# Efficiency İlgili Makaleler

HTX Haber Merkezi, kripto endüstrisindeki piyasa trendleri, proje güncellemeleri, teknoloji gelişmeleri ve düzenleyici politikaları kapsayan "Efficiency" hakkında en son makaleleri ve derinlemesine analizleri sunmaktadır.

Tsinghua University's Special Award Winner, Gu Yuxian, Joins DeepSeek

Tsinghua University's prestigious Graduate Special Scholarship recipient and 2021 Ph.D. candidate, Yuxian Gu, has officially joined DeepSeek. This news coincides with DeepSeek's major recruitment drive and the imminent launch of DeepSeek V4, on whose research paper Gu is listed as an author. A doctoral student in the Conversational AI group under Professor Minlie Huang at Tsinghua, Gu's research focuses on enhancing efficiency throughout the entire lifecycle of large language models. His key contributions span three areas: innovative methods for pre-training data selection (e.g., PDS), advanced knowledge distillation techniques for model compression (notably MiniLLM), and the development of efficient model architectures like Jet-Nemotron. His work has gained significant recognition, with nearly 5,000 citations on Google Scholar. Key publications include the highly cited surveys and papers on pre-trained models and the MiniLLM distillation method. As first author, he has presented at top-tier AI conferences including NeurIPS, ICLR, and ACL. One of his notable achievements is the Jet-Nemotron architecture, which combines Post-Neural Architecture Search (PostNAS) and a novel linear attention module called JetBlock. This model series demonstrates state-of-the-art performance rivaling larger models while achieving substantial efficiency gains in inference. Gu's expertise in creating powerful yet efficient AI systems aligns with industry needs, as evidenced by the adoption of his MiniLLM method by leading tech companies. His move to DeepSeek is anticipated to contribute further advancements in the field.

marsbit11 saat önce

Tsinghua University's Special Award Winner, Gu Yuxian, Joins DeepSeek

marsbit11 saat önce

Dialogue with IOTA Foundation's Jens: From Kenya to the UK, TWIN Propels Global Trade into the '5-Minute' Era

**Summary: IOTA's TWIN Project is Speeding Up Global Trade** Jens Munch Lund-Nielsen, Head of Global Trade & Supply Chain at the IOTA Foundation, discusses how the TWIN (Trade and Logistics Information Network) project is addressing long-standing inefficiencies in global trade. Traditionally, cross-border trade involves numerous parties, extensive paperwork, and delays spanning weeks. TWIN, built on IOTA's decentralized, scalable, and low-cost infrastructure, aims to replace this fragmented, paper-based system with a real-time, trusted, and interconnected digital collaboration layer. As a neutral digital public infrastructure, TWIN allows governments, businesses, and ports to exchange verifiable data and documents without ceding control to a single entity. This solves a core coordination problem that previous, proprietary platforms failed to address due to a lack of trust and neutrality. Key real-world implementations include: * **TLIP in East Africa:** Reduced document retrieval times for flower, coffee, and tea exports from 6-7 hours to ~30 minutes, cutting administrative work by 50-60%. * **UK Ecosystem of Trust Trials:** Provided border agencies with much earlier visibility (up to 20+ hours) of incoming chilled poultry shipments from Poland, enabling better resource planning. * Other pilots focus on supply chain traceability for fruits/vegetables, port efficiency, and critical mineral sourcing. TWIN ensures interoperability across jurisdictions by adopting global standards (e.g., UN/CEFACT data models), leveraging decentralized identity systems (DIDs), and supporting legal frameworks like MLETR. Looking ahead, Jens is optimistic that a trusted "connective layer" like TWIN will scale in the coming years. Its success could fundamentally transform global trade, most notably by helping to close the current $2.5 trillion trade finance gap. By providing lenders with access to verifiable, real-time supply chain data, TWIN could unlock capital and reshape how businesses discover partners and secure financing.

marsbit06/30 02:06

Dialogue with IOTA Foundation's Jens: From Kenya to the UK, TWIN Propels Global Trade into the '5-Minute' Era

marsbit06/30 02:06

The First Large-Scale Generative Model Using Physics as a Computational Primitive, Un-0, is Here: Could AI Energy Consumption Be Reduced by 1000x?

Unconventional AI has unveiled Un-0, a large-scale image generation model that uses the physical dynamics of coupled oscillators as its computational primitive. Described as the "first to use physics as a computation primitive for large-scale generative models," Un-0 aims to demonstrate a path toward dramatically reducing AI's energy consumption—potentially by up to 1000x compared to current GPU-based digital systems. The model operates by training a system of thousands of Kuramoto-like oscillators, where learned coupling strengths and natural frequencies define its behavior. Starting from random phases, the system evolves under its physical dynamics, guided by class-conditioning inputs, and a small traditional decoder then renders the final image. On ImageNet 64x64, Un-0 achieved an FID score of 6.74, performance comparable to early traditional generative models, though not yet state-of-the-art. The project, led by former Databricks AI chief Naveen Rao, represents a foundational step in using physical systems for computation, merging memory and processing in a single dynamic entity to bypass the energy-intensive data movement of von Neumann architectures.

marsbit06/26 10:54

The First Large-Scale Generative Model Using Physics as a Computational Primitive, Un-0, is Here: Could AI Energy Consumption Be Reduced by 1000x?

marsbit06/26 10:54

Matrixdock Featured Again in SBMA’s 《Crucible》: Discussing How Tokenisation Enhances Efficiency in the Precious Metals Market

Matrixdock's research article, titled "Why Tokenisation Matters for the Bullion Industry and How Carrying Costs Fit In," has been featured again in the SBMA's industry publication *Crucible*. Authored by Matrixdock lead Eva Meng, the piece examines how tokenisation enhances the efficiency and utility of the precious metals market. The article argues that tokenisation builds upon the accessibility improvements brought by gold ETFs, not by redefining gold's value but by enabling it to function within digital finance. It extends gold's role beyond a portfolio holding, potentially facilitating instant settlement, digital collateral, and operation in 24/7 markets. A key focus is transparently handling the unavoidable carrying costs (storage, insurance) of physical assets like gold and silver. Matrixdock introduces the Fungible Reserve Standard (FRS) framework, based on an "Economic Purity Principle," which aims to reflect these real-world economic costs clearly within the token mechanism, rather than bundling them opaquely. The platform's practical applications are highlighted, including its gold token XAUm and its silver token XAGm, the first built on the FRS framework. As the tokenised gold market surpassed $6 billion in February 2026, the industry's focus is shifting from initial proofs of reserves to broader concerns of market efficiency and capital utilization. Tokenisation is positioning gold and other precious metals to become active components within the evolving digital financial system.

marsbit06/18 09:20

Matrixdock Featured Again in SBMA’s 《Crucible》: Discussing How Tokenisation Enhances Efficiency in the Precious Metals Market

marsbit06/18 09:20

How Much of the Subscription Fee You Pay to Claude Can Optical Module Companies Get?

How much of your $20 Claude Pro subscription actually goes to AI model companies like Anthropic? A viral breakdown image highlights the fundamental valuation challenge for AI applications versus traditional SaaS. Unlike SaaS with high software margins, AI subscriptions face variable "inference costs": every user query consumes GPU time, power, and cloud resources. This creates a tension between fixed subscription fees and usage-driven expenses. While the specific dollar splits are illustrative, the core question is whether AI revenue can achieve SaaS-like margins as usage scales. Currently, infrastructure providers (cloud platforms, GPU makers like Nvidia, HBM suppliers, power/data centers) capture more certain revenue from growing AI usage. Their financials reflect pricing power and faster earnings validation. The bullish case hinges on efficiency improvements: model optimization, caching, smaller models, and custom chips could lower per-token costs over time. The key debate is whether cost declines can outpace increases in user workload complexity and volume. Ultimately, for AI companies to command high SaaS-like valuations, they must demonstrate not just user growth but also improving gross margins after accounting for inference costs. Investors will scrutinize not just subscriber numbers, but usage patterns, enterprise pricing tiers, and real efficiency gains.

marsbit06/17 03:43

How Much of the Subscription Fee You Pay to Claude Can Optical Module Companies Get?

marsbit06/17 03:43

AI Leads to Layoffs? Research Shows AI Is More Expensive Than the People It Replaces

Title: AI Layoffs? Research Shows AI is More Expensive Than the Workers It Replaces. This year, nearly 50,000 employees have been laid off due to AI, with companies initially believing AI could replace human jobs. However, recent findings indicate that the actual operational costs of AI often exceed the expense of the human labor it was meant to replace. Examples include Uber exhausting its annual AI budget in just four months, Microsoft cutting Claude Code licenses due to high costs, and an Anthropic employee incurring $150,000 in API usage in a single month. A CloudZero survey reveals that 45% of enterprises spend over $100,000 monthly on AI, yet only 8% of S&P 500 companies report any AI-related revenue, and half struggle to measure ROI. Analyst Scott Galloway predicts a shift toward cheaper Chinese AI models, which are 10 to 30 times more affordable than American counterparts. Data shows Chinese models' usage among developers surged from 1% in 2024 to over 60% by mid-2026, with 80% of U.S. AI startups adopting them. This trend may prompt regulatory responses, such as potential restrictions from the Trump administration.

marsbit06/09 02:25

AI Leads to Layoffs? Research Shows AI Is More Expensive Than the People It Replaces

marsbit06/09 02:25

Can DeepSeek Save China One Trillion Dollars?

"DeepSeek and the $1 Trillion Infrastructure Question" The article examines whether DeepSeek's AI optimization breakthroughs could potentially save China $1 trillion in future AI infrastructure costs. The analysis begins with Nvidia's upcoming Vera Rubin AI platform, costing ~$7.8 million, where memory (HBM4/LPDDR5X) constitutes $2 million—a 435% cost increase in one year, highlighting how AI hardware spending is shifting toward expensive memory components. DeepSeek's approach works in the opposite direction. Through three key technical innovations showcased in DeepSeek V4, the company dramatically improves hardware efficiency: 1. **Memory Compression (MLA)**: Re-engineers the attention mechanism to compress long-context memory (KV Cache) by over 90%, drastically reducing expensive HBM usage. 2. **Selective Activation (MoE)**: Employs Mixture-of-Experts architecture where only a small fraction of parameters (e.g., 49B out of 1.6T in V4-Pro) are activated per token, allowing most parameters to reside in cheaper memory/SSD. 3. **Computation Caching**: Reuses previously computed results via cache hits, replacing expensive GPU computations with cheap memory reads. Combined, these optimizations allow the same hardware to produce approximately 4x more tokens, effectively reducing required hardware investment by 75%. DeepSeek's pricing reflects this: a 10-billion token workload costs ~$522 monthly versus ~$9,000-$10,000 for competitors. The $1 trillion savings projection stems from McKinsey's estimate that global AI infrastructure will require ~$5.2 trillion investment by 2030. As China's daily token consumption grows toward quadrillions, even marginal efficiency gains scale massively. With a conservative 4x throughput improvement, China could avoid building tens of thousands of AI data centers equivalent to ~7 trillion RMB ($1 trillion) in saved investment. Critically, this strategy shifts dependency from scarce, expensive GPU/HBM—where China lags—toward more accessible storage, caching, and systems engineering where domestic suppliers like CXMT are gaining strength. Rather than "replacing Nvidia," DeepSeek rebalances AI's value chain away from monolithic hardware dependency. Ultimately, DeepSeek's technical breakthroughs could lower the barrier to AI adoption across Chinese industries by making advanced capabilities affordable at scale—transforming who can access next-generation AI.

marsbit06/03 00:47

Can DeepSeek Save China One Trillion Dollars?

marsbit06/03 00:47

Running MoE on Mobile Phones? Meta Proposes MobileMoE, Speeding Up iPhone 16 Pro by 3.8x

Meta's MobileMoE, a mobile-optimized Mixture-of-Experts (MoE) language model architecture, enables efficient on-device large language model (LLM) inference for the first time on commercial smartphones. Designed for decoder-only Transformers, it replaces dense feed-forward layers with MoE layers. Key design choices include 8 experts with granularity g=8, top-4 routing, and a shared expert. The model undergoes a four-stage training process: pre-training, intermediate training, supervised fine-tuning, and quantization-aware training. Results show MobileMoE models, with similar memory footprint, achieve equal or higher average accuracy across 14 foundational benchmarks while using only 1/2 to 1/4 of the FLOPs compared to dense baselines. After INT4 quantization, they remain competitive. Notably, on an iPhone 16 Pro, MobileMoE-S demonstrates significant speedups: up to 3.8x faster in the prompt phase and 2.2-3.4x faster in per-token generation compared to a dense counterpart, with lower peak memory usage. While MobileMoE establishes a new Pareto frontier for on-device LLMs in accuracy-compute trade-offs, particularly excelling in code and math tasks, it currently lags behind models like Qwen3.5 2B in advanced instruction following and knowledge reasoning. Future work includes improving post-training techniques, exploring NPU deployment, and managing the runtime memory sensitivity of MoE models to varying inputs.

marsbit06/01 06:09

Running MoE on Mobile Phones? Meta Proposes MobileMoE, Speeding Up iPhone 16 Pro by 3.8x

marsbit06/01 06:09

NVIDIA Launches DSX Platform, Expanding into AI Factory Infrastructure

NVIDIA has unveiled the DSX platform at its GTC Taipei event, marking a strategic expansion from GPU sales into comprehensive AI factory infrastructure solutions. The platform addresses challenges like power supply, cooling, and resource orchestration as AI models scale, shifting the industry focus from single-chip performance to overall infrastructure efficiency. DSX integrates NVIDIA's chips, systems, software, and partner technologies to cover the entire AI factory lifecycle—from design and simulation to deployment and operations. It aims to accelerate deployment, improve reliability and operational efficiency, and reduce the cost per generated token in AI inference. The software suite includes DSX MaxLPS, which uses 45°C liquid cooling and rack-level optimization to allow up to 40% more GPUs per megawatt, and DSX OS, an open-source platform for AI factory operations. The platform also encompasses reference designs, digital twin simulation (DSX Sim), dynamic workload adjustment based on grid conditions (DSX Flex), and data exchange between systems. Early adopters include cloud providers like CoreWeave and Lambda. Major hardware partners, including Dell, HPE, Lenovo, and Supermicro, are developing DSX-ready systems. Pilot projects for DSX Flex are underway with energy providers. Strategically, DSX represents NVIDIA's ongoing transition from an AI chip supplier to a full-stack AI infrastructure platform provider, aiming to set industry standards and solidify its market leadership.

marsbit06/01 04:27

NVIDIA Launches DSX Platform, Expanding into AI Factory Infrastructure

marsbit06/01 04:27

After Burning Tens of Billions of Dollars in Tokens, Silicon Valley Giants Start Limiting Employee Token Usage

After burning tens of billions of dollars on AI tokens, major Silicon Valley firms are now restricting employee usage. Companies like Microsoft, Uber, and Salesforce, which heavily promoted AI for "efficiency," are facing a cost crisis. The practice of "tokenmaxxing"—pushing employees to maximize AI tool usage—led to wasteful spending on trivial tasks like checking the weather or writing birthday messages, with studies showing significant hidden costs for bug fixes and code rewrites. The core issue is a misalignment between individual productivity gains and actual business value. While employees use AI to automate tasks they dislike, such as writing reports, this often doesn't translate to increased company revenue or improved core business outcomes. For instance, AI-generated code speeds up development but also sees an 800% increase in "code churn" (code being discarded or rewritten). As a result, only 14% of CFOs report seeing a clear, measurable return on AI investments. Firms are now shifting strategies. Microsoft has revoked most internal licenses for Claude Code, while others are implementing monitoring and cost controls. New tools from companies like Harness and CloudZero aim to track AI spending and tie costs to business results. Some AI vendors, like HubSpot, are moving from token-based pricing to charging based on outcomes, such as "resolved conversations" or "leads generated." This represents a necessary correction in the AI adoption cycle. The challenge now is for companies to move beyond using AI merely to speed up old tasks and instead rethink their workflows and business models fundamentally. The future of enterprise AI depends on proving its value, not just its usage.

marsbit06/01 04:06

After Burning Tens of Billions of Dollars in Tokens, Silicon Valley Giants Start Limiting Employee Token Usage

marsbit06/01 04:06

活动图片

Othersright-arrow

Project Updatesright-arrow