AI 与机器学习2024arxiv.org
Quiet-STaR：语言模型可以自学在开口前思考
Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking
进阶 · 约 60 分钟读完前置知识：Transformer 基础、STaR 论文、思维链推理
导读
Quiet-STaR 研究如何让语言模型在生成每个 token 之前先生成一段内部「思考」（隐式推理），并自我学习保留那些有助于更好预测后续文本的思考。它把 STaR 的自我教学思路扩展到无标注的任意文本，让模型通过试错学会「先想再说」，目标是提升推理表现。
HN 关注点：LLM 能否真正「先思考再说话」，讨论焦点包括 chain-of-thought 与人类双系统思维的类比、模型推理深度的上限，以及对 AI 热潮和 LLM 实际能力的质疑。
HN 评论精选
这几条评论里：支持 1 质疑 1 补充 2
补充把 chain-of-thought 类比为 Kahneman 的双系统理论：LLM 天生只有系统1，step-by-step 提示给了它一个系统2式的思考沙盒。
我前几天想到，chain-of-thought 推理模式让 LLM 系统表现提升，这似乎与 Kahneman 在《思考，快与慢》中的双系统模型平行。书里说我们用「系统1」做低耗、低计算的思考（1+1、天空是___），用「系统2」做有意识的高认知任务（复杂乘法、推理、使用工具）。当我们提示 LLM 一步步思考时，相当于给它一个写下想法的空间，让它在下一个 token 预测时参考，这是一个初级的系统2，像思考的沙盒。
dcrimp查看英文原文
I had this thought the other day that the whole chain of thought reasoning pattern contributing to improved performance in LLM-based systems seems to sit parallel to Kahneman's two-system model of the mind that he covers in 'Thinking, Fast and Slow'. Haven't read it in a few years, but I recall the book suggests that we use one 'System 1' in our brains primarily for low-effort, low computation thinking - like 1+1=? or "the sky is ____". It then suggests that we use a 'System 2' for deliberate, conscious, high-cognitive tasks. Dense multiplication, reasoning problems, working with tools - generally just decision-making. Anything that requires focus or brain power. Our brain escalates tasks from S1 to S2 if they feel complex or dangerous. Maybe I'm being too cute, but it feels like critique that "LLMs aren't intelligent because they are stochastic parrots" is an observation that they are only equipped to use their 'System 1'. When we prompt an LLM to think step-by-step, we allow it a workspace to write down it's thoughts which it can then consider in it's next token prediction, a rudimentary System 2, like a deliberation sandbox. We do a similar thing when we engage our System 2 - we hold a diorama of the world in the front of our mind, where we simulate what the environment will do if we proceed with a given action - what our friend might respond to what we say, how the sheet steel might bend to a force, how the code might break, how the tyres might grip. And we use that simulation to explore a tree of possibilities and decide an action that rewards us the most. I'm no expert, but this paper seems to recognise a similar framework to the above. Perhaps a recurrent deliberation/simulation mechanism will make it's way into models in the future, especially the action models we are seeing in robotics.
支持50 层网络只能做约 50 步推理；人类做 4 位数乘法也需要纸笔，不应指望 LLM 无辅助完成。
在我看来很明显，一个约 50 层深的网络只能对符号问题推理 50 个「步骤」（打引号是因为它并非我们理解的步骤）。人类（至少聪明的人类）似乎能明显超出这些步数，但都得靠真正的思考和反复斟酌，也许还要用记事本。所以，指望 ChatGPT 不动脑、不借助「纸笔」就正确做 4 位数乘法，其实挺离谱的——没多少人类能做到。
anon291查看英文原文
So it seems 'obvious' to me that a network about 50 layers deep (for example) can only reason about symbolic questions for 50 'steps' (in quotes because it's not a step as we think about it). It only seems there's more complexity because it's 50 steps in one or more learned subspaces that the model has been trained in (which might mean the model can accomplish more than one 'human step' in its 'step'). Humans (well intelligent humans at least) seem able to obviously reason beyond those steps, but we all know it requires real thinking and deliberation and perhaps a notepad to be able to do that. It's quite something to, for example, expect ChatGPT to be able to correctly do 4 digit multiplications without any thought or recourse to 'paper' when very few human beings can do that.
补充引用 Dijkstra 轶事：小时候被教导想好结尾再开口，呼应论文「先思考再说话」的主题。
Dijkstra 的英文风格很精确；尽管母语是荷兰语，我觉得他比很多母语者用得更好。在某一篇 EWD 里他回忆，小时候大人教他们：除非已经知道这句话怎么结尾，否则绝不开始说。我打赌这两件事有因果联系。
082349872349872查看英文原文
Edsger Dijkstra had a precise english style; even though his mother tongue was Dutch, I find he made better use of English than many native speakers. In one of the EWD's, he reminisced that, as children, they were taught to never begin to speak a sentence unless they already knew how they were going to finish it. I'd bet these two observations have a causal connection.
质疑质疑：LLM 在非平凡任务上表现很差，怀疑其日常实用性，不解为何 AI 如此受关注。
纯属轶事，我尽量不说，但 HN 首页至少一半是 AI，实在忍不住：ChatGPT 这类 LLM 在我给的非平凡任务上表现糟糕透顶，我认真觉得天天用它的人要么无能，要么领域太简单。从困难的异步 C++ 并行问题到德语语言细节，它都在根本层面搞砸。我不理解为什么这么多技术人如此关注 AI。日常编程真这么琐碎吗？我很想知道用例到底是什么。
lionkor查看英文原文
This is purely anecdotal, and I try to keep it to myself but its very difficult when at least half of the HN homepage is AI related: LLMs like ChatGPT do so utterly terribly at any non-trivial job I throw at it that I seriously consider people who use it daily to either be straight up incompetent, or maybe their domain is so trivial that the LLM actually does well. From asking LLMs to solve a highly difficult async C++ parallelism problem, to german language specifics, it just fucks up at a fundamental level. I understand that LLMs cannot solve these issues and why, but then I do not understand the heavy focus on AI by so many tech people. Is day to day programming job so trivial that LLMs do a good job, while at the same time being too difficult for you to do it yourself? I really, really want to know exactly what the use case is. Do people just throw simple problems at it to validate their own preconceived notion of how cool and useful LLMs are? Whats the deal?
相关资源
从这篇论文所有 HN 讨论帖的全部评论里整理出来的链接。
历次讨论
- 2024Hacker News 讨论帖280 分 · 264 评
同领域还值得读
DeepSeek-R1：通过强化学习激励 LLM 的推理能力
DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via RL
研究如何用强化学习激励大语言模型的推理能力，提出 DeepSeek-R1，并探讨 RL 训练对模型推理表现的影响。
AI 与机器学习2025 · 1,351 分 · 1,056 评 · 引用 30进阶 · 约 80 分钟编辑精选1-bit LLM 时代：面向低成本计算的三值参数
The Era of 1-bit LLMs: ternary parameters for cost-effective computing
提出用三值参数 {-1, 0, 1} 替代 LLM 中的浮点权重，即所谓“1-bit LLM”。
AI 与机器学习2024 · 1,040 分 · 447 评 · 引用 19进阶 · 约 45 分钟编辑精选DeepSeek-V3.2：推进开放大语言模型的前沿
DeepSeek-v3.2: Pushing the frontier of open large language models
这是 DeepSeek-V3.2 的技术报告，延续了 DeepSeek 开源权重模型的路线，目标是让开放模型继续逼近甚至追平前沿闭源模型。
AI 与机器学习2025 · 982 分 · 465 评进阶 · 约 90 分钟编辑精选通过自我对弈与通用强化学习掌握国际象棋和将棋（AlphaZero）
Mastering Chess and Shogi by Self-Play with General Reinforcement Learning
DeepMind 的 AlphaZero 论文：用通用强化学习算法（自我博弈 + MCTS + 神经网络），从零开始仅数小时训练便在国际象棋和将棋上击败了世界顶级引擎 Stock…
AI 与机器学习2017 · 539 分 · 270 评 · 引用 13进阶 · 约 90 分钟编辑精选用强化学习发现更快的矩阵乘法算法
Discovering faster matrix multiplication algorithms with reinforcement learning
论文提出 AlphaTensor，用强化学习搜索矩阵乘法算法。
AI 与机器学习2022 · 518 分 · 111 评专业 · 约 90 分钟编辑精选