{"@context":"https://schema.org","@type":"NewsArticle","generatedAt":"2026-07-28T06:20:51.496Z","headline":"Induction Labs 发布 Photon-1：无动作标签视频预训练学会隐式策略","description":"Induction Labs 发布 Photon-1，一个 106B-A5B 参数的稀疏 MoE Transformer，仅通过无动作标签的原始视频预训练，学会了隐式策略。在内部计算机使用基准上，Photon-1 以约 27 倍更少的预训练算力和约 3 倍更低的推理成本击败 Gemini 3.1 Flash-Lite。该模型目前仅作为研究结果发布，无权重、无 API、无许可证。","url":"https://www.aioga.com/news/cms1lgts700liro05jiwzgotq/","mainEntityOfPage":"https://www.aioga.com/news/cms1lgts700liro05jiwzgotq/","datePublished":"2026-07-26T09:14:22.000Z","dateModified":"2026-07-26T09:14:22.000Z","inLanguage":"zh-CN","publisher":{"@type":"NewsMediaOrganization","name":"Aioga","url":"https://www.aioga.com"},"citation":["https://www.marktechpost.com/2026/07/26/induction-labs-photon-1-simulates-desktops-plays-checkers-and-models-billiard-physics-from-one-pretraining-run","https://aihot.virxact.com/items/cms1lgts700liro05jiwzgotq"],"canonicalUrl":"https://www.aioga.com/news/cms1lgts700liro05jiwzgotq/","directAnswer":{"@type":"Answer","text":"Aioga 编辑摘要：Induction Labs 发布 Photon-1，一个 106B-A5B 参数的稀疏 MoE Transformer，仅通过无动作标签的原始视频预训练，学会了隐式策略。 Aioga 将其归入「论文研究」方向，重点关注它对真实使用和行业竞争的影响。","url":"https://www.aioga.com/news/cms1lgts700liro05jiwzgotq/","dateCreated":"2026-07-26T09:14:22.000Z","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":"marktechpost.com source article","url":"https://www.marktechpost.com/2026/07/26/induction-labs-photon-1-simulates-desktops-plays-checkers-and-models-billiard-physics-from-one-pretraining-run","datePublished":"2026-07-26T09:14:22.000Z","provider":{"@type":"Organization","name":"marktechpost.com","url":"https://www.marktechpost.com/2026/07/26/induction-labs-photon-1-simulates-desktops-plays-checkers-and-models-billiard-physics-from-one-pretraining-run"}},{"@type":"CreativeWork","name":"AIHot archive record","url":"https://aihot.virxact.com/items/cms1lgts700liro05jiwzgotq","datePublished":"2026-07-26T09:14:22.000Z","provider":{"@type":"Organization","name":"AIHot","url":"https://aihot.virxact.com/items/cms1lgts700liro05jiwzgotq"}}],"aggregationSource":"MarkTechPost（RSS）","originalPublisher":{"name":"marktechpost.com","url":"https://www.marktechpost.com/2026/07/26/induction-labs-photon-1-simulates-desktops-plays-checkers-and-models-billiard-physics-from-one-pretraining-run"},"article":{"id":"cms1lgts700liro05jiwzgotq","slug":"cms1lgts700liro05jiwzgotq","url":"https://www.aioga.com/news/cms1lgts700liro05jiwzgotq/","title":"Induction Labs 发布 Photon-1：无动作标签视频预训练学会隐式策略","title_en":"Induction Labs Photon-1 Simulates Desktops， Plays Checkers， and Models Billiard Physics From One Pretraining Run","summary":"Induction Labs 发布 Photon-1，一个 106B-A5B 参数的稀疏 MoE Transformer，仅通过无动作标签的原始视频预训练，学会了隐式策略。在内部计算机使用基准上，Photon-1 以约 27 倍更少的预训练算力和约 3 倍更低的推理成本击败 Gemini 3.1 Flash-Lite。该模型目前仅作为研究结果发布，无权重、无 API、无许可证。","source":"MarkTechPost（RSS）","sourceUrl":"https://www.marktechpost.com/2026/07/26/induction-labs-photon-1-simulates-desktops-plays-checkers-and-models-billiard-physics-from-one-pretraining-run","aiHotUrl":"https://aihot.virxact.com/items/cms1lgts700liro05jiwzgotq","publishedAt":"2026-07-26T09:14:22.000Z","category":"论文研究","score":35,"selected":false,"articleBody":["Most agents that learn from video need to know what action produced each frame. Induction Labs：https://www.inductionlabs.com/ is arguing that this requirement is the bottleneck. Last week, they released imagination models , a foundation model architecture that pretrains on raw video with no action labels at all.","Their test system is Photon-1 , a sparse 106B-A5B mixture-of-experts (MoE) transformer trained on 18 years of computer demonstration video. On an internal computer use benchmark, Induction Labs reports that Photon-1 beats Gemini 3.1 Flash-Lite：https://www.inductionlabs.com/news/scaling-video-pretraining while using far less pretraining compute and costing roughly 3× less to serve.","An imagination model predicts future frames autoregressively using a next-latent-token-prediction objective. It does not generate pixels during pretraining. Everything is modeled in a learned representation space.","The claim that matters is this: predicting future states teaches the model to complete tasks, even though it never sees an action during pretraining. Induction Labs calls this an implicit policy . The model learns concepts of what a person is doing, rather than a label for each mouse click.","The architecture depends on a vision encoder that uses finite scalar quantization (FSQ) . Each frame is compressed into 960 discrete tokens. Each token is an 8-dimensional vector. Each dimension takes one of five values: −1, −1/2, 0, 1/2, 1. That gives a codebook of 5⁸ possible codes.","The resulting encoding is about 2.2 KB per frame. Induction Labs reports over 100× better compression than existing OCR and multimodal-model representations, while preserving text, layout and state changes.","To hit that rate, Photon-1 uses a differential latent encoder . It encodes video frames as pairs, so the latents describe differences between frames rather than frame contents.","The corpus starts from an internal index of 2 billion publicly available videos . Filtering reduces that to roughly 2 million computer screen recordings . An internal keyframe detection model strips redundant frames.","The final dataset is 575 million frames , sampled at 1 frame per second. That equals 552 billion tokens, or about 18 years of video. Photon-1 was pretrained from scratch for a single epoch .","Training the 106B-A5B MoE at 32K context took approximately 30,000 H200 GPU-hours , or 4.4×10²² training FLOPs. The research team implemented training in PyTorch with custom fused kernels for the vision encoder and MoE layers, sustaining 40% end-to-end MFU . Those three figures are mutually consistent: 30,000 H200-hours at 40% MFU lands almost exactly on 4.3×10²².","Induction Labs finetuned Photon-1 on fewer than 35,000 computer use trajectories to teach the action and instruction format. Special computer use tokens let the model emit actions. At inference, Photon-1 predicts the next frame’s state first, then outputs the action that gets there.","Online reinforcement learning follows. Rollouts run in real time on virtual machines at scale, and outcomes are verified programmatically to produce reward. The Linux VMs run five desktop environments (LXQt, Xfce, MATE, GNOME and Plasma), each with a Google account for login-restricted web apps and an internal ChatGPT clone with no rate limits.","*Weighted at a 10:1 input-to-output token ratio, which Induction Labs says matches its computer use tests.","Two caveats belong next to that table. First, the Gemini figure is Induction Labs’ own conservative estimate , assuming 8B active parameters and 25T pretraining tokens. Taken at face value the ratio is about 27×, not the 30× headline; Induction Labs states “at least 30×” on the basis that the true Gemini number is likely higher and the model was likely distilled. Second, the benchmark is internal and unreleased, so the result is not independently reproducible today.","Photon-1’s own breakeven cost on Induction Labs’ hardware is $0.06 per 1M input tokens and $0.60 per 1M output tokens , with no speculative decoding.","This is the more interesting test, because Photon-1 saw only computer use video. The research team finetuned it on domains absent from pretraining and compared against two baselines: a vision encoder baseline with the same architecture and size but no imagination pretraining, and an LLM baseline (Ling-flash-2.0：https://huggingface.co/inclusionAI/Ling-flash-2.0 from Inclusion AI, pretrained on 20T tokens).","On 20,000 tournament checkers games from the Open Checkers Archive 2.0：https://fierz.ch/download.php/checkers.htm, Photon-1 beat both baselines on world simulation and on move quality. On 10,000 synthetically generated billiard games simulated at 5 fps, it produced a mean absolute error of 0.47 against the ground-truth physics engine, versus 1.15 for the LLM baseline and 1.44 for the vision encoder baseline.","Photon-1 also picked up human priors from the pretraining video. After RL, it learned to use the in-VM ChatGPT clone to draft artifacts and answer knowledge questions, steering the LLM the way a person would.","Check out the full technical writeup from Induction Labs：https://www.inductionlabs.com/news/scaling-video-pretraining and the announcement thread on X：https://x.com/induction_labs/status/2080322704973160760. All credit for this research goes to the researchers of this project.","Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.","Build an Agentic Event Venue Operator [Full Codes]：https://pxllnk.co/twdn5","Thanks! Our team will contact you soon 🙌"],"articleImages":[{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2025/07/a-professional-linkedin-headshot-photogr_0jcmb0R9Sv6nW5XK-zkPHw_uARV5VW1ST6osLNlunoVWg-300x300.png","alt":"","afterParagraph":18,"url":"/media/articles/cms1lgts700liro05jiwzgotq/84e64b03066de40c.webp"},{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2026/07/blog6171-4-100x70.png","alt":"FAIRChem v2 UMA for Multidomain Atomistic Simulation across Molecules, Catalysts, Materials, Vibrations, and Molecular Dynamics","afterParagraph":19,"url":"/media/articles/cms1lgts700liro05jiwzgotq/a8de8cc0d3522061.png"},{"sourceUrl":"https://www.marktechpost.com/wp-content/uploads/2026/07/blog6171-1-100x70.png","alt":"Designing High-Performance GPU Kernels with TileLang: Tensor-Core GEMM, Fused Softmax, FlashAttention, and Autotuning","afterParagraph":19,"url":"/media/articles/cms1lgts700liro05jiwzgotq/80925b48ffa65ad2.webp"}],"mediaStatus":"ok","articleBodyZh":["大多数从视频中学习的智能体需要知道每一帧产生的动作。Induction Labs（https://www.inductionlabs.com/）认为这一要求是瓶颈。上周，他们发布了想象模型，一种基础模型架构，可以在完全没有动作标签的原始视频上进行预训练。","他们的测试系统是 Photon-1，一种稀疏 106B-A5B 专家混合（MoE）变换器，在 18 年的计算机演示视频上训练。在内部计算机使用基准测试中，Induction Labs 报告称 Photon-1 在预训练计算量远低于 Gemini 3.1 Flash-Lite（https://www.inductionlabs.com/news/scaling-video-pretraining）的情况下表现更佳，服务成本大约低 3 倍。","想象模型使用下一潜变量标记预测目标自回归地预测未来帧。在预训练期间它不生成像素。一切都在学习到的表示空间中建模。","关键声明是：预测未来状态会教模型完成任务，即使预训练期间从未看到动作。Induction Labs 将其称为隐式策略。模型学习的是人的行为概念，而不是每次鼠标点击的标签。","该架构依赖于使用有限标量量化（FSQ）的视觉编码器。每帧压缩为 960 个离散标记。每个标记是 8 维向量，每个维度取五个值之一：−1, −1/2, 0, 1/2, 1。这给出了 5⁸ 个可能代码的码本。","生成的编码大约每帧 2.2 KB。Induction Labs 报告称与现有的 OCR 和多模态模型表示相比，压缩效果提高超过 100 倍，同时保留文本、布局和状态变化。","为了达到这一速率，Photon-1 使用差分潜变量编码器。它将视频帧编码为对，因此潜变量描述的是帧之间的差异，而不是帧的内容。","语料库起始于内部索引的 20 亿个公开可用视频。筛选后减少到约 200 万个计算机屏幕录制。内部关键帧检测模型去除了冗余帧。","最终数据集为 5.75 亿帧，采样率为每秒 1 帧。这相当于 5520 亿个标记，约 18 年的视频。Photon-1 从零开始预训练了一个单一的 epoch。","在 32K 上下文下训练 106B-A5B MoE 大约消耗了 30,000 H200 GPU 小时，或 4.4×10²² 次训练 FLOPs。研究团队在 PyTorch 中实现了训练，并针对视觉编码器和 MoE 层使用了自定义融合内核，保持了 40% 的端到端 MFU。这三个数字是相互一致的：30,000 H200 小时在 40% MFU 下几乎正好对应 4.3×10²² 次。","Induction Labs 在不到 35,000 次计算机使用轨迹上微调了 Photon-1，以教模型动作和指令格式。特殊的计算机使用标记允许模型输出动作。在推理时，Photon-1 首先预测下一帧的状态，然后输出可以达到该状态的动作。","随后是在线强化学习。Rollout 在规模化的虚拟机上实时运行，结果通过编程方式验证以生成奖励。Linux 虚拟机运行五个桌面环境（LXQt、Xfce、MATE、GNOME 和 Plasma），每个环境都有一个 Google 帐号用于登录受限的网络应用，以及一个内部 ChatGPT 克隆，没有速率限制。","*按 10:1 的输入到输出标记比加权，Induction Labs 表示这与其计算机使用测试相匹配。","该表旁应有两个注意事项。首先，Gemini 的数字是 Induction Labs 自己的保守估计，假设 8B 活跃参数和 25T 预训练标记。按表面价值计算，比率约为 27×，而非 30× 头条；Induction Labs 表示“至少 30×”，基于 Gemini 的真实数字可能更高，并且模型可能经过蒸馏。其次，该基准是内部未发布的，因此结果目前无法独立复现。","Photon-1 在 Induction Labs 硬件上的自身盈亏成本是每 1M 输入标记 $0.06，每 1M 输出标记 $0.60，没有投机性解码。","这是更有趣的测试，因为 Photon-1 只看过计算机使用视频。研究团队在预训练中没有涉及的领域上微调了它，并与两个基线进行了比较：一个具有相同架构和规模但没有想象力预训练的视觉编码器基线，和一个 LLM 基线 (Ling-flash-2.0：https://huggingface.co/inclusionAI/Ling-flash-2.0，由 Inclusion AI 提供，预训练 20T 标记)。","在来自开放跳棋档案 2.0（Open Checkers Archive 2.0：https://fierz.ch/download.php/checkers.htm）的 20,000 场锦标赛跳棋游戏中，Photon-1 在世界模拟和移动质量方面都击败了两个基线。在以 5 帧/秒模拟的 10,000 场合成台球游戏中，它相对于真实物理引擎的平均绝对误差为 0.47，而 LLM 基线为 1.15，视觉编码器基线为 1.44。","Photon-1 还从预训练视频中学习到了人类先验知识。经过强化学习（RL）后，它学会了使用虚拟机内的 ChatGPT 克隆来起草工件和回答知识问题，以类似人的方式引导 LLM。","请查看 Induction Labs 的完整技术报告：https://www.inductionlabs.com/news/scaling-video-pretraining 以及 X 上的公告帖：https://x.com/induction_labs/status/2080322704973160760。对此研究的所有功劳归属于该项目的研究人员。","Michal Sutter 是一名数据科学专业人员，拥有帕多瓦大学（University of Padova）的数据科学硕士学位。凭借在统计分析、机器学习和数据工程方面的坚实基础，Michal 擅长将复杂的数据集转化为可操作的洞察。","构建一个智能事件场地运营代理 [完整代码]：https://pxllnk.co/twdn5","谢谢！我们的团队将很快与您联系 🙌"],"translationStatus":"translated","bodyOrigin":"source-page","editorial":{"summary":"Aioga 编辑摘要：Induction Labs 发布 Photon-1，一个 106B-A5B 参数的稀疏 MoE Transformer，仅通过无动作标签的原始视频预训练，学会了隐式策略。 Aioga 将其归入「论文研究」方向，重点关注它对真实使用和行业竞争的影响。","background":"背景分析：模型与研究类动态需要结合能力边界、开放方式、成本、可用性和真实任务表现判断，单项指标领先不等于已经形成稳定采用。","viewpoint":"Aioga 判断：这条动态更适合作为行业观察信号，当前信息足以建立线索，但不足以推导长期结论。","implications":"影响分析：对相关团队而言，短期应先核对来源、可用范围和实际成本，再判断是否值得接入或跟进。","nextStep":"后续观察：继续观察官方文档、实际可用性、价格变化、开发者反馈和竞品回应。","evidenceRefs":["title","summary","articleBody"],"confidence":"medium","status":"published","aiGenerated":false,"autoApproved":true,"generatedBy":"rule-safe-fallback","generatedAt":"2026-07-28T06:29:10.489Z","sourceHash":"ae00a6b573b78c0f","validation":{"passed":true,"mode":"rule-safe-fallback","checks":["schema","length","source-attribution","no-html"]}},"tags":["论文研究","MarkTechPost（RSS）"],"translations":{"zh-CN":{"title":"Induction Labs 发布 Photon-1：无动作标签视频预训练学会隐式策略","summary":"Induction Labs 发布 Photon-1，一个 106B-A5B 参数的稀疏 MoE Transformer，仅通过无动作标签的原始视频预训练，学会了隐式策略。在内部计算机使用基准上，Photon-1 以约 27 倍更少的预训练算力和约 3 倍更低的推理成本击败 Gemini 3.1 Flash-Lite。该模型目前仅作为研究结果发布，无权重、无 API、无许可证。","category":"论文研究","source":"marktechpost.com","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs 发布 Photon-1：无动作标签视频预训练学会隐式策略 - Aioga AI资讯","description":"Induction Labs 发布 Photon-1，一个 106B-A5B 参数的稀疏 MoE Transformer，仅通过无动作标签的原始视频预训练，学会了隐式策略。在内部计算机使用基准上，Photon-1 以约 27 倍更少的预训练算力和约 3 倍更低的推理成本击败 Gemini 3.1 Flash-Lite。该模型目前仅作为研究结果发布，无权重、无...","url":"https://www.aioga.com/news/cms1lgts700liro05jiwzgotq/"},"en":{"title":"Induction Labs Releases Photon-1: Action-Free Label Video Pretraining Learns Implicit Policies","summary":"Induction Labs released Photon-1, a 106B-A5B parameter sparse MoE Transformer, which learned implicit strategies solely through pretraining on raw videos without action labels. On internal computer usage benchmarks, Photon-1 outperformed Gemini 3.1 Flash-Lite with approximately 27 times less pretraining compute and about 3 times lower inference cost. The model is currently released only as a research result, with no weights, no API, and no license.","category":"Research","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs Releases Photon-1: Action-Free Label Video Pretraining Learns Implicit Policies - Aioga AI News","description":"Induction Labs released Photon-1, a 106B-A5B parameter sparse MoE Transformer, which learned implicit strategies solely through pretraining on raw videos without action labels. On...","url":"https://www.aioga.com/en/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:42:35.234Z"},"ja":{"title":"Induction Labs、Photon-1を発表：動作ラベルなしの動画事前学習で暗黙の戦略を習得","summary":"Induction Labs は Photon-1 を発表しました。これは 106B-A5B パラメータのスパース MoE トランスフォーマーで、アクションラベルなしの生の動画で事前学習するだけで、暗黙の戦略を学習しました。社内のコンピュータベンチマークでは、Photon-1 は Gemini 3.1 Flash-Lite を約 27 倍少ない事前学習計算量と約 3 倍低い推論コストで上回りました。このモデルは現在、研究成果としてのみ公開されており、重み、API、ライセンスは提供されていません。","category":"論文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs、Photon-1を発表：動作ラベルなしの動画事前学習で暗黙の戦略を習得 - Aioga AIニュース","description":"Induction Labs は Photon-1 を発表しました。これは 106B-A5B パラメータのスパース MoE トランスフォーマーで、アクションラベルなしの生の動画で事前学習するだけで、暗黙の戦略を学習しました。社内のコンピュータベンチマークでは、Photon-1 は Gemini 3.1 Flash-Lite を約 27 倍少ない事前学習計算量...","url":"https://www.aioga.com/ja/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:42:43.094Z"},"ko":{"title":"Induction Labs, Photon-1 출시: 동작 레이블 없는 비디오 사전 학습으로 암묵적 정책 학습","summary":"Induction Labs가 Photon-1을 발표했습니다. 이 모델은 106B-A5B 파라미터의 희소 MoE Transformer로, 동작 라벨이 없는 원시 비디오만으로 사전 학습을 통해 암묵적인 전략을 학습했습니다. 내부 컴퓨터 벤치마크에서 Photon-1은 약 27배 적은 사전 학습 연산과 약 3배 낮은 추론 비용으로 Gemini 3.1 Flash-Lite를 능가했습니다. 이 모델은 현재 연구 결과로서만 공개되며, 가중치, API, 라이선스는 제공되지 않습니다.","category":"연구","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs, Photon-1 출시: 동작 레이블 없는 비디오 사전 학습으로 암묵적 정책 학습 - Aioga AI 뉴스","description":"Induction Labs가 Photon-1을 발표했습니다. 이 모델은 106B-A5B 파라미터의 희소 MoE Transformer로, 동작 라벨이 없는 원시 비디오만으로 사전 학습을 통해 암묵적인 전략을 학습했습니다. 내부 컴퓨터 벤치마크에서 Photon-1은 약 27배 적은 사전 학습 연산과 약 3배 낮은 추론 비용으...","url":"https://www.aioga.com/ko/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:43:24.473Z"},"es":{"title":"Induction Labs lanza Photon-1: aprendizaje de estrategias implícitas mediante preentrenamiento de video sin etiquetas de acción","summary":"Induction Labs lanzó Photon-1, un Transformer MoE disperso con 106B-A5B parámetros, que aprendió estrategias implícitas mediante preentrenamiento únicamente con videos crudos sin etiquetas de acción. En los benchmarks internos de computación, Photon-1 superó a Gemini 3.1 Flash-Lite usando aproximadamente 27 veces menos capacidad de preentrenamiento y alrededor de 3 veces menos costo de inferencia. Actualmente, el modelo solo se publica como un resultado de investigación, sin pesos, sin API y sin licencia.","category":"Investigación","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs lanza Photon-1: aprendizaje de estrategias implícitas mediante preentrenamiento de video sin etiquetas de acción - Aioga Noticias de IA","description":"Induction Labs lanzó Photon-1, un Transformer MoE disperso con 106B-A5B parámetros, que aprendió estrategias implícitas mediante preentrenamiento únicamente con videos crudos sin e...","url":"https://www.aioga.com/es/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:43:19.111Z"},"fr":{"title":"Induction Labs publie Photon-1 : pré-formation sur vidéos sans étiquettes d'action pour apprendre des stratégies implicites","summary":"Induction Labs a publié Photon-1, un Transformer MoE épars de 106 milliards de paramètres A5B, qui a appris des stratégies implicites uniquement par pré-entraînement sur des vidéos brutes sans étiquettes d'action. Sur des benchmarks internes sur ordinateur, Photon-1 a surpassé Gemini 3.1 Flash-Lite avec environ 27 fois moins de puissance de calcul pour le pré-entraînement et environ 3 fois moins de coût d'inférence. Ce modèle est actuellement publié uniquement à titre de résultat de recherche, sans poids, sans API et sans licence.","category":"Recherche","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs publie Photon-1 : pré-formation sur vidéos sans étiquettes d'action pour apprendre des stratégies implicites - Aioga Actualités IA","description":"Induction Labs a publié Photon-1, un Transformer MoE épars de 106 milliards de paramètres A5B, qui a appris des stratégies implicites uniquement par pré-entraînement sur des vidéos...","url":"https://www.aioga.com/fr/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:44:02.802Z"},"de":{"title":"Induction Labs veröffentlicht Photon-1: Vortrainiertes Video ohne Aktionsetiketten erlernt implizite Strategien","summary":"Induction Labs hat Photon-1 veröffentlicht, einen sparsamen MoE-Transformer mit 106B-A5B Parametern, der nur durch unbeschriftete Rohvideos vortrainiert wurde und implizite Strategien erlernt hat. Im internen Computernutzungs-Benchmark übertrifft Photon-1 Gemini 3.1 Flash-Lite mit etwa 27-mal weniger Vortrainingsrechenleistung und etwa 3-mal niedrigeren Inferenzkosten. Das Modell wird derzeit nur als Forschungsergebnis veröffentlicht, ohne Gewichte, ohne API und ohne Lizenz.","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs veröffentlicht Photon-1: Vortrainiertes Video ohne Aktionsetiketten erlernt implizite Strategien - Aioga KI-News","description":"Induction Labs hat Photon-1 veröffentlicht, einen sparsamen MoE-Transformer mit 106B-A5B Parametern, der nur durch unbeschriftete Rohvideos vortrainiert wurde und implizite Strateg...","url":"https://www.aioga.com/de/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:43:58.689Z"},"pt-BR":{"title":"Induction Labs lança Photon-1: pré-treinamento de vídeo sem rótulo de ação aprende estratégias implícitas","summary":"A Induction Labs lançou o Photon-1, um Transformer MoE esparso com 106B-A5B parâmetros, que aprendeu estratégias implícitas apenas por pré-treinamento com vídeos brutos sem rótulos de ação. Em benchmarks internos de computadores, o Photon-1 superou o Gemini 3.1 Flash-Lite com aproximadamente 27 vezes menos poder de computação de pré-treinamento e cerca de 3 vezes menor custo de inferência. O modelo atualmente é lançado apenas como resultado de pesquisa, sem pesos, sem API e sem licença.","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs lança Photon-1: pré-treinamento de vídeo sem rótulo de ação aprende estratégias implícitas - Aioga Notícias de IA","description":"A Induction Labs lançou o Photon-1, um Transformer MoE esparso com 106B-A5B parâmetros, que aprendeu estratégias implícitas apenas por pré-treinamento com vídeos brutos sem rótulos...","url":"https://www.aioga.com/pt-BR/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:44:37.982Z"},"ru":{"title":"Induction Labs выпустила Photon-1: предварительное обучение видео без меток действий для освоения неявной стратегии","summary":"Induction Labs выпустила Photon-1, разреженный MoE Transformer с 106B-A5B параметрами, который обучился скрытой стратегии исключительно на необработанных видео без меток действий. По внутренним вычислительным бенчмаркам Photon-1 обошел Gemini 3.1 Flash-Lite, используя примерно в 27 раз меньше вычислительных ресурсов при предобучении и примерно в 3 раза меньшие затраты на вывод. В настоящее время модель опубликована только как исследовательский результат, без весов, API и лицензии.","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs выпустила Photon-1: предварительное обучение видео без меток действий для освоения неявной стратегии - Aioga Новости ИИ","description":"Induction Labs выпустила Photon-1, разреженный MoE Transformer с 106B-A5B параметрами, который обучился скрытой стратегии исключительно на необработанных видео без меток действий....","url":"https://www.aioga.com/ru/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:44:51.368Z"},"ar":{"title":"Induction Labs تصدر Photon-1: التدريب المسبق لمقاطع الفيديو بدون علامات الحركة لتعلم الاستراتيجيات الضمنية","summary":"أعلنت Induction Labs عن Photon-1، وهو محول MoE متناثر بمعاملات 106B-A5B، وقد تعلم استراتيجيات ضمنية فقط من خلال التدريب المسبق على مقاطع الفيديو الخام بدون تسميات للحركة. في الاختبارات الداخلية باستخدام أجهزة الكمبيوتر القياسية، تغلب Photon-1 على Gemini 3.1 Flash-Lite باستخدام قوة تدريب مسبق أقل بحوالي 27 مرة وتكلفة استنتاج أقل بحوالي 3 مرات. النموذج متاح حاليًا فقط كنتيجة بحثية، بدون أوزان، بدون API، وبدون ترخيص.","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs تصدر Photon-1: التدريب المسبق لمقاطع الفيديو بدون علامات الحركة لتعلم الاستراتيجيات الضمنية - Aioga أخبار الذكاء الاصطناعي","description":"أعلنت Induction Labs عن Photon-1، وهو محول MoE متناثر بمعاملات 106B-A5B، وقد تعلم استراتيجيات ضمنية فقط من خلال التدريب المسبق على مقاطع الفيديو الخام بدون تسميات للحركة. في الاختب...","url":"https://www.aioga.com/ar/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:45:37.096Z"},"hi":{"title":"Induction Labs ने Photon-1 जारी किया: बिना कार्रवाई लेबल वाले वीडियो पूर्व-प्रशिक्षण से निहित नीति सीखना","summary":"Induction Labs ने Photon-1 जारी किया, एक 106B-A5B पैरामीटर वाला विरल MoE Transformer, जिसे केवल बिना क्रियात्मक टैग वाले कच्चे वीडियो के प्रीट्रेनिंग के माध्यम से, अंतर्निहित रणनीतियाँ सीखने के लिए प्रशिक्षित किया गया। आंतरिक कंप्यूटर बेंचमार्क पर, Photon-1 ने लगभग 27 गुना कम प्रीट्रेनिंग कंप्यूटिंग पावर और लगभग 3 गुना कम अनुमान लागत के साथ Gemini 3.1 Flash-Lite को हराया। यह मॉडल वर्तमान में केवल शोध परिणामों के रूप में जारी किया गया है, इसके कोई वेट, कोई API, कोई लाइसेंस नहीं है।","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs ने Photon-1 जारी किया: बिना कार्रवाई लेबल वाले वीडियो पूर्व-प्रशिक्षण से निहित नीति सीखना - Aioga AI समाचार","description":"Induction Labs ने Photon-1 जारी किया, एक 106B-A5B पैरामीटर वाला विरल MoE Transformer, जिसे केवल बिना क्रियात्मक टैग वाले कच्चे वीडियो के प्रीट्रेनिंग के माध्यम से, अंतर्निहित रणनीत...","url":"https://www.aioga.com/hi/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:45:30.531Z"},"it":{"title":"Induction Labs rilascia Photon-1: pre-addestramento di video con etichette senza azione per imparare strategie implicite","summary":"Induction Labs ha pubblicato Photon-1, un Transformer MoE sparso con 106 miliardi di parametri A5B, che ha appreso strategie implicite tramite pre-addestramento su video grezzi senza etichette di azione. Nei benchmark interni sui computer, Photon-1 ha superato Gemini 3.1 Flash-Lite con circa 27 volte meno capacità di calcolo per il pre-addestramento e circa 3 volte meno costo di inferenza. Attualmente il modello è stato rilasciato solo come risultato di ricerca, senza pesi, API o licenza.","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs rilascia Photon-1: pre-addestramento di video con etichette senza azione per imparare strategie implicite - Aioga Notizie IA","description":"Induction Labs ha pubblicato Photon-1, un Transformer MoE sparso con 106 miliardi di parametri A5B, che ha appreso strategie implicite tramite pre-addestramento su video grezzi sen...","url":"https://www.aioga.com/it/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:46:15.933Z"},"nl":{"title":"Induction Labs lanceert Photon-1: video-voortraining zonder actielabels leert impliciete strategieën","summary":"Induction Labs heeft Photon-1 uitgebracht, een sparsely MoE Transformer met 106B-A5B parameters, die alleen door middel van ongeëtiketteerde ruwe videogegevens is voorgetraind en impliciete strategieën heeft geleerd. Op interne computerbenchmarks verslaat Photon-1 Gemini 3.1 Flash-Lite met ongeveer 27 keer minder pretrainingscomputerkracht en ongeveer 3 keer lagere inferentiekosten. Het model is momenteel alleen als onderzoeksresultaat beschikbaar en heeft geen gewichten, geen API en geen licentie.","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs lanceert Photon-1: video-voortraining zonder actielabels leert impliciete strategieën - Aioga AI-nieuws","description":"Induction Labs heeft Photon-1 uitgebracht, een sparsely MoE Transformer met 106B-A5B parameters, die alleen door middel van ongeëtiketteerde ruwe videogegevens is voorgetraind en i...","url":"https://www.aioga.com/nl/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:46:15.406Z"},"tr":{"title":"Induction Labs, Photon-1'i yayınladı: Aksiyon etiketli olmayan videolarla ön eğitim ile örtük stratejiler öğrenme","summary":"Induction Labs, yalnızca etiketsiz ham video verileriyle ön eğitim yaparak örtük stratejiler öğrenen, 106B-A5B parametreli seyrek MoE Transformer olan Photon-1'i duyurdu. Dahili bilgisayar kullanım ölçütlerinde, Photon-1, Gemini 3.1 Flash-Lite'ı yaklaşık 27 kat daha az ön eğitim hesaplama gücü ve yaklaşık 3 kat daha düşük çıkarım maliyetiyle geride bırakıyor. Model şu anda yalnızca bir araştırma sonucu olarak yayınlanmış olup, ağırlık, API veya lisans içermemektedir.","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs, Photon-1'i yayınladı: Aksiyon etiketli olmayan videolarla ön eğitim ile örtük stratejiler öğrenme - Aioga AI Haberleri","description":"Induction Labs, yalnızca etiketsiz ham video verileriyle ön eğitim yaparak örtük stratejiler öğrenen, 106B-A5B parametreli seyrek MoE Transformer olan Photon-1'i duyurdu. Dahili bi...","url":"https://www.aioga.com/tr/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:46:56.828Z"},"vi":{"title":"Induction Labs ra mắt Photon-1: Huấn luyện trước video không nhãn hành động học chiến lược ẩn","summary":"Induction Labs phát hành Photon-1, một Transformer MoE thưa thớt với 106B-A5B tham số, chỉ được huấn luyện trước bằng video thô không nhãn hành động và đã học được chiến lược ngầm. Trên các chuẩn so sánh nội bộ của máy tính, Photon-1 đã đánh bại Gemini 3.1 Flash-Lite với khoảng 27 lần ít tài nguyên huấn luyện trước và khoảng 3 lần chi phí suy luận thấp hơn. Mô hình này hiện chỉ được phát hành như kết quả nghiên cứu, không có trọng số, không có API, không có giấy phép.","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs ra mắt Photon-1: Huấn luyện trước video không nhãn hành động học chiến lược ẩn - Tin tức AI Aioga","description":"Induction Labs phát hành Photon-1, một Transformer MoE thưa thớt với 106B-A5B tham số, chỉ được huấn luyện trước bằng video thô không nhãn hành động và đã học được chiến lược ngầm....","url":"https://www.aioga.com/vi/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:47:08.459Z"},"id":{"title":"Induction Labs merilis Photon-1: Pelatihan awal video tanpa label tindakan mempelajari strategi implisit","summary":"Induction Labs merilis Photon-1, sebuah Transformer MoE jarang dengan 106B-A5B parameter, yang hanya melalui pra-pelatihan video mentah tanpa label tindakan, belajar strategi implisit. Di tolok ukur internal komputer, Photon-1 mengalahkan Gemini 3.1 Flash-Lite dengan sekitar 27 kali lebih sedikit daya komputasi pra-pelatihan dan sekitar 3 kali biaya inferensi lebih rendah. Model ini saat ini hanya dirilis sebagai hasil penelitian, tanpa bobot, tanpa API, tanpa lisensi.","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs merilis Photon-1: Pelatihan awal video tanpa label tindakan mempelajari strategi implisit - Berita AI Aioga","description":"Induction Labs merilis Photon-1, sebuah Transformer MoE jarang dengan 106B-A5B parameter, yang hanya melalui pra-pelatihan video mentah tanpa label tindakan, belajar strategi impli...","url":"https://www.aioga.com/id/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:47:46.543Z"},"th":{"title":"Induction Labs เปิดตัว Photon-1: การฝึกวิดีโอแบบไม่มีป้ายกำกับเรียนรู้กลยุทธ์แฝง","summary":"Induction Labs เปิดตัว Photon-1 ซึ่งเป็น Transformer แบบ MoE ที่มีพารามิเตอร์ 106B-A5B แบบบาง ๆ เรียนรู้กลยุทธ์โดยนัยจากการพรีเทรนด้วยวิดีโอดิบโดยไม่มีป้ายกำกับการกระทำ ในมาตรฐานการใช้งานคอมพิวเตอร์ภายใน Photon-1 เอาชนะ Gemini 3.1 Flash-Lite ด้วยกำลังคำนวณพรีเทรนที่น้อยกว่าประมาณ 27 เท่า และต้นทุนการอนุมานต่ำกว่าประมาณ 3 เท่า โมเดลนี้ปัจจุบันเผยแพร่อย่างเดียวในฐานะผลงานวิจัย ไม่มีน้ำหนักโมเดล ไม่มี API และไม่มีใบอนุญาต","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs เปิดตัว Photon-1: การฝึกวิดีโอแบบไม่มีป้ายกำกับเรียนรู้กลยุทธ์แฝง - ข่าว AI Aioga","description":"Induction Labs เปิดตัว Photon-1 ซึ่งเป็น Transformer แบบ MoE ที่มีพารามิเตอร์ 106B-A5B แบบบาง ๆ เรียนรู้กลยุทธ์โดยนัยจากการพรีเทรนด้วยวิดีโอดิบโดยไม่มีป้ายกำกับการกระทำ ในมาตรฐานกา...","url":"https://www.aioga.com/th/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:47:55.944Z"},"pl":{"title":"Induction Labs wydaje Photon-1: wstępne uczenie wideo bez etykiet działań uczy ukrytych strategii","summary":"Induction Labs wydało Photon-1, rzadki transformator MoE o parametrach 106B-A5B, który nauczył się ukrytej strategii wyłącznie poprzez wstępne szkolenie na surowych wideo bez etykiet akcji. W wewnętrznych testach komputerowych Photon-1 pokonał Gemini 3.1 Flash-Lite przy około 27 razy mniejszej mocy obliczeniowej wstępnego szkolenia i około 3 razy niższym koszcie wnioskowania. Model jest obecnie udostępniany jedynie jako wynik badań, bez wag, bez API i bez licencji.","category":"论文研究","source":"MarkTechPost（RSS）","aggregationSource":"MarkTechPost（RSS）","pageTitle":"Induction Labs wydaje Photon-1: wstępne uczenie wideo bez etykiet działań uczy ukrytych strategii - Aioga Wiadomości AI","description":"Induction Labs wydało Photon-1, rzadki transformator MoE o parametrach 106B-A5B, który nauczył się ukrytej strategii wyłącznie poprzez wstępne szkolenie na surowych wideo bez etyki...","url":"https://www.aioga.com/pl/news/cms1lgts700liro05jiwzgotq/","contentTranslated":true,"sourceHash":"5e858677a83d370d","translatedAt":"2026-07-28T04:48:37.992Z"}}}}