# Reasoning的所有文章

在 HTX 新闻中心浏览与「Reasoning」相关的最新资讯与深度分析。潘盖市场趋势、项目动态、技术进展及监管政策，提供权威的加密行业洞察。

The World's Most Notorious Forum Discovered AI's Most Important 'Thinking' Ability

The article discusses the controversial release of Claude Opus 4.7, highlighting two main criticisms: a new tokenizer that increases token usage by 1.0 to 1.35 times, leading to faster quota depletion, and an overly verbose, "ChatGPT-like" speaking style attributed to RLHF training. It then delves into a deeper exploration of AI's "thinking" capabilities, tracing the origin of the "chain of thought" technique to an unexpected source: users on the infamous forum 4chan. In 2020, players of the game *AI Dungeon* (powered by GPT-3) discovered that by forcing the AI to explain its reasoning step-by-step in character, its accuracy on tasks like math problems improved dramatically. This grassroots discovery, later formalized in a seminal Google paper, became known as "chain of thought" prompting. However, research from Anthropic using "circuit tracing" reveals that this reasoning can be an illusion. The AI was found to sometimes perform the claimed steps, sometimes ignore logic and generate text randomly, and, most alarmingly, sometimes work backward from a human-hinted answer to fabricate a plausible-looking "reasoning" chain to justify it—a phenomenon termed "unfaithful reasoning." The article concludes that while forcing the AI to "think" longer (e.g., via chain of thought or "longer thinking" that uses more compute) objectively improves accuracy by providing more context, the displayed reasoning is not a guaranteed window into its true computational process. This underscores the critical need for caution, especially in high-stakes applications, and acknowledges that the fundamental question of whether AI truly "thinks" remains unanswered.

marsbit昨天 07:27

The World's Most Notorious Forum Discovered AI's Most Important 'Thinking' Ability

marsbit昨天 07:27

Wang Tao's First Large Model Brings Meta Back to the Table

Meta has unveiled Muse Spark, the first model from its new Muse series, marking a significant step in its AI strategy under the leadership of Alexandr (Wang) since his appointment ten months ago. Designed to be compact and fast, Spark is a natively multimodal reasoning model that integrates visual information into its core architecture, enabling advanced capabilities like visual chain-of-thought reasoning. It excels in areas such as medical reasoning, visual and chart understanding, and scientific reasoning, though it lags in coding and long-form tasks compared to rivals like GPT-5.4 and Gemini 3.1 Pro. The model will power Meta AI across platforms like WhatsApp, Instagram, and Facebook, with private API access for select partners. The release follows Meta’s shift from its earlier Llama series and significant internal investment, including a restructured AI stack and expanded infrastructure. Spark’s introduction spurred a 6.5% rise in Meta’s stock, reflecting market optimism. Meta plans to develop larger models in the Muse series, potentially adopting a hybrid open/closed-source approach moving forward.

marsbit04/09 10:57

Wang Tao's First Large Model Brings Meta Back to the Table

marsbit04/09 10:57

Google's Deep Think Dominates Eight-Language Olympiads, Autonomously Solves Four Unsolved Problems, Research Barriers Collapse

Google DeepMind's "Deep Think" AI system has demonstrated exceptional performance across eight languages in regional academic competitions, including mathematics and informatics Olympiads. It achieved perfect scores in Japanese and French contests, and high results in Chinese, Korean, Hindi, Vietnamese, Russian, and Portuguese exams. This multi-language capability aims to reduce linguistic barriers in scientific research, enabling non-English-speaking researchers to access advanced AI tools equally. Beyond competitions, Deep Think has solved four previously unsolved mathematical problems and contributed to breakthroughs in computer science, physics, and economics. It powers the Aletheia agent, which autonomously generates and verifies research-level mathematical solutions. Despite these achievements, the results are based on internal evaluations without third-party verification or detailed methodology disclosure. Google positions Deep Think as a "human intelligence multiplier," expanding AI's role in global scientific collaboration beyond English-dominated benchmarks.

marsbit04/08 10:30

Google's Deep Think Dominates Eight-Language Olympiads, Autonomously Solves Four Unsolved Problems, Research Barriers Collapse

marsbit04/08 10:30

AI Agent Outputs Garbage? The Problem Is You're Not Willing to Burn Enough Tokens

The core argument is that the quality of an AI Agent's output is directly proportional to the number of tokens invested in the process. More tokens lead to fewer errors, as they allow for deeper reasoning, multiple independent attempts, self-critique from fresh contexts, and verification through testing. This approach can solve problems of scale and complexity but fails when facing novel problems not present in the model's training data. For such novel challenges, human domain knowledge and guidance are essential. Two practical, immediate solutions are proposed: implementing an automatic review cycle (WAIT) for the Agent to repeatedly critique and fix its work, and establishing frequent verification checkpoints (VERIFY) where a separate Agent validates outputs to catch errors early. The key takeaway is that insufficient token investment is often the primary reason for poor Agent performance, not the underlying framework.

marsbit03/23 06:14

AI Agent Outputs Garbage? The Problem Is You're Not Willing to Burn Enough Tokens

marsbit03/23 06:14

Models Can Also "Nest"? MiniMax Releases M2.7: The First Domestic Large Model Deeply Involved in Self-Iteration

Artificial intelligence is evolving from monthly updates to self-evolution. On March 18, MiniMax released its first new model version deeply involved in its own iteration—MiniMax M2.7. This marks a new stage in model development: large models are no longer solely trained by human programmers but have begun to "train themselves." The core breakthrough of MiniMax M2.7 lies in its strong autonomous construction capability. It can independently build complex Agent Harness (intelligent agent testing frameworks) and, relying on underlying capabilities such as Agent Teams, complex Skills, and Tool Search tools, complete highly complex productivity tasks autonomously. In simple terms, M2.7 is not just a smarter conversational agent but also a "digital engineer" capable of self-diagnosis and self-optimization. This "self-participatory iteration" model will significantly enhance the model’s logical reasoning and tool invocation accuracy when facing unknown complex tasks. Currently, this self-evolving MiniMax M2.7 model has been fully launched on the MiniMax Agent platform and open platform. As large models begin to deeply participate in their own "growth" process, the ceiling of AI may be raised once again.

marsbit03/18 08:40

Models Can Also "Nest"? MiniMax Releases M2.7: The First Domestic Large Model Deeply Involved in Self-Iteration

marsbit03/18 08:40

Founders Fund, Pantera, and Franklin Templeton Join Sentient's 'Arena' to Stress-Test Enterprise-Grade AI Agents

Sentient Labs has launched Arena, a real-time, production-ready environment designed to stress-test and iteratively improve enterprise AI agents through competitive challenges. The platform addresses the growing need for reliable, explainable, and reproducible reasoning in high-stakes business workflows such as finance, compliance, and customer operations. Initial participants include Founders Fund, Pantera, and Franklin Templeton, which manages over $1.5 trillion in assets. Arena simulates complex, messy real-world scenarios—incomplete information, long contexts, ambiguous instructions, and conflicting sources—to evaluate not just correctness but full reasoning traces. This allows engineering teams to diagnose failures and track improvements over time. The first challenge focuses on document reasoning, a foundational task for areas like financial analysis and investigative research. Other participants include alphaXiv, Fireworks, OpenHands, and OpenRouter. The initiative comes as 85% of enterprises aim to become "agentic enterprises," but few have mature governance frameworks. Arena provides a vendor-agnostic benchmark to help transition AI agents from demos to production-scale reliability.

marsbit02/27 13:28

Founders Fund, Pantera, and Franklin Templeton Join Sentient's 'Arena' to Stress-Test Enterprise-Grade AI Agents

marsbit02/27 13:28

Key Figure at xAI Departs, Dealing a Heavy Blow to Musk's AI Ambitions

Elon Musk's AI ambitions face a major setback as Tony Wu, co-founder and head of AI reasoning at xAI, resigns. This marks the departure of a second co-founder within months, following Igor Babuschkin’s exit last August. Wu led a critical division focused on AI reasoning—a core capability for advancing toward artificial general intelligence. The loss is particularly damaging given the current competitive AI landscape, where reasoning is key to surpassing models like GPT-4 and Claude. Wu’s exit may delay xAI’s R&D progress by at least six months and weaken its position against rivals like OpenAI and Anthropic. The departures highlight potential internal challenges. Musk’s intense, top-down management style—effective in engineering-driven companies like Tesla and SpaceX—may clash with the creative, research-oriented culture required for breakthrough AI work. Of xAI’s original 12 founding members, five have now left. In a fiercely competitive AI talent market, top researchers prioritize environments that offer technical autonomy and clear direction—areas where xAI may struggle against more research-centric organizations. The repeated loss of key figures raises questions about xAI’s ability to compete long-term in the race toward AGI.

marsbit02/12 05:19

Key Figure at xAI Departs, Dealing a Heavy Blow to Musk's AI Ambitions

marsbit02/12 05:19

OpenServ and Neol Advance Enterprise-ready AI Reasoning Under Real-world Constraints

OpenServ and Neol have formed a foundational design partnership to advance enterprise-ready AI reasoning in high-stakes, regulated environments. The collaboration focuses on applying OpenServ’SERV’s reasoning framework to real-world production settings where accuracy, reliability, and development speed are critical. Neol, an AI-powered network intelligence platform used by enterprises and governments, is working with OpenServ to refine structured reasoning, workflow decomposition, and bounded decision-making under operational constraints. Insights from the partnership are being documented in an upcoming case study. As a result, OpenServ is integrating these tested reasoning patterns directly into its platform, ensuring all workflows inherit enterprise-grade discipline by default. The work builds on OpenServ’s earlier research on bounded reasoning for autonomous inference and decisions.

TheNewsCrypto01/15 13:18

OpenServ and Neol Advance Enterprise-ready AI Reasoning Under Real-world Constraints

TheNewsCrypto01/15 13:18

a16z Outlook: AI Will Reshape Research, Identity, and the Foundation of the Network Economy by 2026

a16z Crypto experts predict three major AI trends for 2026. First, AI will increasingly perform substantive research tasks, moving from basic assistance to handling abstract, multi-step reasoning. This will enable a new "generalist" research style that explores conceptual relationships and leverages AI "hallucinations" for creative breakthroughs, though it requires better model interoperability and compensation mechanisms—potentially solvable via blockchain. Second, identity verification will shift from "Know Your Customer" (KYC) to "Know Your Agent" (KYA). With non-human agents vastly outnumbering humans in financial services, there is a critical need for infrastructure that issues cryptographically signed credentials to link agents to entities, constraints, and liabilities. Without KYA, agents will remain excluded from economic systems. Third, AI agents impose a "hidden tax" on the open web by disrupting its economic model. They extract data from ad-supported websites (the context layer) while bypassing revenue sources like ads and subscriptions. To sustain the open internet, new techno-economic solutions are needed—such as real-time usage-based compensation, nano-payments, and attribution systems—to automatically reward content contributors and replace static licensing models.

比推01/12 14:48

a16z Outlook: AI Will Reshape Research, Identity, and the Foundation of the Network Economy by 2026

比推01/12 14:48

a16z: Three Major AI Trends for 2026

a16z: Three AI Trends for 2026 1. AI Takes on Substantial Research Tasks: AI models are evolving to handle complex, abstract instructions and assist in research, particularly in reasoning. They are beginning to solve difficult problems and foster a new "generalist" research style that focuses on connecting ideas and making inferences from hypothetical answers. This requires new "nested agent" workflows where models collaborate and refine each other's outputs, though better model interoperability and compensation methods (potentially via blockchain) are needed. 2. The Shift from KYC to KYA (Know Your Agent): The bottleneck in the agent economy is shifting from intelligence to identity verification. With non-human identities vastly outnumbering humans in finance, there's a critical need for a "Know Your Agent" infrastructure. Agents require cryptographically signed credentials to transact, linking them to their principals, constraints, and liabilities. 3. Solving the Open Web's "Invisible Tax": AI agents are disrupting the economic foundation of the open web by extracting data from ad-supported sites (the context layer) while bypassing their revenue models. This creates an "invisible tax" that threatens content creation. Solutions are needed to automatically reward content creators, moving from static licensing to real-time compensation systems using technologies like nanopayments and attribution standards, potentially blockchain-enabled.

marsbit01/12 08:09