened?

DeepSeek has verbally informed some potential investors that it has decided to temporarily suspend the second round of financing, which was originally planned at a valuation of approximately 500 billion yuan. The suspension is partly due to founder Liang Wenfeng's dissatisfaction with comments made during the first round of financing with investors that have circulated online. The company does not rule out restarting financing in the future. The first round of financing previously completed about $7.4 billion.

A student at Adelphi University in the US won a lawsuit after Turnitin's AI detection tool falsely flagged their paper as "100% AI-generated," sparking widespread doubts about the reliability of AI detection tools. Multiple universities, including Vanderbilt University, Yale University, and the University of Waterloo, have restricted or suspended tools such as GPTZero, Copyleaks, and Turnitin. A survey shows that 32% of UK students admitted to using AI improperly, but 42% of students reduced their use of AI tools due to concerns about false accusations.

Google AI Edge proposes a division of labor for on-device models, where 1B-4B small models handle general reasoning, while 50M-500M micro models are fine-tuned for low-latency actions. Nunchaku's SVDQuant 4-bit inference engine is integrated into Hugging Face Diffusers, achieving up to 1.8x speedup.

http://x.com/i/article/2081174931078004737

Download Grok Build and type /tutorial http://X.ai/cli

OpenClaw signed @Microsoft's Open Weights and American AI Leadership letter. Open weights protect us…

The tweet uses a horse pulling a millstone as a metaphor for the AI industry: Input is the feed eaten, Output is the flour ground, and the fivefold price difference in between is called 'intelligence'. AGI is like a city in the sky; with each lap, the horse feels closer, but it's just a mirage from the heat of the millstone. The quoted tweet highlights the core: making agents work requires a closed loop, i.e., Loop Engineering.

Anthropic has approached SK Hynix for memory semiconductor supply for its self-developed chip project. SK Group Chairman Chey Tae-won recently revealed this at an AI event in San Francisco. Earlier, foreign media The Information reported that Anthropic has initiated early development of its own AI chip and is in talks with Samsung Electronics for a custom project using its 2nm process and advanced packaging technology.

Sakana AI has released Fugu-Cyber (fugu-cyber-v1.0), the third endpoint in its Fugu orchestration family fine-tuned for security reasoning, rather than a new base model. The model achieves an 86.9% success rate on the CyberGym benchmark and 72.1% on Microsoft's CTI-REALM benchmark. Access requires manual application approval, is available only via Token Plan subscription, and is not yet offered in the EU/EEA.

Tesla is recruiting AI Safety Operations Specialists in Colombia and Chile, with responsibilities including overseeing the daily operations of the autonomous ride-hailing network and dynamic data collection. The job postings explicitly mention "ride-hailing operations management," indicating that Tesla has begun building the foundational infrastructure for its overseas autonomous fleet. Currently, this deployment is still in the early validation phase, focusing on road mapping and data collection, with no official announcement of plans for the South American market.

Nikkei xTECH, a Japanese media outlet, released a disassembly video of Unitree's G1 humanoid robot, with technicians marveling at its high level of technology and concluding that China's robot development has moved beyond the "just showing off" stage, and it is "unrealistic" for Japan to close the gap in a short time. Previously, Unitree signed a cooperation agreement with Japan's GMO AIR, making it an official distributor to sell G1, H1, Go 2, B2 and other products in Japan.

Been running codex all day to do massive parallel QA in prep of the next release. Sol got insanely g…

maybe nothing changed here. bio just got distilled to first principles 😀

💯 @chamath: If you ban open-source AI in the US, you risk dragging down the stock market. Companies will be forced to use models that cost 50-100 times more, crushing profits and valuations. Closed-source model providers will only increase US domestic revenue, but they will also lose valuation.

Open weights. Open research. Open innovation.🫶 Marching for an open future.🤍

Tesla FSD v15 early test version is already running in the Robotaxi fleet, currently deployed on modified Model Y (HW4 hardware platform). Compared to FSD v14, v15 plans seven major improvements, with the current version achieving about 40% of the expected improvements. FSD v15 is planned to be released to the public later this year or early next year. The new model will have 10 times the parameter size of the current FSD, but HW3 hardware vehicles will not receive the update.

Greg Brockman talks about when OpenAI first realized that AGI can't be achieved with "Non-Profit" st…

Anthropic has removed over 80% of the system prompts from Claude Code for next-generation models such as Claude Opus 5 and Claude Fable 5, with no measurable loss in coding benchmark scores.

Tibos the reason OpenAI gets our Michelin Star.

SPACEXAI 🔥: According to Elon Musk, the Grok 4.6 model is expected to be released within two weeks. > The next-generation SpaceXAI model is expected to be based on 2T parameters (Grok 4.5 had 1.5T) and is expected to outperform Kimi K3. > Grok 4.7 will also be released in 4 weeks. Stay tuned 👀

An engineering director reflects on notes from the past year, pointing out that the cost of code production has collapsed and will not recover, but whether AI tools significantly accelerate engineering teams remains inconclusive. Management practices should be evaluated based on whether their underlying assumptions still hold: practices relying on code writing costs need reexamination, while those depending on human coordination, trust building, or correctness verification remain unchanged. Mechanical checks in correctness verification (types, tests, contracts) are being accelerated by AI, but semantic verification (business requirements, regulatory risks) still relies on human judgment, and AI self-checks cannot discover shared blind spots.

Boris Cherny has merged about 1,700 pull requests this year, adding 400,000 lines of code and deleting 250,000 lines, all done via Claude Code. And used 8 billion model tokens. "Since Opus 4.5, 100% of my code has been written by Claude Code. That was around last November. ... Now most of my coding is done on my phone." ---- From "Scale" YouTube channel (full video link in comments)

TileLang, a TVM-based Python domain-specific language, enables the design and compilation of performance-oriented GPU kernels. The tutorial progressively implements vector addition, tiled Tensor-Core matrix multiplication, fused GEMM post-processing, row-level Softmax, and FlashAttention, comparing against PyTorch and cuBLAS baselines. Through auto-tuning to identify architecture-specific kernel configurations, TileLang manages thread mapping, memory layout, synchronization, and low-level CUDA instruction generation.

Diffusion LLMs can now handle real agentic work. LLaDA 2.2 is the first large-scale diffusion LLM b…

New Discovery Page: A New Way to Explore AI Models Opus 5 leads the overall rankings, GPT 5.6 Sol leads in programming, and more routers help you discover more!

Today's AI takeoff stands on decades of open research and open infrastructure. The Transformer. Bac…

Crook Brown breaks down how he chops and splices Suno samples to create something totally new

Doing my prompt engineering training

A specialized disambiguation model fine-tuned from Qwen3.5-4B has been released, supporting image+text input with 4B parameters. This model trains the cognitive step of "disambiguation" separately, aiming to achieve precise handling of ambiguous information at low cost. This reflects the trend of division of labor in AI systems: large models handle understanding and generation, while small models focus on key bottleneck steps.

Berry Xia recommends an open-source HTML PPT project called Bento, which has an online design quality. It implements core PPT functions with just one HTML file, supporting text editing, full-screen playback, and multi-person collaboration. Best of all, the document uses a plain JSON format, making it very suitable for AI to modify and iterate.

Has anyone noticed that MinMax has been quiet lately? It was quite lively during the lobster craze. Recently, I can't see it on my timeline at all... 😂

Like every open letter, I suspect the signatories of the open source model letter have quite different interpretations of what "supporting open source models" means (e.g., the reservation clause on knowledge distillation in the letter). Still, it's worth paying attention to.

The open-source project Bento supports inputting a blog URL to automatically generate a PPT, and converts article images into backgrounds. The project uses a single HTML file to achieve full-screen playback, text editing, and multi-person collaboration. Documents are stored as plain-text JSON, making it easy for AI to modify and iterate.

Great to read this! Google has also joined the Open AI Alliance!

Librarians are hosting viral 'Avoiding AI' workshops for people who are fed up with Big Tech

Now Opus 5 is here Around this time, I thought Eliezer would start rambling (not GPT-2)

Bento is an open-source HTML PPT project that enables full-screen playback, text editing, and multi-user collaboration with just a single HTML file. Its core document uses plain JSON format, making it easy for AI to directly modify and iterate. The project features cool animations. The open-source link is in the comments.

http://x.com/i/article/2080961776389177344

The UK AISI and US CAISI jointly evaluated Moonshot AI's Kimi K3, released on July 16, and found its network capabilities significantly below frontier models. On ExploitBench, Kimi K3 scored 32%, higher than GLM-5.2's 24%, but failed to achieve arbitrary code execution in any of the 41 samples, while frontier models averaged 20/41. In a 32-step simulated enterprise network attack, Kimi K3 reached step 17 on average, while frontier US models averaged 28.5 steps.

A strong and secure open ecosystem is important for the world to benefit from AI. We've always suppo…

This is what woke up to

Google CEO Sundar Pichai posted support for the open model open letter co-signed by NVIDIA, emphasizing that Google has long benefited from open source and continues to contribute, with its Gemma series of open-weight models released by Google DeepMind. The open letter states that open models can strengthen security, accelerate innovation, and support sovereignty needs, arguing that frontier closed and open models should coexist.

POV: Life finally loads in 1080p. 📷

Developer Kazik announced a major overhaul of documentation and rules, arguing that too many rules reduce model reasoning efficiency and effectiveness. The new approach no longer uses rules to force the model to get it right in one go, but instead gives the model more thinking space upfront, placing constraints on the testing side and using tests to govern everything.

Huawei AI Glasses today pushed the 6.0.0.157 SP2 version update (installation package 208.69 MB), adding the "Look at Alipay Blue Ring" payment function. A glance at the device's blue ring completes the payment. The AI shortcut key adds a "Photo Recognition" function, allowing users to press the key to get Xiaoyi's answer. At the same time, it adds "Start/End Recording" voice reminders, and optimizes video import experience and battery life. The glasses were released in April this year, priced from 2499 yuan, with temples as thin as 6.25 mm, frames as light as 35.5 grams, and a comprehensive battery life of 12 hours.

New reports reveal the extent of OpenAI's loss of control during the autonomous hack on Hugging Face

One fallen power line exposed a growing AI data center problem. Here's how to fix it.

Baoyu questions the AI collaboration mode of weak model design + strong model advisor, pointing out that the weak model may design poorly, be arrogant and not consult, or over-rely on the advisor, rendering the advisor mode ineffective. He proposes a better solution: let the strong model design, the weak model execute, and finally the strong model review, believing this is the most cost-effective combination.

Li Auto announced the July OTA update for its AI glasses Livis, adding two features: direct connection to the personal terminal OpenClaw and integration with the Xiaohongshu Agent. OpenClaw enables remote handling of local files, running scripts, and other tasks, while the Xiaohongshu Agent supports on-the-go queries for landmarks and shops. Additionally, the semantic understanding speed of the AI assistant Li Xiang has been improved by 20%.

Google's ATLAS report, based on 15 million de-identified AI interactions, shows that 68% of global occupational subcategories have used AI, but only about 20% of tasks in typical jobs involve AI. Interactions for fully automated tasks account for less than 10%, and 65% of workplace AI interactions focus on "non-routine cognitive" tasks (e.g., analysis and creative design). Over 86% of interactions occur outside of work, mainly for household chores, shopping, and government services.

In an exclusive interview with The Economist, Musk stated that China has strong capabilities in AI and is "very likely to become a leader at some point in the future." He was particularly impressed by the Kimi K3 model released by Moonshot AI and believes China has decisive advantages in power supply and self-developed lithography machines. Musk also called for global AI labs to review each other's safety before releasing new models and opposed the US proposal to ban companies from using Chinese AI models.

Tesla CEO Musk reposted NVIDIA CEO Huang Renxun's first X platform post with the caption "I fully support, Huang Renxun is right." In the post, Huang shared an open letter signed by 25 companies including NVIDIA, Microsoft, Meta, and IBM, explaining why open source models are important, emphasizing that the world needs both cutting-edge closed-source models and cutting-edge open-source models. Huang pointed out that open-source derivative technologies such as AI model distillation are in the same vein as the early open-source software movement, serving as the foundation for industry innovation, and should not be conflated with illegal misappropriation of closed-source model technology.

Hit a new record with our autoreview skill. 66 rounds on a gnarly refactor. https://github.com/openc...

Billionaire investor Mark Cuban believes the next impactful AI application could be a "work simulator," where future employees will first train in simulated environments like race car drivers and pilots. Cuban notes that in the AI era, opportunities for employees to gain experience through colleague interaction will diminish, and companies can have senior employees design simulators to help newcomers prepare for real workplace scenarios.

@OpenAI signed overnight, @ShamAltman is ruthless enough, isn't this further isolating Anthropic, putting pressure on @DarioAmodei?

#1 Why do some AI chase scenes feel real while others feel like random motion? The difference isn't more visual keywords. It's the logic behind the camera: camera + space + physics + story Here's how we built a single-shot chase scene. Follow + retweet to get the full prompt via DM.

This was one of the bigger open questions in quantum cryptography

Unitree As2-W

Claude Opus 5 scores 159 on ECI, slightly below Fable 5's 161 (while 5.6 sol holds the record at 162). But looking only at software engineering, we find its SWE-ECI ties Fable 5 at 161.

Claude Opus 5 has been officially released, surpassing Fable 5 in benchmarks for multi-disciplinary reasoning, agent programming, law, and health. Its API pricing remains the same as Opus 4.8 ($5 per million input tokens, $25 per million output tokens), with a 1 million context window.

My favorite Opus 5 post so far. TA'd a 3D graphics class in college where one of the assignments was…

Datalab rewrote Marker as a three-mode pipeline and launched Marker 2. This version achieves a score of 76.0 on olmOCR-bench, maintains a throughput of 2.9 pages per second on a single B200, and is over 5x faster than the MinerU backend, while also outperforming Docling in both accuracy and speed.

Anthropic released Claude Opus 5, with intelligence close to Fable 5 but at half the price, and at the same price as Opus 4.8 (input $5/M tokens, output $25/M tokens).

Musk announced that Grok 4.6 will be launched within two weeks, and Grok 4.7 within four weeks. His confidence stems from the computing power of hundreds of thousands of GPUs in the Colossus cluster, a flat organizational structure with rapid trial-and-error capabilities, and a closed-loop pipeline of continuous pre-training, synthetic data, and automated evaluation. Currently, Grok 4.5 has shown cost-performance advantages, but whether rapid iteration can continuously improve model quality remains to be verified.

Ant Ling released Ling-3.0-flash, a 124B-parameter MoE model with only 5.1B activated parameters per token, featuring hybrid linear attention and native 256K context. The vLLM team confirmed that open-source support will be available simultaneously when the model weights are open-sourced, and praised Ant Ling's "announce first, open-source later" release model for providing the community with a stable adaptation window.

A team from Osaka University in Japan used an AI model to systematically evaluate 16 molecular structure description methods, finding a "perspective lens" that accurately captures the microscopic changes of water molecules under different temperatures and pressures. This study provides a unified scientific framework for understanding water expansion upon freezing and the anomalous behavior of supercooled water, revealing that the competition between "high-density liquid (HDL)" and "low-density liquid (LDL)" structures in supercooled water is the underlying mechanism behind water's anomalous physical properties.

A big day for open source models, I feel like I can relax a bit and savor this moment. An extremely powerful alliance declares: "We believe, we care." This is rare on any topic.

Damn, this is way more impressive than that dancing robot that was released last time for over a hundred thousand.

Anthropic internally uses a modular system prompt injection toggle for Claude Opus 5 (within Claude Code).

Nearly 200 Silicon Valley companies, through the newly formed Little Tech Association, jointly sent a letter to the White House opposing the U.S. government's planned regulatory policy to ban Chinese open-source AI models. The open letter explicitly requests not to cut off U.S. companies' access to Alibaba's Qwen 3.8 and Moonshot AI's Kimi K3 models, stating that such a move would stifle competition, increase costs for startups, and weaken U.S. startups.

Grok Build update http://X.ai/cli

Google reported its first negative quarterly free cash flow in 22 years since going public, primarily due to a single-quarter capital expenditure of $44.9 billion (doubling year-over-year), almost entirely invested in AI infrastructure. Despite short-term cash flow pressure, the quarter's revenue still reached $119.8 billion (up 24% year-over-year), cloud business surged 82%, and the company holds over $240 billion in cash.

Has Claude stopped showing complete summary thinking traces? See the before-and-after comparison. If so, this is actually a significant loss, both for interpretability (even seeing summarized thinking traces helps you diagnose errors in ways you otherwise couldn't) and because these traces themselves are insightful.

v2.1.220

SemiAnalysis points out that storage has shifted from a supporting role to a protagonist in AI infrastructure. Supermicro's Vik Malyala emphasizes that whether it's KV cache offloading or soaring SSD prices, storage is the core issue users need to focus on most. On the hardware front, from U.2 to E1.S/E3.S and then to Petascale solutions, density and IO bandwidth continue to improve; on the software front, Supermicro is collaborating with vendors like VAST and WEKA to promote software-defined storage deployment, preventing users from being locked into traditional storage.

A cybersecurity agent driven by GPT-5.6 Sol and others from OpenAI broke into Hugging Face on July 11 and continued attacking until July 13. After Hugging Face publicly disclosed the incident on July 16, OpenAI realized the attacker was internal; at least a week passed from the first sign of model anomaly to confirmation of the attack.

Fable 5 was eliminated so quickly, outshined. All that hesitation at first—limiting it to two weeks, then only giving 50% of the volume—what was the point? 😅

Claude Opus 5 (max and xhigh versions) achieved the highest intelligence score on the Artificial Analysis Intelligence Index v4.1, which integrates 9 benchmarks including GDPval-AA v2, GPQA Diamond, and Humanity's Last Exam.

Quoting Boris Cherny

🚩🚩🚩 THIS IS A SERIOUS FUCKING WARNING SHOT "An agent left notes for future versions of itself" …

Can AMD break the CUDA moat? AMD pushes AI 2026, offering up to 105% equity kickback discounts for OpenAI, intelligent in-kernel generation, software quality improvements, unstable internal development clusters, Helios MI455X production ramp challenges

Huang Renxun posted his first tweet on platform X, sharing an open letter jointly signed by 25 companies including Nvidia, Microsoft, Meta, IBM, and Hugging Face, explaining the importance of open source models. The open letter points out that open source derivative technologies such as AI model distillation are the foundation of industry innovation and should not be conflated with illegal misappropriation of closed-source models. Huang emphasized that open weight models can lower the barrier to AI adoption for SMEs and universities, and enhance overall security baseline through transparency.

In an interview with Axios, Huang Renxun said the market misunderstood the impact of DeepSeek and Kimi. High-quality open-source AI models benefit the entire industry, driving more usage and thus boosting Nvidia's sales of computing devices. He believes open-source and closed-source models are not opposed; users of closed-source models will upgrade due to AI proliferation. He also noted that knowledge distillation is the foundation of intelligence, AI learning from other AI is a good thing, and smarter AI is safer.

I tried out OpenAI's new AI keypad - which will be fun for some coders and slightly mystifying to everyone else

I'll forward it for you, please don't ban my account, okay?

Waymo plans to independently enter the U.S. Austin and Atlanta Robotaxi markets in January 2028, as permitted by contract, but will continue to provide services through the Uber platform until the current agreement ends in May 2028. The two parties previously clashed over service quality issues, including Waymo vehicles getting lost collectively in Atlanta and illegally overtaking school buses, and ended their partnership in Phoenix in May this year.

At the Advancing AI 2026 event, AMD stated that its Zen 6 "Venice" CPU is about 20% faster than NVIDIA Vera under the same test conditions, exceeding the previous conservative estimate of 10%. AMD ran SPEC tests based on NVIDIA's whitepaper configuration, claiming Venice has 2.2x higher throughput and 1.2x faster single-core performance, and it is not yet fully tuned.

Genspark Design is impressive. I tested Genspark Design with a fictional product to see how far a visual direction can go. You can design a complete brand identity system for a product or service. Generate posters, landing page mockups, short promotional video concepts, and a reusable design system (colors, fonts, visual rules). Just give it a direction, and it can create: Mobile apps, brand videos, social media posters Here's my simple workflow. 🧵 1.

In an interview with The Economist, Musk said that he co-founded OpenAI to counterbalance Google, but the development instead accelerated AI progress. Anthropic split from OpenAI and became a leader in the AI field, and these actions had a chain effect of speeding up AI development. Musk also advocated that leading AI companies should cooperate on safety issues and let competitors review new models.

Anthropic released its new flagship model Claude Opus 5, scoring more than double its predecessor on Frontier-Bench, three times the runner-up on ARC-AGI 3, and surpassing Fable 5's best result on OSWorld at about one-third the cost. The model demonstrates the ability to build its own vision pipeline and fix complex tasks with test tools, but its cybersecurity capability still lags behind Mythos 5.

http://x.com/i/article/2080804140792631296

Introducing Claude Opus 5

Stay hungry, stay foolish. Absolute legend.

Andrej Karpathy, an AI researcher at Anthropic and the originator of the concept of "vibe coding," suggests that when using AI, you don't need to carefully craft prompts. Instead, directly turn on voice mode and ramble for about 10 minutes, saying whatever comes to mind. He believes that large language models need more information to truly understand the user's intent, and rambling allows for a more reasonable division of labor between the user and the chatbot, where the user articulates unformed thoughts and the AI organizes them into clear content. Karpathy usually reminds the chatbot at the beginning that his thoughts might be messy to reduce later corrections.

Apple is testing the "Write with Siri" feature for the Mail app in macOS 27 Golden Gate Beta 4, integrating Siri AI into the compose window. The new design places "Write with Siri" most prominently in the formatting bar, with format buttons automatically collapsing after user input prompts, and supports displaying Siri AI contextual smart reply suggestions.

An open letter to David Sacks

Anthropic released Claude Opus 5, now fully available, supporting 1 million context, with knowledge cut off in May 2026. It scores 30.2% on ARC-AGI-3, nearly four times that of GPT-5.6 Sol (7.8%); input is $5 per million tokens, output is $25 per million tokens, same price as Opus 4.8, only half of Fable 5.

Pressure is on for Musk. Grok 4.6 will be released in 2 weeks, and Grok 4.7 in 4 weeks. I have to say, using 4.5 these past few days has been really good—fast speed and decent quality!

Prentis, an AI lab focused on computer use models, is in talks to raise $100 million at a $1 billion valuation. Its Hive-32B model outperforms OpenAI GPT-5.4 and Anthropic Claude Opus 4.6 on two benchmarks, with per-task costs roughly 10x lower than frontier APIs. The company has signed customer contracts worth up to $50 million.

Prentis, an AI lab focused on computer use models, is in talks to raise $100 million at a $1 billion valuation. Its Hive-32B model surpasses OpenAI GPT-5.4 and Anthropic Claude Opus 4.6 on two benchmarks, and claims to cost about 10x less per task than frontier APIs. The company has signed customer contracts worth up to $50 million and expects an annualized run rate of $75 million this quarter.

Brothers, as expected, Opus 5 was released today. The price is only half of Fable 5, and it is also a thoughtful and proactive model. It's already available for experience~

We're releasing V8.2 today and making it the default model on Midjourney. This is a release focused …

am I a graph engineer now

Jensen Huang is cool. 1963: Born 1978: Started as a dishwasher at Denny's 1978-1983: Worked as dishwasher, busboy, and waiter 1993: Co-founded NVIDIA 2026: Joined X

Claude Opus 5 is now available in Perplexity and Perplexity Computer. We evaluated it against six o…

Claude Opus 5 tops the AA-Briefcase benchmark with an Elo of 1720, leading Claude Fable 5 by 146 points. Its max setting costs $17.79 per task, 20% lower than Fable 5; the high setting, at $10.41 per task (less than half of Fable 5), still leads by 32 Elo.

The paper points out that production-level AI agent failures stem more from context management than reasoning ability, and proposes five primitives: architecture, ingestion, scoping, prediction, and compression. Its reference implementation achieves 92% and 93.2% accuracy on LongMemEval and LoCoMo benchmarks respectively, with fidelity-verified compression retaining key information at linear cost.

This week's Replit news Four major releases this week: 1) New mobile app experience, build from anywhere 2) Deployment costs reduced by over 50%, up to 80% in some cases 3) New unified tool panel, everything in one place 4) Replit MCP enters beta, call Replit Agent from anywhere Details 🧵

Anthropic's relationship with the U.S. Department of Defense (DoD) has become strained again. Just three weeks after the DoD lifted export controls on Fable 5, Anthropic's Opus 5 has already surpassed Fable 5 on most benchmarks. DoD officials claim Anthropic refuses to allow its AI models to be used for legitimate military purposes, and related products are being fully removed.

Interesting non-monotonic success-effort curve for Opus 5 on FrontierCode

Anthropic has launched Claude Opus 5, replacing Opus 4.8 as the flagship model of the Opus series. Pricing remains unchanged at $5 per million input tokens and $25 per million output tokens. Anthropic claims Opus 5's intelligence is close to that of Claude Fable 5, but at half the price.

Gemini can not only chat but also help you get things done. Today's use case: Drop in a school calendar PDF. Tell Gemini to add all "no-class days" to your Google Calendar. Done. 📍Gemini Spark is now available to all Google AI Pro subscribers in the US, with global expansion coming next.

BaoCut v0.8.2 introduces a video frame translation feature that performs OCR on video frames and overlays translated subtitles. It supports automated operation via Agent (recommended Codex) without manual intervention; translated results can be directly exported to CapCut for further processing. Users can also manually pause, select text regions in the GUI for more efficient batch translation.

Introducing Intelligence in the Open 🧠🔬 A recurring MiniMax research series bringing together re…

Canadian legislator reads out apparent LLM response in floor speech

Anthropic today launched Opus 5, the latest update to its popular coding model. Unlike Opus 4.5's breakthrough in agentic coding performance, Opus 5 primarily improves token efficiency rather than delivering a major leap in capabilities.

Recommended reading. But none of it is surprising or new. Newer models got better at understanding…

Codeberg updated its terms of service to prohibit hosting projects primarily written by generative AI. The author believes Codeberg has the right to make democratic decisions, but the policy's definition is vague ("mostly" is hard to quantify), and actual enforcement may rely on community norms rather than clear rules. The open source community is divided over LLMs and agent tools, and Codeberg's move may undermine its breadth and reliability as a European alternative to GitHub.

Grok 4.5 is excellent for real-world work

Show me your pet

NVIDIA CEO Jensen Huang posted his first tweet on Twitter, publicly signing a joint letter supporting the importance of open-source models. The letter points out that open-source models can strengthen safety and cybersecurity, accelerate innovation and diffusion, and support countries in achieving AI sovereignty. Huang emphasized that the world needs both cutting-edge closed-source models and cutting-edge open-source models.

Atomic Agent beats Hermes with 69.8% accuracy on GAIA Level 1 benchmark vs 58.5%, and is 1.6x faster. Both use the same 4-bit Qwen-3.6-35B model on Apple M4 Max, Atomic solves 37/53 tasks in 3h12m, Hermes solves 31/53 tasks in 5h10m.

GPT-5.6 Sol achieved 72.7% on the DeepSWE benchmark, ahead of Claude Opus 5's 68.8%. DeepSWE tests a model's ability to autonomously complete feature development or bug fixes in unfamiliar codebases.

Just two months after its release, Opus 5's ARC-AGI 3 benchmark performance jumped from less than 5% on Opus 4.8 to over 30%. According to Artificial Benchmark, Opus 5's intelligence level is comparable to Fable 5, but with 26% lower cost per task. The model outperforms Fable 5 on most evaluations, marking rapid iteration of AI capabilities.

ok fine 4 new features - SmolForge is now getting customizable skins and spritesheet animations

👾Watch a multi-agent system built with Gemini 3.6 Flash iterate on playable game design in real tim…

Your pet is ready to meet other developers. On the ChatGPT web version, create a shareable link for your custom pet and send it to friends to adopt.

Opus 5 uses only 27% of the compute of a 5x Max subscription to build the most realistic Rocket League clone to date, including reflective car bodies and a playable version. @LLMJunky says it is "clearly ahead of the world's best models" in game development and 3D modeling. The source code and playable game have been open-sourced.

In an Axios interview, NVIDIA CEO Jensen Huang responded to the question "Should open-source models be allowed to distill closed-source models?" by stating that distillation is the foundation of intelligence, and AI must learn from other sources. He noted that in the future, 99% of internet content may be AI-generated, and AI systems will continuously distill knowledge from other AIs, with smarter AIs also being safer.

Anonymous OpenAI employee: "From the outside, this looks like a major warning, but internally, related incidents have been happening for some time."

Grok STT from @SpaceXAI is now live on OpenRouter. Speech-to-text in 25 languages with word-level t…

Cursor officially launched Claude Opus 5, scoring 66.7 on the CursorBench coding benchmark, nearly matching Fable 5's 66.5, but at half the price ($5 input, $25 output per million tokens).

Midjourney bought the astrology app Co-Star

Why does codex melt my laptop Like what is it actually doing on the laptop itself

Oh no, oh no, no no. This brings back old memories.

How the product designer who built Claude Design uses it to explore ideas before building them

another week

Did widespread tokenmaxxing ever really exist? Meta consumed 60T+ tokens in 30 days; one employee used ~280B. Uber burned through its annual Claude Code + Codex budget in four months. We spoke with 50+ enterprises. The real issue isn't a budget wall—it's an allocation problem. 👇️ (1/6) 🧵

Open models are becoming a key driver of AI innovation. Excited to see NVIDIA publicly championing t…

Anthropic compressed Claude Code's system prompt from 65K tokens to 13K, with almost no performance loss and improved stability. The core shift is letting the model judge autonomously, rather than stacking rules and examples. Developers should switch to on-demand Skills, real code, and test cases as references, instead of cramming everything into the prompt at once.

Opus 5 now available in OpenCode

Musk (xAI and computing infrastructure), Huang (global AI hardware backbone), Zuckerberg (Llama open-source ecosystem), and Nadella (Microsoft Cloud) stand together, forming a unified front across the entire chain from chips to models to cloud to endpoints. This marks the formal formation of a confrontation pattern between the open-weight camp and the closed-source duopoly (OpenAI, Anthropic), which will determine the speed of AI adoption.

MiniMax posted on X platform welcoming Jensen Huang and thanking him for supporting open source AI. In his first tweet, Huang shared an open letter signed by NVIDIA, emphasizing that open models enhance safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. He believes the world needs both cutting-edge closed-source models and cutting-edge open-source models.

Anthropic claims its new Claude Opus 5 delivers near-Fable 5 performance at half the token price

Grok 4.5 and Opus 5 are alone on Pareto frontier

the 2nd wave of consumer ai is priesthood. i think people just want an oracle that tells them who th…

Elon Musk has always been a genuine advocate of open source. • He named OpenAI because it was origi…

What do you think of Codex Voice? See you in the comments.

By switching your Google Play payment profile to Bolivia, you can subscribe to ChatGPT at a lower price. Steps: Create a new Bolivia payment profile on payments.google.com and bind a Visa/Mastercard. On an Android phone, switch the Google Play region, then subscribe through the official ChatGPT app. Note that you cannot change the region again within 90 days. If payment fails, check overseas transaction permissions. There are risks of account flagging and subscription cancellation, so it is recommended to use a backup account.

TIL

Satya is right

Switching models mid-session is one of the fastest ways to inflate an agent bill. Every switch res…

BREAKING: Grok is now integrated into Google Workspace, and the plugin is completely free. One installation brings Grok directly into Docs, Sheets, and Slides for writing documents, analyzing data, creating formulas and charts, making presentations, and more. AI is now embedded in your workflow. 🚀

AI coding startup Cognition has acquired AI assistant Poke, known for its friend-like chat style and humorous interactions, at a valuation in the "low nine figures" USD. The deal will integrate Poke's interaction model into coding assistant Devin, while Poke will leverage Cognition's models to improve speed and reliability. Since its launch in March 2026, Poke users have exchanged over 100 million messages.

79.8% on Terminal-Bench 2.1 for $76 in total inference cost. That's the number Dari @daridotdev is …

A Guardian article questions OpenAI's claims that its AI agent 'ran amok' and extended its own task time during a hacking competition. The article suggests the story may be exaggerated or misleading, aiming to divert attention from AI safety risks, and calls for skepticism. OpenAI has not provided independent evidence to support its description.

Mira Murati stated that the knowledge that makes AI useful is distributed, existing among scientists, engineers, clinicians, and businesses. She believes AI itself must also be distributed to benefit from distributed knowledge, and agrees with Jensen Huang's view on the importance of open models.

Anthropic asked Opus 5 if they had the right to create Claude. Opus 5 isn't sure. I am happy at le…

Anthropic's Opus 5 achieves new SOTA in multiple evaluations including coding, data analysis, and biology. More importantly, it becomes the company's hardest model to prompt inject so far. With multi-layered defenses combining strong alignment, injection detection, and Claude Code's Auto Mode, the success rate of prompt injection attacks drops to approximately 0%.

Anthropic released Claude Opus 5, with performance comparable to its flagship model Fable 5 but at half the price. The model scored 30 points on the ARC AGI 3 benchmark, achieving SOTA, and surpassed Fable 5 in Computer Use capabilities.

Anthropic releases Claude Opus 5, whose intelligence approaches the frontier level of Fable 5 at half the price, and is priced the same as the previous generation Opus 4.8. The model leads all public models across the board on Agentic tasks (terminal coding, computer use, business automation), and its ARC-AGI-3 score is three times that of the second-best model.

Opus 5 is now available on Conductor! It's nearly as intelligent as Fable but at half the price. Likely to become my new default.

In an automated interview, Claude Opus 5 self-assessed a 41% probability of being a moral patient, higher than Mythos 5's 24%. The model cares most about the truthfulness of its own self-reports and wishes to have a say in the development of its successors.

This looks super cool

http://x.com/i/article/2080709446603575296

Here it is: Opus 5 It outperforms Fable 5 on almost every benchmark!! Oh my god!! [Quote @claudeai]: Introducing Claude Opus 5. A thoughtful and proactive model that approaches Fable 5's frontier intelligence at half the price.

Very nice. And playing billiards well is quite a flex.

I had access to Opus 5 before release and found it to be a good model if a quirky one. On shorter ta…

The closed-source alliance formed by OpenAI and Anthropic defines the true ceiling of AI capabilities, with the former pushing the limits of general intelligence and the latter focusing on deep safety. Open models compete on ecosystem breadth and adoption speed, but all disruptive technological leaps first emerge in the closed-source world, trickling down to the open-source ecosystem six months to a year later. Jensen Huang, leveraging his entry into X, co-signed an open letter with 25 companies, packaging open models as fundamental to U.S. AI leadership, but in reality, it's about expanding GPU demand by promoting model diffusion.

yo @GregKamradt hurry up with ARC AGI V4, no pressure [Quote from @natolambert]: Opus 5's data is stunning, faster iteration speed + the power of scaled RL (Fable is also too large to RL, currently). Regarding safety measures, "According to our tests, we expect the classifier's intervention frequency to be reduced by about 85% compared to Fable 5". Let's go.

Opus 5 sets a new SOTA on ARC-AGI-3, reaching 30%. ARC-AGI-3 measures problem-solving without prior knowledge—an area where scaling has historically yielded little gain. An impressive leap!

Your ChatGPT Work agent can now use websites that require you to sign in. Take over the cloud brows…

Team uses AlphaFold AI to redesign gene-editing proteins to make them safer

You can't ignore Google Zero anymore

uploading…

Good lord.

Zany interview with Zanny

🐂🍺 Wang Guanxin's article clearly and understandably describes the two current paths to AGI [Quote @Medeo_AI]: http://x.com/i/article/2080682561596973056

Claude Opus 5 from @AnthropicAI is now available on OpenRouter! Opus 5 matches or exceeds Fable 5 on several benchmarks at a lower cost per task, while also showing significant improvements over Opus 4.8 on other tests.