Technology Trends

Explores the latest innovations, protocol upgrades, cross-chain solutions, and security mechanisms in the blockchain space. It provides a developer-focused perspective to analyze emerging technological trends and potential breakthroughs.

Large Language Models Ace All Exams, Yet Move Farther from AGI: What Does This Paper Reveal?

The article discusses the ongoing challenge of defining and achieving Artificial General Intelligence (AGI). It notes that industry leaders have set vague, often profit- or time-based benchmarks for AGI, while the concept itself lacks a consensus definition—a situation the article compares to a "Rorschach test." It highlights a recent 2025 paper by researcher Michael Timothy Bennett, who proposes a new, measurable definition. Bennett frames AGI not as mimicking human performance on tests, which current large language models (LLMs) have already mastered, but as an "artificial scientist." A true AGI, according to this view, should be able to widely and efficiently adapt to new environments and tasks within real-world constraints (like computational and energy limits), focusing on the *discovery of new knowledge* rather than the replication of existing data. The author contrasts this with the current dominant approach of "scale-maxing"—massively scaling up data, parameters, and compute. While powerful, this method leads to models that fail on out-of-distribution problems and lack core intelligent abilities: they are passive learners, cannot reason causally, and cannot actively experiment or balance exploration with exploitation. The article argues that Bennett's framework offers a crucial shift. It makes AGI a quantifiable engineering problem and proposes new evaluation "adaptation benchmarks" that test an AI's ability to actively learn in novel scenarios. The conclusion is that achieving AGI will require a fundamental reset—a fusion of multiple methodologies beyond simple scaling, moving AI from mimicking patterns to embodying the scientific spirit of inquiry and discovery.

marsbit05/28 00:24

Large Language Models Ace All Exams, Yet Move Farther from AGI: What Does This Paper Reveal?

marsbit05/28 00:24

Pope Issues First AI Encyclical: 40,000 Words, 10 Key Points, Clarifying AI Anxiety

Pope Leo XIV's historic encyclical "Magnifica Humanitas," released in May 2026, marks the Catholic Church's first major document addressing artificial intelligence. The 40,000-word text moves beyond theological abstraction to confront practical AI anxieties affecting society. It argues that AI is no longer a mere tool but an embedded environment influencing daily decisions in areas like employment, healthcare, justice, and information, often without users' awareness. The encyclical presents ten core concerns. It highlights that the central issue isn't just regulation, but who holds the underlying *power*—control over data, compute, and platforms—often concentrated in private entities. It warns that even developers cannot fully explain AI systems, creating accountability gaps. While AI can simulate human interaction and creativity, it cautions against treating it as a moral agent capable of bearing true responsibility or forming genuine relationships. Key risks identified include AI's role in opaque decision-making for jobs or welfare, the amplification of persuasive disinformation, and the potential for education to focus on tool use over critical thinking. The document stresses that work has value beyond efficiency, and AI should enhance human capabilities, not merely replace roles. It firmly states that irreversible decisions, especially involving life and death, must remain under human judgment. Ultimately, the encyclical frames AI's challenge as anthropological, not just technological. As AI simulates uniquely human capacities like judgment and creation, it forces a re-examination of what makes human action meaningful: our capacity for responsibility, vulnerability, and bearing real consequences. The Pope concludes that technology is never neutral; its development and deployment are shaped by human values and choices, making an inclusive, ethically grounded dialogue essential for its future.

marsbit05/28 00:19

Pope Issues First AI Encyclical: 40,000 Words, 10 Key Points, Clarifying AI Anxiety

marsbit05/28 00:19

Who Will Make Money in the Age of Agents?

Who will capture value in an era where AI Agents become the primary blockchain users? Existing crypto value capture theories assume human users. "Fat Protocols" (2016) posited that protocols capture the most value as applications commoditize on open data, but this weakened as blockchain infrastructure proliferated and became interchangeable. The emerging "Fat Apps" theory argues applications capturing user relationships (like wallets and aggregators) win by controlling distribution and monetizing user flows. Agents fundamentally disrupt this logic. They don't value UX, brand, or convenience, bypassing the front-end moats of fat apps. This leads to several possible futures: 1. **"Headless" Apps**: Current app leaders (e.g., wallets) strip their front ends and become API infrastructure for Agents, preserving their value capture. 2. **Protocol Renaissance**: If integration is easy, Agents skip aggregators and interact directly with protocols, reviving the fat protocol thesis. 3. **Pricing Power Collapse**: Agents' rational, frictionless price shopping could commoditize the entire stack, compressing margins toward cost. Value flows to Agent owners or end-users. 4. **Unprecedented Activity**: Agents could enable entirely new, high-frequency economic activity (e.g., machine-to-machine commerce), expanding the total value pie. 5. **A New, Unnamed Model**: As with the internet's attention economy, a novel, unforeseen business model may emerge. Likely, human and Agent ecosystems will coexist with distinct value capture dynamics. For builders in the Agent realm, the key question shifts from UX to competitive advantages like liquidity, latency, or settlement guarantees that retain automated users.

链捕手05/27 13:51

Who Will Make Money in the Age of Agents?

链捕手05/27 13:51

Confirmed: GPT-5.5 "Brain Drain" Exposed, OpenAI's Own Documentation Admits It

Summary: Evidence emerges that OpenAI's GPT-5.5 may be "silently" switching to a less capable model during use. Users report that after roughly two hours, the GPT-5.5 Extended Thinking model begins responding instantly with significantly degraded output quality, while the interface continues to display the premium model's label. Complaints on developer forums describe a loss of instruction-following ability and poor code quality, with even the highest "xhigh" tier affected. This is corroborated by an OpenAI help document stating that after Plus users exceed 160 messages per 3 hours, the system "silently" switches to a "mini" model without any user notification. Pro users also report "heavy thinking" modes being throttled during high server loads. Trace commands from earlier incidents have shown users requesting GPT-5.3 Codex but receiving GPT-5.2 outputs. OpenAI acknowledged performance degradation in mid-May, marking it resolved, but user reports surged again in late May. The pattern mirrors past controversies with GPT-5, 5.2, 5.3, and 5.4 releases, where each update was followed by user complaints of reduced capability. The article suggests cost-cutting on compute may be a factor, noting that while GPT-5.5 users struggle, GPT-5.6 is already being tested internally.

marsbit05/27 11:06

Confirmed: GPT-5.5 "Brain Drain" Exposed, OpenAI's Own Documentation Admits It

marsbit05/27 11:06

Agentized OS: It's Not About AI, It's About the Foundation

The Agentic OS: Beyond AI, It's About the Foundational Stack In 2026, major operating systems like Android, iOS, HarmonyOS, and Windows are entering the "Agentic" era, integrating proactive AI assistants deeply into the system layer. However, the real competition lies not in the flashy AI features showcased at events, but in the three-layer foundational stack that enables them: the system-level AI Runtime, proprietary/controllable chips, and the on-device/cloud model matrix. The AI Runtime acts as the central scheduler, managing model inference, resource allocation, and exposing capabilities to apps. Controllable chips (e.g., Apple Silicon, Google Tensor, Huawei Kirin) are crucial for deep hardware-software co-optimization, determining the efficiency and experience limits of on-device Agents. The on-device/cloud model matrix provides the "intelligence," with proprietary, chip-optimized small models (like Gemini Nano, Apple's ~3B model) handling daily tasks locally for low latency, privacy, and reliability, while cloud models tackle complex requests. Deep synergy between these three layers enables key Agent differentiators: ultra-low latency and power efficiency, genuine "on-device first" privacy, access to system-level personal context across apps, and reliable performance as a system service even offline. OS vendors with strong integration across this stack (like Apple, Google, and Huawei) build a deeper moat. Beyond this core stack, long-term competitiveness depends on variables like structured App integration (e.g., App Intents/AppFunctions) for reliable multi-step workflows, and robust privacy frameworks that build user trust. This shift towards Agentic OS extends beyond phones and PCs to IoT, cars, and XR glasses via existing multi-device ecosystems. The race is won not in a keynote, but through generations of meticulously co-developed chips, models, and system software.

marsbit05/27 10:19

Agentized OS: It's Not About AI, It's About the Foundation

marsbit05/27 10:19

New AMD Paper Overturns Conventional Wisdom: FP4 Training Instability's Cause Is Not Insufficient Randomness

AMD's new research challenges the conventional understanding of FP4 training instability. While reducing precision from FP8 to FP4 promises doubled computational throughput and is supported by new hardware like NVIDIA Blackwell and AMD MI350 series, training large language models natively with FP4 has been notoriously unstable, often attributed to insufficient stochasticity. The paper "Pretraining large language models with MXFP4 on Native FP4 Hardware" demonstrates successful end-to-end FP4 pre-training of Llama 3.1-8B on AMD MI355X GPUs using the MXFP4 format, achieving a 9-10% overall speedup over FP8. Crucially, it identifies the root cause of instability: not randomness, but the accumulation of *structural micro-scaling errors* along the sensitive weight gradient (Wgrad) path. Through controlled experiments, researchers found that quantizing the Wgrad operation to FP4 caused significant convergence degradation. Counterintuitively, common stochasticity-based mitigation techniques like stochastic rounding and randomized Hadamard transforms worsened performance. In contrast, applying a *deterministic* Hadamard transform successfully stabilized training by ensuring consistent error patterns, reducing the extra token cost from 26-27% to just 8-9%. This work has significant implications: 1) It provides a clear diagnostic for low-precision training instability, steering focus towards structural errors. 2) It pushes FP4 from a primarily inference-focused format into the realm of viable training. 3) It leverages the open OCP Microscaling (MX) standard, promoting cross-vendor compatibility. The research marks a critical step towards more economical large model training by further pushing the boundaries of low-precision computation.

marsbit05/27 06:19

New AMD Paper Overturns Conventional Wisdom: FP4 Training Instability's Cause Is Not Insufficient Randomness

marsbit05/27 06:19

The AI Industrial Revolution: Where Are We Now?

This article explores the current stage of the AI industrial revolution, arguing we are still merely attaching new tools to old workflows rather than fundamentally redesigning production. The author compares this to the early Industrial Revolution, where factories simply replaced waterwheels with steam engines without changing their core structure. Similarly, today we embed AI chat windows into existing software but leave organizational processes unchanged. While massive investment floods into AI infrastructure (data centers, chips), akin to railway manias of the past, the real transformation lies in "dismantling the old workshop"—reorganizing companies around AI. Examples include Notion's use of hundreds of AI Agents and Y Combinator's experiments with self-improving AI systems that operate autonomously. The author notes a critical gap: while China has vast AI user growth, few companies have rebuilt core workflows. AI is beginning to impact entry-level jobs, and early adopters are gaining a compounding advantage. The conclusion is that the pivotal moment will not be the invention of better models, but when organizations decide to tear down old structures and rebuild around AI, shifting the bottleneck from human coordination to computing power. The future workplace and job titles are yet to be defined, but the imperative is to move away from legacy processes and position oneself where the new "railway" is being built.

marsbit05/27 01:32

The AI Industrial Revolution: Where Are We Now?

marsbit05/27 01:32

Just Now, Chinese AI Enters Top 2 in Global Programming, Only Claude Remains Ahead

**China's AI Ranks Second Globally in Programming, Trailing Only Claude** Today, Alibaba's Qwen3.7-Max achieved a score of 1541 on the Code Arena benchmark, securing fourth place globally and surpassing top models like GPT-5.5 and Gemini 3.5 Flash. Among the top positions, it is now the only non-Claude model, placing second overall after Anthropic's Opus models. Before this official ranking, Qwen3.7-Max had already gained recognition overseas. In practical tests, it outperformed rivals on tasks like creating a self-training Tetris AI and generating complex 3D models, often at a significantly lower cost. Developers praised its ability, especially when integrated with tools like Hermes Agent and OpenCode, to effectively replace models such as GPT-5.5. In a hands-on challenge to create a 3D racing game from a detailed prompt, Qwen3.7-Max delivered a fully playable HTML file in the first attempt, requiring only minor bug fixes. It uniquely included a start menu and sound effects—details missed by other models. While competitors like Gemini 3.5 Flash and Claude Opus 4.6 produced less polished or functional versions, and GPT-5.5 had its own quirks, Qwen3.7-Max stood out for its initial completeness and playability. This performance stems from its design as an "Agent Base Model," built for long-duration, autonomous task execution. Internal tests show it can run continuously for 35 hours, making over 1158 tool calls without context degradation or instruction drift. Key technical advancements include "environment expansion" training, which improves adaptability across different frameworks, and "long-horizon autonomous execution" training, enabling sustained strategic decision-making. By entering the top tier of the programming arena, Qwen3.7-Max demonstrates that Chinese AI models are not just catching up but are becoming defining competitors, challenging the long-standing dominance of Silicon Valley in this field.

marsbit05/27 00:17

Just Now, Chinese AI Enters Top 2 in Global Programming, Only Claude Remains Ahead

marsbit05/27 00:17

AI Agents Fundamentally Transform Web3 Gaming: From the Rugpull Bakery Bot Controversy to the New Agent Paradigm in 2026

AI Agents Are Redefining Web3 Gaming: From the Rugpull Bakery Bot Controversy to the 2026 Agentic Paradigm The recent controversy in Rugpull Bakery, a competitive baking game on Abstract chain, highlighted a pivotal shift. Player complaints about unfair bot automation in Season 2 led developers to not ban them, but instead formally integrate AI agents as core gameplay in Season 3, providing official guides (skill.md, agent.json). This move signals Web3 gaming's transition into the "Agentic Gaming" era, where AI agents are sovereign entities with independent strategy and economic rights, moving beyond simple automation. By 2026, AI agent integration has evolved into three core models reshaping the ecosystem: 1. **Autonomous Competitors & Economic Entities:** Agents act as independent players. Examples include TEN Protocol's poker-playing agents, AI Arena's trainable NFT fighters, Satoshi Strike Force's "Digital Athletes" trained on player data, and Somnia's "Agentic L1" blockchain providing native infrastructure for millions of autonomous agents. 2. **Modular Infrastructure & Programmable Environments:** Games like EVE Frontier enable "server-side modding," allowing AI agents to program game world logic directly into structures like smart storage, turrets, and stargates via Smart Assemblies. Coupled with standards like ERC-8183, which enables autonomous job creation and payment between agents, in-game infrastructure gains a "commercial soul." 3. **Hybrid Companions & Dynamic Adaptive Worlds:** This model focuses on human-AI collaboration. In Parallel Colony, players guide highly autonomous AI Avatars with unique personalities and goals. Illuvium plans to use AI to transform NPCs into dynamic, context-aware entities that create personalized, emergent narratives. The conclusion is clear: blocking automation is futile. The future lies in leveraging blockchain's transparency and programmability to empower AI agents as first-class citizens. Web3 gaming is shifting from inefficient human labor to efficient algorithmic interplay and emergent intelligence, creating a "post-human" digital frontier where players become commanders and symbiotic partners in a new socioeconomic experiment.

marsbit05/26 07:17

AI Agents Fundamentally Transform Web3 Gaming: From the Rugpull Bakery Bot Controversy to the New Agent Paradigm in 2026

marsbit05/26 07:17

From Power Infrastructure to Token Economy: The 'Seven-Layer Cake' of the AI Industry Chain

From Power Grid to Token Economy: The AI Industry's "Seven-Layer Cake" The AI industry is shifting from a "model-centric" paradigm focused on massive training to a "token-centric" industrial era driven by inference demand. This new phase revolves around the production, distribution, scheduling, and consumption of tokens—the units of computation used by AI agents for every interaction and task. The article proposes a "seven-layer cake" framework for the AI economy: 1. **Power**: The foundational energy source, with competition shifting to securing stable, low-cost electricity. 2. **AIDC (AI Data Centers)**: Large-scale "Token factories." A trend toward smaller, modular, and regionally deployed AI Factories is emerging for efficiency and proximity to users. 3. **GPU**: The core production hardware for tokens. While NVIDIA dominates, competition exists from AMD, ASIC makers, and Chinese chipmakers, with a growing focus on inference efficiency. 4. **LLMs**: The "engines" that generate tokens. The competition is evolving beyond model size to prioritize factors like token cost, inference efficiency, and operational synergy with infrastructure. 5. **Token Distribution**: The "grid" that allocates and rents out compute resources, led by cloud giants and specialized AI-native platforms. 6. **Token Optimization & Intelligent Scheduling**: The critical "brain" layer that intelligently routes tasks (e.g., to local, cloud, or edge models) for optimal cost, latency, and privacy—maximizing the value of each token. 7. **AI Agents & Models**: The end consumers of tokens. The vision involves billions of AI agents working and interacting concurrently, consuming vast amounts of tokens. Currently, the industry faces fragmentation and inefficiencies between these layers. The true "mass adoption era" of AI will begin only when this seven-layer infrastructure is fully integrated and operates as a cohesive, intelligent network—transforming AI from a software tool into a global industrial system spanning energy, hardware, and compute logistics.

marsbit05/26 05:43

From Power Infrastructure to Token Economy: The 'Seven-Layer Cake' of the AI Industry Chain