# Training Articoli collegati

Il Centro Notizie HTX fornisce gli articoli più recenti e le analisi più approfondite su "Training", coprendo tendenze di mercato, aggiornamenti sui progetti, sviluppi tecnologici e politiche normative nel settore crypto.

DeepSeek Secretly Builds AI Chip, Specializing in Inference, Project Started a Year Ago with No Public Recruitments

DeepSeek, the Chinese AI company known for its algorithmic models, is secretly developing its own AI chip to reduce dependence on Nvidia, according to a Reuters report. The chip is designed specifically for AI inference, not training, and the project began approximately a year ago. Currently in early stages, DeepSeek is reportedly in discussions with chip design firms, foundries, and memory suppliers. The company, historically focused on algorithmic efficiency, has been discreetly hiring chip design engineers without public job postings. This move aligns with a global trend where major AI model companies like OpenAI and Anthropic are also pursuing custom chip development. DeepSeek founder Liang Wenfeng previously highlighted chip shortages as a challenge. While the company initially trained models on Nvidia H800s and later adapted to Huawei's Ascend chips, it now seeks greater control over its hardware foundation. Designing a competitive AI chip is a significant challenge, requiring years and substantial investment with no guarantee of success. However, DeepSeek's efforts are backed by a recent major funding round of approximately 51 billion RMB (about $7.4 billion) raised in June 2026. The funds are designated for expanding data centers based on domestic chips, developing proprietary AI chips, and recruiting top global talent. Infrastructure plans are also advancing, with job postings for data center design engineers, including projects in locations like Ulanqab, Inner Mongolia. The company remains characteristically low-key, with sources speaking anonymously and no official comment from DeepSeek itself. Nevertheless, this initiative marks a strategic expansion from software algorithms into the hardware layer that powers its AI systems.

marsbit4 h fa

DeepSeek Secretly Builds AI Chip, Specializing in Inference, Project Started a Year Ago with No Public Recruitments

marsbit4 h fa

Dwarkesh Patel: The Next Generation of AI May Be Built Through Actual Work

In his latest podcast, Dwarkesh Patel explores the next paradigm for AI training. While current progress in fields like coding and math relies on Reinforcement Learning with Verifiable Rewards (RLVR), which requires tasks that are both verifiable and highly scalable ("grindable"), Patel questions whether this is sufficient for complex real-world objectives like starting a business, winning a legal case, or managing an organization. These tasks provide verifiable outcomes but lack the resetable, parallelizable environments needed for efficient RLVR training. Patel argues the key limitation of current models is their inability to convert valuable in-context learning from real deployment into permanent weight updates—a process he terms "learning back to the weights." He proposes two potential solutions: On-Policy Self-Distillation (OPSD), where a model distills knowledge from long, task-specific sessions back into its base weights, and "dreaming," where an AI constructs simulated environments from real-world observations to practice and refine strategies. Ultimately, Patel envisions a future training paradigm where AI advances not just through pre-training on static datasets but through continual, post-deployment learning from real-world experience. This shift would enable AI to move beyond "grindable" tasks and develop robust, generalizable agent capabilities for complex, real-world challenges.

marsbit06/28 23:49

Dwarkesh Patel: The Next Generation of AI May Be Built Through Actual Work

marsbit06/28 23:49

First Long-Horizon Doc2Repo Training Dataset: Code Agents Move Beyond Bug Fixing and Begin Creating Repositories

With the advancement of LLM Code Agents, the research focus is shifting towards long-horizon, real-world tasks, moving beyond simple bug fixes to full repository generation. To address this, researchers from Renmin University of China introduced the DeNovoSWE dataset. This dataset focuses on long-term software engineering tasks, specifically the "document-to-repository" challenge—generating an entire, executable code repository from a task description. The DeNovoSWE construction method employs a Divide & Conquer approach. It breaks down target repositories into core capabilities and uses a multi-agent Draft-Critic-Repair workflow to automatically generate high-quality, evaluation-aligned task documents. The dataset also implements difficulty-aware filtering to balance quality and diversity. The result is a high-quality, anti-leakage dataset of 4,818 instances. Experiments show that models trained on DeNovoSWE achieve significant improvements in long-horizon repository generation. For instance, Qwen3-30B-A3B-Instruct's performance on the BeyondSWE-Doc2Repo benchmark increased from 5.8% to 47.2%, and on NL2RepoBench from 4.3% to 23.0%. Similar gains were observed with stronger backbones, demonstrating that dedicated long-horizon training data is crucial for advancing Code Agents from maintainers to architects capable of planning and building complete software projects from scratch.

marsbit06/25 08:51

First Long-Horizon Doc2Repo Training Dataset: Code Agents Move Beyond Bug Fixing and Begin Creating Repositories

marsbit06/25 08:51

The Computing Power Dilemma in the Sino-US AI Rivalry

The Sino-US AI rivalry faces a fundamental bottleneck: the widening compute power gap. While Chinese AI chip companies have seen investment surges, their current focus remains largely on the less demanding inference market. The real challenge lies in the high-end training chip sector, crucial for developing cutting-edge large language models (LLMs), where Nvidia holds a near-monopoly. The compute disparity is stark. US tech giants like Meta, Google, and xAI command massive GPU clusters, enabling them to train trillion-parameter models rapidly. Estimates suggest US data center count and total compute capacity significantly outstrip China's. This "brute force" advantage allows for faster model iteration and exploration of larger parameter scales, with top US models reportedly leading their Chinese counterparts by 8 to 15 months. Chinese alternatives, such as Huawei's Ascend and others from companies like Moore Thread and Biren, are emerging. They show promise in inference and some training scenarios, closing the performance gap with mid-range Nvidia products. However, the core hurdle extends beyond raw chip performance to the entrenched software ecosystem, exemplified by Nvidia's CUDA platform. The path forward involves "walking on two legs": navigating import restrictions while heavily investing in the domestic chip industry. Though still in a catch-up phase, China's vast market, talent pool, and capital are fostering progress. The ultimate test is whether Chinese firms can build a competitive hardware-software ecosystem to power the next generation of AI.

marsbit06/22 10:21

The Computing Power Dilemma in the Sino-US AI Rivalry

marsbit06/22 10:21

Robots Begin to 'Consume Data': The Hidden Production Chain from Indian Data Factories to Billion-Dollar Humanoid Robots

Robots have started to 'consume data,' driving the formation of a new industrial supply chain focused on producing training data for embodied AI. Unlike large language models, which are trained on vast internet text corpora, embodied AI models face a 'data desert' in the physical world. This has created a massive demand for first-person perspective video data (Ego Data), captured by workers wearing cameras in places like Indian garment factories. Companies like Neocambrian AI are establishing 'data factories' where workers perform standardized tasks (e.g., sorting clothes, kitchen organization) to generate thousands of hours of video. Research, such as NVIDIA's EgoScale, demonstrates that scaling this human demonstration data predictably improves robot performance, particularly for dexterous manipulation. This has validated a training path combining large-scale human data for pre-training with smaller amounts of robot-specific data for fine-tuning. The value of different data types varies significantly, forming a 'data pyramid.' The base consists of low-cost, large-scale internet and Ego Data. Higher layers include more expensive motion-capture data (e.g., from data gloves), simulation/synthetic data, and the most costly and scarce layer: real robot teleoperation data. This demand has spawned a layered ecosystem of data suppliers: low-cost data factories, motion capture and alignment specialists, robot-native teleoperation service providers, simulation data companies, and platforms aiming for data standardization. Robot companies themselves are adopting a 'layered procurement' strategy: outsourcing generic Ego Data while building in-house capabilities for robot-specific adaptation data and the critical deployment/failure data generated in real-world applications. The industry is shifting focus from hardware and basic mobility to the data pipelines required for general-purpose capability. While parallels exist to data labeling companies like Scale AI in the LLM boom, the physical complexity of robot data—involving action success ambiguity and sim-to-real gaps—requires more integrated solutions for data collection, annotation, and a continuous feedback loop. The race is on to build the data engines that will teach robots to operate reliably in the unstructured real world.

marsbit06/13 03:32

Robots Begin to 'Consume Data': The Hidden Production Chain from Indian Data Factories to Billion-Dollar Humanoid Robots

marsbit06/13 03:32

The 2026 Landscape of Decentralized AI: Why Blockchain is the Inevitable 'Antidote' for AI?

Decentralized AI 2026 Landscape: Why Blockchain is AI's Essential "Antidote" Centralized AI faces structural bottlenecks—expensive compute, concentrated control, unverifiable outputs, and difficult data access—that cannot be solved by capital or code alone. Blockchain offers a path to make intelligence open, verifiable, and economically accessible. The decentralized AI stack comprises: * **Infrastructure:** The foundation with compute, verifiable inference, distributed training, data/storage, and privacy/verification layers. Projects like Akash, Render, and Filecoin provide cheaper, decentralized alternatives for raw resources. * **Middleware:** The coordination layer for agent discovery, identity, and commerce. Key players include Bittensor (a network of specialized AI subnets), Virtuals (an agent economy OS), and frameworks providing agent identity and tooling. * **Applications & Services:** Dominated by Agentic Finance (AI agents executing on-chain actions based on natural language) and Agentic Payments (machine-to-machine transactions using blockchain as a settlement layer). Projects like Giza, Infinit Labs, and x402 are enabling these use cases. Key trends for 2026-2027 show AI demand outgrowing infrastructure, compute becoming an asset class, and tokenomics emerging as a structural advantage for coordinating capital, compute, and data. While still early—with adoption uneven and revenue often trailing token incentives—projects like Bittensor, NEAR, and Venice demonstrate decentralized AI is evolving from a narrative into a new model for coordinating intelligence.

Foresight News06/11 10:02

The 2026 Landscape of Decentralized AI: Why Blockchain is the Inevitable 'Antidote' for AI?

Foresight News06/11 10:02

Crypto Criminals In Crosshairs As Chainalysis Partners With South Korean Police

South Korea's national police have partnered with blockchain analytics firm Chainalysis to strengthen their investigative capabilities against crypto-related crimes. The agreement, formalized through a memorandum of understanding, provides Korean investigators with access to specialized training and certification programs. This collaboration aims to build institutional expertise for tackling a broad spectrum of virtual asset crimes, including money laundering and state-sponsored hacking. The move follows the recent creation of a dedicated multi-agency task force to eradicate crypto-based money laundering. While North Korea-linked crypto theft, which saw massive losses in 2025, is a significant concern, officials emphasize the partnership is designed to enhance overall investigative capacity for handling complex, cross-border crypto crime cases.

bitcoinist06/11 10:01

Crypto Criminals In Crosshairs As Chainalysis Partners With South Korean Police

bitcoinist06/11 10:01

Is AI Creating a New Class of 'Information Poor'?

AI is generating a new kind of "information poverty." The core issue isn't that AI denies answers to the poor; it's that it provides abundant, cheap, and plausible-sounding answers to everyone. This availability shifts the true scarcity from obtaining answers to possessing the **judgment to evaluate them** and the access to turn them into real-world opportunities. New information poverty thus describes those who have AI tools and outputs, but lack the complementary skills, authorization, and contextual experience to critically assess and act on them. Research reveals a multi-layered divide: access to AI is stratified by income and platform design (e.g., premium vs. free, embedded tools). In workplaces, usage heavily favors higher-paid, more experienced, or formally trained employees, with AI often automating entry-level tasks that were traditional stepping stones. Crucially, the heaviest users are often mid-career professionals whose existing expertise allows them to effectively judge and leverage AI outputs, while novices risk over-relying on them without building judgment. While controlled experiments show AI can significantly boost low-skilled workers' performance, real-world adoption and benefit are constrained by unequal social and organizational structures. Historically, general-purpose technologies first reward those with existing complementary capital. AI, by affecting judgment-based work, may accelerate and deepen this initial inequality gap, even if it narrows over decades. The danger lies in the illusion of competence it creates, potentially stunting the very critical thinking needed in an era where judgment is paramount.

marsbit06/08 11:38

Is AI Creating a New Class of 'Information Poor'?

marsbit06/08 11:38

Grok's Latest Model Gets a 'Bonus' from Cursor, Musk: Significant Improvement in Coding

Elon Musk announced that xAI's new Grok foundation model, V9-Medium (1.5T), has completed training and is expected to launch in 2-3 weeks. Key improvements focus on programming capabilities, achieved by supplementing the model's training with substantial data from Cursor. Musk emphasized a "huge improvement" in coding performance. Concurrently, xAI's new AI programming agent tool, Grok Build, has entered an early beta testing phase. It functions as a CLI tool/agent for developers, featuring planning modes, parallel subagents, and integration capabilities. This move aligns with xAI's broader push into the AI coding space, following its $60 billion acquisition of Cursor last month and the recruitment of key Cursor engineering talent. The integration of Cursor's data and expertise is seen as a major boost for Grok's real-world engineering and coding proficiency.

marsbit05/26 04:18

Grok's Latest Model Gets a 'Bonus' from Cursor, Musk: Significant Improvement in Coding

marsbit05/26 04:18

When Computing Power Becomes Commoditized, How Long Until a GPU Futures Market Emerges?

"When Will GPU Futures Arrive? A Framework for Assessing Compute as a Commodity" The article explores the potential for a robust futures market for compute power (GPUs), arguing that such a market is not yet mature but may emerge. It analyzes the landscape using a five-part framework developed for new commodity futures markets. The analysis scores the current state: * **Fragmented Supply (Red)**: Supply is highly concentrated among hyperscale cloud providers (AWS, Azure, GCP, Oracle), limiting the need for price discovery. * **Price Volatility (Green)**: GPU pricing is already highly volatile due to uncertain supply and surging demand. * **Physical Settlement Infrastructure (Green)**: Early infrastructure exists via OTC brokers and price indices (e.g., Ornn, Silicon Data) standardizing contracts. * **Standardized Unit (Red)**: A lack of standardized, tradable units hinders markets; a GPU instance hour varies by region, configuration, and contract terms. * **Lack of Alternatives (Yellow)**: Large players hedge internally via vertical integration, while smaller players bear spot market risk. Overall, the market shows promise (volatility, early infrastructure) but lacks the fragmented supply and standardization needed for large-scale futures trading. Most activity remains OTC. Key open questions and hypotheses: 1. Supply is expected to fragment moderately in 1-2 years, driven by new cloud providers, cheap power locations, and demand from non-frontier labs and AI startups using open-source models. 2. Standardization is most likely to emerge around inference workloads (forecast to be >65% of AI compute demand by 2029), which have simpler, more homogeneous hardware needs than training. Widespread adoption of open-source model weights could accelerate this by democratizing inference and creating demand for optimized, standardized infrastructure. 3. The primary traded unit will likely be the **"chip instance hour"** (akin to electricity, traded regionally), not the physical chip or the downstream AI output (tokens).

marsbit05/18 09:09

When Computing Power Becomes Commoditized, How Long Until a GPU Futures Market Emerges?

marsbit05/18 09:09

活动图片

Market Analysisright-arrow

Industry Newsright-arrow