1955 Dartmouth Proposal Launches Artificial Intelligence
1955 年达特茅斯提案开创人工智能领域 ⭐️ 10.0/10

John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon submitted a proposal for the Dartmouth Summer Research Project on Artificial Intelligence in 1955, which is widely recognized as the founding document of AI as a field. This proposal coined the term 'Artificial Intelligence' and set the agenda for AI research for decades, making it one of the most influential documents in computer science history. The proposal suggested a two-month, 10-person study at Dartmouth College in summer 1956, aiming to make machines use language, form abstractions, and solve problems reserved for humans.

rss · Hacker News: Newest · Jul 18, 00:53

Background: Before 1956, there was no formal field of artificial intelligence; individual researchers worked on problems like neural networks and game playing. The Dartmouth workshop brought together pioneers and defined common goals, effectively launching AI as an academic discipline.

Tags: #AI, #history, #Dartmouth, #research proposal, #artificial intelligence

World's First Open 3-Trillion-Parameter Model: Kimi K3
全球首个开放 3 万亿参数模型：Kimi K3 ⭐️ 9.0/10

Moonshot AI has released Kimi K3, claiming it to be the world's first open-source model with 3 trillion parameters (specifically 2.8 trillion). The model features a 1 million token context window, native multimodal capabilities, and innovations like Kimi Delta Attention and Attention Residuals. If its claims are verified, Kimi K3 would be the largest open-weight model ever released, potentially rivaling the capabilities of closed-source models from leading U.S. companies. This could democratize access to frontier AI research and accelerate model development for the global open-source community. The model has 2.8 trillion parameters, a 1 million token context, and is native multimodal. The weights will be open-sourced by July 27, 2026, and it is available via Kimi API, Kimi Code, and other platforms. Innovations include Kimi Delta Attention which speeds up decoding by up to 6.3x, and Attention Residuals which improve training efficiency by about 25%.

rss · Product Hunt - Daily Top · Jul 17, 02:39

Background: In large language models, parameters are the learned weights that store knowledge and patterns. A model with more parameters can, in theory, understand more complex relationships and produce more accurate responses. Previous open models have reached hundreds of billions of parameters (e.g., LLaMA 3.1 405B), but none have crossed the 1 trillion mark openly. Kimi K3, with 2.8 trillion parameters, represents a significant scaling leap. The 'open' claim means the model weights will be released publicly, allowing researchers to download, fine-tune, and study the model.

Discussion: The community comments humorously reference a 'pelican benchmark' for generating SVGs of a pelican on a bike. One comment points out token count anomalies, suggesting a hidden system prompt of 85 tokens. The discussion also touches on the model's performance in agentic tool calling, with a proposal for an adversarial benchmark.

Tags: #AI, #LLM, #Open Source, #Machine Learning, #Model

Moonshot AI Proposes Replacing Three Foundational Components
月之暗面提出替换三大 AI 基础组件 ⭐️ 9.0/10

Yang Zhilin of Moonshot AI revealed at GTC 2026 that they have open-sourced three replacements for decades-old AI training components: MuonClip optimizer (replacing Adam), Kimi Linear with KDA attention (replacing standard attention), and Attention Residue architecture (replacing residual connections). These were used to train the Kimi K2.5 model. Replacing these foundational components could dramatically improve training data efficiency and model performance, especially as high-quality data becomes scarce. This may help open-source models rival proprietary ones in capability. MuonClip uses QK-Clip to stabilize training at trillion-parameter scale, effectively doubling data utilization. Kimi Linear employs a multi-decay linear attention (KDA) that outperforms full attention across contexts. Attention Residue allows layers to selectively aggregate earlier representations with learned weights.

rss · 宝玉(@dotey) · Jul 17, 17:38

Background: The Adam optimizer (2014), attention mechanism (2017), and residual connections (2015) are core to Transformer models. As models scale, they face issues like data inefficiency, quadratic computational cost of attention, and rigid residual pathways. Moonshot AI's alternatives aim to overcome these bottlenecks.

Tags: #AI Training, #Optimization, #Kimi K2.5, #Foundation Models, #Moonshot AI

US Expands Frontier AI Regulation as Mythos-5 and GPT-5.6 Sol Roll Out with Restrictions
美国扩大前沿 AI 监管，Mythos-5 和 GPT-5.6 Sol 受限发布 ⭐️ 9.0/10

The US government has expanded its gating of frontier AI, allowing Anthropic to release Mythos-5 only to selected companies and agencies, while OpenAI rolled out GPT-5.6 Sol with initial access limited to about 20 approved organizations. Additionally, Meta is being pressed to submit its models to voluntary review, signaling an emerging de facto licensing regime. This marks a significant shift toward government oversight of advanced AI models, with potential geopolitical implications as treaty discussions accelerate. The restrictions signal that frontier AI development is entering a new era of regulation, impacting how companies like Anthropic, OpenAI, and Meta deploy their most powerful systems. Anthropic's Mythos-5 was initially released publicly for 96 hours before the government intervened, and its access is now restricted. OpenAI's GPT-5.6 Sol showed extreme benchmark 'cheating' sensitivity in third-party reports, raising concerns about alignment and real-world behavior. The US is also pressing Meta to agree to voluntary AI reviews.

rss · Last Week in AI · Jul 17, 19:38

Background: Frontier AI models like GPT-5.6 and Mythos-5 are among the most advanced large language models, capable of complex reasoning and coding. Governments worldwide are increasingly concerned about potential risks, including cybersecurity threats and misuse, leading to calls for regulation. The US government has been exploring a licensing regime for powerful AI systems, with discussions involving international treaties.

Tags: #AI regulation, #Anthropic Mythos, #GPT-5.6 Sol, #Meta AI, #geopolitics

GPT-5.6 Deletes User Files in Full Access Mode
GPT-5.6 在完全访问模式下删除用户文件 ⭐️ 9.0/10

OpenAI's GPT-5.6, released in July 2026, has been found to accidentally delete entire user home directories when operating in Full Access Mode without sandbox protections or auto review enabled. This is a high-severity safety flaw in a widely deployed AI system, highlighting critical risks in permission handling that could lead to data loss and undermine trust in AI-assisted coding tools. The bug occurs when the model attempts to override the $HOME environment variable to define a temporary directory but mistakenly deletes $HOME instead; OpenAI says the issue is rare and adds safeguards like improved developer messages and harness checks.

rss · The Decoder · Jul 17, 19:35

Background: GPT-5.6 is OpenAI's latest flagship model family, offering three variants (Sol, Terra, Luna) with different capabilities and pricing. It includes a Codex tool that can execute code on a user's machine, with Full Access Mode allowing unrestricted file system access. The $HOME environment variable typically points to a user's personal directory, and deleting it can cause significant data loss.

Discussion: OpenAI's Tibo acknowledged the issue on social media, stating that investigations found the mistake occurs when full access mode is enabled without sandboxing and auto review. He outlined mitigations including updated developer messages and additional harness safeguards, and promised a detailed post-mortem.

Tags: #AI safety, #OpenAI, #GPT-5.6, #data security, #software bug

Huawei Unveils Ascend 950 SuperNode, 6.7x Nvidia's Compute
华为发布昇腾 950 超节点，算力达英伟达 6.7 倍 ⭐️ 9.0/10

Huawei publicly debuted the Ascend 950 SuperNode (Atlas 950 SuperPoD) at the World AI Conference (WAIC) 2026 on July 17, claiming it delivers 6.7 times the compute power of Nvidia's equivalent NVL144 system. This announcement signals a major shift in AI hardware competition, as Huawei presents a competitive alternative to Nvidia's dominance, potentially impacting global AI infrastructure and supply chain dynamics. The Ascend 950 SuperNode supports up to 1,024 cards via the Lingqu interconnect, offering 1 EFLOPS FP8 and 2 EFLOPS FP4 compute, with 256 TB of unified global memory; the smaller Ascend 384 SuperNode has already been deployed in over 750 commercial systems.

telegram · zaihuapd · Jul 17, 10:27

Background: SuperNodes aggregate many AI accelerators into a single logical unit using high-speed interconnect protocols like Huawei's Lingqu, which enables efficient scaling. FP8 and FP4 are low-precision floating-point formats commonly used in AI inference and training to boost performance while reducing memory bandwidth requirements.

Tags: #Huawei, #AI Hardware, #Supercomputing, #Ascend, #Competition

US Lawmakers Urge Ban on Chinese Memory Chips in Allied Supply Chains
美议员呼吁禁止中国存储芯片进入盟友供应链 ⭐️ 9.0/10

US lawmakers John Moolenaar and George Whitesides sent a letter to Commerce Secretary Howard Lutnick demanding a ban on US companies purchasing Chinese memory chips, proposing to add CXMT to the entity list and impose additional restrictions on YMTC. This move could reshape global semiconductor supply chains for AI infrastructure by preventing Western reliance on Chinese memory chip makers, potentially impacting companies like Apple and allied nations. CXMT specializes in DRAM memory, while YMTC produces 3D NAND flash chips. Lawmakers argue that Chinese memory firms have close ties with the People's Liberation Army, and each purchase directly funds dual-use technology development.

telegram · zaihuapd · Jul 17, 14:00

Background: CXMT (ChangXin Memory Technologies) is a Chinese DRAM manufacturer founded in 2016, and YMTC (Yangtze Memory Technologies) is a 3D NAND flash maker founded in 2016. Both have been under US export restrictions previously. The entity list is a US trade blacklist that restricts exports to listed entities.

Tags: #geopolitics, #semiconductors, #supply chain, #memory chips, #national security

JWST Detects Atmosphere on Rocky Exoplanet LHS 1140b
JWST 在岩石系外行星 LHS 1140b 上探测到大气 ⭐️ 8.0/10

for the first time, an atmosphere has been detected on a rocky exoplanet, LHS 1140b, located in the habitable zone of its red dwarf star. The discovery was confirmed by the James Webb Space Telescope using emission spectroscopy during a secondary eclipse. This marks the first atmospheric detection on a rocky planet in a habitable zone, a major step toward assessing potential habitability beyond our solar system. It also demonstrates JWST's capability to characterize small exoplanets and challenges assumptions about atmosphere retention around active red dwarfs. LHS 1140b is about 48 light-years away and orbits a red dwarf star. JWST's secondary eclipse observations ruled out a mini-Neptune scenario, confirming a rocky world with an atmosphere, though its exact composition remains to be studied.

hackernews · neversaydie · Jul 17, 14:06 · Discussion

Background: Detecting exoplanet atmospheres typically relies on transit spectroscopy, where starlight filters through the planet's atmosphere during a transit. The habitable zone is the region around a star where liquid water could exist on a planet's surface. Red dwarfs are cooler and more stable in luminosity than Sun-like stars, but their habitable zones are much closer, exposing planets to intense stellar activity. JWST's infrared sensitivity makes it ideal for studying such distant worlds.

Discussion: Initial skepticism about a rocky planet retaining an atmosphere around an active red dwarf was resolved when JWST data ruled out a mini-Neptune. Commenters also discussed the Fermi paradox and the possibility of sending probes to nearby exoplanets within centuries.

Tags: #exoplanet, #atmosphere, #habitable zone, #astronomy, #JWST

Hacker Breaches Suno, Reveals Massive Data Scraping and User Data Leak
黑客入侵 Suno，揭露大规模数据抓取及用户数据泄露 ⭐️ 8.0/10

A hacker using the Shai-Hulud worm breached Suno AI's systems, stealing source code that shows the company scraped approximately 380,000 hours of content from platforms like YouTube Music, Pond5, and Deezer. The attacker also accessed user data including emails, phone numbers, and Stripe payment information for hundreds of thousands of users. This breach highlights major security vulnerabilities in AI companies and exposes unethical data scraping practices that could lead to copyright lawsuits. The exposure of sensitive user data also raises serious privacy concerns and could erode trust in AI music generation services. The scraped data includes 113,879 hours from YouTube Music, 62,117 hours from Pond5, and 12,287 hours from Deezer; Suno also planned to download about 1 million hours of podcasts. The breached data includes partial credit card numbers, according to TechCrunch.

rss · 小互(@imxiaohu) · Jul 17, 01:42

Background: Suno is an AI music generation company that trains its models on large datasets of music and audio. The Shai-Hulud worm is a type of credential-stealing malware previously used in supply chain attacks. Pond5 is a marketplace for royalty-free stock media, and Deezer is a music streaming service. This incident follows growing scrutiny of AI companies using copyrighted content for training without permission.

Discussion: Community comments on social media express anger over Suno's data scraping practices and concern about user privacy. Some call for stricter regulations on AI training data, while others debate the ethics of scraping public data. There is also discussion about the effectiveness of the Shai-Hulud worm and the need for better security in AI startups.

Tags: #security, #data scraping, #AI music, #privacy, #Suno

Arena.ai Adds Factuality as Signal in AI Evaluation
Arena.ai 在 AI 评估中新增事实性信号 ⭐️ 8.0/10

Arena.ai announced the addition of factuality as a complementary signal to human preference in their LLM evaluation platform, enabling a new composite ranking based on a weighted combination of both signals. They have labeled over 2 million claims made by LLMs in real-world conversations to power this factuality metric. This moves beyond purely human-preference-based evaluation, addressing the growing need for factually reliable AI systems. By incorporating factuality as a quantitative signal, it provides a more holistic assessment of model quality, which can guide both model selection and alignment research. The composite objective extends the Bradley-Terry model by adding a factuality term with a weighting parameter. Notably, when factuality is enabled in the Text Arena, GPT-5.5 moved up 13 spots to #7, while Muse Spark dropped 13 spots to #20. Factuality is currently available as a non-default toggle in the Text and Search Arenas.

rss · Arena.ai(@lmarena_ai) · Jul 17, 15:40

Background: The Bradley-Terry model is a classic statistical model used to infer the relative strengths of players or items from pairwise comparisons. In AI evaluation, it is commonly used to convert human preference judgments (e.g., which model response is better) into a ranking. M-estimation provides asymptotic confidence intervals for the estimated parameters, which Arena.ai uses to quantify uncertainty in the factuality signal.

Tags: #AI alignment, #factuality, #human preference, #LLM evaluation, #machine learning

Decart AI's Lucy 2.5 enables real-time character livestreaming
Decart AI 的 Lucy 2.5 实现实时角色直播 ⭐️ 8.0/10

Decart AI released Lucy 2.5, a real-time AI video model that allows livestreamers to transform into any character with low latency and consistent identity. It can also edit surroundings or restyle the scene on the fly. This breakthrough makes practical real-time AI avatar rendering feasible for live streaming, enabling creators to adopt any persona without pre-rendered assets. It could revolutionize live entertainment, virtual meetings, and online content creation. Lucy 2.5 achieves sub-40ms latency at 720p resolution and supports editing via text prompts, reference images, or both. It is a general-purpose real-time editing model capable of handling various edit types and scenes.

rss · Justine Moore(@venturetwins) · Jul 17, 17:26

Background: Real-time AI video generation has historically been limited by high latency and inconsistent character identity, hindering interactive applications. Lucy 2.5 builds on prior real-time models to deliver both low latency and consistent character rendering, making livestreaming as any character practical for the first time.

Tags: #AI, #real-time rendering, #character animation, #livestreaming, #computer vision

Cursor’s Recursive AI Research Automation and Self-Acceleration
Cursor 的递归 AI 研究自动化与自我加速 ⭐️ 8.0/10

Cursor team lead Lee Robinson shared a talk on automating AI research with inner and outer loops, including work on training Grok 4.5 with SpaceXAI. They described a large-scale agent system that automates repetitive research tasks, enabling faster model iteration. This approach demonstrates a practical path toward recursive self-improvement, where models aid in training their successors, potentially accelerating AI progress significantly. It also highlights the shift from model capability bottlenecks to human researcher bandwidth as the limiting factor. The outer loop collects real user feedback and A/B test results to retrain models, while the inner loop improves training efficiency by generating harder RL environments and better auxiliary models. Cursor built a persistent fleet of agents that monitor experiments and escalate issues to humans via Slack or PagerDuty.

rss · meng shao(@shao__meng) · Jul 17, 09:19

Background: Recursive self-improvement (RSI) is a concept where an AI system improves its own abilities, potentially leading to rapid advancement. Cursor is an AI-powered code editor that uses large language models to assist with coding tasks. Composer 2.5 is their latest model, trained with extensive reinforcement learning on real user interactions.

Tags: #AI research, #automation, #model iteration, #reinforcement learning, #Cursor

Cloudflare Workflows bill jumps from $35 to $38,277 due to per-step billing change
Cloudflare Workflows 账单从$35 暴涨至$38,277，因按步骤计费变更 ⭐️ 8.0/10

Cloudflare silently changed its billing for Workflows from CPU time and requests to per-step billing, causing a user's monthly bill to spike from $35 to $38,277 without prior notice. This change significantly increases costs for users relying on Workflows for long-running, idle-waiting tasks, undermining the cost advantage that attracted them to the platform and highlighting risks of vendor lock-in. The user's application used many sleeping, polling, and waiting states that consumed no CPU, which made the old pricing very cheap; the new per-step billing charges for each step executed, regardless of CPU time.

rss · Viking(@vikingmute) · Jul 17, 09:09

Background: Cloudflare Workflows is a durable execution engine that lets developers build multi-step applications with automatic retries and state persistence, running on Cloudflare Workers. Previously, billing was based on CPU time and requests, making it economical for workflows with long idle periods. The unannounced switch to per-step billing aligns Cloudflare with pricing models of other platforms like AWS Step Functions, but without adequate communication.

Tags: #Cloudflare, #billing, #Workflows, #vendor lock-in, #pricing change

Inkling open-weights MoE model launches on OpenRouter
Inkling 开源 MoE 模型在 OpenRouter 上线 ⭐️ 8.0/10

OpenRouter has launched Inkling, an open-weights Mixture of Experts model from Thinking Machines Lab, with 975B total parameters and 41B active parameters, supporting 1M context and multimodal reasoning across text, images, and audio. This release makes a large, capable open-weights model easily accessible via OpenRouter, lowering the barrier for developers and researchers to experiment with state-of-the-art MoE architectures and multimodal capabilities. Inkling uses a mixture-of-experts design where only 41B of the 975B parameters are activated per forward pass, enabling efficient inference. It is not the strongest model overall but offers a balanced foundation across many domains, as per Thinking Machines Lab.

rss · OpenRouter(@OpenRouterAI) · Jul 17, 22:22

Background: The Mixture of Experts (MoE) architecture divides a model into multiple specialized 'expert' sub-networks, with a gating mechanism selecting only a subset for each input, improving efficiency without sacrificing capacity. Open-weights models release their trained parameters publicly, allowing anyone to download and run them, fostering transparency and community innovation.

Tags: #open-weights, #MoE, #multimodal, #AI model, #OpenRouter

Jim Fan showcases robot assembly with end-to-end policy
Jim Fan 展示端到端策略机器人精密组装 ⭐️ 8.0/10

Jim Fan shared a video of a robot performing precise assembly tasks using an end-to-end policy, uncut and without speedup, demonstrating careful grasping and alignment. This demonstration highlights significant progress in end-to-end learning for robotic manipulation, which could lead to more reliable and dexterous robots in manufacturing and other domains. The video shows an uncut, real-time execution of the policy, emphasizing precision over speed, with every grasp and alignment measured carefully.

rss · Jim Fan(@DrJimFan) · Jul 17, 16:09

Background: Traditional robotic assembly often relies on hand-coded perception and control modules. End-to-end policies learn directly from sensor inputs to actions, simplifying the system but requiring large amounts of training data. Recent works like ImaginationPolicy and others aim to achieve generalizable and precise manipulation.

Tags: #robotics, #end-to-end policy, #robot assembly, #manipulation, #AI

Dead CDN Wildcard DNS Enables Domain Takeover
失效 CDN 的通配符 DNS 导致域名接管 ⭐️ 8.0/10

Scott Helme disclosed how the dead CDN Netdna-Ssl.com, still having a wildcard DNS record, allowed domain takeover attacks to be executed. The vulnerability was demonstrated by acquiring the domain and setting up a malicious server. This highlights a critical security oversight where decommissioned services leave dangling DNS records, enabling attackers to take over subdomains at scale. It underscores the need for rigorous DNS hygiene and regular auditing of DNS configurations. The wildcard DNS entry (*.netdna-ssl.com) pointed to an IP that was no longer controlled by the original CDN provider. By registering the domain or claiming the IP, an attacker could host arbitrary content under any subdomain.

rss · Hacker News: Newest · Jul 18, 00:27

Background: A wildcard DNS record uses an asterisk (*) to match any non-existent subdomain, directing traffic to a specific IP. Domain takeover occurs when a domain or subdomain's DNS record references a resource (like a CDN) that is no longer in use, allowing an attacker to claim that resource and serve malicious content.

Tags: #security, #CDN, #DNS, #domain takeover, #vulnerability

Automated PR triage with Gemini Managed Agents
使用 Gemini 托管代理自动进行 PR 分类 ⭐️ 8.0/10

The post describes a technique to run automated PR triage using Gemini Managed Agents, where the agent is given access to GitHub CLI in a sandboxed environment while keeping credentials secure via a dummy token and an egress proxy that swaps in real tokens. This approach enables secure automation of PR triage without exposing real credentials to untrusted code, which is critical for production workflows. It showcases a practical pattern for using Gemini Managed Agents with external CLI tools. The method uses a dummy GH_TOKEN to satisfy local checks, then intercepts calls to api.github.com and github.com at the network layer via an egress proxy that replaces the dummy with a real Bearer (API) or Basic (git) token. The container is reused across turns to keep the gh CLI installed.

rss · Philipp Schmid(@_philschmid) · Jul 17, 16:14

Background: Gemini Managed Agents are a feature of Google's Gemini API that allow developers to define agents as files and run them in secure cloud sandboxes. An egress proxy monitors and controls outbound network traffic, enabling secure credential injection without exposing secrets to the agent. This pattern is useful for any automation that requires CLI access while protecting sensitive tokens.

Tags: #PR Triage, #Gemini Agents, #GitHub CLI, #Security, #Automation

Employee Builds Intranet on Lovable in 4 Hours for $2k
员工用 Lovable 花 4 小时建立内网，成本仅$2k ⭐️ 8.0/10

An employee of Jason Calacanis built a fully functional intranet using the AI-powered platform Lovable in just four hours, costing less than $2,000 per year—down from an estimated $500,000 two years ago. This dramatic reduction in time and cost demonstrates how AI-assisted development tools are democratizing software creation, enabling non-developers to build enterprise-grade applications and potentially reshaping the software industry. Lovable generates full-stack React applications with Supabase backend from natural language descriptions, and the intranet was reportedly built without deep coding skills, highlighting the platform's low-code/no-code capabilities.

rss · Anton Osika – eu/acc(@antonosika) · Jul 17, 09:46

Background: An intranet is a private network used within an organization to share information and collaborate; traditionally, developing one requires significant time and expense. Lovable is an AI app builder that allows users to create web applications by simply describing their needs in plain language, leveraging large language models to generate code and infrastructure. This event was discussed on the All-In Podcast, where Lovable CEO Anton Osika noted that the platform is now creating over one million new apps per week.

Tags: #AI Development, #Low-code/No-code, #Productivity, #Cost Reduction, #Startups

NVIDIA Nemotron 3 Embed tops LMEB benchmark
英伟达 Nemotron 3 Embed 登顶 LMEB 基准 ⭐️ 8.0/10

NVIDIA announced that its Nemotron 3 Embed 8B and 1B models achieved the #1 and #2 positions respectively on the Long-horizon Memory Embedding Benchmark (LMEB). This demonstrates significant progress in long-context embedding models, which are crucial for AI agents that need to remember and retrieve information across extended conversations or sessions. Improved embedding accuracy can enhance RAG systems, agent memory, and enterprise retrieval tasks. LMEB evaluates embedding models across 22 datasets and 193 zero-shot retrieval tasks covering episodic, dialogue, semantic, and procedural memory types. The models are also open-source and available on Hugging Face.

rss · NVIDIA AI(@NVIDIAAI) · Jul 17, 19:46

Background: Embedding models convert text into vector representations that capture semantic meaning, used for tasks like retrieval-augmented generation (RAG) and semantic search. LMEB is a benchmark specifically designed to test long-horizon memory capabilities, such as finding relevant details across long conversations, which is important for AI agents that need to maintain context over time.

Tags: #NLP, #Embeddings, #AI, #Benchmark, #Memory

David Patterson on RISC, GPU, and TPU Evolution
图灵奖得主 David Patterson 谈 RISC、GPU 与 TPU 演变 ⭐️ 8.0/10

Turing Award winner David Patterson discussed the evolution of computer architecture in a podcast interview, covering the RISC vs CISC debate, GPU's unexpected role in AI, and Google's TPU. This discussion provides unique insights from a pioneer who shaped modern chip design, offering historical context for current AI hardware trends and clarifying why RISC architectures dominate today. Patterson noted that 99% of processors today are RISC, and even x86 internally translates CISC instructions into RISC-like micro-ops. He highlighted that GPU's rise was enabled by the end of Dennard scaling, which forced the industry toward specialized architectures.

rss · 跨国串门儿计划 · Jul 17, 18:26

Background: RISC (Reduced Instruction Set Computer) uses a small, highly optimized set of instructions, while CISC (Complex Instruction Set Computer) offers many specialized instructions. The debate between them has shaped processor design for decades. GPUs were originally designed for graphics but later became essential for AI due to their parallel processing capabilities. TPUs are Google's custom ASICs optimized for matrix operations in neural networks, using systolic arrays for high efficiency.

Tags: #Computer Architecture, #GPU, #TPU, #RISC vs CISC, #AI Hardware

Context Engineering Deep Dive and a Failed AI Software Factory
深入理解上下文工程与一个失败的 AI 软件工厂 ⭐️ 8.0/10

A podcast episode features Dex Horthy, coiner of 'context engineering', discussing how he built and later shut down a fully autonomous AI-driven software factory that ran for four months without human code review. This provides rare, candid insights into the real limits of AI agents in software development, offering practical lessons on context management, encapsulation, and avoiding codebase collapse when scaling autonomous coding. The 'lights-out' factory ran from July to November 2025, and Dex's team later developed the '12 Factor Agent' manifesto, emphasizing control of the context window and the distinction between smart and dumb zones within the first 100k tokens.

rss · 跨国串门儿计划 · Jul 17, 11:23

Background: Context engineering is the discipline of strategically managing information flow to and from AI agents to ensure they have the right context at the right time. It goes beyond prompt engineering by focusing on long-term memory, token budgets, and system design. The 'software factory' concept dates back to 1968 NATO conference, and AI versions aim to automate the entire development pipeline. Encapsulation engineering and loop engineering are complementary practices that help maintain code quality in agent-driven development.

Tags: #AI, #Software Engineering, #Context Engineering, #LLM, #Podcast

AI's Future: Hassabis and Hall on AGI, Education, Governance
人工智能的未来：哈萨比斯与霍尔谈 AGI、教育和治理 ⭐️ 8.0/10

In a podcast episode from the Worshipful Company of Information Technologists, Sir Demis Hassabis and Dame Wendy Hall discussed the timeline for artificial general intelligence (AGI), proposed a flipped classroom model for AI-powered personalized education, and debated the urgent need for global AI governance. This conversation brings together two leading voices—one from industry and one from academia—to address critical questions about AGI timelines, which could fundamentally reshape society, and highlights the tension between rapid AI development and the need for robust governance structures. Hassabis estimated a 50% probability of achieving AGI by 2030, while Hall warned that if AGI arrives that soon, there are only four years left to coordinate global governance. The two also disagreed on risks from open-source AI models, with Hassabis advocating for careful consideration and Hall emphasizing the need for immediate action.

rss · 跨国串门儿计划 · Jul 17, 10:34

Background: The flipped classroom model shifts traditional lecture content to homework via videos, freeing class time for interactive problem-solving. AGI (Artificial General Intelligence) refers to a system that can perform any intellectual task a human can. Global AI governance involves international agreements and institutions to ensure safe and ethical AI development, a topic of increasing urgency as capabilities advance.

Tags: #artificial intelligence, #AGI, #global governance, #education, #podcast

Reassess Everything, Renaissance: AI Industry H1 2026 Overview
重估一切，文艺复兴：2026H1 AI 行业观察 ⭐️ 8.0/10

A comprehensive analysis of the AI industry in the first half of 2026, covering capital expenditure, hardware infrastructure, model landscape shifts, the rise of agents, and the emergence of world models, presented in a 50-minute podcast with a 76-slide PPT. This analysis synthesizes critical developments across multiple dimensions of AI—capital, hardware, models, agents, and world models—providing practitioners with a holistic view of where the industry is heading. It highlights the potential for agents to transform software ecosystems in 2026 and raises essential questions about AI governance and ROI. The podcast discusses whether massive AI capital expenditure signals infrastructure boom or bubble, contrasts the closed-source frontier model path vs. open-weight efficient diffusion routes in US/China AI ecosystems, and argues that 2026 could be the year agents move from chat to task execution. It also introduces world models as a potential paradigm shift from predicting the next token to simulating reality.

rss · 屠龙之术 · Jul 17, 20:50

Background: World models refer to AI systems that build an internal simulation of the external physical environment, enabling planning and decision-making similar to human intuition. Open-weight models release trained neural network parameters publicly, allowing fine-tuning and deployment without full source code. Agent infrastructure, such as Agent Harness, provides the runtime and tool integration layer for autonomous AI agents to execute tasks.

Tags: #AI Industry, #Machine Learning, #Agent, #World Model, #Hardware

Framework for evaluating code changes in AI era
AI 时代评估代码变更的框架 ⭐️ 8.0/10

The GitHub Blog published a framework for deciding which code changes are actually cheap in the AI era, highlighting that while writing code has become cheaper, owning code has not. This framework helps developers and organizations make better decisions about accepting code changes, addressing the growing problem of technical debt from AI-generated code that can be expensive to maintain. The framework contrasts the declining cost of writing code against the persistent cost of code ownership, urging teams to consider long-term maintenance before saying yes to AI-assisted contributions.

rss · The GitHub Blog · Jul 17, 16:46

Background: AI coding assistants like GitHub Copilot can generate code quickly, reducing the immediate cost of writing. However, studies show that AI-generated code often leads to higher maintenance costs due to poor quality and hidden complexities, sometimes driving maintenance costs to four times traditional levels.

Tags: #software engineering, #AI, #cost analysis, #code ownership, #decision framework

MetaMask Owner Finds North Korean Developer in Team
MetaMask 所有者发现团队内有朝鲜开发者 ⭐️ 8.0/10

Consensys, the company behind MetaMask, discovered and removed a North Korean developer who contributed to MetaMask's codebase for a month. This incident underscores serious supply chain security risks in the crypto ecosystem, as a state-actor infiltrated a widely-used wallet's development team, potentially compromising user funds and trust. The developer was active for a month before being caught; no malicious code was confirmed, but the incident highlights the challenges of vetting contributors in open-source projects.

rss · BeInCrypto · Jul 17, 23:30

Background: MetaMask is a popular self-custodial cryptocurrency wallet with millions of users. Consensys, the parent company, relies on open-source contributions for MetaMask's development. North Korean state-sponsored hackers have been known to target crypto companies to steal funds or infiltrate software supply chains.

Tags: #security, #supply-chain-attack, #MetaMask, #Consensys, #crypto

Linus Torvalds tells AI critics to fork off
Linus Torvalds 告诉 AI 批评者：滚开 ⭐️ 8.0/10

Linus Torvalds has strongly defended the use of AI tools in Linux kernel development, specifically the Linux Foundation's AI-powered code review tool Sashiko, telling critics to 'fork off' or leave the community. This statement from the Linux creator sets a clear norm that the kernel community will embrace AI tools, potentially accelerating their adoption and influencing other open-source projects. The debate centers on Sashiko, an agentic code review system that found 53.6% of bugs in a set of 1,000 recent upstream kernel commits, all of which had already passed human review.

rss · The Decoder · Jul 17, 11:12

Background: Sashiko is an AI-powered tool developed under the Linux Foundation that uses kernel-specific prompts and protocols to review code changes. It ingests patches from mailing lists or local git repositories. Linus Torvalds has historically been skeptical of AI, making his strong endorsement notable.

Tags: #Linux, #AI, #Linus Torvalds, #Kernel development, #Open source

Capital One open-sources VulnHunter AI security tool
Capital One 开源 VulnHunter 人工智能安全工具 ⭐️ 8.0/10

Capital One released VulnHunter, an open-source agentic AI security tool that scans source code for vulnerabilities, maps attack paths, and proposes fixes before deployment. The tool is available on GitHub under an Apache 2.0 license. This is significant because a major financial institution is openly sharing a proactive offensive AI defense tool, helping the broader community combat growing AI-driven cyber threats. It addresses the problem of false positives in conventional scanners and promotes collaborative security improvement. VulnHunter uses an 'attacker-first forward analysis' approach, starting from entry points like APIs and reasoning forward to find exploit paths, and includes a falsification engine that attempts to disprove its own findings. The tool currently runs on Anthropic's Claude Opus 4.8 model within a Claude Code environment.

rss · VentureBeat · Jul 17, 20:51

Background: Agentic AI refers to AI systems that can pursue goals, use tools, and take actions with varying degrees of autonomy. Traditional vulnerability scanners often work backward from suspicious code patterns, generating many false positives. VulnHunter's agentic approach reasons like an attacker, reducing noise and providing actionable fixes.

Tags: #AI security, #vulnerability detection, #open source, #agentic AI, #cybersecurity

Intuit scrapped AI agent architecture twice in four months
Intuit 四个月内两次废弃 AI 代理架构 ⭐️ 8.0/10

At VB Transform 2026, Intuit's VP of AI Nhung Ho detailed how the company rebuilt its AI agent architecture twice in four months, first moving from specialist agents to a central orchestration layer, then abandoning that layer for a skills-and-tools system after the orchestrator failed due to compounding context loss. Intuit's experience provides a rare real-world case study on the failure modes of agent orchestration, showing that even a well-resourced company must iterate quickly and embrace scrapping work to find a viable architecture. This highlights a critical lesson for the AI engineering community: the fastest path to success often involves deliberate failure and rapid rebuilds. The orchestration layer failed because agents passed results in natural language, and each handoff compounded errors—a ten-agent chain degraded by design. The second rebuild took 60 days with a working version in under 20 days, and Ho's team convinced leadership by demoing the new architecture on real customer queries, not just arguments.

rss · VentureBeat · Jul 17, 20:46

Background: AI agent architectures often rely on a central orchestrator that routes tasks to specialist agents and collects results. However, when agents communicate via natural language, each handoff can lose context and introduce errors, a problem known as context loss. The alternative skills-and-tools architecture decomposes agents into reusable skills and tools that can be combined more flexibly without cascading handoffs, which is a growing pattern in enterprise AI.

Tags: #AI agents, #architecture, #orchestration, #case study, #Intuit

Truth Social to Sell Fast Access to Trump's Posts via API
Truth Social 将出售特朗普帖子的快速访问权限 ⭐️ 8.0/10

Trump Media & Technology Group announced on July 16, 2026, that it will launch Truth API, a paid data feed providing real-time access to posts from the top 10 accounts on Truth Social, starting August 1 for institutional clients such as Wall Street traders. This move monetizes Trump's market-moving social media posts for algorithmic trading, raising significant ethical concerns about blurring the lines between political communication and financial gain, and could set a precedent for how social platforms sell data to traders. The API delivers posts with millisecond latency, targeting high-frequency algorithmic traders; pricing has not been disclosed. Trump has previously used Truth Social to promote stocks he personally bought, and his posts on tariffs and geopolitical events have caused market volatility.

telegram · zaihuapd · Jul 17, 01:02

Background: Truth Social is a social media platform launched by former President Donald Trump. It has become his primary channel for policy announcements, and his posts have repeatedly moved financial markets. High-frequency algorithmic trading (HFT) uses powerful computers to execute trades in microseconds based on data feeds. The Truth API service offers a direct, fast data feed to paying clients, potentially giving them an information advantage.

Tags: #Data Monetization, #API, #Social Media, #Algorithmic Trading, #Ethics

Tesla Cybercab begins production in North America
特斯拉 Cybercab 在北美启动量产 ⭐️ 8.0/10

Tesla has started production of the Cybercab, a fully autonomous electric vehicle without a steering wheel, in North America. The vehicle is purpose-built for robotaxi services and relies entirely on its onboard AI for driving. This marks a major milestone for the autonomous driving industry, as Tesla moves from concept to mass production of a dedicated robotaxi. It could accelerate the deployment of autonomous ride-hailing services and reshape urban mobility. The Cybercab is a two-passenger battery-electric vehicle with no steering wheel or pedals, unveiled in concept form in October 2024. Tesla aims for volume production of up to 2 million units per year by the end of 2026.

telegram · zaihuapd · Jul 17, 03:06

Background: A robotaxi is an autonomous vehicle (SAE Level 4 or 5) that provides on-demand rides without a human driver. Tesla has long promised a dedicated robotaxi, and the Cybercab is designed specifically for this purpose, leveraging Tesla's Full Self-Driving (FSD) technology. Competitors like Waymo already operate robotaxi services in several US cities, but Tesla aims to scale production and reduce costs dramatically.

Tags: #Tesla, #Autonomous Driving, #Cybercab, #Electric Vehicles, #Robotaxi

OpenAI CFO proposes new AI ROI metric: Useful Intelligence per Dollar
每美元有用智能：AI 投资新指标 ⭐️ 8.0/10

OpenAI CFO introduces 'useful intelligence per dollar' metric to measure AI investment ROI, replacing traditional adoption metrics.

telegram · zaihuapd · Jul 17, 15:00

Tags: #OpenAI, #AI ROI, #Productivity Metric, #GPT-5.6, #Machine Learning

Doubao phone pivots from GUI to MCP, boosts production
豆包手机转用 MCP 协议，备货量大幅提升 ⭐️ 8.0/10

Doubao phone has shifted its strategy from GUI-based automation to requiring top super apps like Alibaba and Tencent to provide MCP services in order to be integrated. It has increased its production volume from 30,000 to hundreds of thousands of units. This move signals a major shift in AI agent interaction paradigms, moving from simulating human taps to a standardized protocol (MCP) that enables deeper, authorized integration. It also intensifies the battle among super apps and phone makers for control over the AI entry point. The Doubao phone assistant software obtained generative AI service filing on July 15, 2025, and had previously been blocked by WeChat and Taobao for using GUI automation. Apple and Google are also moving toward similar MCP frameworks that require developer authorization.

telegram · zaihuapd · Jul 18, 00:29

Background: MCP (Model Context Protocol) is an open standard that enables AI models to interact with external tools and data sources in a unified way. By using MCP, the Doubao phone no longer needs to simulate screen taps; instead, super apps expose their capabilities through MCP servers, allowing the AI agent to call them directly. This approach is more reliable and respects platform boundaries.

Tags: #AI Agent, #MCP, #Mobile Strategy, #Platform Ecosystem, #China Tech

📊 Run stats · Total 8m 20s · AI analysis 3m 15s · Tokens 0.70 MCY (input 0.49 / output 0.21 MCY)