LearnByDoing With Steven

•

Canal • 4 programas

Educación

Educación

Cada semana
Educación

Educación

Cada semana
Educación

Educación

Cada semana
Educación

Educación

Cada semana

5 may

Stanford CS336 Lecture 6: Mastering GPU Programming Models, Performance, and Triton Kernels

🚀 Deep dive into GPU architecture! Just summarized Stanford CS336 Lecture 6: Mastering GPU Programming Models, Performance, and Triton Kernels. ⚡️ Key takeaways: Memory hierarchy: Registers > Shared Memory > HBM Kernel Fusion to beat the Memory Wall Tiling strategies for MatMul Why Triton is a game-changer for custom kernels Full video in my channels. linktr.ee/learnbydoingwithsteven #GPUProgramming #TritonKernels #StanfordCS336 #DeepLearning #CUDA #PerformanceOptimization #LanguageModeling #LearnByDoingWithSteven #StevenDataTalk #AIInfrastructure #LLM
13 jul

1 ACP Agent Blueprint w

1 ACP Agent Blueprint w
18 mar

Connecting Agents and Data: A Guide to A2A and MCP

The provided text outlines two primary frameworks, A2A and MCP, designed to solve the communication challenges inherent in isolated AI agents. A2A (Agent-to-Agent) is an open protocol that facilitates direct interaction between different AI entities, allowing them to exchange tasks and status updates regardless of their underlying vendors. In contrast, the Model Context Protocol (MCP) provides a standardized way for individual agents to access external data sources like databases, file systems, and code repositories. While A2A focuses on the collaborative relationship between multiple agents, MCP serves as a bridge between an agent and its necessary tools or information. Both protocols utilize JSON RPC to ensure structured data exchange, but they serve different roles in the AI ecosystem. Ultimately, the source illustrates how these technologies complement each other to create a fully integrated environment where agents can both work together and interact with existing digital infrastructure.
29/12/2025

Steven数据漫谈EP15 | Mickey | AI产品出海，GTM秘诀，海内外AI产品GTM与迭代的差异，哪些市场信息及市场需求值得关注

网易9年老兵谈AI出海：如何跨越文化鸿沟与建立品牌信任？对话 Grow Max 创始人 MK本期视频介绍：欢迎来到 Steven 数据慢谈。本期节目我们邀请到了 Grow Max 的创始人 MK (Mickey)。MK 曾在网易负责海外业务长达九年，完整经历了游戏产品从研发到宣发的全生命周期，服务过米哈游、莉莉丝等顶尖客户。现在她转型创立了 AI 营销公司，致力于帮助中国 AI 产品出海，连接全球用户。为什么从成熟的游戏行业转向 AI 创业？中国 AI 产品在海外面临哪些水土不服？如何通过讲好品牌故事来实现用户增长？MK 将结合她丰富的实战经验，为我们带来深度的行业洞察。本期内容简介：在本期播客中，MK 分享了她从游戏出海转型做 AI 产品全球化营销的心路历程。她指出，虽然中国工程师在开发通用型 AI 产品上具有优势，但在进军欧美市场时，往往面临三大核心挑战：文化鸿沟与细分人群覆盖、数据安全与信任问题、以及品牌叙事（Storytelling）能力的缺失。节目中，MK 详细对比了中美产品迭代逻辑的巨大差异：硅谷团队倾向于内容先行，通过社区反馈来驱动产品功能的迭代，而国内团队往往陷入功能堆砌的误区。她还分享了一个大厂 AI 产品出海先败后胜的真实案例，揭示了深入挖掘用户痛点（冰山之下的需求）的重要性。此外，我们还探讨了 AI 创业的未来机会，包括 AI+社交/情感陪伴以及针对老年群体的银发经济。在节目的最后，MK 从母亲的角度，分享了在 AI 时代关于教育和培养孩子深度思考能力的独到见解。时间戳：00:00 嘉宾介绍：从网易游戏出海到 AI 营销创业02:15 行业选择：为什么从游戏转战 AI 出海赛道？04:45 核心痛点：中国 AI 产品出海面临的三大挑战（文化、信任、品牌）10:30 竞争优势：Grow Max 如何服务创业团队与资源稀缺性13:20 逻辑差异：硅谷的内容驱动迭代 vs 国内的功能驱动迭代16:40 实战案例：某大厂产品如何通过深度调研扭转 40 万美金的亏损20:15 未来赛道：AI + 社交、情感陪伴与心理咨询的前景24:15 创业脑洞：如果有 50 万美金，我会做 AI + 银发经济27:50 行业思考：马斯克预言与 AI 训练 AI 的未来32:30 教育焦虑：AI 时代如何培养孩子的深度思考与辨别能力#AI出海 #人工智能 #SaaS营销 #创业 #GrowMax #产品增长 #品牌故事 #数据合规 #AI教育 #银发经济 #Steven数据慢谈 #Podcast
28 jun

🚀 Stanford CS336 Lecture 13: The Evolution of Language Model Data. -Notebooklm Summary

Stanford CS336 Lecture 13 focuses on the critical evolution of language model training data. While model architectures are widely disclosed, dataset details remain highly proprietary due to commercial competition and copyright considerations. The lifecycle of language model training spans pre-training, mid-training, and post-training. The mid-training phase curates high-quality datasets to enhance specific capabilities like coding, mathematics, and long-context reasoning. Ultimately, data curation shifts from high-volume, low-quality web crawls to low-volume, high-quality specialized datasets. Data filtering methodologies have evolved from basic rule-based heuristics and language identification to advanced model-based approaches. Modern curation pipelines leverage large language models to assess the educational value of documents, rewrite low-quality texts, and synthesize high-quality QA pairs to scale data efficiently. Legal and compliance challenges, including copyright and fair use, remain central to data acquisition. As models risk memorizing training text, developers navigate the balance between direct commercial licensing and fair use arguments. Key Takeaways: Mid-training acts as a crucial bridge, refining models for targeted reasoning tasks.Advanced LLM-driven filtering and synthesis scale high-quality data while avoiding rule-based bias.Copyright compliance and memorization concerns limit public dataset disclosures.All my links: https://linktr.ee/learnbydoingwithsteven #learnbydoingwithsteven #AI #DeepLearning #Research #TechSummary #MachineLearning #LLM #DataScience #DataCuration #Copyright #ArtificialIntelligence
3 may

How GPUs Actually Drive LLM Scaling: Insights from Stanford CS336 L5 2026

How GPUs Actually Drive LLM Scaling: Insights from Stanford CS336 Ever wondered why the "Memory Wall" is the biggest hurdle in AI training? Stanford's CS336 (Lecture 5) dives deep into the hardware foundations that make today’s large language models possible. Key takeaways on system-level optimization: Compute vs. Memory: GPU throughput is outpacing HBM bandwidth. Modern AI engineering is more about managing memory movement than raw calculation.The Power of Low-Precision: Moving to FP8 and FP4 isn't just about saving space; it's about maximizing hardware utilization through specialized matrix units.FlashAttention's Secret: It’s not just a faster algorithm; it’s a masterclass in tiling and operator fusion that avoids the quadratic memory bottleneck.Understanding the underlying hardware—from SMs to warps to shared memory—is essential for anyone building or scaling next-gen AI systems. All my links: https://linktr.ee/learnbydoingwithsteven #learnbydoingwithsteven #AI #GPU #Hardware #DeepLearning #FlashAttention #Stanford #CS336 #LLM #SystemOptimization #ComputerArchitecture #AIInfrastructure
20 feb

The 2025-2026 AI Technology Stack Panorama Report

The provided text outlines the comprehensive evolution of the global AI technology stack between 2025 and 2026, transitioning from experimental tools to integrated autonomous systems. It details a shift toward specialized hardware and advanced cooling solutions like liquid cooling to manage the immense power demands of modern computing. The report emphasizes the critical role of refined data engineering and architectural shifts, such as the rise of sparse Mixture of Experts (MoE) models, to improve efficiency and reasoning. Beyond infrastructure, it explores the optimization of inference engines and the rise of intelligent agents capable of multi-step planning and tool usage. Furthermore, the sources contrast divergent global regulatory landscapes, highlighting the strict risk-based approach of the European Union against the innovation-focused policies in the United States. Ultimately, the documentation predicts a future where AI becomes deeply embedded in operating systems and physical robotics, eventually leading to self-evolving research and development.
20 jul

1 Agents You Can t Answer For w_learnbydoingwithsteven.github.io_linktr.ee_learnbydoingwithsteven

1 Agents You Can t Answer For w_learnbydoingwithsteven.github.io_linktr.ee_learnbydoingwithsteven
19 mar

Agentic Storage: Solving AI State and Persistence with MCP

Standard AI agents currently operate like volatile memory, meaning they lose all progress once a session concludes. While technologies like Retrieval Augmented Generation allow agents to read existing data, they lack a dedicated method for saving original work or code. Agentic storage addresses this limitation by providing a persistent layer where autonomous systems can permanently store and manage outputs. To simplify this interaction, the Model Context Protocol (MCP) acts as a universal translator, allowing agents to communicate with various storage types through a standardized interface. Because giving AI write access involves risks, this architecture includes safety features like immutable versioning, sandboxing, and intent validation to prevent data loss or unauthorized actions. Ultimately, this evolution transforms AI from a temporary chatbot into a functional worker with a reliable, long-term memory.
2 may

Stanford CS336 2026 L4: Linear Time Attention and Sparse Architectural Alternatives

How do we scale LLMs beyond current limits? This lecture explores the transition from quadratic attention to linear alternatives and the rise of sparse Mixture of Experts (MoE). Topics Covered: The fundamental bottleneck of Transformers.RNN-like inference speed with linear attention.How MoE partitions parameters for efficiency.Optimizing for hardware with shared experts and MLA.Full videos in youtube, tiktok, substack, etc Subscribe for more SOTA AI research summaries! All my links: https://linktr.ee/learnbydoingwithsteven #CS336 #StanfordAI #MoE #Architecture #LearnByDoingWithSteven #StevenDataTalk #数能生智

Learn By Doing With Steven 数能生智