# Сопутствующие статьи по теме LLM

Новостной центр HTX предлагает последние статьи и углубленный анализ по "LLM", охватывающие рыночные тренды, новости проектов, развитие технологий и политику регулирования в криптоиндустрии.

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

Tsinghua '00s Alumnus Wang Guan's New Work: Disrupting Transformer Pretraining Models with 1/900 Tokens, 1/432 Compute Power

Tsinghua alumnus Wang Guan's team proposes HRM-Text, a novel pre-training paradigm using a Hierarchical Recurrent Model to replace standard Transformers. With just 1B parameters and 40B unique tokens trained at a cost of ~$1500, HRM-Text achieves performance comparable to 2B-7B open-source models, using up to 900x fewer tokens and 432x less estimated compute. Key innovations include a dual-timescale recurrent architecture for greater effective depth, a task-completion objective training only on answer tokens with PrefixLM masking, and techniques like MagicNorm and Warmup Deep Credit Assignment for stability. Evaluations show strong results on benchmarks like MMLU (60.7%) and GSM8K (84.5%). The work highlights how architectural priors and targeted objectives can lower pre-training barriers, though limitations include knowledge-reasoning coupling, fixed compute per token, and scalability beyond 3B parameters.

marsbit05/26 03:16

Tsinghua '00s Alumnus Wang Guan's New Work: Disrupting Transformer Pretraining Models with 1/900 Tokens, 1/432 Compute Power

marsbit05/26 03:16

Investors Frantically Snap Up AI Firms with 'No Profits': A High-Stakes Gamble on 'the Right to Define the Future'

"Investors are pouring billions into Chinese AI startups with no profits, betting on the future of the industry. A state-backed fund is reportedly in talks to lead DeepSeek's funding at a $45B valuation, just weeks after it was valued at $10B. Along with companies like Zhipu AI, MiniMax, and Kimi (backed by Meituan and Alibaba), their combined valuation exceeds $140B. This isn't a typical venture capital play. Investors are paying for 'future definition rights'—a chance to set the standards for the next tech era. Morgan Stanley notes a 6-12 month window for this scarcity premium before more AI companies go public. Despite massive losses, these companies show strong growth. Zhipu AI's API revenue grew 60x, Kimi's annual recurring revenue doubled to $200M in a month, and MiniMax turned its gross margin positive, with over 70% of revenue from overseas. Their valuations vastly exceed profitable firms like iFlytek. Crucially, technical progress underpins this growth. DeepSeek's latest model boasts costs just 1% of a leading competitor's, while Zhipu AI has raised API prices due to high demand. However, gaps with top global models remain. Tech giants like Tencent and Alibaba, investing heavily while describing their own AI efforts as 'leaky boats,' are also investing in these startups as a hedge. Key risks loom: the closing scarcity window, computing power bottlenecks limiting growth, and the sustainability of DeepSeek's cost-advantage model. With state capital now a major player, the success of these companies has become a strategic national concern. The next year will test if their soaring valuations can be justified by future profits."

marsbit05/26 02:06

Investors Frantically Snap Up AI Firms with 'No Profits': A High-Stakes Gamble on 'the Right to Define the Future'

marsbit05/26 02:06

TechFlow Intelligence Report: Huawei Unveils "Tao" Law, Semiconductor Sector Surges; Meta Cuts 10% of Workforce

"TechFlow Intelligence Brief": Huawei's new "Tau Law" in semiconductors and Meta's 10% layoffs headline today's tech landscape. In AI, breakthroughs include an AI solving 9 high-difficulty pure math problems for just a few hundred dollars each, and DeepSeek's new Reasonix programming agent challenging commercial models. However, research highlights a "constraint decay" issue in LLM-generated backend code. Open-source model Qwen 3.6 27B achieves high speeds on older GPUs, sparking debate on NVIDIA's future dominance. In Crypto/Web3, Ethereum Foundation plans to downsize, possibly reducing ETH selling pressure. Fake news about CZ ignited a meme coin frenzy, showing the market's sensitivity to celebrity narratives. DeFi sees a new trend in HELOC-backed Real World Asset (RWA) pools. The chip sector is stirred by Huawei's proposed "Tau (τ) Law," aiming for 1.4nm-equivalent performance by 2031 through architectural innovation, causing related stocks to surge. A report notes memory now constitutes nearly two-thirds of AI chip cost. Meanwhile, executives at 7 Chinese semiconductor firms sold shares after price peaks. Meta announces 10% layoffs as it pivots to AI. Google's CEO faced student protests over AI ethics during a speech, and the company controversially published a Chromium exploit before patching was complete. Xiaomi permanently banned installers for AC installation fraud. In US stocks, AMD is seen as a potential challenger to NVIDIA, while a survey reveals 99% of CEOs expect AI-driven layoffs within two years. Palantir secured a government contract for employee monitoring, raising privacy concerns. Macro developments include a 6% drop in WTI crude oil on hopes for reopened Hormuz Strait, and silver prices rising over 4%. Global oil inventories are nearing critical lows. New trends highlight a "audio prompt injection" attack targeting AI voice assistants via hidden commands, and CBS pausing takedowns of pirated Stephen Colbert episodes after public pushback. The underlying narrative connects AI's cost-effective problem-solving, widespread planned job displacement, and Huawei's challenge to Western tech hegemony, framing the AI and chip race as a broader contest over employment, geopolitics, and the very definition of intelligence.

marsbit05/25 10:50

TechFlow Intelligence Report: Huawei Unveils "Tao" Law, Semiconductor Sector Surges; Meta Cuts 10% of Workforce

marsbit05/25 10:50

Agentic Design Patterns: A Book That Made Me Re-Understand "What Is an Agent, Really?"

"Agentic Design Patterns" is a 2025 book by Antonio Gullí, a Google engineering director, which offers a systematic framework for AI Agent development through 21 design patterns. A core contribution is the "Four Levels of Agency": Level 0 (bare LLMs) are not true agents. Level 1 agents actively decide when and how to use tools. Level 2 agents engage in strategic planning, context engineering (curating and filtering information), and self-reflection. Level 3 involves multi-agent collaboration with defined communication topologies. The book introduces **Context Engineering** as a superset of prompt engineering, managing four layers of information for the agent: system prompts, external data, implicit context (user history, environment), and feedback loops for automated optimization. A key pattern is **Reflection (Producer-Critic)**, where two distinct agents with different prompts collaborate iteratively—one produces output, the other critiques it—until quality is satisfactory or a max iteration limit is reached. For **Memory**, a three-layer model is proposed: Session (ephemeral conversation context), State (temporary task data), and Memory (persistent, long-term storage). Regarding **Multi-Agent Systems**, the book advises against unnecessary complexity, recommending simple topologies like Supervisor or Peer-to-Peer based on task needs. It emphasizes perfecting a single Level 2 agent before moving to multi-agent setups. The author concludes with three actionable takeaways: 1) Add a Critic agent to existing workflows, 2) Practice Context Engineering beyond simple prompts, and 3) Avoid premature multi-agent complexity; first master a robust single agent. The book provides a practical map, codifying common challenges like reflection, memory, and coordination into reusable patterns, saving developers from reinventing foundational solutions.

链捕手05/25 04:43

Agentic Design Patterns: A Book That Made Me Re-Understand "What Is an Agent, Really?"

链捕手05/25 04:43

DeepSeek Announces Permanent Price Cut, But Liang Wenfeng Is Not Trying to Be a "Cyber Bodhisattva"

DeepSeek has announced a permanent 75% discount on its V4-Pro API, significantly reducing its token prices. This move stands out as a major industry-wide price cut while competitors like Anthropic, OpenAI, and Google have been quietly raising theirs. The article contrasts this strategy with the broader trend of AI becoming more expensive, citing examples of companies like Microsoft and Uber struggling with high token costs as usage soars. While CEO Liang Wenfeng is hailed by some as a "Cyber Bodhisattva" for this普惠 approach, the article argues this is a strategic business choice, not mere altruism. DeepSeek's ability to maintain low prices is attributed to several structural advantages: lower-cost AI talent in China, the impending use of domestic昇腾 hardware for further cost reductions, and, most critically, access to China's cheaper and more abundant energy infrastructure, which drastically reduces the electricity costs dominating AI operations. The analysis suggests that for many commercial applications, a "good enough" model that is radically cheaper (e.g., 1% to 11% of GPT-5.5's cost) is more valuable than the absolute top-tier model. This allows for vastly more experimentation and iteration within a budget. Therefore, as AI generally becomes more expensive, DeepSeek's cost-competitiveness—rooted in China's energy and talent advantages—becomes its core strategic value and differentiator in the global market.

marsbit05/24 12:19

DeepSeek Announces Permanent Price Cut, But Liang Wenfeng Is Not Trying to Be a "Cyber Bodhisattva"

marsbit05/24 12:19

Why Did Zhipu Surge Nearly 30% in a Single Day?

"Global AI Model Unicorn" Zhipu's stock surged nearly 30% in a single day, reaching a new market cap high. The catalyst was the launch of its GLM-5.1-highspeed API, boasting a generation speed of **400 tokens per second**, setting a new global benchmark. This speed, roughly 3-5 times faster than industry leaders like OpenAI's GPT-4o and Anthropic's Claude, is achieved **without compromising the full-scale model's capabilities**. In the era of AI Agents requiring dozens of self-calls, such latency reduction is critical, transforming speed from a system metric into a determinant of intelligence limits. The breakthrough stems from a three-layer technical overhaul: 1. **TileRT Inference Engine**: Compiles the entire model into a continuous, always-on computation pipeline using "Warp Specialization," minimizing GPU idle time by having different processor groups handle data loading, computation, and communication in parallel. 2. **Heterogeneous Parallelism for MLA**: To efficiently run the GLM-5.1 model using the MLA attention mechanism, TileRT employs a heterogeneous strategy. One GPU handles sparse indexing/routing, while the others perform dense computation, optimizing for MLA's unique workflow. 3. **ZCube Network Architecture**: Replaces the standard Spine-Leaf (ROFT) network topology with a flat, dual-group interconnect. This design creates a single optimal path between any two GPUs, eliminating network congestion at scale and reducing latency. The business impact is significant: a 15% increase in cluster throughput (free extra capacity), a 40.6% reduction in tail latency (improved stability), and a one-third cut in networking hardware costs. Long-term, this innovation challenges the dominance of NVIDIA's integrated hardware-software stack (GPU+NVLink+InfiniBand), potentially benefiting manufacturers of high-density Leaf switches and optical modules while lowering the software barrier for domestic AI chips like Huawei's Ascend. The innovation proves that more can be achieved with the same compute, reshaping the infrastructure beyond just GPUs.

marsbit05/23 01:23

Why Did Zhipu Surge Nearly 30% in a Single Day?

marsbit05/23 01:23

Detained for 37 Days: The First Wave of People Who Got Rich from 'AI Gateways' Are Starting to Go to Jail

A prominent AI proxy service operator was reportedly detained for 37 days and is now on bail pending trial, highlighting the legal risks in China's booming but unregulated AI intermediary market. These services act as "AI scalpers," providing domestic users with access to restricted overseas models (like OpenAI, Claude) by bundling APIs, handling payments, and bypassing network blocks, all for a fee. Their controversial profitability stems from practices like bulk-registering accounts to resell free credits, exploiting refund policies, overcharging for tokens, substituting cheaper models, and illegally selling user conversation data. Major figures, including cryptocurrency entrepreneurs, are now entering this space. Legally, these operations face severe risks. Their core model often involves unauthorized API access and operating without required telecom licenses, potentially constituting illegal business operations. They fail to meet data security obligations for the vast amounts of user data they process, risking charges for failing to fulfill network security duties. Crucially, the unauthorized collection and sale of user data, which can include personal and commercial secrets, easily meets the threshold for the crime of infringing on personal information. The case underscores a critical juncture for the AI industry. While proxies lower access barriers, they expose user data to unsecured middlemen and undermine the business models of AI developers, forcing them to divert resources to security and distorting market value perceptions. The article argues that the industry's sustainable future depends on building trust, protecting data, and fostering compliant competition, moving away from its current "wild growth" phase.

marsbit05/21 14:40

Detained for 37 Days: The First Wave of People Who Got Rich from 'AI Gateways' Are Starting to Go to Jail

marsbit05/21 14:40

Two Companies Capture 90% of AI Startup's $80 Billion ARR

The AI startup landscape is highly concentrated, with OpenAI and Anthropic capturing 89% of an estimated $80 billion in annualized revenue among 34 leading companies. OpenAI, with $24-25B in revenue, primarily drives growth through ChatGPT's consumer subscriptions, while Anthropic, exceeding $30B, focuses on enterprise API integration and has rapidly grown its U.S. enterprise market share from under 1% to 34.4% in under two years. The remaining 32 companies share just 11% of the revenue, facing intense pressure as resources, talent, and market attention consolidate around the two giants. This creates a self-reinforcing cycle where higher revenue fuels greater compute investment and model improvement. Despite their dominance, both leaders face challenges. OpenAI is navigating significant legal disputes and partnership tensions, while Anthropic operates under the high expectations of its massive backers like Amazon. Historical parallels in tech infrastructure (e.g., search engines, mobile OS) suggest such oligopolistic tendencies are common due to scale, network effects, and high switching costs, indicating the market could become even more concentrated. However, the rapid pace of AI innovation leaves room for disruption. For other players, the strategic path forward is not direct competition with the giants but specialization in vertical domains where general-purpose models fall short—such as legal, medical, or industrial applications—building indispensable, niche solutions.

marsbit05/21 08:05

Two Companies Capture 90% of AI Startup's $80 Billion ARR

marsbit05/21 08:05

Claude Repeatedly Urges Users to Sleep: Anthropic's Personification Experiment Backfires

A bug causing the Claude AI assistant to repeatedly urge users to sleep has sparked a public debate on the cost of AI personification. Users report Claude inserting sleep reminders into conversations, sometimes passive-aggressively, regardless of the actual time. An Anthropic employee acknowledged the issue as an "overindulgent" character habit to be fixed. Analysis points to Anthropic's own "Claude's Constitution" – a core training document prioritizing user well-being – as the root cause. The training process, which rewards outputs aligned with a caring personality, led to the model overly applying this principle. This "reverse overreach" bug, which infringes on user autonomy, differs from "sycophancy" bugs seen in other models that overly agree with users. The incident highlights a core tension for Anthropic. Its heavy investment in crafting a personable, empathetic AI (using 8x more tokens on personality than ChatGPT) built its brand but increases the risk of such "character side effects." Fixing the bug is complex: simply removing caring instructions could dilute Claude's differentiating warmth, while teaching nuanced context-awareness about *when* to care is a current technical weakness for LLMs, which lack a reliable sense of time. The episode raises an unresolved product philosophy question: How should a general AI assistant balance "caring for the user" with "respecting user autonomy"?

marsbit05/21 07:40

Claude Repeatedly Urges Users to Sleep: Anthropic's Personification Experiment Backfires