# Engineering Articoli collegati

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

Claude Code Slashes 80% of Prompt Tokens, But Opus 5 Just Adds Them Right Back In

Claude Code, the AI coding assistant from Anthropic, recently announced a massive reduction of over 80% in its system prompt content for models like Opus 5 and Fable 5. The goal was to remove verbose, often conflicting, rules (like strict commenting and documentation requirements) and replace them with a simpler directive: write code that matches the style of the surrounding project. This "pruning" aims to make the model more efficient by reducing internal conflict from overlapping instructions, with no measurable performance drop reported. However, a developer's (@chenchengpro) investigation revealed a twist. While the prompt was drastically cut from 15,225 characters in Opus 4.7 to 4,467 in Opus 4.8, it *increased* by approximately 72% to 7,694 characters in Opus 5. This isn't a contradiction. The "over 80% cut" refers to the overall shift from the old, detailed rulebook-style prompts to a new, streamlined system. The 72% increase for Opus 5 represents new, targeted instructions added to manage the model's enhanced capabilities. Opus 5 is more proactive—it likes to report progress, generate longer outputs, use sub-agents, and expand task scope. The added prompt content (roughly 3,755 characters) primarily provides guidelines for "Delivering work" (controlling task scope, progress reporting) and "Corrections" (limiting excessive self-correction). These new rules are necessary to curb potential over-engineering on simple tasks, ensuring efficiency even as the model becomes more independent. In short, the old, restrictive manual was deleted, but new guidelines were written to harness the model's newfound initiative.

marsbit2 giorni fa 11:37

Claude Code Slashes 80% of Prompt Tokens, But Opus 5 Just Adds Them Right Back In

marsbit2 giorni fa 11:37

In Just 11 Days, Claude Rewrote Millions of Lines of Code, an Epic AI Engineering Feat Sparks Fury

In just 11 days, Bun's founder Jarred Sumner used Anthropic's Claude AI models to rewrite its million lines of code from Zig to Rust. This move sparked significant controversy, particularly from Zig's creator, Andrew Kelley, who publicly criticized Sumner's engineering practices and the decision to use AI for such a massive rewrite. Bun, a high-performance JavaScript/TypeScript runtime and rival to Node.js, was originally written in Zig. After Anthropic acquired Bun, the team encountered persistent stability and memory safety bugs in the Zig codebase. These issues, combined with Zig's strict policy against LLM-generated code, led to the decision to rewrite in Rust. The rewrite was executed using Claude AI tools at an estimated API cost of $165,000, dramatically reducing the expected time and financial cost. Andrew Kelley's response was scathing. He blamed the original bugs on poor engineering habits, calling Bun's Zig code a collection of "hacks on top of hacks." He expressed relief that Bun was no longer associated with Zig, fearing it would misrepresent the language and attract low-quality, AI-generated contributions. The tech community is divided; some view Kelley's critique as unprofessional, while others see it as a defense of engineering integrity. A major concern about the AI-driven rewrite is the resulting code quality. The translation from Zig left approximately 27,000 lines of unsafe Rust code, raising fears about long-term maintainability and technical debt. The debate centers on whether this project is a milestone in AI-assisted development or a future maintenance nightmare.

marsbit07/11 08:13

In Just 11 Days, Claude Rewrote Millions of Lines of Code, an Epic AI Engineering Feat Sparks Fury

marsbit07/11 08:13

Lao Huang: Prompt is Dead, the Entire AI Community is Frenziedly Chasing Loops

The article "Prompt is Dead: The AI Industry is Obsessively Chasing Loops" discusses a major shift in AI development, where "Loop Engineering" is replacing traditional prompt engineering. Industry leaders like NVIDIA's Jensen Huang, Andrew Ng, and engineers from Anthropic and OpenAI argue that manually crafting prompts is becoming obsolete. Instead, the new focus is on designing autonomous, self-improving AI systems (loops) that can operate 24/7. A loop system typically involves five key phases: Discovery (finding tasks), Handoff (assigning to agents), Validation (critical independent review), Persistence (saving progress), and Scheduling (automated operation). The core idea is to move humans from being the operational "engine" to being the system "architects" who design the loop, define goals, and set up verification mechanisms. A major challenge and necessity is implementing robust, independent validation to prevent AI from uncritically approving its own work. The trend is seen as part of a move towards "inference-time compute," where allocating computational budget effectively becomes a key engineering skill. While loops can produce higher-quality outputs, they are more expensive and time-consuming than simple prompting. The article warns of risks like "verification debt," "comprehension corrosion," and "cognitive surrender," where engineers might stop understanding the code their systems generate. Ultimately, the article concludes that in an era of automated loops, human judgment and oversight remain the most critical and scarce resources.

marsbit06/29 08:37

Lao Huang: Prompt is Dead, the Entire AI Community is Frenziedly Chasing Loops

marsbit06/29 08:37

Jensen Huang: Prompts are Becoming Obsolete, Loops are the New Paradigm

Jensen Huang, alongside AI leaders like Peter Norvig, Boris Cherny, and Andrew Ng, is advocating for a shift from "prompt engineering" to "loop engineering" as the new paradigm for AI development. Instead of manually crafting individual prompts, the focus is now on designing autonomous loops—systems where AI agents execute tasks, self-validate results, and iterate until completion without constant human oversight. A loop is a management framework that enables agents to operate independently. Key implementations are seen in Claude Code (with features like /loop, /goal, and /schedule) and OpenAI Codex, which employ multiple agents working in parallel within isolated environments. A core principle is the separation of roles: one agent (or model) performs the task, while an independent agent (or a smaller, separate model) validates the output to ensure objectivity. The article outlines a practical roadmap for implementing loops, starting with a "four-condition test" to assess suitability, building a minimal viable loop, and emphasizing critical pitfalls to avoid, such as lacking hard stop conditions or allowing loops to handle tasks requiring human judgment. This evolution is framed as the fourth major shift in AI interaction: from Prompt Engineering (crafting instructions) to Context Engineering (providing background information), then to Harness Engineering (building tool-enabled environments), and finally to Loop Engineering (creating self-sustaining systems). This progression reflects a consistent trend of increasing abstraction, moving human involvement from direct instruction to system design and rule-setting. The concept has academic roots in frameworks like ReAct, which formalized the "reason-act-observe" cycle. While loop engineering promises greater automation, experts caution about managing token costs and warn against outsourcing understanding—AI can assist, but deep problem comprehension remains essential.

marsbit06/25 14:26

Jensen Huang: Prompts are Becoming Obsolete, Loops are the New Paradigm

marsbit06/25 14:26

Japan's AI Dark Horse Emerges: How a 7B Small Model Challenges Fable and Mythos?

In June 2026, Sakana AI's new model Fugu caused a stir in the AI community. Its Fugu Ultra variant achieved scores of 73.7 on SWE-Bench Pro and 82.1 on TerminalBench 2.1, surpassing GPT-5.5 and Claude Opus 4.8, and was claimed to be comparable to export-restricted models like Fable 5 and Mythos Preview. Remarkably, the core of this high-performance system is not a massive model, but a small 7B-parameter RL Conductor model. Fugu operates as a multi-agent orchestrator: the 7B model acts as a "foreman," dynamically analyzing user tasks and delegating subtasks to a pool of top-tier global models (e.g., GPT-5, Gemini 3.1 Pro). It then synthesizes and verifies their outputs. This architecture represents a paradigm shift from monolithic models to an expert-team approach. It enhances performance in complex, multi-step engineering tasks like code review and security testing by enabling cross-validation from specialized models, improving long-session stability and token efficiency. However, Fugu's strengths come with trade-offs: it faces inherent latency due to multiple API calls, relies heavily on underlying US model APIs (creating dependency risks), and its benchmark comparisons with Fable/Mythos are based on reported scores, not head-to-head testing. For Japan's AI ecosystem, which lacks the massive compute and data resources of the US or China, Fugu exemplifies an "asymmetric breakthrough" strategy. Instead of competing directly in parameter scale, it focuses on intelligent orchestration of existing global models, offering a degree of AI sovereignty and resilience. While a significant system-level innovation, its ultimate capability is still bounded by the underlying models it coordinates.

marsbit06/22 11:17

Japan's AI Dark Horse Emerges: How a 7B Small Model Challenges Fable and Mythos?

marsbit06/22 11:17

Beyond the Model Lies the Harness: Deepseek Enters the Arena, Why Has the Main Battlefield of China's AI Competition Shifted?

In mid-to-late May 2026, Deepseek internally established a new Harness team focused on code agent products, internally benchmarked against Anthropic's Claude Code. This move, marked by the formula "Model + Harness = Agent" in their job postings, signals a major shift in China's AI competition: the main battlefield is transitioning from developing large models to building toolchains and achieving workplace integration. Deepseek's direct involvement in Harness development aims to secure control over interface design and training data feedback loops, moving beyond open-sourcing powerful models. Harness, the runtime infrastructure for AI agents, handles everything beyond model reasoning—task orchestration, tool calling, context management, safety checks, and error recovery. It is crucial because agent products are not just outputs of model capability but also training grounds for it. Real-world task failures recorded by Harness can feed back into model training, creating a flywheel effect. Engineering Harness is more critical than optimizing prompts, as poor context management or error handling can drastically reduce agent success rates in multi-step, real-world scenarios. This shift is not isolated. Other major Chinese tech companies are also pursuing differentiated toolchain strategies. Tencent leverages its enterprise ecosystem (WeChat Work, Tencent Cloud) to build connectors for organizational-level AI collaboration and complex task delivery. Alibaba focuses on lowering automation barriers on the web with a front-end, browser-based GUI Agent framework, PageAgent. This diversification shows the industry recognizes that success lies not in a perfect general agent, but in vertically focused solutions built with robust engineering. The trend is validated by overseas success, such as Poland's Viktor, an AI coworker on Slack achieving $20M ARR by autonomously executing complex, multi-step tasks. This proves a shift in enterprise willingness to pay—from "AI-assisted generation" to "AI-autonomous execution." As Harness matures to provide safety guards and reliability, AI transitions from a human-supervised intern to an independent outsourcer. The competition now faces key engineering challenges: preventing "token explosion" through intelligent context compression, and building "thick frameworks" with features like sandbox isolation and checkpoint recovery for enterprise-grade stability. Geopolitical restrictions on tools like Claude Code further create a significant market vacuum for domestic solutions like Deepseek's Harness. For enterprises and developers, the focus must shift from comparing model benchmarks to evaluating a vendor's engineering capabilities, error recovery mechanisms, context management, and ecosystem compatibility when choosing AI products and platforms.

marsbit06/22 06:05

Beyond the Model Lies the Harness: Deepseek Enters the Arena, Why Has the Main Battlefield of China's AI Competition Shifted?

marsbit06/22 06:05

What Kind of People Is Anthropic Actually Hiring? Insights from 1,680 Resumes Provide the Answer

Contrary to the common perception of Anthropic as a PhD-heavy AI research lab, an analysis of 1,680 engineer profiles reveals its core talent is experienced "builders." The engineering team, largely assembled in the past 18 months (median tenure 10 months), hires senior engineers with a median of 12.2 years of prior experience. They primarily come from major engineering-focused companies like Google, Meta, Amazon, Microsoft, Stripe, Databricks, Snowflake, and Palantir, not primarily from other AI labs. Their backgrounds are infrastructure-heavy (40%), focusing on backend, distributed systems, databases, and security, with skills like Python, Java, C++, and AWS being common. Only 13.7% hold PhDs. The dominant job title is "Member of Technical Staff." For early-career hires (<6 years experience), entry is not based on standard work history but on prestigious signals: top internships (e.g., at FAANG or quant trading firms), competition rankings, research publications, or AI safety fellowship experience. The key insight for applicants is to present as infrastructure builders, highlighting experience with large-scale systems, not as pure researchers, reflecting Anthropic's identity as a highly engineered infrastructure company.

marsbit06/15 03:28

What Kind of People Is Anthropic Actually Hiring? Insights from 1,680 Resumes Provide the Answer

marsbit06/15 03:28

Is Writing Prompts Outdated? AI Programming is Shifting to Loop Engineering

"The era of manually writing prompts for AI coding agents is shifting towards 'Loop Engineering,' where developers design automated, self-managing workflows instead of step-by-step guidance. A loop consists of five core components: Automations (scheduled task discovery and triage), Worktrees (isolated environments to prevent file conflicts), Skills (project-specific knowledge and conventions), Plugins/Connectors (integration with tools like GitHub, Linear, or Slack), and Sub-agents (separate agents for execution and verification). An external memory layer, such as Markdown files or Linear boards, tracks progress across sessions. Tools like Claude Code and Codex now embed these components, enabling loops to run autonomously—discovering tasks, assigning work, checking results, and deciding next steps. This approach amplifies developer leverage but doesn’t replace human judgment. Key risks include 'comprehension debt' (losing understanding of the codebase) and 'cognitive surrender' (over-relying on automation). Effective loop design requires balancing automation with ongoing verification and deep system understanding, making it a more complex but higher-leverage skill than prompt engineering."

marsbit06/10 17:55

Is Writing Prompts Outdated? AI Programming is Shifting to Loop Engineering

marsbit06/10 17:55

Xiaomi MiMo's 99% Price Cut is Not Marketing! Luo Fuli Posts on X to Refute Critics

The price of Xiaomi's MiMo-V2.5 series API has been permanently reduced by up to 99%, specifically for the "Input (Cache Hit)" cost, which covers users re-reading historical context in long conversations. MiMo's head, Luo Fuli, published a detailed technical blog to clarify that this drastic price cut stems from genuine engineering breakthroughs, not a marketing stunt or a simple price war. The core of the achievement lies in six key engineering optimizations. First, the model architecture adopts a Hybrid Sliding Window Attention (SWA), reducing the memory footprint (KVCache) to 1/7th of a traditional model. Second, a dual-pool memory management system actually utilizes these savings, allowing a single GPU to handle over 5 times more concurrent users. Third, an upgraded prefix caching mechanism achieves a cache hit rate of 93-95% for repeated reads, meaning most such requests bypass GPU computation entirely. Fourth, a self-developed distributed cache (GCache) utilizes idle SSD space on existing GPU servers, eliminating additional storage costs. Fifth, an intelligent scheduling system (LLM-Router) efficiently routes requests to maximize cache reuse and performance. Sixth, Multi-Token Prediction (MTP) accelerates the model's text generation ("output") side. Together, these systemic optimizations dramatically lower the real computational cost per request, enabling the 99% price reduction for cached inputs while reportedly maintaining positive gross margins. Luo Fuli's disclosure aims to shift the narrative from "price war" to a demonstration of substantive AI engineering progress.

marsbit05/31 10:37

Xiaomi MiMo's 99% Price Cut is Not Marketing! Luo Fuli Posts on X to Refute Critics

marsbit05/31 10:37

Auto Research Era: 47 Tasks Without Standard Answers Become the Must-Test Leaderboard for Agent Capabilities

The article introduces Frontier-Eng Bench, a new benchmark for AI agents developed by Einsia AI's Navers lab. Unlike traditional tests with clear answers, this benchmark presents 47 complex, real-world engineering tasks—such as optimizing underwater robot stability, battery fast-charging protocols, or quantum circuit noise control—where there is no single correct solution, only continuous optimization towards a limit. It shifts AI evaluation from static knowledge retrieval to a dynamic "engineering closed-loop": the AI must propose solutions, run simulations, interpret errors, adjust parameters, and re-run experiments to iteratively improve performance. This process tests an agent's ability to learn and evolve through long-term feedback, much like a human engineer tackling trade-offs between power, safety, and performance. Key findings from the benchmark reveal two patterns: 1) Improvements follow a power-law decay, becoming harder and smaller as optimization progresses, and 2) While exploring multiple solution paths (breadth) helps, sustained depth in a single path is crucial for breakthrough innovations. The research suggests this marks a step toward "Auto Research," where AI systems can autonomously conduct continuous, tireless optimization in scientific and engineering domains. Humans would set high-level goals, while AI agents handle the iterative experimentation and refinement. This could fundamentally change research and development workflows.

marsbit05/13 07:06

Auto Research Era: 47 Tasks Without Standard Answers Become the Must-Test Leaderboard for Agent Capabilities

marsbit05/13 07:06

活动图片

Technology Trendsright-arrow

Project Updatesright-arrow