From 21 items, 10 important content pieces were selected

1. Anthropic’s Mythos Preview Discovers AES Möbius Bridge via Chain of Thought ⭐️ 8.0/10
2. Claude AI Used to Discover Cryptographic Weaknesses ⭐️ 8.0/10
3. OlmoEarth: Planetary-Scale Geospatial AI Platform ⭐️ 8.0/10
4. Wall Street Giants Endorse the Clarity Act ⭐️ 8.0/10
5. Kimi Delta Attention Explained with Bra-Ket Notation ⭐️ 8.0/10
6. Ondo Abandons Public Blockchain for Private Trading Network ⭐️ 7.0/10
7. Coinbase STX Delayed Sends and Receives ⭐️ 7.0/10
8. SBCL 2.6.7 adds SIMD support for ARM64 and AVX512 ⭐️ 7.0/10
9. OpenAI Reports Scientists Use AI Coding Agents in Genomics ⭐️ 7.0/10
10. Google Expands Gemini API Managed Agents with 3.6 Flash, Hooks, Triggers ⭐️ 7.0/10

Anthropic’s Mythos Preview Discovers AES Möbius Bridge via Chain of Thought ⭐️ 8.0/10

Anthropic published a preview showing that Claude Mythos Preview, using chain-of-thought reasoning, autonomously discovered a new cryptographic technique called the AES Möbius Bridge, which accelerates the best known attack on 7-round AES-128 by 200 to 800 times. This breakthrough highlights the growing capability of AI models to perform independent cryptographic research, potentially advancing both AI interpretability and practical cryptography. It also raises important considerations about the security implications of AI-driven vulnerability discovery. The Möbius Bridge technique was discovered by Claude Mythos Preview over approximately three days of autonomous work, using about a billion output tokens and significant API compute. The resulting attack improves upon a previous 2^105 chosen-plaintext and 2^99-time algorithm by Derbez, Fouque, and Jean.

rss · Anthropic News · Jul 29, 01:28

Background: AES (Advanced Encryption Standard) is a widely used symmetric encryption algorithm. The AES S-Box is not a random permutation and has algebraic structure that can be exploited in cryptanalysis. Chain-of-thought reasoning is a technique where AI models break down complex problems into step-by-step intermediate steps, improving their ability to solve novel challenges.

Tags: #AI, #research, #cryptography, #chain-of-thought, #Anthropic

Claude AI Used to Discover Cryptographic Weaknesses ⭐️ 8.0/10

Anthropic researchers used Claude Mythos Preview to discover improved methods for attacking cryptographic algorithms, demonstrating advanced AI application in security research. This breakthrough shows that state-of-the-art AI can actively uncover vulnerabilities in foundational security systems, potentially influencing how cryptographic protocols are designed and evaluated. The discovery specifically targets cryptographic weaknesses in algorithms used to protect online data, leveraging Claude’s reasoning capabilities to find novel attack vectors.

rss · Anthropic News · Jul 29, 01:17

Background: Claude is a large language model developed by Anthropic, trained using Constitutional AI for safety. Cryptographic algorithms are mathematical methods that encrypt data to ensure privacy and security online. AI-assisted vulnerability discovery is an emerging field where models help identify potential weaknesses in code or protocols.

OlmoEarth: Planetary-Scale Geospatial AI Platform ⭐️ 8.0/10

Allen AI announced the OlmoEarth platform on November 4, 2025, an open end-to-end system for planetary-scale geospatial inference using frontier foundation models. This platform democratizes access to powerful geospatial AI for non-profits and NGOs, enabling real-time, decision-ready insights from multi-sensor Earth data at an unprecedented scale. OlmoEarth provides an end-to-end solution including data management, fine-tuning, embeddings, and production deployment, with source code, training data, and pre-trained weights openly available.

rss · Hugging Face Blog · Jul 29, 00:27

Background: Geospatial inference applies AI to extract insights from satellite imagery and other Earth observation data. Planetary-scale AI involves coupling large models with global sensing networks to understand and reason about the Earth in real time. OlmoEarth builds on recent foundation models for geospatial tasks, such as the Population Dynamics Foundation Model (PDFM) from Google.

Wall Street Giants Endorse the Clarity Act ⭐️ 8.0/10

BlackRock, Fidelity, and other major Wall Street firms have publicly endorsed the Clarity Act, a U.S. crypto regulatory bill that aims to define jurisdiction between the SEC and CFTC. This endorsement signals significant institutional support for regulatory clarity, which could accelerate mainstream adoption of digital assets and reduce legal uncertainty for market participants. The Clarity Act, while providing a framework, would still require months or years of rulemaking by the SEC and CFTC before full implementation, meaning immediate certainty is limited.

rss · CoinDesk · Jul 29, 00:08

Background: The Clarity Act is a bipartisan U.S. legislative proposal designed to resolve the jurisdictional dispute between the Securities and Exchange Commission (SEC) and the Commodity Futures Trading Commission (CFTC) over digital assets. It aims to classify cryptocurrencies as either securities or commodities, thereby clarifying regulatory oversight. The bill has gained renewed momentum following a compromise between lawmakers.

Kimi Delta Attention Explained with Bra-Ket Notation ⭐️ 8.0/10

A technical blog post by doubleword.ai explains the Kimi Delta Attention (KDA) mechanism using bra-ket notation, making the complex delta-rule linear attention algorithm more accessible to researchers and practitioners. This explanation lowers the barrier to understanding a state-of-the-art attention mechanism that improves upon architectures like Gated DeltaNet and Mamba, potentially accelerating adoption in efficient long-context AI inference. The blog post uses Dirac notation to illustrate how KDA stores summed outer products in a fixed-size state S_t instead of storing every past key and value, enabling efficient recurrent updates with channel-wise forgetting.

hackernews · AnhTho_FR · Jul 29, 00:02 · Discussion

Background: Kimi Delta Attention is a delta-rule linear attention mechanism that combines channel-wise forgetting with efficient chunkwise recurrent updates. It is part of the Kimi Linear architecture, which introduces a hybrid linear attention approach for scalable long-context processing. Bra-ket notation, originally from quantum mechanics, is a compact way to represent vectors and linear operators, and it is used here to simplify the mathematical exposition.

Discussion: Community comments express skepticism about the accessibility, with one user noting they could not have come up with KDA despite the explanation. Others appreciate the explicit notation explanation but debate potential LLM authorship of the post.

Ondo Abandons Public Blockchain for Private Trading Network ⭐️ 7.0/10

Ondo Finance has abandoned its plans to build a public blockchain for tokenized real-world assets, instead opting to develop a private, high-speed trading network aimed at institutional investors. This pivot signals a shift in the tokenization industry toward private, permissioned infrastructure that prioritizes speed and regulatory compliance over decentralization, potentially accelerating institutional adoption of digital assets. The new private network will focus on high-speed trading of tokenized assets, moving away from the public blockchain model that Ondo originally pursued. Details on the technology stack and timeline have not been disclosed.

rss · CoinDesk · Jul 29, 03:44

Background: Ondo Finance is a platform that tokenizes real-world assets (RWAs) such as U.S. Treasuries, offering yield-bearing products like USDY and OUSG. Asset tokenization uses blockchain to represent ownership of physical assets as digital tokens, enabling fractional ownership and automated trading. Many tokenization projects have struggled with public blockchain scalability and regulatory hurdles, leading some to explore private networks.

Tags: #tokenization, #Ondo, #blockchain, #institutional adoption, #trading network

Coinbase STX Delayed Sends and Receives ⭐️ 7.0/10

Coinbase identified an issue causing delayed sends and receives on the Stacks network for STX cryptocurrency, and a fix is being implemented as of July 28. This directly affects STX users on Coinbase who may experience delays in transferring funds, though buys, sells, and fiat transactions remain unaffected. It highlights ongoing operational challenges in layer-2 blockchain networks. The issue was first investigated at 07:59 PDT and then identified at 10:18 PDT on July 28. Coinbase assured users that funds are safe.

rss · Coinbase Status · Jul 29, 01:18

Background: Stacks is a layer-1.5 blockchain that connects to Bitcoin using Proof of Transfer (PoX) consensus and enables smart contracts via the Clarity language. STX is the native cryptocurrency of the Stacks network, used for fees and participation in decentralized applications. Coinbase is a major cryptocurrency exchange where users can buy, sell, and transfer STX.

SBCL 2.6.7 adds SIMD support for ARM64 and AVX512 ⭐️ 7.0/10

Steel Bank Common Lisp version 2.6.7 introduces SIMD support for ARM64 via the SB-SIMD contrib and adds AVX512 instruction support on x86-64. These improvements enable vectorized operations for better performance on modern hardware. This update significantly boosts the performance of Common Lisp programs on modern CPUs, making SBCL more competitive for computationally intensive tasks like scientific computing and data processing. It also reflects the continued evolution of Lisp implementations to leverage cutting-edge hardware features. The ARM64 SIMD support was contributed by Sylvia Harrington, while AVX512 support came from Robert Smith and Arthur Miller. Users must explicitly invoke SIMD intrinsics; the release does not mention auto-vectorization, so manual use is required.

hackernews · tmtvl · Jul 29, 01:11 · Discussion

Background: Steel Bank Common Lisp (SBCL) is a high-performance, open-source Common Lisp compiler. SIMD (Single Instruction, Multiple Data) allows a single instruction to process multiple data points simultaneously, crucial for performance in numerical and multimedia applications. AVX512 is Intel’s advanced 512-bit SIMD extension, offering significant speedups for parallel workloads. The name ‘Steel Bank’ originates from Carnegie Mellon University, referencing Andrew Carnegie (steel) and Andrew Mellon (banking).

Discussion: Community members are excited about the new SIMD features, with some inquiring about usage and whether auto-vectorization is supported. Others discussed historical naming origins and comparisons with other Lisp implementations like Clozure Common Lisp (CCL). One user requested documentation for the memory arena feature.

Tags: #Common Lisp, #SBCL, #SIMD, #Programming Languages, #Performance

OpenAI Reports Scientists Use AI Coding Agents in Genomics ⭐️ 7.0/10

OpenAI published a field report documenting how scientists are leveraging AI coding agents to modernize scientific computing, significantly accelerating software development and genomic discoveries. This report demonstrates a practical, high-impact application of agentic AI in a critical scientific domain, potentially speeding up research and enabling new discoveries in genomics and other fields. The report highlights the use of AI coding agents to automate and optimize computational workflows, reduce development time, and handle complex data analysis tasks in genomics research.

rss · OpenAI Blog · Jul 29, 01:00

Background: Agentic AI refers to AI systems that can pursue goals, use tools, and take actions with varying degrees of autonomy. In scientific computing, researchers often write custom code for simulations and data analysis. AI coding agents can assist by generating, debugging, and optimizing code, thereby accelerating the research pipeline.

Google Expands Gemini API Managed Agents with 3.6 Flash, Hooks, Triggers ⭐️ 7.0/10

Google announced new features for Gemini API Managed Agents, including support for the Gemini 3.6 Flash model, hooks, and triggers, enabling more flexible agent behaviors and event-driven integrations. These enhancements significantly improve the capabilities of managed agents, allowing developers to build more autonomous, event-responsive AI applications with lower latency and greater control. Gemini 3.6 Flash is a natively multimodal reasoning model optimized for speed; hooks and triggers allow agents to execute custom code at specific lifecycle points or respond to external events.

rss · Google AI Blog · Jul 29, 00:00

Background: Managed Agents in the Gemini API provide a config-driven, REST-first environment for building autonomous agents in a fully managed Linux sandbox. With a single API call, developers can spin up agents that reason, use tools, and execute code in isolated, ephemeral environments. The new features build on this foundation to enable more sophisticated agent workflows.

From 63 items, 11 important content pieces were selected

1. Moonshot Releases 2.8T Parameter Kimi K3 Open-Weight Model ⭐️ 9.0/10
2. Morgan Stanley launches ETH, SOL exchange-traded products ⭐️ 8.0/10
3. Core Scientific signs AMD AI deal, exits bitcoin mining ⭐️ 8.0/10
4. Apple Sued Over Fake Bitcoin Wallet After $875K Theft on App Store ⭐️ 8.0/10
5. Coinbase to Suspend ACX Trading July 28, 2026 ⭐️ 8.0/10
6. First Implicit Tactile World Action Model Trained on 500k Hours of Video ⭐️ 8.0/10
7. World’s First Agentic Diffusion Model with 128K Context ⭐️ 8.0/10
8. A $500 RL Fine-Tune of 9B Open Model Beats Frontier Models ⭐️ 8.0/10
9. Anthropic Publishes Stance on Open-Weights AI Models ⭐️ 8.0/10
10. Documentation for self-contained Python builds ⭐️ 8.0/10
11. macOS Tahoe 26.6 Security Content Released ⭐️ 7.0/10

Moonshot Releases 2.8T Parameter Kimi K3 Open-Weight Model ⭐️ 9.0/10

Moonshot has released the weights for their 2.8 trillion parameter Kimi K3 model on Hugging Face, with the 1.56TB checkpoint now available under a new license that restricts commercial use by large Model-as-a-Service providers. OpenRouter already offers K3 from seven providers at the same pricing as Moonshot. This release represents a major milestone in open-weight AI, as a 2.8 trillion parameter model is among the largest ever made publicly available. The shift from a modified MIT license to a more restrictive one for large commercial entities may set a precedent for how future open-weight models balance openness with commercial protection. The new license for K3 requires Model-as-a-Service businesses exceeding $20 million in annual revenue to enter a separate agreement with Moonshot, whereas previous K2 license only required attribution for entities with over 100 million MAUs or $20M monthly revenue. Moonshot consistently uses the term ‘open weight’ rather than ‘open source’ to describe this release.

rss · Simon Willison · Jul 28, 07:39

Background: Model weights are the learned parameters that define how an AI model processes inputs to produce outputs, typically resulting from training on large datasets. In large language models, parameters include billions or trillions of adjustable values that encode the model’s knowledge and behavior. Open-weight models allow users to download and run the weights locally, but may come with usage restrictions that distinguish them from fully open-source software.

Morgan Stanley launches ETH, SOL exchange-traded products ⭐️ 8.0/10

Morgan Stanley is launching exchange-traded products (ETPs) for ether (ETH) and solana (SOL), following the success of its earlier bitcoin fund. This marks a significant expansion of its institutional crypto offerings. This move by a major traditional financial institution like Morgan Stanley signals growing institutional acceptance of cryptocurrencies beyond bitcoin. It could increase mainstream investor access and market legitimacy for ETH and SOL. The ETPs will provide exposure to ether and solana, allowing investors to trade them on traditional stock exchanges. The exact structure (ETF or ETN) and launch date have not been specified in the provided content.

rss · CoinDesk · Jul 28, 22:30

Background: Exchange-traded products (ETPs) are investment vehicles traded on stock exchanges that track an underlying asset. Morgan Stanley’s previous success with a bitcoin fund paved the way for expanding into other cryptocurrencies. Solana is a high-throughput blockchain platform known for fast transactions and low fees, while ether is the native token of the Ethereum network, the leading smart contract platform.

Tags: #institutional adoption, #ETP, #Morgan Stanley, #ether, #solana

Core Scientific signs AMD AI deal, exits bitcoin mining ⭐️ 8.0/10

Core Scientific has signed a deal with AMD to provide AI infrastructure, while simultaneously winding down its bitcoin mining operations. This marks a major strategic pivot for one of the largest Bitcoin miners, signaling a broader industry shift from crypto mining to AI compute. The deal involves Core Scientific’s data centers being repurposed for AMD GPU-based AI workloads, as the company phases out ASIC-based bitcoin mining.

rss · CoinDesk · Jul 28, 20:46

Background: Core Scientific was originally a major player in Bitcoin mining, operating large-scale data centers with specialized ASIC miners. However, the company has been shifting focus to AI and cloud services, leveraging its data center expertise. This deal with AMD solidifies that transition.

Tags: #bitcoin mining, #artificial intelligence, #Core Scientific, #AMD, #industry pivot

Apple Sued Over Fake Bitcoin Wallet After $875K Theft on App Store ⭐️ 8.0/10

A new lawsuit alleges that Apple kept a fake Bitcoin wallet on the App Store even after being notified of an $875,000 theft from a user. The lawsuit claims Apple was negligent in its app review process. This incident underscores critical security flaws in Apple’s App Store vetting for cryptocurrency wallets, directly causing significant financial losses. It raises urgent questions about platform liability and user protection in the crypto space. The fake wallet mimicked legitimate Bitcoin wallet apps and reportedly led to a theft of $875,000. Apple allegedly ignored reports from the victim and continued hosting the app for some time.

rss · CoinDesk · Jul 28, 19:59

Background: Fake crypto wallets are malicious apps designed to steal users’ cryptocurrency by providing a backdoor to fraudsters. Apple claims to review all apps for security, but this case shows that fraudulent apps can still bypass checks. Users often rely on App Store’s reputation for safety, making such incidents particularly damaging.

Discussion: No community comments were provided in the input.

Tags: #Security, #Apple, #Fake Wallet, #Lawsuit, #Bitcoin

Coinbase to Suspend ACX Trading July 28, 2026 ⭐️ 8.0/10

Coinbase announced it will suspend trading of the ACX token on July 28, 2026, as the project team is winding down the token and offering conversion options to holders. The exchange has already moved ACX order books to limit-only mode ahead of the suspension. This suspension removes a major trading venue for ACX, affecting liquidity and access for holders who wish to trade or exit their positions. It highlights the risks of token wind-downs and the importance of following project communications for conversion opportunities. Trading on Coinbase will be suspended from July 28 to August 3, 2026, and holders must move their ACX to a self-custodial wallet to participate in the project’s conversion options. The ACX token is used for governance, staking, and fee payments within the Across Protocol ecosystem.

rss · Coinbase Status · Jul 28, 15:00

Background: The ACX token is the native token of Across Protocol, a cross-chain bridge that allows token transfers between different blockchains. It is used for governance voting, staking to become a relayer, and paying transaction fees. A token wind-down occurs when a project decides to discontinue its token, often offering holders a conversion to another asset or stablecoins. Limit-only mode means traders can only place limit orders, not market orders, which is a common step before a full trading suspension.

First Implicit Tactile World Action Model Trained on 500k Hours of Video ⭐️ 8.0/10

Researchers have trained the first implicit tactile world action model using 500,000 hours of video, enabling AI systems to understand and simulate tactile interactions. This marks a significant step in embodied AI and multimodal learning, allowing robots to manipulate objects and interact with the environment more naturally, with profound implications for automation, healthcare, and home robotics. The model employs implicit representation learning, likely combining world models with tactile sensing data, and was trained on an enormous dataset of 500,000 hours of video. Specific architectural details have not been publicly disclosed, but this is the first attempt to integrate tactile sensing into a world model.

rss · QbitAI 量子位 · Jul 28, 13:25

Background: A world model is an internal simulation of the environment that allows AI to predict consequences of actions. Tactile sensing technology aims to replicate human touch. Previous world models relied primarily on visual and textual inputs, lacking the tactile modality.

World’s First Agentic Diffusion Model with 128K Context ⭐️ 8.0/10

Researchers have released the first agentic diffusion model capable of real-time action and error correction, achieving a 128K context length that matches autoregressive models. This breakthrough extends diffusion models to long-horizon agent tasks, potentially enabling more robust and interactive AI agents that can plan and adapt without full retraining. The model supports 128K token context, matching the context window of leading autoregressive models, and incorporates an agentic loop that iteratively refines outputs.

rss · QbitAI 量子位 · Jul 28, 12:18

Background: Diffusion models traditionally excel at generating high-quality images and audio by gradually denoising random noise. Autoregressive models have dominated long-context tasks like text generation and planning. This work bridges the gap by enabling diffusion models to handle agentic tasks requiring step-by-step decision-making and error correction.

A $500 RL Fine-Tune of 9B Open Model Beats Frontier Models ⭐️ 8.0/10

A team fine-tuned a 9 billion parameter open-weight model using reinforcement learning for under $500, achieving better performance on catalog review tasks than much larger frontier models like GPT-4 or Claude. This demonstrates that small, fine-tuned models can be highly cost-effective for specific tasks, challenging the economic rationale for massive frontier models and the associated infrastructure spending. The fine-tuning process utilized reinforcement learning to optimize the model for catalog review accuracy. The specific model, dataset, and exact training cost breakdown were detailed in the blog post, though the author notes potential overfitting concerns.

hackernews · ilreb · Jul 28, 10:18 · Discussion

Background: Reinforcement learning fine-tuning (RLFT) is a technique where a language model is trained to maximize a reward signal based on its outputs, aligning it with specific goals. Catalog review involves automatically evaluating product descriptions, metadata, or listings for quality and consistency. Frontier models like GPT-4 and Claude are very large, general-purpose models that are expensive to train and run.

Discussion: Commenters largely praised the cost-efficiency demonstration, with some noting that most use cases don’t need massive models. However, concerns were raised about overfitting due to lack of held-out test data and that such specialized benchmarks don’t accurately compare to frontier models’ generalized capabilities.

Anthropic Publishes Stance on Open-Weights AI Models ⭐️ 8.0/10

Anthropic published an official statement outlining its position on open-weights AI models, advocating for mandatory safety testing for all sufficiently capable models rather than a blanket ban. The statement clarifies that Anthropic has never called for banning open-weights models but emphasizes the need for regulation as capabilities advance. This statement fuels the ongoing debate between open and closed AI development, as open-weights models lower barriers to access but raise safety concerns. Anthropic’s position could influence regulatory policies, especially as governments consider how to balance innovation and risk. Anthropic’s CEO Dario Amodei wrote the statement, which also supports banning chip sales to China and cracking down on smuggling of chips to prevent advanced AI capabilities from proliferating. The statement explicitly opposes a ban on open-weights models but supports mandatory safety testing that critics argue could effectively function as a ban.

hackernews · surprisetalk · Jul 28, 06:03 · Discussion

Background: Open-weights AI models are neural network models whose trained parameters (weights and biases) are publicly released for anyone to download, modify, and run. This contrasts with closed models where only an API is accessible. The open-weights approach enables wider access and customization but also raises concerns about misuse, as developers cannot control how the model is used after release.

Discussion: Commenters largely criticized Anthropic for perceived hypocrisy, pointing out contradictions between supporting chip bans while opposing open-weights bans, and questioning the feasibility of safety testing regimes. Some accused Anthropic of virtue signaling and protecting its commercial interests under the guise of safety concerns.

Documentation for self-contained Python builds ⭐️ 8.0/10

Documentation for python-build-standalone, a set of self-contained portable Python distributions, has been published, detailing how these builds are used by tools like uv, pipx, and Hatch to provide Python installations without system dependencies. This infrastructure simplifies Python deployment across platforms, enables bundling Python into desktop applications, and reduces friction for users who need a specific Python version without system-level installation. The distributions are maintained by Astral (the company behind uv), support many platforms (Linux, macOS, Windows), and are available for multiple Python versions. They are designed to be fully relocatable and have no external dependencies.

hackernews · jcbhmr · Jul 28, 02:43 · Discussion

Background: Standard Python distributions are often tied to a specific operating system or require compiling from source. python-build-standalone provides pre-built, relocatable Python binaries that can be extracted and used anywhere, making them ideal for packaging tools like package managers or embedding in applications. They are produced using a custom build process that strips out system-specific baggage.

Discussion: Comments highlight widespread adoption: charliermarsh (uv lead) confirms uv uses these distributions and notes significant engineering effort in keeping up with CPython. simonw praises them as excellent for bundling Python into apps, and zie points out an alternative approach using Cosmopolitan cross-platform binaries. rsyring mentions PyOxy as a sister project that creates single-file executables from these distributions.

macOS Tahoe 26.6 Security Content Released ⭐️ 7.0/10

Apple released security content for macOS Tahoe 26.6, addressing multiple vulnerabilities including memory handling issues. This update is important for macOS users as it fixes critical vulnerabilities that could lead to sandbox escapes and code execution. It also highlights the ongoing challenge of memory safety in operating systems. The security content includes fixes for path parsing vulnerabilities and other issues, with many patches citing improved bounds checking and memory management.

hackernews · andor · Jul 28, 17:45 · Discussion

Background: macOS security updates are released periodically to fix vulnerabilities. Memory safety issues are common in systems programming languages like C and C++, leading to bugs such as buffer overflows. Apple’s updates often include improved memory handling to mitigate such risks.

Discussion: Commenters noted the high number of memory-related fixes, with some debating the monetary cost of using unsafe languages like C. Others observed a similar update for macOS 15 and questioned the prevalence of path parsing vulnerabilities.

Tags: #security, #macOS, #Apple, #vulnerabilities, #memory safety