SGLang v0.5.12 Adds Full Inference Support for DeepSeek V4
SGLang v0.5.12 增加对 DeepSeek V4 的完整推理支持 ⭐️ 9.0/10

SGLang v0.5.12 introduces comprehensive inference support for DeepSeek V4, including tensor, expert, context, and data parallel attention, as well as custom kernels like DeepGemm and FlashMLA for MegaMoE. This release enables efficient serving of the state-of-the-art DeepSeek V4 model on a wide range of NVIDIA and AMD hardware, significantly advancing LLM inference optimization and GPU utilization. Key additions include HiSparse for offloading inactive KV cache to CPU, HiCache with UnifiedRadixTree, W4A4 MegaMoE kernels, and a unified Docker image for all NVIDIA GPUs.

github · Fridge003 · May 16, 18:23

Background: DeepSeek V4 is a large language model that uses mixture-of-experts (MoE) and sparse attention, requiring specialized parallelism and kernel optimizations for efficient inference. Expert parallelism distributes different experts across GPUs, while HiSparse offloads inactive KV cache to host memory to reduce GPU memory pressure.

Tags: #SGLang, #DeepSeek V4, #LLM Inference, #GPU Optimization, #Open Source

Video Comparison of OpenClaw vs Hermes Agent on Qwen 35B Sparks Founders' Debate
Qwen 35B 本地任务对比视频引发 OpenClaw 与 Hermes Agent 创始人辩论 ⭐️ 9.0/10

Atomic Bot published a video comparing OpenClaw and Hermes Agent on a real-world task using the local Qwen 35B model, showing OpenClaw completing the task in 12 minutes with 203k tokens versus Hermes Agent's 33 minutes and 257k tokens, sparking a debate between the founders. This debate underscores the importance of rigorous benchmarking in AI agent development, as single-run comparisons can be misleading; the outcome may shape developer trust and adoption of these open-source agent systems. Hermes co-founder Teknium criticized the video's single-run methodology, noting that the Qwen 35B model is known for high variance, and cited Hermes' superior performance on standard benchmarks like Wildclawbench and WolfBench, as well as higher real-world token throughput.

rss · meng shao(@shao__meng) · May 17, 00:48

Background: OpenClaw and Hermes Agent are two prominent open-source AI agent systems capable of performing multi-step tasks such as web scraping and building dashboards. The Qwen 35B model used in the video is a Mixture-of-Experts (MoE) model with 35 billion total parameters and only 3 billion active parameters, designed for efficient local deployment.

Discussion: The discussion features a direct back-and-forth: OpenClaw founder Peter Steinberger (@steipete) shared the video to highlight performance advantages, while Hermes co-founder Teknium (@Teknium) argued that proper evaluation requires repeated trials and cited Hermes' lead in public benchmarks and user token throughput, with some commenters noting the need for standardized testing.

Tags: #AI Agents, #Qwen, #OpenClaw, #Hermes Agent, #Real-world Benchmark

Rebuilding AlphaGo from Scratch: Insights for LLM RL
从零重建 AlphaGo：对 LLM 强化学习的启示 ⭐️ 9.0/10

Former DeepMind scientist Eric Jang reimplemented AlphaGo using modern open-source tools and low budget, and discussed how its core mechanisms—Monte Carlo Tree Search, policy and value networks—illuminate fundamental challenges in large language model reinforcement learning such as variance and credit assignment. This work bridges classical reinforcement learning and modern LLM training, offering practical lessons for improving sample efficiency and credit assignment in self-play and RL fine-tuning of language models. It also demonstrates that complex AI systems can now be reproduced at a fraction of the original cost, democratizing advanced research. Eric Jang's reconstruction used ResNet instead of Transformers for the neural network backbone, noting that ResNet was more cost-effective for the small-budget setup. He also highlighted that initializing the network with human expert games via supervised learning is crucial before self-play, a technique that contrasts with LLM RL which often starts from pre-trained models.

rss · 跨国串门儿计划 · May 17, 02:58

Background: AlphaGo is a Go-playing AI developed by DeepMind that combined deep neural networks with Monte Carlo Tree Search (MCTS). It uses a policy network to propose promising moves and a value network to evaluate board positions, then uses MCTS to simulate games with those networks guiding the search. AlphaGo famously defeated world champion Lee Sedol in 2016, demonstrating the power of combining search with learned priors.

Tags: #AlphaGo, #Reinforcement Learning, #Large Language Models, #Monte Carlo Tree Search, #AI Research

NVIDIA's SANA-WM: 2.6B open-source world model for 1-min 720p video
NVIDIA 发布 SANA-WM：2.6B 参数开源世界模型，可生成 1 分钟 720p 视频 ⭐️ 8.0/10

NVIDIA released SANA-WM, a 2.6B parameter open-source world model that can generate 1-minute 720p videos with 6-DoF camera control. The model is built on the SANA-Video codebase and uses a diffusion transformer architecture. SANA-WM represents a significant step toward efficient long-video generation with precise camera control, potentially impacting gaming, robotics, and simulation industries. However, the community notes that model weights are not yet publicly available, raising concerns about true openness. The model employs hybrid linear attention, combining Gated DeltaNet with softmax attention for memory-efficient long-context modeling. Code is licensed under Apache 2.0, while the model weights fall under the NVIDIA Open Model License, but weights are promised 'soon' and currently unavailable.

hackernews · mjgil · May 16, 12:06 · Discussion

Background: A world model in AI is a machine learning system that builds an internal representation of an environment and predicts its dynamics over time. 6-DoF (six degrees of freedom) camera control allows movement along three translational axes (forward/back, left/right, up/down) and three rotational axes (yaw, pitch, roll). SANA-WM is built on the SANA-Video architecture, a diffusion transformer designed for video generation.

Discussion: The community is excited about the technical achievement but skeptical due to missing weights. Some commenters note the synthetic training data resembles Unreal Engine renders, while others highlight the heavy bandwidth consumption of the demo page. The lack of immediate open weights has led to debates about the definition of 'open-source.'

Tags: #world model, #video generation, #open-source, #NVIDIA, #AI

Moving away from Tailwind to structured CSS
从 Tailwind 转向结构化 CSS ⭐️ 8.0/10

The author Julia Evans shares her personal experience transitioning from Tailwind CSS to writing semantic HTML and structured CSS, detailing her learning process and trade-offs. This post contributes to the ongoing debate about CSS methodologies, encouraging developers to reconsider utility-first approaches and invest in core CSS skills. The article reflects on the trade-offs between Tailwind's utility classes and traditional CSS, emphasizing the importance of semantic HTML and maintainability.

hackernews · mpweiher · May 16, 09:14 · Discussion

Background: Tailwind CSS is a utility-first CSS framework that provides low-level utility classes to build designs directly in HTML. It has gained popularity for its rapid development but also criticism for cluttering HTML and discouraging semantic markup. In contrast, structured CSS involves organizing styles with meaningful class names and separating content from presentation.

Discussion: Commenters generally praise Julia's honest writing. Some agree that Tailwind inverts the order of thinking about HTML and CSS, while others recommend CSS Modules as a simpler alternative. There is also criticism that Tailwind advocates often lack deep CSS knowledge.

Tags: #CSS, #Tailwind, #web development, #semantic HTML, #frontend

Accelerando (2005) Predicted AI Agents and Tech Dependency
《加速》预言 AI 代理与技术依赖 ⭐️ 8.0/10

A high-scoring community post highlights that Charles Stross's 2005 novel Accelerando accurately predicted the rise of AI agents and users' total dependency on them. Commenters note that current technology is already mirroring the novel's scenarios. This validates the novel's prescience and underscores how science fiction can anticipate real-world technological shifts, influencing discussions on AI safety, human dependency, and the trajectory of the technological singularity. In Accelerando, characters use AI agents via glasses to autonomously perform tasks, and losing the glasses renders them incapable of functioning. This mirrors current dependence on smartphones and AI assistants like ChatGPT.

hackernews · eamag · May 16, 11:36 · Discussion

Background: Accelerando is a 2005 science fiction novel by British author Charles Stross, released as a free e-book under a Creative Commons license. It explores themes of the technological singularity, AI agents, and the transformation of humanity in a rapidly accelerating world.

Discussion: Commenters express that the novel's prophecies are 'coming true' and finding it 'scary.' One user highlights a scene where an AI agent runs on glasses to help with tasks, and losing the glasses causes total helplessness. Others praise the novel's plausible depiction of future weirdness.

Tags: #science fiction, #AI, #predictions, #technology, #singularity

AI Has Broken the Open CTF Format
AI 已打破开放 CTF 格式 ⭐️ 8.0/10

Frontier AI models can now solve most Capture The Flag (CTF) challenges automatically, diminishing the hands-on problem-solving and community collaboration that were central to these competitions. This threatens the educational value of CTFs, which are key training grounds for cybersecurity professionals and enthusiasts, and may erode the collaborative spirit that defined the community. The author argues that AI reduces the need for deep analytical effort, and challenges that once required hours of teamwork can now be solved by querying a language model, altering the fundamental learning process.

hackernews · frays · May 16, 07:01 · Discussion

Background: Capture The Flag (CTF) is a cybersecurity competition format where participants solve puzzles, reverse-engineer code, or break into systems to find hidden flags. Traditionally, CTFs involve open formats where challenges are built by the community and solved manually through creative thinking and collaboration, fostering both technical skills and camaraderie.

Discussion: Comments express frustration that AI removes the collaborative, hours-long problem-solving experience, with one user noting the 'do it for me' temptation as hard to resist. Some suggest an arms race in challenge difficulty, while others question the fundamental shift in education.

Tags: #AI, #CTF, #Cybersecurity, #Education, #Community

δ-mem: Delta-Rule Memory Compression for LLMs
δ-mem：基于 Delta 规则的大语言模型内存压缩 ⭐️ 8.0/10

A new method called δ-mem compresses past information into a fixed-size state matrix using delta-rule learning, enabling efficient online memory for large language models without linearly growing cache. This approach could drastically reduce memory usage for long-context LLM inference, enabling models to run with limited GPU memory and potentially supporting infinite context windows in practice. δ-mem uses a delta-rule update to incorporate new tokens into a fixed-size state, approximating the compression limit of context windows. It does not eliminate fundamental capacity constraints, as similar input queries can produce diverse activations, making association difficult.

hackernews · 44za12 · May 16, 09:30 · Discussion

Background: Large language models typically use a key-value (KV) cache that grows linearly with sequence length, consuming huge memory for long contexts. Delta-rule learning adjusts weights based on the error between desired and actual output, commonly used in neural network training. A fixed-size state matrix, as in linear attention models, compresses history into a constant memory footprint, enabling constant-time inference.

Discussion: Commenters expressed mixed opinions: some praised the fixed-size state idea for enabling unlimited context and efficient GPU packing, while others doubted it solves the capacity problem because slight input variations cause large activation differences, making caching difficult. Practical concerns about memory cost and missing token associations were also raised.

Tags: #LLM, #memory compression, #AI efficiency, #online learning, #deep learning

OpenAI Codex Performance Boost: 75% Less Re-rendering
OpenAI Codex 性能提升：重渲染减少 75% ⭐️ 8.0/10

OpenAI announced significant performance improvements to Codex, including approximately 75% less re-rendering when switching threads and 10-50x faster Git operations in large repositories. These improvements make Codex much more responsive for developers working on large codebases, reducing background churn and enabling faster coding sessions. Some streaming paths now have zero unnecessary re-renders, and expensive Git operations are reduced by 10-50x depending on the operation. The update also improves time to usefulness at startup and first interaction.

rss · OpenAI Developers(@OpenAIDevs) · May 16, 18:32

Background: Codex is OpenAI's AI coding agent integrated with ChatGPT, launched as a research preview in May 2025. It helps developers write features, fix bugs, and review codebases. The Codex app provides a command center with worktrees and cloud environments for agentic coding.

Tags: #Codex, #OpenAI, #AI coding, #performance

What Hands-On IC Work Should Managers Do in the AI Era?
AI 时代，管理者该做什么亲自动手的工作？ ⭐️ 8.0/10

Julie Zhuo identifies four categories of individual contributor work suitable for engineering managers in the age of AI, based on discussions with senior Bay Area tech managers who are now expected to be hands-on. This signals a paradigm shift from the traditional separation of manager and IC roles, as AI tools lower the cost of hands-on work and demand that managers understand the technology to make informed decisions. The four categories are: internal efficiency tools, quality-of-life improvements, celebration stories, and vision pieces. A critical rule is to avoid critical-path product work, as managers' fragmented schedules can become a bottleneck.

rss · meng shao(@shao__meng) · May 16, 14:01

Background: Traditionally, engineering careers split into Individual Contributor (IC) and Management tracks, with managers focusing on coordination and people leadership. The rise of AI has made it cheaper and faster to build, pressuring managers to stay technically engaged to evaluate team efficiency and direction.

Tags: #engineering management, #AI, #hands-on, #tech trends, #leadership

GitHub Launches Copilot Desktop App for Agent-Driven Dev
GitHub 推出面向 Agent 驱动开发的 Copilot 桌面应用 ⭐️ 8.0/10

GitHub has officially released a dedicated Copilot desktop application designed for agent-driven development workflows. It enables running multiple AI agent workflows simultaneously and natively integrates with GitHub Issues, Pull Requests, and CI pipelines. This marks a significant evolution in AI-assisted development, shifting from single-agent coding assistants to multi-agent orchestration that covers the entire development lifecycle. It promises to drastically accelerate development speed and streamline collaboration between human developers and AI agents. The app covers the complete development lifecycle from coding to Pull Request merge by integrating with GitHub Issues, Pull Requests, and CI/CD pipelines. It is available at github.com/github/app.

rss · Geek(@geekbb) · May 16, 09:33

Background: Agent-driven development (ADD) is an emerging paradigm where AI agents act as primary executors of development tasks such as planning, coding, testing, and deployment, while humans provide oversight and strategic direction. Multi-agent AI workflows involve using multiple specialized agents in parallel for different stages of the software development lifecycle, enabling faster iteration and more autonomous task execution. GitHub's Copilot has evolved from an inline code completion tool into a platform supporting these advanced agentic workflows.

Tags: #GitHub, #Copilot, #AI Agent, #Developer Tools, #CI/CD

X algorithm open-sourced, reveals Grok-powered recommendation
X 算法开源，揭示 Grok 驱动的推荐机制 ⭐️ 8.0/10

Elon Musk has published the latest X (formerly Twitter) algorithm on GitHub, and a subsequent analysis using Grok Build confirms that the recommendation engine is now heavily driven by Grok, an AI chatbot from xAI. This marks a paradigm shift where a major social media platform's feed is no longer based on simple engagement signals but on an advanced AI model, setting a new standard for content recommendation and transparency. The For You system simultaneously pulls posts from followed accounts and the entire network, then uses a Grok-based transformer to predict which posts a user is most likely to interact with. The open-source repository allows developers to inspect the algorithm's code.

rss · AI Will(@FinanceYF5) · May 16, 08:20

Background: Grok is a generative AI chatbot developed by xAI, launched in November 2023 and integrated with the X social network. Grok Build is a terminal-based AI coding agent that can analyze codebases and provide insights. Open-sourcing the recommendation algorithm represents a major step toward transparency in how X curates content.

Tags: #X algorithm, #Grok, #Recommendation Engine, #Open Source, #Social Media

Cognition-to-Action Gap in Tool-Using Models
工具使用模型中认知到行动的差距 ⭐️ 8.0/10

A new study probes hidden states of tool-using agents and finds that large language models correctly recognize when they should call a tool but fail to execute the call 26-54% of the time, with the gap concentrated entirely in the cognition-to-action transition. This finding reveals a fundamental limitation beyond prompt engineering—late-layer geometry can rotate the internal signal nearly orthogonal to the required action, explaining persistent tool-use failures even when the model 'knows' the correct behavior. The internal probe direction is decodable, but the late-layer last-token regime rotates the signal nearly orthogonal to the action it should produce, which prevents successful tool calling despite correct cognition.

rss · elvis(@omarsar0) · May 16, 20:40

Background: Hidden state probing is an interpretability technique that analyzes neural network internal representations. Tool-using agents are LLMs extended with the ability to call external tools (e.g., APIs). The cognition-to-action gap refers to the model knowing it should call a tool but not actually generating the call.

Tags: #interpretability, #tool-using agents, #hidden state probing, #cognition-action gap

Tokens as Universal Input for Problem Solving
令牌正成为解决问题的通用输入 ⭐️ 8.0/10

A prominent AI figure tweeted that tokens are rapidly becoming the universal input modality for solving problems, signaling a shift toward using tokenized representations for diverse tasks. This matters because if tokens can unify different data types (text, code, images), it could accelerate progress toward artificial general intelligence by enabling models to handle any problem through a single interface. The tweet received high engagement with over 800 likes and 130 replies, indicating strong community interest, although it lacks technical specifics on how tokenization extends beyond text to other modalities.

rss · Greg Brockman(@gdb) · May 16, 23:25

Background: Tokenization is a fundamental process in natural language processing that breaks text into smaller units like words or subwords, enabling machines to process language. Advances in AI have extended tokenization to other modalities such as images and audio, making tokens a potential universal representation for problem-solving.

Tags: #AI, #LLM, #tokens, #problem-solving, #AGI

Codex Skill Analyzes Code Complexity Hotspots
Codex 技能分析代码复杂度热点 ⭐️ 8.0/10

Greg Brockman tweeted about a Codex skill created by Kappaemme that automatically scans codebases for computational complexity hotspots, such as O(n²) loops and repeated lookups, and suggests safe optimizations. This demonstrates a practical application of AI-driven coding agents to improve software performance, potentially reducing manual effort in profiling and optimization. It highlights Codex's capability to assist with real-world engineering tasks beyond code generation. The skill is open-source and can be installed via one command: npx --yes codex-complexity-optimizer. It provides before/after complexity estimates, risk levels, and testing recommendations, operating in report-only mode by default to avoid breaking changes.

rss · Greg Brockman(@gdb) · May 16, 13:49

Background: OpenAI Codex is an AI coding agent that automates software engineering tasks, from planning to deployment. It can be used as a CLI tool or integrated into IDEs. Computational complexity analysis helps identify inefficient code patterns that cause performance bottlenecks.

Discussion: The tweet garnered significant engagement with over 1385 likes and 56 replies, indicating strong interest. Community comments likely praise the tool's utility; however, no specific comments are provided in the input.

Tags: #AI, #Codex, #computational complexity, #optimization, #Greg Brockman

Cerebras Files for $60B IPO
Cerebras 提交 600 亿美元 IPO 申请 ⭐️ 8.0/10

Cerebras Systems is reportedly filing for a $60 billion initial public offering, marking one of the largest tech IPOs in recent years. This IPO underscores the growing demand for specialized AI hardware and could provide Cerebras with capital to scale production of its unique wafer-scale chips, potentially reshaping the AI semiconductor landscape. Cerebras is known for its Wafer Scale Engine (WSE), an enormous single-chip processor that is used in high-performance computing and AI training, and the company has secured deals with the U.S. Department of Energy and other institutions.

rss · Latent.Space · May 16, 04:36

Background: Cerebras Systems designs and manufactures the Wafer Scale Engine, the largest computer chip ever built, which uses an entire silicon wafer as a single processor. Unlike traditional chips that are cut from wafers, Cerebras leaves the wafer intact, allowing massive parallel processing for AI workloads. The company's chips are used in supercomputers like the Doe system, which features over 400,000 AI cores.

Tags: #AI Hardware, #IPO, #Cerebras, #Semiconductors

Anil Seth: Conscious AI is a projection, not inevitable
阿尼尔·塞思：有意识的 AI 是一种投射，并非必然 ⭐️ 8.0/10

Neuroscientist Anil Seth argues that the belief in inevitable conscious AI is a human tendency to project consciousness onto machines, similar to seeing faces in clouds. This challenges the widespread assumption among many AI experts that consciousness will eventually emerge in machines, influencing both AI development and ethical debates. Seth emphasizes that advanced AI like large language models are brilliant mimics but not sentient beings, and that consciousness is biologically grounded in living organisms.

rss · TED · May 16, 19:00

Background: Anil Seth is a neuroscientist known for his theory that consciousness arises from the brain's predictive processes, rather than being a separate property. He distinguishes between the 'real problem' of explaining how consciousness is generated and the 'hard problem' of why it feels like something. Seth argues that progress on the real problem may dissolve the hard problem, and he applies this view to AI, suggesting machines lack the biological basis for genuine consciousness.

Tags: #AI, #consciousness, #neuroscience, #philosophy

Thinking Machines Labs Unveils Real-Time AI Interaction Models
思考机器实验室推出实时 AI 交互模型 ⭐️ 8.0/10

Thinking Machines Labs demonstrated a new AI system called Interaction Models, capable of real-time translation while a person is still speaking and understanding conversational nuances like timing, pauses, and posture. This represents a significant step beyond current AI assistants, enabling more natural, human-like interactions that can interrupt and respond with full context, potentially transforming customer service, translation, and collaboration tools. The model processes continuous audio streams rather than chunked segments, allowing it to produce translated output as speech unfolds, and it also incorporates visual cues from posture and facial expressions.

rss · Matt Wolfe · May 16, 23:48

Background: Traditional AI models are optimized for autonomous tasks and respond only after a user finishes speaking. Interaction models are designed for real-time, two-way communication where the AI can listen while speaking and adapt to the flow of conversation, much like human dialogue.

Tags: #AI, #Interaction Models, #Real-time Translation, #Thinking Machines Labs, #AI Demos

Top 10 HackerNews Stories: AI Critique, Gaming Bill, Privacy, Open Source
HackerNews 十大热门：AI 批评、游戏法案、隐私与开源 ⭐️ 8.0/10

A curated list of 10 top HackerNews stories from May 17, 2026, covering AI industry critique, California's online game preservation bill, privacy concerns over data subpoenas, JavaScript ecosystem vulnerabilities, and Nvidia's open-source video generation model SANA-WM. This aggregation highlights key debates and developments in technology from AI over-reliance and gaming preservation to open-source advances, reflecting the community's priorities and concerns. Notable items include Mitchell Hashimoto's warning about 'AI psychosis', a California bill requiring game refunds or offline modes, the DOJ subpoena of over 100,000 user data, a satire on npm supply chain attacks, and Nvidia's SANA-WM model capable of generating one-minute 720p videos on a single GPU.

rss · HackerNews每日摘要 on SuperTechFans · May 16, 23:38

Background: HackerNews is a social news website focusing on computer science and entrepreneurship where the community votes on stories. This curated list aggregates top stories from May 17, 2026, providing a snapshot of current tech discourse.

Discussion: In the top story on AI psychosis, the community expressed concerns about AI hype, potential job displacement, and the risk of economic demand collapse. Some argued that AI will augment rather than replace human work, while others highlighted the challenge of redistributing productivity gains.

Tags: #HackerNews, #AI, #开源, #数据隐私, #技术新闻

DOJ Demands Apple, Google Identify 100k Car Tuning App Users
美司法部要求苹果谷歌识别 10 万汽车改装应用用户 ⭐️ 8.0/10

The U.S. Department of Justice issued subpoenas to Apple, Google, and Amazon demanding that they identify over 100,000 users of the car tuning app EZ Lynk, as part of an investigation into potential Clean Air Act violations. This case underscores the tension between law enforcement's need for user data and individual privacy rights, especially when requests are broad. The outcome could set a precedent for how tech companies handle mass government data demands. The subpoenas were issued in March and April 2026. Apple and Google are reportedly planning to challenge the request as overly broad, arguing it exceeds the scope of the investigation and raises privacy concerns.

telegram · zaihuapd · May 16, 05:34

Background: The Clean Air Act requires vehicle manufacturers to certify emissions controls. Aftermarket devices that disable these controls are illegal. EZ Lynk sells hardware and software that allows users to modify engine parameters, potentially bypassing emissions systems. The DOJ initially sued EZ Lynk in 2021. The current data demand seeks to identify customers who may have used the product for illegal emissions tampering.

Tags: #privacy, #legal, #tech companies, #government surveillance, #car tuning

Google Bans Manipulation of AI Search Results in Spam Policy
Google 将操纵 AI 搜索结果列入垃圾政策 ⭐️ 8.0/10

Google updated its search spam policy to explicitly prohibit manipulating generative AI search responses, including AI Overviews and AI Mode, effectively targeting practices like Generative Engine Optimization (GEO). This policy shift establishes clear rules for SEO in the AI era, threatening sites that use GEO tactics with ranking penalties or removal from search results, and signals Google's commitment to maintaining the integrity of its AI-powered search features. The policy equates manipulating AI search responses with traditional search ranking manipulation, and violations can lead to site demotion or complete removal from Google's index. Common GEO techniques include bulk-generating biased 'best of' content or embedding prompts to influence model citations.

telegram · zaihuapd · May 16, 06:31

Background: AI Overviews is an AI feature in Google Search that generates concise summaries from multiple sources, but has been criticized for accuracy and reducing site traffic. Generative Engine Optimization (GEO) emerged as a practice to improve content visibility in these AI-generated responses, often through tactics like mass-producing favorable content or hidden instructions. Google's updated spam policy now explicitly bans these manipulation techniques, treating them similarly to traditional SEO spam.

Tags: #Google, #AI search, #spam policy, #SEO, #GEO