Artículos Relacionados con Deep Learning

El Centro de Noticias de HTX ofrece los artículos más recientes y un análisis profundo sobre "Deep Learning", cubriendo tendencias del mercado, actualizaciones de proyectos, desarrollos tecnológicos y políticas regulatorias en la industria de cripto.

Richard Sutton, the 69-Year-Old Father of Reinforcement Learning, Launches Startup: Building a 20-Watt Human-Brain-Level Agent

Reinforcement Learning pioneer Richard Sutton, 69, has co-founded Oak Lab after departing Keen Technologies. The new venture aims to build a trillion-parameter intelligent agent capable of real-time learning and planning with a power consumption of only 20 watts, matching the human brain. Sutton, a foundational figure in reinforcement learning and co-author of its seminal textbook, believes current deep learning approaches are inefficient and require a fundamental paradigm shift. Oak Lab's core thesis is that intelligence stems from runtime experience. The proposed OaK (Options and Knowledge) architecture seeks to enable agents to learn continuously from their own interactions, abstracting sequences of actions into reusable skills without storing or replaying past data batches. This direction builds upon Sutton's long-held views, including those expressed in "The Bitter Lesson," and marks a move towards what he calls the "age of experience"—where AI learns from self-generated experience rather than static, human-curated datasets. Sutton's first public appearance following the launch will be at the Shanghai WAIC conference.

marsbit07/14 12:30

Richard Sutton, the 69-Year-Old Father of Reinforcement Learning, Launches Startup: Building a 20-Watt Human-Brain-Level Agent

marsbit07/14 12:30

Deforming the Transformer, LLMs Become Smarter

A new research paper proposes "Tapered Language Models (TLMs)," a method that improves large language model performance without adding any parameters. It challenges the standard Transformer design where each layer has the same number of parameters ("feed-forward network" width). Building on evidence that layers are not equally important—earlier layers handle foundational information like grammar, while later layers often reinforce existing judgments—the researchers suggest reallocating model capacity from later to earlier layers. The core idea is to make the layer width taper off monotonically from start to end, keeping total parameters and compute constant. Experiments compared linear, cosine, and sigmoid tapering curves on a 440M parameter model. The cosine curve (e.g., starting width 1.5x baseline, ending 0.5x) achieved the best result, reducing perplexity by 1.84 points compared to the uniform baseline—a significant gain at zero cost. This finding proved robust across four different model architectures (including gated attention and memory-augmented models) and at larger scales (760M and 1.3B parameters), consistently improving performance on commonsense reasoning and language modeling tasks without harming long-context retrieval ability. The work highlights a long-overlooked design dimension: optimal parameter allocation across depth. It offers a "free lever" for efficiency, potentially applicable beyond language models to vision Transformers and diffusion models. The study was conducted by researchers from Mila, Cornell University, and the University of Montreal.

marsbit06/29 12:53

Deforming the Transformer, LLMs Become Smarter

marsbit06/29 12:53

Introduction to the Concept of World Models: A Story from Psychology to the Main Battlefield of AI

**World Models: From Psychology to AI's Core Concept** "World model" is a trending but often confusing term in AI, describing a system that allows machines to internally simulate, predict, and rehearse potential outcomes before taking real-world action—like a mental "sandbox." While definitions vary—Yann LeCun emphasizes physical understanding, OpenAI's Sora is a video-based "world simulator," Google DeepMind's Genie 3 creates interactive 3D environments, and companies like Alibaba and Tesla focus on practical applications—the core goal is consistent: reduce reliance on vast real-world data by creating an internal, predictive model for safer and more efficient AI. The concept has deep roots, tracing back to psychologist Kenneth Craik (1943). In AI, it was revitalized by researchers like David Ha and Jürgen Schmidhuber (2018). Major technical approaches include: 1) generative video models (e.g., Sora) for visual realism; 2) abstract predictive models (e.g., LeCun's JEPA) for efficiency and physical reasoning; and 3) explicit 3D simulators (e.g., NVIDIA Omniverse) for precision. Fei-Fei Li proposes a classification based on the AI action loop: renderers (output observations), simulators (output world states), and planners (output actions). The emerging "World Action Model" (WAM) paradigm aims to unify future prediction and action generation. An industry framework is forming: upstream (data, compute, sensors), midstream (general and vertical platforms), and downstream applications (autonomous driving, robotics, gaming, etc.). Autonomous driving is currently the most mature use case. The current lack of a unified definition reflects the field's early, dynamic stage, similar to past tech revolutions. Different approaches—focusing on pixels, physics, or behavior—represent parallel explorations of how best to compress and understand the world. This diversity, while seemingly chaotic, signals that world models have moved from an academic idea to a critical industrial battleground, ultimately aiming to give machines the ability to understand, imagine, and reason about the world.

marsbit06/29 05:09

Introduction to the Concept of World Models: A Story from Psychology to the Main Battlefield of AI

marsbit06/29 05:09

How to Conduct Deep Research Using Claude's Dynamic Workflows

The article "How to Use Claude's Dynamic Workflows for Deep Research" discusses overcoming the pitfalls of technical research, where both humans and AI can get overwhelmed by information, leading to vague conclusions. It introduces Claude Code's new "Dynamic Workflows" feature, which automatically designs and executes task-specific workflows before starting a task, unlike simpler "planning modes." This approach incorporates validation, result convergence, and adversarial verification from the outset. The core of Dynamic Workflows is six predefined scheduling patterns that address how to decompose tasks and synthesize results: 1. **Classify-and-Act (Routing):** An agent classifies the task and routes it to the most suitable specialist agent for execution. It's precise and efficient but struggles with ambiguous tasks. 2. **Fan-out & Merge:** The task is split into parallel, independent subtasks whose results are later merged. It's fast and isolates contexts but is more expensive and challenging to synthesize. 3. **Adversarial Verification:** Multiple "challenger" agents critique a worker agent's conclusion, requiring majority approval. This counters confirmation bias and self-assessment errors but relies on verifiable facts. 4. **Generate & Filter:** Multiple agents generate many candidate solutions, which are then filtered against a rubric to output only the best. It fosters diversity but depends heavily on the filter's quality. 5. **Tournament:** Multiple agents compete on the same task, with pairwise comparisons eliminating contestants over rounds to select the best. This offers stable relative judgment but is complex. 6. **Loop:** An agent iteratively attempts a task, learning from errors and adjusting until a stop condition is met. It handles tasks with unknown scope but risks infinite loops without proper design. The author compares their own custom deep-research system, which involved multi-agent analysis and deduplication but lacked goal-oriented convergence, to Claude's built-in workflow. The official workflow adds critical layers: initial problem decomposition, credibility assessment of sources, cross-agent voting to delete weak conclusions (not just averaging), and output tightly focused on the user's original goals and actionable recommendations. This structurally addresses common AI issues like goal drift, premature stopping, context pollution, and output bias. In summary, Dynamic Workflows represent a shift from smarter single conversations to a structured research process, compressing what used to require many dialogues into 3-4 interactions, albeit at higher token cost. The author notes remaining challenges for their specific domain (blockchain research): the need for fact-based verification over official documentation, depth in truly novel interdisciplinary thinking, the practical validation of proposed solutions, and tailoring information density to the audience.

marsbit06/09 03:07

How to Conduct Deep Research Using Claude's Dynamic Workflows

marsbit06/09 03:07

Turing Award Laureate Sutton's New Work: Using a Formula from 1967 to Solve a Major Flaw in Streaming Reinforcement Learning

New research titled "Intentional Updates for Streaming Reinforcement Learning" (arXiv:2604.19033v1), involving Turing Award laureate Richard Sutton, addresses a core challenge in deep reinforcement learning (RL): the "stream barrier." Current deep RL methods typically rely on replay buffers and batch training for stability, failing catastrophically when learning online from single data points (streaming). The authors propose a fundamental shift: instead of prescribing how far to move parameters (a fixed step size), their "Intentional Updates" method specifies the desired change in the function's output (e.g., a 5% reduction in value prediction error). It then calculates the step size needed to achieve that intent. This idea is inspired by the Normalized Least Mean Squares (NLMS) algorithm from 1967. Applied to value and policy learning, this yields algorithms like Intentional TD(λ) and Intentional AC. The method inherently stabilizes learning by adapting the step size based on the local gradient landscape, preventing overshooting/undershooting. In experiments on MuJoCo continuous control and Atari discrete tasks, Intentional AC achieved performance rivaling batch-based algorithms like SAC in a streaming setting (batch size=1, no replay buffer), while being ~140x more computationally efficient per update. The work demonstrates significant robustness, reducing reliance on numerous stabilization tricks. A remaining challenge is bias in policy updates due to action-dependent step sizes. Overall, this approach advances efficient, online, "learn-as-you-go" RL, enabling adaptive systems without massive data buffers or compute clusters.

marsbit05/10 06:28

Turing Award Laureate Sutton's New Work: Using a Formula from 1967 to Solve a Major Flaw in Streaming Reinforcement Learning

marsbit05/10 06:28

Farewell to Brute Force Computing: Reconstructing the Valuation Logic of AI for Science through HKUST's "GrainBot"

In 2026, Hong Kong's AI sector is rapidly transitioning from infrastructure development to deep application deployment. A key example is GrainBot, an AI tool developed by a team led by Prof. Guo Yike at HKUST, which represents a significant shift from general-purpose AI to specialized scientific discovery. GrainBot addresses critical challenges in materials science, particularly in analyzing microstructures like grain boundaries in materials used in semiconductors, batteries, and solar panels. Traditionally, this required manual, time-consuming, and error-prone analysis of microscopy images. GrainBot automates this process using computer vision and deep learning to accurately identify, segment grains, and quantify geometric features. It also correlates microstructural data with macro-material properties, as demonstrated in its application to perovskite solar cell research. This breakthrough highlights a broader trend in AI for Science (AI4S), where value is measured not by user metrics but by accelerated R&D cycles and novel discoveries. GrainBot’s potential to drastically shorten development timelines or uncover new materials with superior properties underscores a new valuation logic centered on industrial intellectual property. Hong Kong’s strength in combining domain expertise (e.g., materials science, chemistry) with AI capabilities creates a competitive advantage, positioning it as a hub for "autonomous labs" that integrate AI analysis with robotic experimentation. This model enables high-value patent output through fully automated, data-driven R&D, supporting a "Hong Kong R&D + Bay Area manufacturing" framework. However, challenges remain, particularly regarding data scarcity and silos in scientific research. High-quality, annotated datasets are limited, and data sharing barriers must be overcome through secure mechanisms like privacy computing for broader commercialization. GrainBot symbolizes a convergence of algorithmic innovation and scientific rigor, redirecting investment focus from sheer compute power to AI’s ability to solve real-world physical challenges. Hong Kong’s progress in AI4S signals emerging opportunities in a trillion-dollar AI-driven discovery market.

marsbit03/05 09:42

Farewell to Brute Force Computing: Reconstructing the Valuation Logic of AI for Science through HKUST's "GrainBot"

marsbit03/05 09:42

活动图片

Othersright-arrow

Ethereumright-arrow