AI 与机器学习2025arxiv.org
Mercury：基于扩散模型的超快语言模型
Mercury: Ultra-fast language models based on diffusion
进阶 · 约 60 分钟读完前置知识：扩散模型基础、Transformer/语言模型基础、文本生成
导读
这篇论文介绍基于扩散模型的超快语言模型 Mercury，目标是大幅提升文本生成速度。它采用扩散式、非逐 token 自回归的生成路线，通过迭代细化来产出文本；评论区的 playground 实测也普遍称其速度极快。
HN 关注点：扩散式 LLM 能否把推理速度提升到实用级别，并讨论它对 agent/CI 工作流、上下文窗口和专用硬件的影响。
HN 评论精选
这几条评论里：支持 1 质疑 1 补充 1 亲历 1
支持LLM agent 会让本就缓慢的 CI/测试成为更大瓶颈，因此 Mercury 的速度值得关注。
我一直在跟同事提醒：有了 LLM agent，测试性能会更快成为 CPU/CI 瓶颈；很多团队在 LLM 出现之前就已经卡在 CI 速度上。agent 写代码再快，如果每次改动要一小时测试也没意义。我见过的项目里，大量开发时间耗在等 PR 变绿，排队数小时、flake 重跑；未来 agent 把简单 ticket 变成绿色 PR 并自行修测试失败，只会更糟。测试设施有不少低垂果实，但多年几乎没进展，大家习惯了 CI 慢而贵；为 hermetic 构建去掉跨运行缓存、从本地机器搬到 I/O 慢的云 VM 反而更慢。Mercury 快得离谱，我做的几次快速测试……
mike_hearn查看英文原文
A good chance to bring up something I've been flagging to colleagues for a while now: with LLM agents we are very quickly going to become even more CPU bottlenecked on testing performance than today, and every team I know of today was bottlenecked on CI speed even before LLMs. There's no point having an agent that can write code 100x faster than a human if every change takes an hour to test. Maybe I've just got unlucky in the past, but in most projects I worked on a lot of developer time was wasted on waiting for PRs to go green. Many runs end up bottlenecked on I/O or availability of workers, and so changes can sit in queues for hours, or they flake out and everything has to start again. As they get better coding agents are going to be assigned simple tickets that they turn into green PRs, with the model reacting to test failures and fixing them as they go. This will make the CI bottleneck even worse. It feels like there's a lot of low hanging fruit in most project's testing setups, but for some reason I've seen nearly no progress here for years. It feels like we kinda collectively got used to the idea that CI services are slow and expensive, then stopped trying to improve things. If anything CI got a lot slower over time as people tried to make builds fully hermetic (so no inter-run caching), and move them from on-prem dedicated hardware to expensive cloud VMs with slow IO, which haven't got much faster over time. Mercury is crazy fast and in a few quick tests I did, created good and correct code. How will we make test execution keep up with it?
质疑实测 playground 出现失控生成：模型写 regex 测试停不下来直至上下文耗尽，后期判断也出错。
我试了 playground，得到了奇怪的回答。我要一个 regex 模式，模型先给自己列了个小计划，然后写出模式并开始写测试；但它一直没停，测试越写越大，直到达到上下文限制、回答被取消。每个测试它还注释应该 pass 还是 fail，但大约第 30 个测试后开始判断错，把本应 pass 的说成 fail；到第 120 个左右测试已经是无意义字符，直到回答被截断。它给出的 regex 也是错的，不过我觉得停不下来这个问题更有意思。
true_blue查看英文原文
I tried the playground and got a strange response. I asked for a regex pattern, and the model gave itself a little game-plan, then it wrote the pattern and started to write tests for it. But it never stopped writing tests. It continued to write tests of increasing size until I guess it reached a context limit and the answer was canceled. Also, for each test it wrote, it added a comment about if the test should pass or fail, but after about the 30th test, it started giving the wrong answer for those too, saying that a test should fail when actually it should pass if the pattern is correct. And after about the 120th test, the tests started to not even make sense anymore. They were just nonsense characters until the answer got cut off. The pattern it made was also wrong, but I think the first issue is more interesting.
补充关注点从知识/准确率转向上下文窗口和推理速度，并追问 dLLM 能否与 Groq 硬件结合、训练方式是否不同。
喜欢 playground 的 UI，让我想起 Qwen chat。生成式 AI 的瓶颈已经不是知识或准确率，而是上下文窗口和速度。Google（还有 Meta？）把上下文推到约 100 万 token 很惊人，但今天多数选项仍卡在每轮约 128k，之后就开始忘。另一个问题是推理和思考耗时；dLLM 是有意思的路线，另外还有 Groq 硬件。我好奇这能否和 Groq 硬件结合，响应会不会瞬时？playground 每轮能处理多少 token？我没找到太多信息。推理用的是哪个模型？dLLM 的训练和标准自回归 LLM 一样吗，还是权重和模型完全不同？
Alifatisk查看英文原文
Love the ui in the playground, it reminds me of Qwen chat. We have reached a point where the bottlenecks in genAI is not the knowledge or accuracy, it is the context window and speed. Luckily, Google (and Meta?) has pushed the limits of the context window to about 1 million tokens which is incredible. But I feel like todays options are still stuck about ~128k token window per chat, and after that it starts to forget. Another issue is the time time it takes for inference AND reasoning. dLLMs is an interesting approach at this. I know we have Groqs hardware aswell. I do wonder, can this be combined with Groqs hardware? Would the response be instant then? How many tokens can each chat handle in the playground? I couldn't find so much info about it. Which model is it using for inference? Also, is the training the same on dLLMs as on the standardised autoregressive LLMs? Or is the weights and models completely different?
亲历实测速度极快，并讨论 diffusion 可视化是否真实反映从噪声向量逐步细化到确定 token 的过程。
用了免费 playground 链接，确实极快。“diffusion mode”开关作为可视化也挺有意思，但我不确定它有多准确——它先渲染成线噪声再逐步细化，而现实中大概是 token 从状态空间里某个不精确的向量逐渐变精确，直到成为确定单词，对吧？
chc4查看英文原文
Using the free playground link, and it is in fact extremely fast. The "diffusion mode" toggle is also pretty neat as a visualization, although I'm not sure how accurate it is - it renders as line noise and then refines, while in reality presumably those are tokens from an imprecise vector in some state space that then become more precise until it's only a definite word, right?
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
- 2025Hacker News 讨论帖10 分 · 2 评
- 2025Hacker News 讨论帖576 分 · 242 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