{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-09-22T18:00:52.144Z","headline":"Show HN：LLM 注意力可视化","description":"一个基于 Transformers.js 的 React 应用，可点击或悬停生成的 token，查看影响其生成的历史 token 的注意力权重。可视化通过缩放值向量幅度并跨所有注意力头和层聚合实现，能展示模型从原文复制信息及组合多短语语义的过程。应用使用 6 亿参数模型，并预生成提示词以便即时加载。","url":"https://www.aioga.com/news/cmttautvn02parofpcrrra5u4/","mainEntityOfPage":"https://www.aioga.com/news/cmttautvn02parofpcrrra5u4/","datePublished":"2026-09-08T22:55:57.000Z","dateModified":"2026-09-08T22:55:57.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://ishamf.dev/p/llm-attention-visualizer","https://aihot.news/items/cmttautvn02parofpcrrra5u4"],"canonicalUrl":"https://www.aioga.com/news/cmttautvn02parofpcrrra5u4/","directAnswer":{"@type":"Answer","text":"该项目是一个基于 Transformers.js 的 React 应用，用于可视化大语言模型生成阶段的注意力机制。用户点击或悬停生成的 token，可查看影响其生成的历史 token 及相应可视化结果。","url":"https://www.aioga.com/news/cmttautvn02parofpcrrra5u4/","dateCreated":"2026-09-08T22:55:57.000Z","author":{"@type":"Organization","@id":"https://www.aioga.com/authors/aioga-editorial/#editorial-team","name":"Aioga Editorial Team","url":"https://www.aioga.com/authors/aioga-editorial/"}},"evidence":[{"@type":"CreativeWork","name":"ishamf.dev source article","url":"https://ishamf.dev/p/llm-attention-visualizer","datePublished":"2026-09-08T22:55:57.000Z","provider":{"@type":"Organization","name":"ishamf.dev","url":"https://ishamf.dev/p/llm-attention-visualizer"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.news/items/cmttautvn02parofpcrrra5u4","datePublished":"2026-09-08T22:55:57.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.news/items/cmttautvn02parofpcrrra5u4"}}],"aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","originalPublisher":{"name":"ishamf.dev","url":"https://ishamf.dev/p/llm-attention-visualizer"},"geoDeepAnswer":null,"article":{"id":"cmttautvn02parofpcrrra5u4","slug":"cmttautvn02parofpcrrra5u4","url":"https://www.aioga.com/news/cmttautvn02parofpcrrra5u4/","title":"Show HN：LLM 注意力可视化","title_en":"","summary":"一个基于 Transformers.js 的 React 应用，可点击或悬停生成的 token，查看影响其生成的历史 token 的注意力权重。可视化通过缩放值向量幅度并跨所有注意力头和层聚合实现，能展示模型从原文复制信息及组合多短语语义的过程。应用使用 6 亿参数模型，并预生成提示词以便即时加载。","source":"Hacker News 热门（buzzing.cc 中文翻译）","sourceUrl":"https://ishamf.dev/p/llm-attention-visualizer","aiHotUrl":"https://aihot.news/items/cmttautvn02parofpcrrra5u4","publishedAt":"2026-09-08T22:55:57.000Z","category":"行业动态","score":58,"selected":false,"articleBody":["A visualization of the attention mechanism in LLMs.","One interesting thing about transformer-based large language models are that, during the generation phase, it is able to draw information from any of its previous tokens. But it needs to be selective; if every token affects the generation equally, it won't be very effective. This process needs a mechanism to decide how much a token affects the next token.","Turns out, we can visualize this mechanism!","You can tap or hover over any of the generated tokens to see the past tokens that affected* the generation.","* \"Affected\" might not be fully accurate, as this visualization is highly simplified. It's calculating the attention weight, scaled by the magnitude of the value vector, aggregated across all attention heads, and summed across all layers. This is then used to control the opacity of the previous tokens. The largest values always have an opacity of 1 and the rest are interpolated.","A lot of information had to be thrown away to limit the visualization to just one numeric value per past token. Because of that, when I started implementing this, I actually thought it might not be comprehensible. But it actually can produce some interesting patterns!","For example, in the default \"Office Move Summary\" prompt, you can hover over the text that are copied verbatim like the address and dates. You can then see the original data stand out quite a bit, because the generated token takes up a lot of the information from the source data.","This addresses one thing that I've previously found unintuitive about LLMs. If they work by predicting the next tokens probabilistically, why are they somehow so good at copy-pasting stuff? Won't they eventually make a mistake just by random chance?","But with this mechanism, you can see that it doesn't predict the entire sequence from some limited internal states. Since it has access to all past tokens, it can just decide which past tokens to draw from when copying, and so the probability of errors can be very low. In the \"Debugging an Average Function\" example, you can see that this quite small model (600 million parameters) can easily reproduce an entire JS function except for the intended modification. (Although it's not actually capable of finding the issue by itself, so it needed some hints.)","Another interesting part is when you hover over the \"remain\" in \"Existing access cards and phone numbers remain \" in the \"Office Move Summary\" prompt. You can see that it draws from \"work\" in \"Existing employee access cards will work \" and \"stay the same\" in \"company phone numbers will stay the same \". So it's kind of combining the information from the words in both phrases, which I find quite cool.","The visualization itself is a pretty basic React app using Transformers.js to generate the text. But, since we need to pull more data out of the model to visualize it, it can't use the regular generation loop. I had to vibe-code the generation loop in the app so we can actually keep track of the values to visualize.","Despite using a smaller model for this, it's still hundreds of megabytes, and waiting for it to download before showing anything just won't work. So I pre-generated a bunch of prompts that can be loaded and viewed instantly.","Another tricky thing is that some of the things in the visualization are not actually meant to be read, so they're not defined as outputs. I suppose if you're implementing this using Python ML libraries, it would still be easy to access them. But Transformers.js uses .onnx files that contains the entire computation graph. The model loading and computation logic is implemented in wasm, so there's no easy way to access anything other than the predefined outputs, as far as I can tell.","In the end, I used a small script to modify the onnx file just enough to expose those internal values. But that means I can't just use the regular .onnx model. Since I want to have a browser-based generation feature, I have to upload a separate instrumented model to my own Hugging Face repo：https://huggingface.co/ishamf/Qwen3-0.6B-ONNX-Instrumented and point the app there.","You can find the code in the GitHub repo：https://github.com/ishamf/llm-visualizer."],"articleImages":[],"mediaStatus":"none","articleBodyZh":["大型语言模型中注意力机制的可视化。","基于 Transformer 的大型语言模型有一个有趣的特点：在生成阶段，它能够从之前的任意一个 token 中提取信息。但它需要有选择性；如果每个 token 对生成都有相等影响，效果就不会很好。这个过程需要一个机制来决定某个 token 对下一个 token 的影响程度。","事实证明，我们可以把这个机制可视化！","你可以点击或悬停在任何生成的 token 上，查看影响*该生成的过去 token。","*“影响”可能不完全准确，因为这个可视化被高度简化。它计算注意力权重，再乘以值向量的大小，聚合所有注意力头，并在所有层中求和。然后用它来控制前面 token 的透明度。最大值的透明度始终为 1，其余的则进行插值处理。","为了将可视化限制为每个过去 token 仅一个数值，许多信息必须被舍弃。正因为如此，当我开始实现这个功能时，我实际上认为它可能不易理解。但它实际上可以产生一些有趣的模式！","例如，在默认的“办公室搬迁总结”提示中，你可以悬停在像地址和日期这样的逐字复制的文本上。你会发现原始数据非常突出，因为生成的 token 从源数据中获取了很多信息。","这解释了我之前觉得大型语言模型不直观的一点。如果它们通过概率预测下一个 token 来工作，为什么它们在复制粘贴内容方面表现得如此出色？难道它们最终不会仅凭随机机会而犯错吗？","但通过这个机制，你可以看到它并不是从一些有限的内部状态预测整个序列。由于它可以访问所有过去的标记，它可以决定从过去的哪些标记中复制，因此错误的概率可以非常低。在“调试平均函数”示例中，你可以看到这个相当小的模型（6亿参数）可以轻松复现整个 JS 函数，除了预期的修改部分。（虽然它实际上无法自己找到问题，所以需要一些提示。）","另一个有趣的部分是，当你将鼠标悬停在“办公室搬迁总结”提示中的“remain”的时候，你会看到它从“Existing employee access cards will work”中的“work”和“company phone numbers will stay the same”中的“stay the same”中提取信息。所以它有点像是将两个短语中的信息结合起来，我觉得这非常酷。","可视化本身是一个相当基础的 React 应用，使用 Transformers.js 来生成文本。但是，由于我们需要从模型中提取更多数据以进行可视化，它无法使用常规的生成循环。我不得不在应用中自定义编写生成循环，这样我们才能真正跟踪数值以进行可视化。","尽管使用了较小的模型，它仍然有数百兆字节，等待它下载完再显示任何内容是不行的。因此，我预生成了一些提示，可以立即加载和查看。","另一个棘手的问题是，可视化中的一些内容实际上并不打算被读取，因此它们未被定义为输出。我想如果你使用 Python ML 库实现它，仍然很容易访问这些内容。但 Transformers.js 使用包含完整计算图的 .onnx 文件。模型加载和计算逻辑是在 wasm 中实现的，所以据我所知，除了预定义的输出之外，没有简单的方法可以访问其他内容。","最后，我使用了一个小脚本来修改 onnx 文件，仅足以暴露那些内部值。但这意味着我不能仅使用常规的 .onnx 模型。由于我想要一个基于浏览器的生成功能，我必须将一个单独的插装模型上传到我自己的 Hugging Face 仓库：https://huggingface.co/ishamf/Qwen3-0.6B-ONNX-Instrumented，并在应用中指向那里。","您可以在 GitHub 存储库中找到代码：https://github.com/ishamf/llm-visualizer。"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"该项目是一个基于 Transformers.js 的 React 应用，用于可视化大语言模型生成阶段的注意力机制。用户点击或悬停生成的 token，可查看影响其生成的历史 token 及相应可视化结果。","background":"应用使用一个 6 亿参数模型，并预生成提示词以实现即时加载。可视化将注意力权重按值向量幅度缩放，再跨所有注意力头和层聚合，最终以历史 token 的透明度呈现相对数值。","viewpoint":"Aioga 判断：该工具把抽象的注意力权重转化为可交互的视觉线索，有助于观察模型在示例中提取原文信息或组合短语语义的模式，但不应将其简化结果直接等同于完整的生成因果解释。","implications":"可能影响：这类可视化可能辅助研究者和用户形成对复制文本现象的直观理解；但由于信息被压缩为每个历史 token 一个数值，且结果聚合了多层与多头，不足以单独证明某个 token 对输出的确定性因果作用。","nextStep":"后续观察：需要关注该工具在不同提示词和生成内容中的表现，并核对可视化结果与原始模型计算之间的差异。建议将其作为探索性观察工具，而非完整的模型解释或准确性验证方法。","evidenceRefs":["title","summary","articleBody","source"],"status":"published","aiGenerated":true,"autoApproved":true,"generatedBy":"aioga-editorial:gpt-5.6-sol","reviewedBy":"aioga-editorial-review:gpt-5.6-sol","generatedAt":"2026-09-09T00:04:31.373Z","sourceHash":"caa69ad4069ada7f","review":{"approved":true,"groundedness":95,"clarity":94,"duplicationRisk":8,"blockingIssues":[],"notes":["“影响其生成”沿用了来源中的简化表述，候选内容已通过“不应将其简化结果直接等同于完整的生成因果解释”等表述明确其局限。","“可能辅助研究者和用户形成直观理解”属于合理推断，并以“可能”限定，未冒充来源中的确定事实。"]},"validation":{"passed":true,"mode":"ai-auto","revisions":0,"checks":["schema","length","source-attribution","editorial-labels","inference-boundary","low-source-overlap","no-html","independent-ai-review"]}},"tags":["行业动态","Hacker News 热门（buzzing.cc 中文翻译）"],"translations":{"zh-CN":{"title":"Show HN：LLM 注意力可视化","summary":"一个基于 Transformers.js 的 React 应用，可点击或悬停生成的 token，查看影响其生成的历史 token 的注意力权重。可视化通过缩放值向量幅度并跨所有注意力头和层聚合实现，能展示模型从原文复制信息及组合多短语语义的过程。应用使用 6 亿参数模型，并预生成提示词以便即时加载。","category":"行业动态","source":"ishamf.dev","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN：LLM 注意力可视化 - Aioga AI资讯","description":"一个基于 Transformers.js 的 React 应用，可点击或悬停生成的 token，查看影响其生成的历史 token 的注意力权重。可视化通过缩放值向量幅度并跨所有注意力头和层聚合实现，能展示模型从原文复制信息及组合多短语语义的过程。应用使用 6 亿参数模型，并预生成提示词以便即时加载。","url":"https://www.aioga.com/news/cmttautvn02parofpcrrra5u4/","articleBody":["大型语言模型中注意力机制的可视化。","基于 Transformer 的大型语言模型有一个有趣的特点：在生成阶段，它能够从之前的任意一个 token 中提取信息。但它需要有选择性；如果每个 token 对生成都有相等影响，效果就不会很好。这个过程需要一个机制来决定某个 token 对下一个 token 的影响程度。","事实证明，我们可以把这个机制可视化！","你可以点击或悬停在任何生成的 token 上，查看影响*该生成的过去 token。","*“影响”可能不完全准确，因为这个可视化被高度简化。它计算注意力权重，再乘以值向量的大小，聚合所有注意力头，并在所有层中求和。然后用它来控制前面 token 的透明度。最大值的透明度始终为 1，其余的则进行插值处理。","为了将可视化限制为每个过去 token 仅一个数值，许多信息必须被舍弃。正因为如此，当我开始实现这个功能时，我实际上认为它可能不易理解。但它实际上可以产生一些有趣的模式！","例如，在默认的“办公室搬迁总结”提示中，你可以悬停在像地址和日期这样的逐字复制的文本上。你会发现原始数据非常突出，因为生成的 token 从源数据中获取了很多信息。","这解释了我之前觉得大型语言模型不直观的一点。如果它们通过概率预测下一个 token 来工作，为什么它们在复制粘贴内容方面表现得如此出色？难道它们最终不会仅凭随机机会而犯错吗？","但通过这个机制，你可以看到它并不是从一些有限的内部状态预测整个序列。由于它可以访问所有过去的标记，它可以决定从过去的哪些标记中复制，因此错误的概率可以非常低。在“调试平均函数”示例中，你可以看到这个相当小的模型（6亿参数）可以轻松复现整个 JS 函数，除了预期的修改部分。（虽然它实际上无法自己找到问题，所以需要一些提示。）","另一个有趣的部分是，当你将鼠标悬停在“办公室搬迁总结”提示中的“remain”的时候，你会看到它从“Existing employee access cards will work”中的“work”和“company phone numbers will stay the same”中的“stay the same”中提取信息。所以它有点像是将两个短语中的信息结合起来，我觉得这非常酷。","可视化本身是一个相当基础的 React 应用，使用 Transformers.js 来生成文本。但是，由于我们需要从模型中提取更多数据以进行可视化，它无法使用常规的生成循环。我不得不在应用中自定义编写生成循环，这样我们才能真正跟踪数值以进行可视化。","尽管使用了较小的模型，它仍然有数百兆字节，等待它下载完再显示任何内容是不行的。因此，我预生成了一些提示，可以立即加载和查看。","另一个棘手的问题是，可视化中的一些内容实际上并不打算被读取，因此它们未被定义为输出。我想如果你使用 Python ML 库实现它，仍然很容易访问这些内容。但 Transformers.js 使用包含完整计算图的 .onnx 文件。模型加载和计算逻辑是在 wasm 中实现的，所以据我所知，除了预定义的输出之外，没有简单的方法可以访问其他内容。","最后，我使用了一个小脚本来修改 onnx 文件，仅足以暴露那些内部值。但这意味着我不能仅使用常规的 .onnx 模型。由于我想要一个基于浏览器的生成功能，我必须将一个单独的插装模型上传到我自己的 Hugging Face 仓库：https://huggingface.co/ishamf/Qwen3-0.6B-ONNX-Instrumented，并在应用中指向那里。","您可以在 GitHub 存储库中找到代码：https://github.com/ishamf/llm-visualizer。"]},"en":{"title":"Show HN: LLM Attention Visualization","summary":"A React app based on Transformers.js that allows you to click or hover over generated tokens to see the attention weights of historical tokens influencing their generation. The visualization is achieved by scaling the value vectors and aggregating across all attention heads and layers, showing how the model copies information from the original text and combines multiple phrase semantics. The app uses a 600 million parameter model and pre-generates prompts for instant loading.","category":"Industry","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN: LLM Attention Visualization - Aioga AI News","description":"A React app based on Transformers.js that allows you to click or hover over generated tokens to see the attention weights of historical tokens influencing their generation. The vis...","url":"https://www.aioga.com/en/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:13.776Z"},"ja":{"title":"Show HN：LLM 注意力の可視化","summary":"Transformers.js に基づく React アプリで、生成されたトークンをクリックまたはホバーすると、その生成に影響を与えた履歴トークンの注意力ウェイトを見ることができます。可視化は、値ベクトルの振幅を拡大し、すべての注意力ヘッドと層で集約することによって実現され、モデルが原文から情報をコピーしたり、複数のフレーズの意味を組み合わせたりする過程を示すことができます。アプリは6億パラメータのモデルを使用し、即座に読み込めるように事前にプロンプトを生成します。","category":"業界動向","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN：LLM 注意力の可視化 - Aioga AIニュース","description":"Transformers.js に基づく React アプリで、生成されたトークンをクリックまたはホバーすると、その生成に影響を与えた履歴トークンの注意力ウェイトを見ることができます。可視化は、値ベクトルの振幅を拡大し、すべての注意力ヘッドと層で集約することによって実現され、モデルが原文から情報をコピーしたり、複数のフレーズの意味を組み合わせたりする過程を示...","url":"https://www.aioga.com/ja/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:14.939Z"},"ko":{"title":"Show HN：LLM 주의력 시각화","summary":"Transformers.js 기반의 React 애플리케이션으로, 생성된 토큰을 클릭하거나 마우스를 올리면 해당 토큰 생성을 영향을 준 이전 토큰들의 주의력 가중치를 확인할 수 있습니다. 시각화는 값 벡터의 크기를 스케일링하고 모든 주의력 헤드와 레이어에서 집계하여 수행되며, 모델이 원문에서 정보를 복사하고 여러 구문의 의미를 결합하는 과정을 보여줄 수 있습니다. 애플리케이션은 6억 파라미터 모델을 사용하며, 즉시 로드할 수 있도록 미리 프롬프트를 생성합니다.","category":"업계 동향","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN：LLM 주의력 시각화 - Aioga AI 뉴스","description":"Transformers.js 기반의 React 애플리케이션으로, 생성된 토큰을 클릭하거나 마우스를 올리면 해당 토큰 생성을 영향을 준 이전 토큰들의 주의력 가중치를 확인할 수 있습니다. 시각화는 값 벡터의 크기를 스케일링하고 모든 주의력 헤드와 레이어에서 집계하여 수행되며, 모델이 원문에서 정보를 복사하고 여러 구문의 의...","url":"https://www.aioga.com/ko/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:18.092Z"},"es":{"title":"Show HN: Visualización de atención de LLM","summary":"Una aplicación React basada en Transformers.js que permite hacer clic o pasar el cursor sobre los tokens generados para ver los pesos de atención de los tokens históricos que influyeron en su generación. La visualización se realiza escalando la magnitud de los vectores de valor y agregando a través de todas las cabezas y capas de atención, mostrando cómo el modelo copia información del texto original y combina múltiples frases semánticas. La aplicación utiliza un modelo de 600 millones de parámetros y pre-genera indicaciones para una carga instantánea.","category":"Industria","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN: Visualización de atención de LLM - Aioga Noticias de IA","description":"Una aplicación React basada en Transformers.js que permite hacer clic o pasar el cursor sobre los tokens generados para ver los pesos de atención de los tokens históricos que influ...","url":"https://www.aioga.com/es/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:17.254Z"},"fr":{"title":"Show HN : Visualisation de l'attention des LLM","summary":"Une application React basée sur Transformers.js, permettant de cliquer ou de survoler les tokens générés pour voir les poids d'attention des tokens historiques qui ont influencé leur génération. La visualisation est réalisée en mettant à l'échelle l'amplitude des vecteurs et en agrégeant à travers toutes les têtes et couches d'attention, montrant ainsi comment le modèle copie des informations du texte source et combine la sémantique de plusieurs phrases. L'application utilise un modèle de 600 millions de paramètres et pré-génère des prompts pour un chargement instantané.","category":"Industrie","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN : Visualisation de l'attention des LLM - Aioga Actualités IA","description":"Une application React basée sur Transformers.js, permettant de cliquer ou de survoler les tokens générés pour voir les poids d'attention des tokens historiques qui ont influencé le...","url":"https://www.aioga.com/fr/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:20.946Z"},"de":{"title":"Show HN: Visualisierung der Aufmerksamkeit von LLM","summary":"Eine auf Transformers.js basierende React-Anwendung, mit der man auf erzeugte Tokens klicken oder über sie hovern kann, um die Aufmerksamkeit der historischen Tokens, die ihre Generierung beeinflussen, zu sehen. Die Visualisierung wird durch Skalierung der Vektorgrößen und Aggregation über alle Aufmerksamkeitsschichten und -köpfe erreicht und zeigt, wie das Modell Informationen aus dem ursprünglichen Text kopiert und mehrere Phrasensemantiken kombiniert. Die Anwendung verwendet ein 600-Millionen-Parameter-Modell und generiert im Voraus Aufforderungswörter für sofortiges Laden.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN: Visualisierung der Aufmerksamkeit von LLM - Aioga KI-News","description":"Eine auf Transformers.js basierende React-Anwendung, mit der man auf erzeugte Tokens klicken oder über sie hovern kann, um die Aufmerksamkeit der historischen Tokens, die ihre Gene...","url":"https://www.aioga.com/de/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:20.706Z"},"pt-BR":{"title":"Show HN: Visualização de Atenção do LLM","summary":"Um aplicativo React baseado em Transformers.js, que permite clicar ou passar o mouse sobre os tokens gerados para ver os pesos de atenção dos tokens históricos que influenciam sua geração. A visualização é realizada escalando a magnitude dos vetores e agregando através de todas as cabeças de atenção e camadas, podendo mostrar como o modelo copia informações do texto original e combina múltiplas semânticas de frases. O aplicativo usa um modelo de 600 milhões de parâmetros e pré-gerou prompts para carregamento imediato.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN: Visualização de Atenção do LLM - Aioga Notícias de IA","description":"Um aplicativo React baseado em Transformers.js, que permite clicar ou passar o mouse sobre os tokens gerados para ver os pesos de atenção dos tokens históricos que influenciam sua...","url":"https://www.aioga.com/pt-BR/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:24.136Z"},"ru":{"title":"Показать HN：Визуализация внимания LLM","summary":"React-приложение на основе Transformers.js, позволяющее кликать или наводить на сгенерированные токены, чтобы увидеть веса внимания исторических токенов, повлиявших на их генерацию. Визуализация достигается путем масштабирования векторов значений и агрегации по всем головам и слоям внимания, что демонстрирует процесс копирования информации из исходного текста и комбинирования значений нескольких фраз. Приложение использует модель с 600 миллионами параметров и предсоздает подсказки для мгновенной загрузки.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Показать HN：Визуализация внимания LLM - Aioga Новости ИИ","description":"React-приложение на основе Transformers.js, позволяющее кликать или наводить на сгенерированные токены, чтобы увидеть веса внимания исторических токенов, повлиявших на их генерацию...","url":"https://www.aioga.com/ru/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:23.738Z"},"ar":{"title":"عرض HN: تصور انتباه LLM","summary":"تطبيق React يعتمد على Transformers.js، يمكن من خلاله النقر أو التحويم على الرمز المولد لرؤية أوزان الانتباه للرموز السابقة التي أثرت في توليده. يتم التصور عن طريق تكبير مقدار القيم وتجميعها عبر جميع رؤوس ومستويات الانتباه، مما يظهر عملية نسخ النموذج للمعلومات من النص الأصلي ودمج معاني العبارات الصغيرة المتعددة. يستخدم التطبيق نموذجًا يحتوي على 600 مليون معلمة، ويقوم بإعداد كلمات الإشارة مسبقًا لتحميلها الفوري.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"عرض HN: تصور انتباه LLM - Aioga أخبار الذكاء الاصطناعي","description":"تطبيق React يعتمد على Transformers.js، يمكن من خلاله النقر أو التحويم على الرمز المولد لرؤية أوزان الانتباه للرموز السابقة التي أثرت في توليده. يتم التصور عن طريق تكبير مقدار القيم...","url":"https://www.aioga.com/ar/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:27.284Z"},"hi":{"title":"Show HN: LLM ध्यान विज़ुअलाइज़ेशन","summary":"एक Transformers.js आधारित React एप्लिकेशन, जिसमें आप जेनरेट किए गए टोकन पर क्लिक या होवर करके यह देख सकते हैं कि इसे जेनरेट करने वाले ऐतिहासिक टोकन पर ध्यान के वज़न क्या हैं। विज़ुअलाइज़ेशन को वेक्टर मैग्नीट्यूड के स्केलिंग और सभी ध्यान हेड और लेयर में एग्रीगेशन के माध्यम से किया गया है, जो मॉडल द्वारा मूल पाठ से जानकारी कॉपी करने और कई वाक्यांशों के अर्थ को संयोजित करने की प्रक्रिया को दिखा सकता है। एप्लिकेशन 600 मिलियन पैरामीटर मॉडल का उपयोग करता है, और तात्कालिक लोडिंग के लिए प्री-जनरेटेड प्रॉम्प्ट शब्दों का उपयोग करता है।","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN: LLM ध्यान विज़ुअलाइज़ेशन - Aioga AI समाचार","description":"एक Transformers.js आधारित React एप्लिकेशन, जिसमें आप जेनरेट किए गए टोकन पर क्लिक या होवर करके यह देख सकते हैं कि इसे जेनरेट करने वाले ऐतिहासिक टोकन पर ध्यान के वज़न क्या हैं। विज़ु...","url":"https://www.aioga.com/hi/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:26.793Z"},"it":{"title":"Mostra su HN：Visualizzazione dell'attenzione LLM","summary":"Un'app React basata su Transformers.js, in cui è possibile cliccare o passare il mouse sui token generati per vedere i pesi di attenzione dei token storici che hanno influenzato la loro generazione. La visualizzazione si realizza scalando l'ampiezza dei vettori dei valori e aggregando attraverso tutte le teste e i livelli di attenzione, mostrando come il modello copi informazioni dal testo originale e combini semantiche di più frasi. L'app utilizza un modello da 600 milioni di parametri e precostruisce i prompt per un caricamento immediato.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Mostra su HN：Visualizzazione dell'attenzione LLM - Aioga Notizie IA","description":"Un'app React basata su Transformers.js, in cui è possibile cliccare o passare il mouse sui token generati per vedere i pesi di attenzione dei token storici che hanno influenzato la...","url":"https://www.aioga.com/it/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:30.606Z"},"nl":{"title":"Show HN: Visualisatie van LLM-aandacht","summary":"Een React-applicatie gebaseerd op Transformers.js, waarmee je op gegenereerde tokens kunt klikken of eroverheen kunt hooveren om de aandacht gewichten van historische tokens die hun generatie beïnvloeden te bekijken. De visualisatie wordt bereikt door de amplitude van de schaalwaarden te vergroten en over alle aandachtshoofden en lagen te aggregeren, waardoor het model laat zien hoe het informatie uit de originele tekst kopieert en meerdere zinnen semantisch combineert. De applicatie gebruikt een model van 600 miljoen parameters en genereert vooraf prompts voor onmiddellijke laden.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN: Visualisatie van LLM-aandacht - Aioga AI-nieuws","description":"Een React-applicatie gebaseerd op Transformers.js, waarmee je op gegenereerde tokens kunt klikken of eroverheen kunt hooveren om de aandacht gewichten van historische tokens die hu...","url":"https://www.aioga.com/nl/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:29.689Z"},"tr":{"title":"Show HN: LLM Dikkat Görselleştirmesi","summary":"Transformers.js tabanlı bir React uygulaması; üretilen tokenlere tıklayabilir veya üzerinde durabilirsiniz ve geçmiş tokenlerin üretim üzerindeki dikkat ağırlıklarını görebilirsiniz. Görselleştirme, değer vektörü büyüklüğünün ölçeklendirilmesi ve tüm dikkat başlıkları ve katmanlar boyunca birleştirme yolu ile yapılır; modelin orijinal metinden bilgi kopyalama ve birden fazla ifade anlamını birleştirme sürecini gösterebilir. Uygulama 600 milyon parametreli bir model kullanmakta olup, anında yükleme için önceden oluşturulmuş ipuçları içermektedir.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN: LLM Dikkat Görselleştirmesi - Aioga AI Haberleri","description":"Transformers.js tabanlı bir React uygulaması; üretilen tokenlere tıklayabilir veya üzerinde durabilirsiniz ve geçmiş tokenlerin üretim üzerindeki dikkat ağırlıklarını görebilirsini...","url":"https://www.aioga.com/tr/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:33.269Z"},"vi":{"title":"Show HN: Trực quan hóa chú ý của LLM","summary":"Một ứng dụng React dựa trên Transformers.js, có thể nhấp hoặc di chuột lên các token được tạo ra để xem trọng số chú ý của các token lịch sử ảnh hưởng đến việc tạo token đó. Việc trực quan hóa được thực hiện bằng cách phóng đại vector giá trị và tổng hợp qua tất cả các đầu và lớp chú ý, có thể hiển thị quá trình mô hình sao chép thông tin từ văn bản gốc và kết hợp ngữ nghĩa của nhiều cụm từ. Ứng dụng sử dụng mô hình 600 triệu tham số và đã tiền sinh các từ gợi ý để tải ngay lập tức.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Show HN: Trực quan hóa chú ý của LLM - Tin tức AI Aioga","description":"Một ứng dụng React dựa trên Transformers.js, có thể nhấp hoặc di chuột lên các token được tạo ra để xem trọng số chú ý của các token lịch sử ảnh hưởng đến việc tạo token đó. Việc t...","url":"https://www.aioga.com/vi/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:33.483Z"},"id":{"title":"Tampilkan HN: Visualisasi Perhatian LLM","summary":"Sebuah aplikasi React berbasis Transformers.js, di mana token yang dihasilkan dapat diklik atau diarahkan kursor, untuk melihat bobot perhatian token historis yang memengaruhi pembuatannya. Visualisasi dilakukan dengan memperbesar amplitudo vektor nilai dan menggabungkan di seluruh kepala perhatian dan lapisan, untuk menampilkan proses model menyalin informasi dari teks asli serta menggabungkan makna beberapa frase. Aplikasi ini menggunakan model 600 juta parameter dan praproduksi prompt agar dapat dimuat secara instan.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Tampilkan HN: Visualisasi Perhatian LLM - Berita AI Aioga","description":"Sebuah aplikasi React berbasis Transformers.js, di mana token yang dihasilkan dapat diklik atau diarahkan kursor, untuk melihat bobot perhatian token historis yang memengaruhi pemb...","url":"https://www.aioga.com/id/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:36.020Z"},"th":{"title":"แสดง HN: การมองเห็นความสนใจของ LLM","summary":"แอป React ที่สร้างขึ้นบน Transformers.js สามารถคลิกหรือเลื่อนเมาส์เหนือโทเค็นที่สร้างขึ้นเพื่อดูน้ำหนักความสนใจของโทเค็นประวัติศาสตร์ที่มีผลต่อการสร้างของมัน การมองเห็นทำผ่านการปรับขนาดขนาดเวกเตอร์และรวมกันในทุกหัวข้อและทุกชั้น ทำให้สามารถแสดงกระบวนการที่โมเดลคัดลอกข้อมูลจากข้อความต้นฉบับและรวมความหมายของวลีหลาย ๆ อัน แอปใช้โมเดลขนาด 600 ล้านพารามิเตอร์ และมีการสร้างคำแนะนำล่วงหน้าเพื่อให้โหลดได้ทันที","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"แสดง HN: การมองเห็นความสนใจของ LLM - ข่าว AI Aioga","description":"แอป React ที่สร้างขึ้นบน Transformers.js สามารถคลิกหรือเลื่อนเมาส์เหนือโทเค็นที่สร้างขึ้นเพื่อดูน้ำหนักความสนใจของโทเค็นประวัติศาสตร์ที่มีผลต่อการสร้างของมัน การมองเห็นทำผ่านการปรั...","url":"https://www.aioga.com/th/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:36.925Z"},"pl":{"title":"Pokaż HN: Wizualizacja uwagi LLM","summary":"Aplikacja React oparta na Transformers.js, pozwala kliknąć lub najechać kursorem na wygenerowany token, aby zobaczyć wagi uwagi historycznych tokenów, które wpłynęły na jego wygenerowanie. Wizualizacja jest realizowana poprzez skalowanie wartości wektorów i agregację przez wszystkie głowy uwagi i warstwy, co pozwala pokazać, jak model kopiuje informacje z oryginalnego tekstu i łączy semantykę wielu fraz. Aplikacja używa modelu o 600 milionach parametrów i wstępnie generowanych podpowiedzi do natychmiastowego ładowania.","category":"行业动态","source":"Hacker News 热门（buzzing.cc 中文翻译）","aggregationSource":"Hacker News 热门（buzzing.cc 中文翻译）","pageTitle":"Pokaż HN: Wizualizacja uwagi LLM - Aioga Wiadomości AI","description":"Aplikacja React oparta na Transformers.js, pozwala kliknąć lub najechać kursorem na wygenerowany token, aby zobaczyć wagi uwagi historycznych tokenów, które wpłynęły na jego wygene...","url":"https://www.aioga.com/pl/news/cmttautvn02parofpcrrra5u4/","contentTranslated":true,"sourceHash":"09ef5f2299387a4e","translatedAt":"2026-09-08T23:41:40.938Z"}},"evidenceTier":"verified-news","reviewStatus":"automated-ingest","indexable":true,"editorialCover":"/page-visuals/topic-timeline.png"}}