Codex in ChatGPT App Enables Phone-Based Development
ChatGPT 应用中的 Codex 支持手机端开发 ⭐️ 9.0/10

Greg Brockman, President of OpenAI, announced that Codex is now integrated into the ChatGPT app, enabling users to build applications directly from their smartphones. This move democratizes software development by allowing anyone with a smartphone to create applications, potentially expanding coding access to millions without a traditional computer. Codex is powered by a version of OpenAI's o3 model optimized for software engineering, and while the mobile interface may have limitations compared to desktop, it offers unparalleled convenience for on-the-go coding.

rss · Greg Brockman(@gdb) · May 17, 05:24

Background: OpenAI Codex is an AI-powered coding assistant that automates repetitive tasks, assists with debugging, and streamlines feature implementation. It was previously only available via API or desktop interfaces, but the new ChatGPT app integration brings its capabilities to mobile devices, making coding accessible from anywhere.

Tags: #OpenAI, #Codex, #ChatGPT, #Mobile Development, #AI Coding

Rebuilding AlphaGo: Self-Play, RL, and LLM Insights
复现 AlphaGo：自我对弈、强化学习与 LLM 启示 ⭐️ 9.0/10

Former DeepMind scientist Eric Jang rebuilt AlphaGo from scratch using modern open-source tools on a low budget, and discussed how its core mechanisms (MCTS, policy and value networks) relate to reinforcement learning in large language models. This connects classical RL success (AlphaGo) to current LLM challenges (high variance, credit assignment), offering insights for improving RL in language models and potentially integrating search into LLM reasoning. The project used a 10-layer ResNet instead of Transformers due to budget constraints, and highlighted the importance of initialization with human games, self-play, and knowledge distillation from MCTS into the network.

rss · 跨国串门儿计划 · May 17, 02:58

Background: AlphaGo combined deep neural networks with Monte Carlo Tree Search (MCTS) to master the game of Go. The policy network suggests promising moves, the value network evaluates board positions, and MCTS balances exploration and exploitation. Self-play generates training data for iterative improvement of the networks.

Tags: #AlphaGo, #reinforcement learning, #LLM, #Monte Carlo tree search, #deep learning

Semble: Code Search for Agents Using 98% Fewer Tokens
Semble：为 AI 代理减少 98%令牌的代码搜索工具 ⭐️ 8.0/10

Semble is an open-source tool that combines static Model2Vec embeddings with BM25 to perform code search, reducing token usage by 98% compared to grep+read. It achieves 99% of the retrieval quality of a 137M-parameter transformer while being ~200x faster. This tool addresses a critical bottleneck for AI agents like Claude Code that rely on code search, drastically reducing token costs and improving speed. It enables more efficient code understanding without needing GPUs or API keys, potentially lowering barriers for agent-based development workflows. Semble uses Model2Vec's static embedding model potion-code-16M, runs entirely on CPU, and indexes a typical repo in ~250ms with ~1.5ms per query. It provides an MCP server for integration with Claude Code, Cursor, Codex, and OpenCode.

hackernews · Bibabomas · May 17, 15:37 · Discussion

Background: Static embeddings like Model2Vec convert text into fixed-size vectors without using a transformer at runtime, making them much faster and smaller. Reciprocal Rank Fusion (RRF) combines multiple ranking lists by weighting each item's reciprocal rank, producing a unified result set. BM25 is a classic text retrieval algorithm based on term frequency and document length.

Discussion: Commenters expressed interest in seeing actual agent benchmarks instead of retrieval-only evaluations, and questioned how Semble compares to existing LSPs. Some noted that while token savings are impressive, the probabilistic nature of semantic search might miss critical code patterns that grep would catch. Others pointed out that semantic code search could also benefit human developers.

Tags: #code search, #AI agents, #embeddings, #open source

I don't think AI will make your processes go faster
AI 不会加快流程速度 ⭐️ 8.0/10

Argues that AI's ability to speed up software processes is limited by the fundamental need for precise requirements, which remains the primary bottleneck.

hackernews · TheEdonian · May 17, 12:13 · Discussion

Tags: #AI, #software engineering, #productivity, #requirements, #bottleneck

Tesla Solar Roof Fails Economically, Shifts to Panels
特斯拉太阳能屋顶因经济不佳转向传统面板 ⭐️ 8.0/10

Tesla is reportedly pivoting from its Solar Roof product to traditional solar panels due to the Solar Roof's high costs and long payback periods, making it economically unviable for most consumers. This strategic shift highlights the challenges of integrating solar technology into building materials at a competitive price, affecting Tesla's renewable energy product lineup and consumer adoption of solar energy. An average Tesla Solar Roof costs approximately $106,000 before incentives, compared to $60,000 for a traditional roof replacement with conventional panels—a $46,000 premium, with a payback period of 15-25 years versus 7-12 years for panels.

hackernews · celsoazevedo · May 17, 04:09 · Discussion

Background: Tesla Solar Roof, introduced in 2016, aimed to replace traditional roofing with solar tiles that blend into the roofline. However, high installation complexity and cost hindered adoption, while traditional solar panels became cheaper and more efficient. The pivot reflects Tesla's focus on profitability after its energy division saw revenue growth but deployment declines.

Discussion: Comments highlight the cost disparity, with some noting the aesthetic appeal but questioning economic viability. One commenter accuses Tesla of using the product to pump stock. Others express hope for solar shingles if costs can be reduced.

Tags: #Tesla, #Solar Energy, #Renewable Energy, #Business Strategy, #Tech

Native text rendering woes: WebView may be faster
原生文本渲染困境：WebView 可能更快 ⭐️ 8.0/10

A recent article argues that for text-heavy rendering tasks in native iOS/macOS apps, using WebKit web views can be more performant than native APIs like SwiftUI Text or TextKit. This challenges the conventional wisdom that native rendering is always faster, suggesting developers should reconsider their technology choices for text-intensive features like editors or chat views. The article highlights specific pain points with SwiftUI text rendering and notes that modern browser engines have mature GPU acceleration and decades of optimization for complex text layouts.

hackernews · dive · May 17, 11:49 · Discussion

Background: In iOS/macOS development, developers often choose native frameworks like SwiftUI or TextKit for rendering text in apps, assuming they offer better performance than embedding a web view. However, browser engines like WebKit have been heavily optimized for text rendering over decades, including features like line breaking, bidirectional text, and advanced typography, which can be challenging to replicate natively.

Discussion: Commenters are divided: some agree that WebKit can outperform native for text-heavy tasks, citing its mature rendering pipeline; others present native alternatives like TextKit 2 achieving sub-8ms keystroke styling, or SwiftUI markdown renderers that work well, suggesting the article may overlook recent native improvements.

Tags: #native performance, #text rendering, #WebKit vs SwiftUI, #iOS development, #performance comparison

AI Should Be Embedded Technology, Not Standalone Product
AI 应是嵌入技术而非独立产品 ⭐️ 8.0/10

A commentary argues that AI should be treated as an enabling technology integrated into products, exemplified by Apple's approach, rather than sold as a standalone product. This reframing challenges the current AI industry trend of selling AI as a separate product, potentially influencing product strategy and user adoption by prioritizing seamless integration. The article draws parallels to the "Dropbox is a feature, not a product" argument, noting that successful AI integration should feel invisible to users and focus on improving existing experiences.

hackernews · ch_sm · May 17, 13:11 · Discussion

Background: Many tech companies currently market AI models or chatbots as standalone products. However, historically, lasting technologies become embedded infrastructure—like cloud computing or GPS. Apple’s strategy aims to embed AI into its ecosystem, making features like Siri smarter without branding AI as a separate offering.

Discussion: Commenters largely agree, with one noting that Apple should just make Siri work seamlessly. Another quotes Steve Jobs' advice to work backwards from customer experience, and a third draws the Dropbox analogy, warning that AI companies risk being commoditized.

Tags: #AI, #Apple, #product strategy, #technology, #user experience

GDS Advises NHS to Keep Open Source Repositories Open
GDS 建议 NHS 保持开源仓库开放 ⭐️ 8.0/10

The UK Government Digital Service (GDS) published guidance on May 14th, 2026, recommending that public sector organizations keep open source repositories open by default, directly countering the NHS's recent decision to close its open source repositories after vulnerabilities were disclosed through Anthropic's Project Glasswing. This underscores a significant policy rift within the UK government between security concerns and the benefits of open source, with GDS advocating for openness as a default posture to maintain scrutiny and reduce costs. The GDS guidance, titled 'AI, open code and vulnerability risk in the public sector,' does not explicitly name the NHS but is interpreted by civil service watchers as a direct rebuke, as noted by Terence Eden.

rss · Simon Willison · May 17, 15:59

Background: The Government Digital Service (GDS) is a UK government unit responsible for digital transformation and setting standards for public sector online services. The NHS recently decided to close its open source repositories in response to vulnerabilities discovered by Anthropic's Project Glasswing, which used AI to find security flaws in public sector code. This move drew criticism from open source advocates, leading to GDS's intervention.

Tags: #open source, #government policy, #NHS, #security, #GDS

Warning: Don't Outsource Your Learning to AI
AI 时代工程师警示：勿将学习外包给 AI ⭐️ 8.0/10

Addy Osmani published a detailed warning against engineers outsourcing their learning to AI, citing three studies (Anthropic 2026, MIT 'Your Brain on ChatGPT', and CHI 2026) that demonstrate cognitive decline and reduced understanding when relying on AI without verification. As AI coding assistants become ubiquitous, engineers risk losing deep understanding and problem-solving skills, which are essential for handling novel bugs, architectural changes, and market shifts. This warning highlights a growing cognitive debt that could undermine long-term career growth. The warning identifies five scenarios where pure delegation to AI inevitably fails: debugging, when AI is confidently wrong, during underlying changes, for outlier problems, and when the labor market revalues skills. Osmani also provides six actionable adjustments, such as forming hypotheses before asking, demanding explanations before code, and occasionally rewriting AI-generated code from scratch.

rss · meng shao(@shao__meng) · May 18, 00:51

Background: Large Language Models (LLMs) like GPT-4 and AI agents are increasingly used to generate code and solve programming problems. While they boost productivity, studies show they can reduce cognitive engagement and learning. The MIT study 'Your Brain on ChatGPT' used EEG to show weaker brain coupling in LLM users, and the CHI 2026 study found that early LLM use anchors the problem space, degrading later decisions.

Tags: #AI, #Software Engineering, #Learning, #Cognitive Skills, #LLM

Claude Team's Playbook for AI-Native Startups
Claude 团队发布 AI 原生初创公司操作手册 ⭐️ 8.0/10

Anthropic's Claude team released 'The Founder's Playbook: Building an AI-Native Startup', a stage-by-stage guide covering Idea, MVP, Launch, and Scale, with practical exercises and frameworks for founders. As AI-native startups become a major trend, this playbook from a leading AI company provides authoritative, actionable advice to help founders avoid common pitfalls and leverage AI effectively at each stage, potentially lowering the barrier to building successful AI-native companies. The playbook emphasizes that AI-native means redesigning each stage of the startup, not just using AI tools, and that founders should shift from doers to orchestrators of AI agents, tools, and small teams. It also warns against technical debt caused by context drift in AI systems.

rss · meng shao(@shao__meng) · May 17, 09:56

Background: An AI-native startup is designed from inception around AI as a core component of its architecture, data model, and operations, unlike AI-powered companies that simply add AI features. Anthropic, the company behind Claude, was founded by former OpenAI members and focuses on AI safety; Claude is trained using constitutional AI.

Tags: #AI-native startups, #founder playbook, #Anthropic, #team building, #entrepreneurship

HiDream-O1-Image: Open Source Rivals Closed-Source Image Gen
HiDream-O1-Image 开源模型媲美闭源 ⭐️ 8.0/10

HiDream-O1-Image, an open-source image generation model released under the MIT license, achieves performance comparable to closed-source alternatives like Nano Banana, excelling in photorealism, text rendering, and editing tasks. This model democratizes high-quality image generation by offering a free, open-source alternative to proprietary systems, potentially accelerating innovation and reducing costs for developers and creators. The 8B parameter variant outperforms all open-source baselines and matches the proprietary Nano Banana model, while a larger 200B version achieves state-of-the-art results. It uses a Pixel-level Unified Transformer (UiT) without external VAEs.

rss · Paul Couvert(@itsPaulAi) · May 17, 22:26

Background: Image generation models like DALL-E and Midjourney are typically closed-source and require API subscriptions. Open-source alternatives have historically lagged in quality. HiDream-O1-Image narrows this gap by leveraging a unified transformer architecture that natively processes pixels, text, and task conditions.

Tags: #open-source, #image-generation, #AI, #deep-learning, #model-comparison

Logan Kilpatrick: 'The model is the product'
Logan Kilpatrick：'模型即产品' ⭐️ 8.0/10

Logan Kilpatrick tweeted that the AI model itself is the product, reflecting a shift in how tech companies view AI. This perspective redefines product strategy in AI, where models become the primary value proposition rather than just features, signaling a move toward model-centric business models. The tweet gained high engagement with 1086 likes and 150 comments, indicating strong community resonance. It was posted by Logan Kilpatrick, a key figure in AI.

rss · Logan Kilpatrick(@OfficialLoganK) · May 17, 20:20

Background: In traditional software, the product is the application or service, with AI models often hidden as components. The phrase 'the model is the product' captures a paradigm where models like GPT-4 or BERT are directly monetized or serve as the core offering, as seen with foundation models and API services.

Tags: #AI, #machine learning, #product strategy, #technology trends

AI Agent Darwinism: Tech Rebuild, Multi-Model, and Future Software
AI Agent 的达尔文时刻：技术重构、多模型与未来软件格局 ⭐️ 8.0/10

Tasklet CEO Andrew Lee describes rebuilding the entire product stack in six months, shifting from workflow automation to a horizontal AI agent platform, and adopting a multi-model strategy including OpenAI, Google, and open-source models. He predicts only three types of software companies will survive: horizontal platforms, headless API providers, and solution companies. This discussion provides a rare inside look at how an AI agent platform is evolving in real-time, including architectural innovation in context management and cost optimization. Andrew's predictions about the software landscape offer strategic insight for startups and enterprises navigating the AI transition. Tasklet introduced a filesystem-based context management that replaces long chat histories, significantly reducing token costs. They also implement multi-model orchestration with a 'harness' approach that treats models as interchangeable components, allowing small models to dispatch tasks to larger ones.

rss · 跨国串门儿计划 · May 17, 20:34

Background: AI agent platforms like Tasklet aim to automate business processes by enabling users to describe tasks in natural language. The 'harness' concept refers to a middleware layer that abstracts away differences between various AI models, allowing companies to switch between providers easily. Multi-model strategies have become important as no single model excels at all tasks, and cost optimization is critical for real-world deployment.

Tags: #AI Agent, #Multi-Model, #Software Engineering, #Product Strategy, #Startup

OpenAI Open-Sources Symphony Agent Orchestrator for Autonomous Coding
OpenAI 开源 Symphony 自主编码代理编排器 ⭐️ 8.0/10

OpenAI has open-sourced Symphony, a specification (SPEC.md) that acts as an agent orchestrator, using project management tools like Linear as a control plane to manage and assign tasks to autonomous coding agents. By open-sourcing Symphony, OpenAI is enabling developers to leverage autonomous coding agents at scale, potentially transforming software engineering workflows by shifting from interactive pair programming to asynchronous task management. Symphony watches project boards (e.g., Linear) for open tasks, spawns dedicated Codex agents to work on each task, and automatically generates pull requests for human review. The approach decouples agent execution from developer attention, allowing agents to work autonomously.

rss · InfoQ · May 17, 20:00

Background: Agent orchestrators coordinate multiple AI agents to accomplish complex tasks. In software engineering, autonomous coding agents can write code independently, but managing them at scale requires a 'control plane' to assign work, monitor progress, and integrate results. Symphony uses existing project management tools as that control plane, reducing the need for custom infrastructure.

Tags: #open-source, #AI agents, #orchestration, #coding, #OpenAI

Monzo Builds Governed Data Mesh Across 100 Teams, 12K dbt Models
Monzo 构建跨 100 个团队和 12000 个 dbt 模型的有治理数据网格 ⭐️ 8.0/10

Monzo redesigned its data warehouse as a governed data mesh, supporting over 100 teams and 12,000 dbt models, resulting in a 40% reduction in warehouse costs and a 25% improvement in data delivery speed. This case demonstrates a successful large-scale data mesh implementation with quantifiable cost and speed benefits, providing a valuable blueprint for enterprises struggling with data silos and scalability. Monzo's 'meshy' approach balanced domain ownership with centralized governance using dbt models for transformations, ensuring reproducibility and version control across the organization.

rss · InfoQ · May 17, 11:02

Background: Data mesh is a decentralized data architecture where domain teams own and treat data as products, with federated governance. dbt (data build tool) enables analysts to write SQL transformations as modular, tested models. Monzo, a UK neobank, migrated from a monolithic data warehouse to this mesh to overcome scaling bottlenecks.

Tags: #data mesh, #dbt, #data engineering, #Monzo, #data warehouse

Salesforce CEO to Spend $300M on Anthropic AI This Year
Salesforce CEO 今年计划在 Anthropic AI 上投入 3 亿美元 ⭐️ 8.0/10

Marc Benioff, CEO of Salesforce, stated in a video that his company will likely spend $300 million on Anthropic's AI services this year. This massive financial commitment from a top enterprise software leader underscores the accelerating adoption of AI in business operations and signals Anthropic's growing importance in the enterprise AI market. The comment was made on the All-In Podcast and shared as a short video; no specific products or timeline were mentioned beyond the annual spend estimate.

rss · All-In Podcast · May 17, 15:30

Background: Anthropic is an AI safety and research company that develops large language models like Claude. Salesforce is a leading customer relationship management (CRM) platform that has been integrating AI into its products. This level of spending indicates deep integration of Anthropic's technology into Salesforce's offerings.

Tags: #AI, #Anthropic, #Salesforce, #Enterprise Software, #Investment

Caitlin Kalinowski on AI Hardware Boom and Robotics
凯特琳·卡利诺夫斯基谈 AI 硬件热潮与机器人 ⭐️ 8.0/10

In an in-depth podcast conversation, former OpenAI, Meta, and Apple hardware leader Caitlin Kalinowski shares insights on the AI hardware boom, VR's evolution, robotics challenges, and leadership lessons from Steve Jobs, Mark Zuckerberg, and Sam Altman. Kalinowski's experience spanning Apple, Meta, and OpenAI makes her perspective highly authoritative on hardware trends; her warnings about memory price shocks and supply chain bottlenecks are actionable for startups. Additionally, her discussion on humanoid robot limitations clarifies the gap between prototype and mass deployment. Kalinowski advises startups to pre-buy memory due to an upcoming price shock driven by AI demand, notes that VR technologies have been repurposed for military applications, and explains that humanoid robots remain prototypes because of high cost, limited energy density, and lack of reliable actuators.

rss · Lenny's Podcast · May 17, 12:30

Background: The AI hardware boom refers to the surge in demand for specialized hardware to run AI models, including GPUs, HBM memory, and custom chips. VR/AR technologies like those developed at Meta for Quest and Orion are now being adapted for military training and simulation. Humanoid robots face challenges in cost, power, and dexterity, with most not yet ready for mass deployment.

Tags: #AI Hardware, #Robotics, #VR/AR, #Tech Leadership, #Startups

Changxin Technology files for STAR Market IPO, Q1 revenue surges 719%
长鑫科技科创板 IPO：一季度营收暴增 719% ⭐️ 8.0/10

Changxin Technology has filed its IPO prospectus on the Shanghai Stock Exchange's STAR Market, reporting Q1 2026 revenue of 50.8 billion yuan, a 719% increase year-over-year, and turning profitable after a loss last year. As a leading domestic DRAM manufacturer, this IPO marks a significant milestone for China's memory chip industry and capital markets, reflecting strong demand and supply constraints in the global DRAM market. The company also guided H1 2026 revenue between 110 billion and 120 billion yuan, up 612% to 677% year-over-year, with net profit expected to be 52-58 billion yuan (excluding non-recurring items).

telegram · zaihuapd · May 17, 11:05

Background: DRAM (Dynamic Random Access Memory) is a type of volatile memory widely used in computers, servers, and consumer electronics. Changxin Technology is a key Chinese player in the DRAM industry, which has experienced severe supply shortages and price surges globally. The STAR Market is China's Nasdaq-style board designed to support high-tech companies.

Tags: #DRAM, #科创板, #半导体, #IPO, #芯片