Core Viewpoint - The discussion around VLA (Vision-Language-Action model) is intensifying, with contrasting opinions on its short-term feasibility and potential impact on the automotive industry [2][12]. Group 1: VLA Technology and Development - The Li Auto i8 is the first vehicle to feature the VLA driver model, positioning it as a key selling point [2]. - Bosch's president for intelligent driving in China, Wu Yongqiao, expressed skepticism about the short-term implementation of VLA, citing challenges in multi-modal data acquisition and training [2][12]. - VLA is seen as an "intelligent enhanced version" of end-to-end systems, aiming for a more human-like driving experience [2][5]. Group 2: Comparison of Driving Technologies - There are two main types of end-to-end technology: modular end-to-end and one-stage end-to-end, with the latter being more advanced and efficient [3][4]. - The one-stage end-to-end model simplifies the process by directly mapping sensor data to control commands, reducing information loss between modules [3][4]. - VLA is expected to outperform traditional end-to-end models by integrating multi-modal capabilities and enhancing decision-making in complex scenarios [5][6]. Group 3: Challenges and Requirements for VLA - The successful implementation of VLA relies on breakthroughs in three key areas: cross-modal feature alignment, world model construction, and dynamic knowledge base integration [7][8]. - Current automotive chips are not designed for AI large models, leading to performance limitations in real-time decision-making [9][11]. - The industry is experiencing a "chip power battle," with companies like Tesla and Li Auto developing their own high-performance AI chips to meet VLA's requirements [11][12]. Group 4: Future Outlook and Timeline - Some industry experts believe 2025 could be a pivotal year for VLA technology, while others suggest it may take 3-5 years for widespread adoption [12][13]. - Initial applications of VLA are expected to be in controlled environments, with broader capabilities emerging as chip technology advances [14]. - Long-term projections indicate that advancements in AI chip technology and multi-modal alignment could lead to significant breakthroughs in VLA deployment by 2030 [14][15].

Core Viewpoint - The article discusses the potential of the VLA (Vision-Language-Action) architecture in autonomous driving, emphasizing its long-term viability and alignment with human cognitive processes [2][12]. Summary by Sections VLA Architecture and Technical Potential - VLA has strong technical potential, transitioning from manual to AI-driven autonomous driving, and is expected to support urban driving scenarios [2]. - The architecture is inspired by robotics and embodied intelligence, suggesting it will remain relevant even after the proliferation of robots [2]. Performance Metrics and Chip Capabilities - The Thor-U chip currently operates at 10Hz, with potential upgrades to 20Hz or 30Hz through optimizations [2]. - The VLA model is designed to be platform-agnostic, ensuring consistent performance across different hardware [2]. Language Integration and Cognitive Abilities - Language understanding is crucial for advanced autonomous driving capabilities, enhancing the model's ability to handle complex scenarios [2]. - VLA's ability to generalize and learn from experiences is likened to human learning, allowing it to adapt to new situations without repeated failures [2]. Model Upgrade and Iteration - The 3.2B MoE vehicle model has a structured upgrade cycle, focusing on both pre-training and post-training updates to enhance various capabilities [3]. User Experience and Trust - The article highlights the importance of user trust and experience, noting that different user groups will gradually accept the technology [2]. - Future iterations aim to improve driving speed and responsiveness, addressing current limitations in specific scenarios [5][12]. Competitive Landscape and Differentiation - The company is closely monitoring competitors like Tesla, aiming to differentiate its approach through gradual iterations and a focus on full-scene autonomous driving [12]. - VLA's architecture is designed to support unique product experiences, setting it apart from competitors [13]. Safety Mechanisms - The AEB (Automatic Emergency Braking) function is emphasized as a critical safety feature, ensuring high frame rates for emergency scenarios [14].

Core Viewpoint - The VLA architecture has significant technical potential and is seen as a long-term framework for autonomous driving, evolving from end-to-end systems to a more robust model that can support urban driving scenarios [1][4]. Group 1: VLA Architecture and Technical Potential - The VLA architecture is derived from robotics and embodied intelligence, emphasizing the need for visual and action capabilities, and is expected to evolve alongside advancements in robotics [1]. - VLA's ability to generalize is not solely dependent on data input but is enhanced through reinforcement learning, allowing it to autonomously address new challenges [5]. - The VLA model is designed to support various platforms without differentiation, ensuring consistent performance across different hardware [2][3]. Group 2: Performance Metrics and Future Enhancements - The current operational speed of the Thor-U chip is 10Hz, with potential upgrades to 20Hz and 30Hz through optimizations in data and algorithm architecture [2]. - The VLA model's upgrade cycle includes both pre-training and post-training updates, allowing for continuous improvement in capabilities such as spatial understanding and language processing [6]. - The VLA architecture aims to achieve L4 autonomous driving capabilities within a year, with a focus on rapid iteration and simulation-based testing [12]. Group 3: User Experience and Interaction - Language understanding is deemed essential for future autonomous driving, enhancing the model's ability to handle complex scenarios and improving overall driving experience [4]. - The VLA system is designed to adapt to user preferences, allowing for different driving styles based on individual needs and enhancing user trust in the technology [19]. - Features such as remote vehicle summoning and real-time monitoring of the vehicle's surroundings are being developed to improve user interaction and experience [13]. Group 4: Competitive Landscape and Strategic Decisions - The company is currently utilizing NVIDIA chips for model deployment, focusing on maintaining versatility and avoiding being locked into specific architectures [3]. - The company is closely monitoring competitors like Tesla, aiming to learn from their advancements while prioritizing a gradual and comprehensive approach to achieving full autonomous driving capabilities [12]. - The VLA architecture is positioned as a differentiating factor in the market, leveraging reinforcement learning to enhance driving logic and user experience [20].

Core Viewpoint - The article discusses the reasoning capabilities of large language models (LLMs) and emphasizes the importance of intermediate reasoning steps in enhancing model confidence and accuracy in problem-solving [5][10][34]. Group 1: Importance of Reasoning in LLMs - Reasoning in LLMs refers to the intermediate thought processes that occur before arriving at a final answer, which can significantly improve the model's ability to solve complex problems [5][11]. - Introducing a chain of thought (CoT) allows LLMs to tackle inherently serial problems without needing to expand the model size, thus bridging the gap between Transformers and Turing machines [12][13]. - The presence of reasoning steps increases the accuracy and reliability of answers, reducing the likelihood of random guessing [14][17]. Group 2: Enhancing Model Confidence - Answers derived from reasoning processes lead to greater confidence in the model's outputs, as they are based on logical deductions rather than mere guesses [19][20]. - Denny Zhou highlights that pre-trained models possess reasoning capabilities even without fine-tuning, although these outputs may not be prioritized in greedy decoding [21][24]. Group 3: Methods to Improve Reasoning - The CoT-decoding method selects reasoning paths from top-k alternatives, enhancing performance on reasoning tasks and approaching the effectiveness of instruction-tuned models [26]. - Supervised fine-tuning (SFT) involves training models on human-written step-by-step problems, but it may lack generalization across new scenarios [27][28]. - Reinforcement learning fine-tuning has emerged as a powerful method for eliciting reasoning, focusing on generating longer responses and improving model performance through iterative training [31]. Group 4: Future Directions - Denny Zhou identifies key areas for future breakthroughs, including addressing tasks with non-unique verifiable answers and developing practical applications beyond benchmark testing [35][40].

Group 1 - The recent incident involving DeepSeek highlights the issue of AI hallucinations, where the model fabricated events and referenced non-existent legal documents, raising concerns about the increasing hallucination rates in AI models [1][2] - OpenAI's o3 model has shown a significant increase in hallucination rates, with 33% of responses exhibiting hallucinations, nearly double that of its predecessor o1, and even higher rates in other models like o4-mini at 48% [1][2] - The phenomenon of hallucinations is linked to over-optimization in reinforcement learning (RL), where models may produce correct answers but through flawed reasoning processes, leading to a disconnect between output and logical reasoning [2][3] Group 2 - Experts suggest that the increase in hallucinations is indicative of a broader issue in understanding what humans truly want from AI, as models optimized for specific tasks may neglect the quality of their reasoning processes [3][4] - The reinforcement learning paradigm primarily rewards final outcomes, which can lead to models developing incorrect but efficient strategies, contributing to the hallucination phenomenon [3][4] - Current reinforcement learning methods, such as GRPO, have not effectively addressed the need for regularization in the reasoning process, resulting in models that may produce correct answers while lacking logical coherence [4][5] Group 3 - The design of reward functions in reinforcement learning remains a critical challenge, as it is difficult to create effective supervisory signals for the reasoning processes of large models [6][7] - There is a need for more sophisticated reward models that can provide feedback on the reasoning process itself, rather than solely on the final output, to mitigate hallucination issues [5][6] - The exploration of non-scalar feedback mechanisms, such as language-based feedback, could enhance the training of models by allowing them to adjust based on qualitative assessments rather than just numerical rewards [7][8] Group 4 - The current benchmarks for evaluating model reasoning capabilities are limited, as they often rely on fixed datasets that do not capture the flexibility of large language models [9][10] - The ability of models to generalize and perform well on varied tasks is still under scrutiny, with evidence suggesting that many models rely heavily on memorization rather than true reasoning [10][11] - Future advancements in model training will require a focus on dynamic interactions with complex environments to foster genuine learning and reasoning capabilities beyond mere imitation of human behavior [15][16]

Core Insights - Technological development does not always exhibit exponential growth; after initial breakthroughs, growth tends to slow down [2][4] - As the gap in foundational models narrows, the focus of industry competition shifts from "technological leadership" to "product experience," creating opportunities for startups [2][6] - A product that fails to establish a strong data barrier or user experience moat is vulnerable to being integrated or replaced by foundational models [2][13] - AI will not change fundamental human needs but has the potential to reshape service delivery methods and service logic, leading to richer interactions and greater system extensibility [2][14] Industry Dynamics - The initial optimism surrounding technologies like ChatGPT has given way to caution as the industry faces pre-training bottlenecks, similar to past expectations in autonomous driving [4][5] - The current stage of AI development can be likened to the mobile internet's evolution, where the emergence of open-source models parallels the explosive growth of the Android platform [8][9] - Companies that enhance existing demand efficiency with new technologies are more likely to succeed than those that create demand for new technologies [9][11] - The infrastructure evolution, such as the rollout of 4G, significantly impacts the growth of applications, similar to how AI's development is currently unfolding [9][11] Competitive Landscape - Major companies are rapidly positioning themselves in key areas of the foundational model chain, which may limit opportunities for startups [10] - AI's ability to enhance business efficiency and penetrate deeply into various sectors suggests that its impact will surpass that of the mobile internet era [11][12] - The phrase "model equals application" highlights the fundamental shift in the competitive landscape, where model upgrades can quickly render certain startup projects obsolete [13][14] Service Innovation - AI's general capabilities are often insufficient for practical applications, revealing limitations that can become entry points for new innovations [14][15] - AI can fundamentally reconstruct service logic rather than merely digitizing existing processes, allowing for personalized service strategies with minimal marginal costs [15]

本期内容是拾象 CEO 李广密对大模型公司阶跃星辰首席科学家张祥雨的访谈。 张祥雨专注于多模态领域，他提出了 DreamLLM 多模态大模型框架，这是业内最早的图文生成理解 一体化的多模态大模型架构之一，基于这个框架，阶跃星辰发布了中国首个千亿参数原生多模态大 模型 Step-1V。此外，他的学术影响力相当突出，论文总引用量已经超过了 37 万次。 一直以来，业界都相当期待一个理解、生成一体化的多模态，但直到今天这个模型还没出现，如何 才能达到多模态领域的 GPT-4 时刻？这一期对谈中，祥雨结合自己在多模态领域的研究和实践历 程，从纯粹的技术视角下分享了自己对多模态领域关键问题的全新思考，在他看来，虽然语言模型 领域的进步极快，但多模态生成和理解的难度被低估了： • 接下来 2-3 年，多模态领域会有两个 GPT-4 时刻：多模态推理和自主学习； • o1 范式的技术本质在于激发出 Meta CoT 思维链：允许模型在关键节点反悔、重试、选择不同分 支，使推理过程从单线变为图状结构。 目录 01 研究主线： 重新回归大模型 • 多模态生成理解一体化难以实现的原因在于，语言对视觉的控制能力弱，图文对齐不精确， ...

Core Viewpoint - The article introduces a novel reasoning paradigm called "Diffusion Chain of Lateral Thought," which enhances the reasoning capabilities of large models by treating intermediate results in the reverse diffusion process as steps in the reasoning process, optimizing the final output's correctness through reinforcement learning [1][34]. Group 1: Introduction of the Concept - The "Diffusion Chain of Lateral Thought" is a new reasoning paradigm proposed by Professor Qi Guojun's team at West Lake University MAPLE Lab, emphasizing the importance of divergent thinking in large model training and inference [1][6]. - This method allows for non-linear generation of responses, contrasting with traditional linear reasoning chains, thereby encouraging more creative and exploratory reasoning paths [1][7]. Group 2: Application and Results - The method has been successfully applied to two representative diffusion language models, showing significant improvements in mathematical reasoning and code generation tasks, surpassing existing models [2][30]. - The team trained the "Ordered Mask Generation Diffusion Language Model" (LLaDOU) based on the LLaDA model, achieving superior performance in complex reasoning tasks compared to other diffusion language models [2][31]. Group 3: Experimental Validation - Experiments demonstrated that the DCoLT approach outperformed traditional methods like Chain of Thought (CoT) and DoT in tasks such as Sudoku solving and mathematical reasoning, achieving a 57.0% accuracy on the GSM8K-Aug dataset [30]. - The LLaDOU model achieved an accuracy of 88.1% in mathematical reasoning tasks, significantly higher than other models, indicating the effectiveness of the proposed reasoning paradigm [32]. Group 4: Theoretical Implications - The research highlights that traditional autoregressive models are not the only choice for generating answers, suggesting that optimizing the order of token generation can lead to more effective reasoning processes [2][34]. - The findings provide important insights into the training and inference of foundational large models, advocating for a shift from linear to non-linear reasoning paradigms in AI [2][6].

Core Insights - The article discusses the latest paradigms in AI, particularly focusing on the concept of "test-time compute" and how large language models (LLMs) can enhance their reasoning capabilities through various methods [3][12][26]. Group 1: AI Paradigms - The blog systematically organizes the latest paradigms in AI, emphasizing "test-time compute" [3]. - LLMs exhibit similarities to human thought processes, drawing parallels with Daniel Kahneman's "Thinking, Fast and Slow" [4][5]. - The reasoning process in LLMs can be likened to human cognitive systems, where "System 1" represents quick, intuitive responses, and "System 2" denotes slower, analytical thinking [6][7]. Group 2: Enhancing Reasoning in LLMs - The concept of "Chain of Thought" (CoT) allows models to allocate variable computational resources based on problem complexity, particularly beneficial for complex reasoning tasks [9]. - Reinforcement learning (RL) has been scaled up in reasoning, with significant changes initiated by OpenAI's developments [14]. - The training process of models like DeepSeek R1 involves parallel sampling and sequential improvement, enhancing the reasoning capabilities of LLMs [15][16]. Group 3: External Tool Utilization - The use of external tools during the reasoning process can improve efficiency and accuracy, such as employing code interpreters for complex calculations [19]. - OpenAI's recent models, o3 and o4-mini, emphasize the importance of tool usage, which marks a paradigm shift in AI development [20][21]. Group 4: Future Research Directions - The article raises open questions for future research, such as improving RNNs to dynamically adjust computation layers and enhancing Transformer architectures for better reasoning [28]. - It also discusses the challenge of training models to generate human-readable CoTs that accurately reflect their reasoning processes while avoiding reward hacking [29][30].

Core Insights - The article discusses the concepts of "Test-time Compute" and "Chain-of-Thought" (CoT) as methods to significantly enhance model performance in artificial intelligence [1][2][6] Group 1: Motivation and Theoretical Background - Allowing models to think longer before providing answers can be achieved through various methods, enhancing their intelligence and overcoming current limitations [2][8] - The core idea is deeply related to human thinking processes, where humans require time to analyze complex problems, aligning with Daniel Kahneman's dual-system theory from "Thinking, Fast and Slow" [10][11] - By consciously slowing down and reflecting, models can engage in more rational decision-making, akin to human System 2 thinking [11][12] Group 2: Computational Resources and Model Architecture - Deep learning views neural networks as capable of accessing computational and storage resources, optimizing their use through gradient descent [13] - In Transformer models, the computational load (flops) for each generated token is approximately double the number of parameters, with sparse models like Mixture of Experts (MoE) utilizing only a fraction of parameters during each forward pass [13] - CoT allows models to perform more computations for each token based on the difficulty of the problem, enabling variable computational loads [13][18] Group 3: CoT and Learning Techniques - Early improvements in CoT involved generating intermediate steps for mathematical problems, with subsequent research showing that reinforcement learning can significantly enhance CoT reasoning capabilities [19][20] - Supervised learning on human-written reasoning paths and appropriate prompts can greatly improve the mathematical abilities of instruction-tuned models [21][23] - The effectiveness of CoT prompts in increasing success rates for solving mathematical problems is more pronounced in larger models [23] Group 4: Sampling and Revision Techniques - The fundamental goal of test-time computation is to adaptively modify the model's output distribution during reasoning [24] - Parallel sampling methods are straightforward but limited by the model's ability to generate correct solutions in one go, while sequential revision requires careful execution to avoid introducing errors [24][25] - Combining both methods can yield optimal results, with simpler problems benefiting from sequential testing and more complex problems performing best with a mix of both approaches [24][25] Group 5: Advanced Techniques and Future Directions - Various advanced algorithms, such as Best-of-N and Beam Search, are employed to optimize the search process for high-scoring samples [29][30] - The RATIONALYST system focuses on synthesizing reasoning based on vast unannotated data, providing implicit and explicit guidance for generating reasoning steps [32][33] - Future challenges include enhancing computational efficiency, integrating self-correction mechanisms, and ensuring the reliability of reasoning outputs [47][50]