LearnByDoing With Steven

•

Kanal • 4 program

Utbildning

Utbildning

Varje vecka
Utbildning

Utbildning

Varje vecka
Utbildning

Utbildning

Varje vecka
Utbildning

Utbildning

Varje vecka

2025-12-06

Miras: Sculpting Memory and Retention in Sequence Architectures

The text outlines the Miras framework, a generalized structure for designing efficient sequence models inspired by human associative memory and cognitive processes. Miras unifies and categorizes neural architectures, including Transformers and recurrent neural networks, based on four critical design choices: the memory architecture, the attentional bias objective, the retention gate, and the memory learning algorithm. The authors observe that standard models often rely on limited objectives and propose alternative biases, such as $l_p$ losses and Huber loss, and sophisticated retention gates rooted in concepts like Bregman and KL divergence, to improve memory management. This new perspective leads to the introduction of three novel sequence models—Moneta, Yaad, and Memora—which are shown to outperform existing state-of-the-art models in language modeling and recall-intensive benchmarks. Ultimately, the framework provides guidelines for developing a new generation of expressive and robust architectures capable of handling long contexts.

2025-12-06

•

7 min
2025-12-06

Model Moats, Scaling Limits, and Agent Wars

The source material, a transcript from an IBM panel discussion, reviews recent major large language model releases, including Mistral 3, DeepSeek-V3.2, and Claude Opus 4.5. The experts note that while all new models are highly capable, labs are increasingly forced to differentiate by choosing specialties, such as DeepSeek’s focus on tool-calling and reasoning or Mistral’s native multimodal capabilities. The conversation also addresses the persistence of AI scaling laws, debating whether enormous infrastructure, exemplified by Google’s use of TPUs, is the primary driver of capability increases. Rather than simple scaling, the panelists suggest that faster algorithmic experimentation and strategic parameter tuning are likely the main sources of current quality improvement. Finally, the discussion shifts to the business challenges facing the technology, specifically analyzing how Amazon is initiating "turf wars" by blocking the ChatGPT shopping agent. This commercial friction may fragment the market and undermine the original concept of a single, all-encompassing agent for users.

2025-12-06

•

16 min
2025-12-06

Developer Leaks, Supply Chains, and Poetic AI Exploits

The IBM Security Intelligence podcast transcript features experts analyzing a range of evolving cybersecurity challenges affecting both consumers and enterprises. A major focus includes the highly disruptive return of the Shai Hulud worm, an automated malware strain that compromises the software supply chain by infecting NPM and Maven repositories and wiping data if it fails to steal developer secrets. The discussion also addresses how developers are unintentionally causing security failures by leaking sensitive credentials onto public code formatting sites and examines the far-reaching impact of the Gainsight breach, which exploited system-to-system trust to compromise over 200 customer organizations. Regarding emerging threats, the experts detail how cheap, legitimate-looking Android streaming devices are secretly leveraging consumer bandwidth for malicious botnets and note the surprising discovery that malicious poems are effectively breaking AI guardrails on major large language models. Across all topics, the panel emphasizes the crucial need for organizational collaboration and heightened personal vigilance to combat these increasingly complicated attack vectors.

2025-12-06

•

7 min
2025-12-06

Data Integration: Agents, Canvas, and Code

The IBM Technology transcript offers a structured analysis of distinct approaches used to design and manage data integration pipelines. The material employs an extended cooking analogy—comparing ordering takeout, utilizing a meal kit, and cooking from scratch—to clarify the functionality of three primary authoring experiences. These methods include No Code (using AI agents), which is quick and accessible for business users, Low Code (visual drag-and-drop), which strikes a balance between speed and execution control, and Pro Code (Python SDKs), which provides maximum customization for developers. The discussion thoroughly outlines the strengths and inherent trade-offs of each system, noting that ease of use generally correlates inversely with control and scalability. Ultimately, the source asserts that successful data teams must possess the flexibility to utilize all three authoring styles based on the technical skills of the user and the specific requirements of the project.

2025-12-06

•

14 min
2025-12-06

Titans: Long-Term Neural Memory for Scaling Context

The provided text details the development of Titans, a new family of neural architectures designed to overcome the fixed context window and computational scaling issues inherent in modern Transformers. The core of this system is a novel neural long-term memory module that operates as a meta in-context learner, dynamically updating its weights based on a calculated "surprise" metric that includes both momentum and an adaptive forgetting mechanism. This mechanism allows the module to store persistent, historical information, thereby acting as long-term memory while the attention component handles immediate context as short-term memory. The paper presents three distinct architectural ways to integrate this new component: Memory as a Context (MAC), Memory as a Gate (MAG), and Memory as a Layer (MAL). Empirical testing across diverse tasks, including language modeling and extreme-length reasoning benchmarks, demonstrates that Titans achieve superior performance and efficiency compared to baseline models, successfully scaling to context window sizes exceeding two million tokens.

2025-12-06

•

16 min
6 JAN.

Architecting Private Agentic AI Flows for Sensitive Data

This source explains how organizations can leverage agentic AI while maintaining strict data privacy through isolated infrastructure. While traditional AI models often rely on public APIs that pose significant risks to sensitive information, private flows allow agents to reason and act entirely behind a company's firewall. The text outlines a three-layer architecture consisting of a foundation layer for hosting models, an augmentation layer for private data retrieval, and an action layer for executing tasks. To mitigate internal risks and meet regulatory requirements like HIPAA or GDPR, the author suggests utilizing data anonymization, rigorous access controls, and data minimization strategies. These private systems are currently being implemented across highly regulated sectors, including healthcare, finance, and defense, to ensure security. Ultimately, the source argues that for any industry handling protected data, transitioning to private agentic flows is an essential evolution rather than an optional feature.

6 jan.

•

17 min
6 JAN.

Architecting Private Agentic AI Flows for Sensitive Data

This source explains how organizations can leverage agentic AI while maintaining strict data privacy through isolated infrastructure. While traditional AI models often rely on public APIs that pose significant risks to sensitive information, private flows allow agents to reason and act entirely behind a company's firewall. The text outlines a three-layer architecture consisting of a foundation layer for hosting models, an augmentation layer for private data retrieval, and an action layer for executing tasks. To mitigate internal risks and meet regulatory requirements like HIPAA or GDPR, the author suggests utilizing data anonymization, rigorous access controls, and data minimization strategies. These private systems are currently being implemented across highly regulated sectors, including healthcare, finance, and defense, to ensure security. Ultimately, the source argues that for any industry handling protected data, transitioning to private agentic flows is an essential evolution rather than an optional feature.

6 jan.

•

6 min
2025-12-05

DeepSeek-V3.2: Gold-Medal AI via Sparse Attention

The paper introduces the DeepSeek-V3.2 Large Language Model (LLM) framework, explicitly designed to bridge the performance gap between open-source and proprietary systems. A key technical advancement is the DeepSeek Sparse Attention (DSA) mechanism, which significantly improves efficiency by reducing computational complexity for processing long-context sequences. The model's reasoning and agentic proficiencies were enhanced through a scalable reinforcement learning framework that allocates substantial post-training compute and a novel synthesis pipeline for generating large-scale agentic tasks. DeepSeek-V3.2 achieves performance parity with closed-source models like GPT-5 on standard benchmarks, demonstrating strong tool-use and generalization capabilities. Notably, the high-compute variant, DeepSeek-V3.2-Speciale, achieved gold-medal performance in the 2025 International Mathematical Olympiad (IMO) and other top-tier competitions, reaching capability parity with models such as Gemini-3.0-Pro. Overall, the work establishes a new performance milestone for open LLMs, though the authors note challenges in world knowledge and token efficiency remain.

2025-12-05

•

7 min
2025-10-23

Steven数据漫谈E08 对话Aria Yang - 创业，营销，PR

在本期《Steven数据漫谈》中，主持人Steven采访了金融科技公司营销主管Aria，她分享了关于初创公司市场营销和公共关系的专业见解。核心要点包括：创始人主导的营销至关重要，创始人必须亲自担任首席“故事官”，以找到“信息与市场的契合点”。Aria详细介绍了科技公关策略，强调创始人应将自己打造为思想领袖，并与记者建立长期的、基于价值的关系。她还指出了线下活动在建立信任和推动销售转化方面的强大作用。最后，Aria给职场人的三大建议是：与同辈建立牢固的关系、勇敢地迈出第一步、在听取建议的同时最终相信自己的直觉。 #播客推荐 #Steven数据漫谈 #创业 #市场营销 #创始人IP #科技公关 #职场干货 #个人成长 🎙️ 《免责声明 / Podcast Disclaimer》本播客节目及相关频道（包括但不限于《Steven数据漫谈》《数能生智》《Steven Data Talk》《Learn by Doing with Steven》）中所表达的所有观点、立场和意见，均仅代表主持人及受邀嘉宾的个人观点，仅供学习、分享与交流之用。节目中提及的任何信息、案例、公司、产品或机构，不构成任何形式的投资、商业、法律、财务、医疗或其他专业建议，也不代表主持人、嘉宾或其所属机构、合作方的官方立场。主持人及嘉宾力求内容准确与时效，但不保证信息的完整性或适用性。听众在依据节目内容采取任何决策前，应自行进行独立判断与核实。《Steven数据漫谈》《数能生智》及其制作团队不对因使用或依赖节目内容所产生的任何直接或间接损失承担责任。除非另有说明，节目内容之著作权归原作者所有，未经授权不得转载、改编或二次传播。

2025-10-23

•

34 min
29 JAN.

The Juggling Act: Defining Success on Your Own Terms

In this presentation, Harvard Professor Howard Stevenson shares personal reflections on defining and achieving a meaningful life after surviving a near-fatal cardiac arrest. He challenges traditional views of success, arguing that it is a dynamic juggling act rather than a finish line to be crossed. Stevenson breaks success into four distinct domains: achievement, significance, happiness, and legacy, noting that these categories are often uncorrelated and require different emotional drivers. Central to his philosophy is the concept of "enough," which serves as a necessary limit to prevent one area of life from overshadowing others. By viewing life through a spiral of continuous choices rather than a sequential path, he encourages individuals to focus on the "falling balls" of their own lives to maintain balance. Ultimately, he advises his audience to live life forward and prioritize the lasting ripples of their actions over temporary splashes of fame or wealth.

29 jan.

•

14 min

Learn By Doing With Steven 数能生智

Leverantör