LearnByDoing With Steven

•

Kênh • 4 chương trình

Giáo dục

Giáo dục

Hằng tuần
Giáo dục

Giáo dục

Hằng tuần
Giáo dục

Giáo dục

Hằng tuần
Giáo dục

Giáo dục

Hằng tuần

10 THG 3

Steven 数据漫谈 EP17｜欧洲 AI 生态、AI 泡沫与创新环境

（ NotebookLM 生成的中文总结版本） Steven 数据漫谈 EP17｜与 Innovator Coffee Podcast 的合作对谈：欧洲 AI 生态、AI 泡沫与创新环境本期 Steven Data Talk 为一次跨播客合作内容。我与 Innovator Coffee Podcast 团队进行了一场深入对谈，围绕当前欧洲的人工智能生态展开讨论。作为常驻意大利米兰的从业者，同时也是Steven Data Talk（一档关注数据、AI 与科技趋势的中英双语播客）的主持人，我从欧洲创业生态的视角分享了对 AI 发展的观察。我们讨论了欧洲与美国在 AI 发展路径上的差异，包括融资环境、基础设施条件，以及文化因素如何影响创新节奏。对话中还涉及当前广泛讨论的 “AI 泡沫”问题、欧洲在算力与数据中心基础设施方面的挑战，以及 AI 创业者在欧洲面临的现实环境。同时，我也分享了一些值得关注的欧洲 AI 初创公司案例，并探讨在 AI 快速发展的背景下，为什么教育、风险意识以及负责任的发展方式变得越来越重要。本期内容为一次播客合作交流，希望通过两个社区之间的互动，带来更多关于全球 AI 发展的不同视角。主要讨论内容个人背景介绍：金融行业经历、在米兰的工作生活，以及创办Steven Data Talk的契机欧洲社会如何看待 AI：公众态度以及不同国家之间的成熟度差异欧洲工作文化与创新速度：工作生活平衡与劳动制度对创业生态的影响 AI 是否存在泡沫：从模型层、应用层到企业落地的不同视角 AI 基础设施问题：欧洲算力与数据中心资源不足对创业者的影响欧洲值得关注的 AI 初创公司：AI Agent 工作流、Excel 与 LLM 的结合、AI 驱动的制造技术欧洲 AI 创新的驱动力与限制：融资规模、政策支持、官僚体系与人才流动欧洲创业现实：税收、运营成本与行政复杂度全球社区建设：欧洲与美国 AI 生态之间的连接与合作未来五年的 AI 发展：Transformer 的局限、世界模型、边缘设备上的小模型以及潜在范式变化如果你关注 AI 创业、欧洲科技生态以及人工智能的全球发展趋势，这一期合作对谈会提供一个来自欧洲视角的深入讨论。我的链接:⁠⁠https://linktr.ee/learnbydoingwithsteven⁠

10 thg 3

•

21 phút
28 THG 3

Securing the Autonomous Frontier: Agentic Security Runtime

As engineers scale AI agents, security can't be an afterthought. Hardcoded static credentials are a major risk. The solution? Agentic Runtime Security. Strategic Pillars: Dynamic Credentials: Just-in-time, session-bound access that is automatically revoked.Context-Aware Identity: Integrating Okta/IBM Verify through OAuth 2.0 to verify user intent.CIBA (Backchannel Auth): Requiring mobile approval for high-risk actions (onboarding/offboarding), bypassing the browser entirely.Don't let a prompt injection compromise your enterprise data. Build with zero-standing-privilege. 🛡️ All my links: https://linktr.ee/learnbydoingwithsteven #AI #Cybersecurity #AgenticSecurity #EnterpriseSecurity #OAuth #Innovation #learnbydoingwithsteven

28 thg 3

•

5 phút
24 THG 3

Benchmarking Open Source VLM and OCR Model Performance

These sources explore the evolving landscape of open-source Vision-Language Models (VLMs) and their specialized application in Optical Character Recognition (OCR). The first source introduces DeepSeek-OCR, a model utilizing a unique DeepEncoder architecture to compress high-resolution visual data into a minimal number of tokens while maintaining high text extraction accuracy. It highlights the potential for optical compression to solve computational challenges associated with processing long document contexts in large language models. The second source provides a comparative benchmark of prominent open-source models, such as Qwen 2.5 VL and Gemma-3, evaluating their ability to perform structured JSON extraction from documents. This report identifies Qwen 2.5 VL as a top performer, matching the accuracy of leading closed-source models like GPT-4o. Together, the texts demonstrate a shift toward end-to-end neural architectures that replace traditional, multi-step OCR pipelines with more efficient and integrated visual-textual processing.all my links: linktree learn by doing with steven数能生智 https://linktr.ee/learnbydoingwithsteven

24 thg 3

•

6 phút
12 THG 2

ADK: Building Autonomous AI Agents Beyond LLMs

IBM's Agent Development Kit (ADK) enables autonomous AI agents to sense, reason, and act. Unlike passive LLMs, these tools use sensors and APIs to operate independently. Developers follow a six-step process, prioritizing safety, fairness, and trust in automation.

12 thg 2

•

15 phút
19 THG 3

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.

19 thg 3

•

40 phút
27/11/2025

Steven数据漫谈 | EP13 | AI泡沫？硅谷裁员潮？创业者焦虑？为什么在欧洲做 AI 却一点都不慌？

本期节目是转载自《创·见｜Build Up!》，我是本期节目的嘉宾。 #025-AI泡沫？硅谷裁员潮？创业者焦虑？为什么在欧洲做 AI却一点都不慌？ - 创·见｜Build Up！全球创始人访谈 | 小宇宙 - 听播客，上小宇宙 “The best way to predict the future is to invent it.” —— Alan Kay本期节目，我们与《创·见｜Build Up!》主播 Lydia Shang 一起，在硅谷与欧洲的交汇点上重新审视当下 AI 浪潮：是泡沫前夜，还是结构性重塑？是焦虑竞速，还是松弛创新？我们讨论了硅谷的“效率革命”、欧洲的合规红利、创业者的节奏选择，以及播客这种“慢媒介”在长期主义时代反而越来越重要。🎧 本期看点｜What’s Inside硅谷的高压与裁员潮，是泡沫还是重置？为什么欧洲创业者“不慌”？AI Act/GDPR 如何重塑创新？构建品牌与内容的长期主义：播客为什么值得一做再做？自动化、数据库优化、端侧模型、本地化 AI 的新机会创作者与创业者如何在“快与慢”之间找到节奏？🎙️ 本期主播｜HostLydia ShangFounder @ 繁星联众 Astra X Ventures《创·见｜Build Up with Lydia》创办人👤 本期嘉宾｜GuestSteven Wang 王健数据科学家、内容创作者，《Steven数据漫谈》《Steven Data Talk》主理人坐标米兰。专注欧洲 AI 创业生态、AI 应用、商业模式与内容创新。⌚ 时间索引为什么在米兰做 AI 与播客从体制到技术：自动化与数据库优化的“内生动力”品牌进化：从教程到“数据漫谈”的内容聚焦“欧洲版硅谷101”的初心与碰壁GDPR + AI Act：合规的复杂与创业的边界逆境中的机会：信息不对称与资源重配长期主义与播客：可做一生事业的媒介硅谷与西雅图：扩张与裁员的两条主线结构性收缩：从成本压力到产品化加速AI 泡沫？更像估值与模式的周期性清洗欧洲式创新：政府数据、大学 spin-off 与“匠心”欧洲年轻华人创业者状态：稳定偏好与破局者为什么从播客出发：亲密度与留存度的价值榜样力量：Lex Fridman / Huberman 的训练法则YC 课精讲的方法：时间拉长、原则依旧跨时区与工具栈：把技术门槛变成节奏优势Gap 的意义：停下，是为了走更远未来选择：打工与创业可以并行快问快答：榜样、遗憾与“何为自由”🌐 关于《Steven数据漫谈》用最清晰的方式，聊最前沿的人工智能 / 科技 / 创新 / 商业 / 创业思考。现已登陆 Spotify、Apple Podcasts、YouTube、Amazon Music、小宇宙、喜马拉雅等平台。所有链接 & 支持创作者：https://linktr.ee/learnbydoingwithsteven通用免责声明Disclaimer⁠：https://mp.weixin.qq.com/s/CSX358HspNIBQiGsGtQBdA

27/11/2025

•

1 giờ 17 phút
11 THG 4

Demystifying the Foundations of Neural Networks with MIT

Demystifying the Foundations of Neural Networks with MIT The revolution in artificial intelligence over the past decade has been nothing short of extraordinary. From solving insurmountable scientific challenges to hyper-realistic content generation, deep learning has moved from a computation-expensive luxury to a highly commoditized technology accessible to everyone. MIT 6.S191 (Lecture One) provides a foundational systematic review of the core intelligence hierarchy. By defining intelligence as information processing for decision-making, it clarifies the distinct roles of AI, Machine Learning, and Deep Learning. The growth we see today is driven by three pillars: massive datasets, efficient GPUs, and mature open-source software. Key takeaways include: The Layered Hierarchy: AI (Science) -> ML (Learning from Data) -> DL (Neural Networks).The Perceptron: The fundamental building block using dot products and non-linear activation.Training Mechanics: Using Loss Functions and Backpropagation for weight optimization.Generalization: Combatting overfitting with techniques like Dropout and early stopping.All my links: ⁠https://linktr.ee/learnbydoingwithsteven⁠ #DeepLearning #MIT #ArtificialIntelligence #MachineLearning #NeuralNetworks #LearnByDoingWithSteven #StevenDataTalk #AIResearch #ComputerScience #TechInnovation

11 thg 4

•

7 phút
19 THG 2

Comparative Analysis of Large Model Inference Optimization Frameworks

This report provides a comparative analysis of specialized large language model (LLM) inference frameworks designed to overcome hardware limitations and high computational costs. It distinguishes between high-throughput server solutions like vLLM, which uses PagedAttention to eliminate memory fragmentation, and SGLang, which optimizes complex, multi-turn interactions through RadixAttention and structured generation. For local deployment, the text evaluates Ollama and LM Studio, highlighting how they leverage llama.cpp and the GGUF format to run models on consumer-grade hardware. The study further explores critical performance-enhancing technologies such as quantization, speculative decoding, and continuous batching. Ultimately, the sources serve as a guide for selecting the right infrastructure based on specific needs, ranging from cloud-scale API services to private local assistants.

19 thg 2

•

17 phút
8 THG 4

Building a Personal Knowledge Management (PKM) system

Stop just reading—start compounding your knowledge. Here is how LLMs are transforming the "Second Brain." Building a Personal Knowledge Management (PKM) system is no longer about manual filing. By leveraging frontier LLMs, we can now automate the ingestion of articles, papers, and datasets into a structured Markdown wiki. Using Obsidian as the frontend, research becomes a continuous process of autonomous indexing, summarization, and connection-finding. Key Workflow Insights: Data Ingest: Automated indexing of raw documents and images.AI-Managed Wiki: Incremental compilation with zero manual intervention.Complex Queries: Autonomous research agents that query hundreds of articles.Visual Output: Auto-generated slide presentations (Marp) and data visualizations.As repositories grow, the future lies in synthetic data and fine-tuning to internalize this knowledge directly into model weights. #LearnByDoingWithSteven #SecondBrain #Obsidian #PersonalKnowledgeManagement #AI #LLM #Productivity #KnowledgeBase #DataScience #AIAgents All my links: https://linktr.ee/learnbydoingwithsteven

8 thg 4

•

4 phút
9 THG 4

Summary of public notes regarding Claude Code source code

The veil has been lifted on Anthropic's Claude Code. Here is what we found in the 512k lines of leaked source code. A recent configuration error led to the exposure of nearly 1,900 TypeScript files, revealing the inner workings of one of the most powerful coding agents. From hidden experimental features like KAIROS and BUDDY to the sophisticated coordination of over 40 internal tools, the architecture is a masterclass in modern agentic design. Key technical takeaways: Runtime: Built on Bun with a React-based terminal interface.Scaling: A massive 30,000-token initial payload for high-precision tool execution.Privacy: An "Undercover Mode" that automatically scrubs AI metadata for employees.Caching: Intelligent static/dynamic prompt splitting to optimize cost and speed.This leak provides a rare glimpse into the future of LLM-driven engineering tools. While the caching bug discovered by the community has already been patched, the implications for agentic security and telemetry are profound. #LearnByDoingWithSteven #ClaudeCode #Anthropic #AIArchitecture #SoftwareEngineering #AgenticAI #TechNews #LLM #Coding #OpenSource All my links: https://linktr.ee/learnbydoingwithsteven

9 thg 4

•

5 phút

Learn By Doing With Steven 数能生智