AI News AI资讯 14h ago Updated 12h ago 更新于 12小时前 50

Induction Labs Photon-1 Simulates Desktops, Plays Checkers, and Models Billiard Physics From One Pretraining Run Induction Labs Photon-1 从一次预训练中模拟桌面、玩跳棋并建模台球物理

Induction Labs introduces "imagination models," a foundation architecture that learns implicit policies from raw video without action labels by predicting future latent states. The system, Photon-1, utilizes Finite Scalar Quantization (FSQ) to compress frames into 960 discrete tokens, achieving over 100x better compression than existing multimodal representations. Trained on 575 million frames (18 years of video) with only 30,000 H200 GPU-hours, Photon-1 outperforms Gemini 3.1 Flash-Lite on inte Induction Labs发布Photon-1,一种无需动作标签即可从原始视频预训练的“想象模型”,通过预测未来状态学习隐式策略。 采用有限标量量化(FSQ)和差分潜在编码器,将每帧压缩至约2.2KB,实现比现有OCR/多模态表示高100倍的压缩率。 在30,000 H200 GPU-hours的训练成本下,Photon-1在内部计算机使用基准测试中击败Gemini 3.1 Flash-Lite,预训练算力消耗仅为后者的约1/27。 尽管仅使用桌面屏幕录像预训练,微调后在跳棋模拟和台球物理引擎预测中均优于LLM基线,展现出跨领域泛化能力。 该模型目前仅为研究成果,未开源权重、API或许可证,

72
Hot 热度
68
Quality 质量
75
Impact 影响力

Analysis 深度分析

TL;DR

  • Induction Labs introduces "imagination models," a foundation architecture that learns implicit policies from raw video without action labels by predicting future latent states.
  • The system, Photon-1, utilizes Finite Scalar Quantization (FSQ) to compress frames into 960 discrete tokens, achieving over 100x better compression than existing multimodal representations.
  • Trained on 575 million frames (18 years of video) with only 30,000 H200 GPU-hours, Photon-1 outperforms Gemini 3.1 Flash-Lite on internal computer use benchmarks at significantly lower cost and compute.
  • Despite being pretrained exclusively on desktop interactions, the model demonstrates strong generalization to physical simulations like checkers and billiard physics after minimal fine-tuning.
  • The model employs an autoregressive next-latent-token prediction objective, allowing it to infer task completion concepts without explicit supervision of mouse clicks or keyboard inputs.

Why It Matters

This research challenges the prevailing paradigm that high-performance AI agents require massive datasets of labeled human actions (supervised learning). By demonstrating that predicting future visual states is sufficient for learning complex behaviors, Induction Labs offers a scalable path to developing agents that can generalize across domains without the bottleneck of manual annotation. For practitioners, this highlights the potential of self-supervised world models to reduce reliance on expensive, curated interaction data while improving efficiency and cost-effectiveness in deployment.

Technical Details

  • Architecture & Objective: Photon-1 is a sparse 106B-A5B mixture-of-experts (MoE) transformer trained using a next-latent-token-prediction objective. It does not generate pixels during pretraining but operates entirely within a learned representation space to predict future states autoregressively.
  • Compression Technique: The vision encoder uses Finite Scalar Quantization (FSQ) with a differential latent encoder. Frames are encoded as pairs to capture differences, resulting in 960 tokens per frame (each an 8-dimensional vector with values from {-1, -1/2, 0, 1/2, 1}). This yields ~2.2 KB per frame, preserving text and layout with >100x better compression than OCR-based methods.
  • Training Data & Compute: The model was pretrained from scratch on a single epoch of 575 million frames sampled from 2 million filtered computer screen recordings. Training required approximately 30,000 H200 GPU-hours (4.4×10²² FLOPs) with a 40% end-to-end Model Flops Utilization (MFU).
  • Inference & Fine-Tuning: Post-pretraining, the model was fine-tuned on fewer than 35,000 computer use trajectories to learn action formats. Inference involves predicting the next frame's state first, then emitting the corresponding action. Online reinforcement learning is applied via rollouts in Linux VMs with programmatic reward verification.

Industry Insight

  • Shift to Self-Supervised World Models: The industry should prioritize research into unsupervised or self-supervised video prediction as a viable alternative to supervised imitation learning. This approach could drastically reduce the cost and time required to train robust agents for new environments.
  • Efficiency in Agent Deployment: The significant reduction in pretraining compute (approx. 27x less than comparable models) and inference costs suggests that future AI agents can be deployed more economically. Practitioners should evaluate FSQ and similar compression techniques to optimize token usage in vision-language-action pipelines.
  • Generalization Potential: The ability of a desktop-trained model to simulate physical laws (billiards) and game logic (checkers) indicates that "implicit policies" captured from human video contain rich priors about causality and physics. This implies that agents trained on diverse observational data may require less domain-specific fine-tuning than previously thought.

TL;DR

  • Induction Labs发布Photon-1,一种无需动作标签即可从原始视频预训练的“想象模型”,通过预测未来状态学习隐式策略。
  • 采用有限标量量化(FSQ)和差分潜在编码器,将每帧压缩至约2.2KB,实现比现有OCR/多模态表示高100倍的压缩率。
  • 在30,000 H200 GPU-hours的训练成本下,Photon-1在内部计算机使用基准测试中击败Gemini 3.1 Flash-Lite,预训练算力消耗仅为后者的约1/27。
  • 尽管仅使用桌面屏幕录像预训练,微调后在跳棋模拟和台球物理引擎预测中均优于LLM基线,展现出跨领域泛化能力。
  • 该模型目前仅为研究成果,未开源权重、API或许可证,且性能对比基于内部未公开基准,需保持审慎态度。

为什么值得看

本文展示了一种突破性的视频理解范式,证明模型无需显式动作标注即可从纯视觉数据中学习复杂的任务规划和世界模拟能力,大幅降低了多模态智能体的数据获取门槛。对于AI从业者而言,其高效的压缩技术和低算力预训练路径为构建低成本、高能效的具身智能或自动化代理提供了极具参考价值的新思路。

技术解析

  • 核心架构与目标:Photon-1是一个稀疏混合专家(MoE)Transformer(106B总参数,5B激活参数),采用自回归的“下一潜在令牌预测”目标。它在预训练阶段不生成像素,而是在学习到的表示空间中预测未来帧的状态,从而形成对任务执行的“隐式策略”。
  • 高效视觉编码:利用有限标量量化(FSQ)技术,每个视频帧被压缩为960个离散令牌(8维向量,每维5个取值之一)。结合差分潜在编码器(编码帧间差异而非内容),实现了极高的压缩效率(约2.2 KB/帧),同时保留文本、布局和状态变化细节。
  • 数据与训练规模:预训练数据源自20亿公开视频,经筛选和关键帧检测后得到5.75亿帧(约5520亿令牌,相当于18年视频)。模型从头预训练一个epoch,耗时约30,000 H200 GPU-hours,端到端机器利用率(MFU)达到40%。
  • 微调与强化学习:使用少于35,000条计算机操作轨迹进行监督微调以引入动作指令格式。推理时先预测下一帧状态再输出动作。随后通过在线强化学习(RL)在Linux虚拟机环境中进行大规模 rollout,并通过程序化验证结果获得奖励信号。

行业启示

  • 去标注化的数据利用潜力:证明了大量无动作标签的历史视频数据具有巨大价值,可转化为智能体的世界模型和策略基础,减少对昂贵、细粒度标注数据的依赖。
  • 能效与成本优化新路径:通过创新的压缩算法(FSQ)和高效的MoE架构,显著降低了预训练算力需求。这提示行业关注底层表示效率和硬件利用率,而非单纯堆砌参数和算力。
  • 通用世界模型的泛化性:仅在桌面交互数据上预训练的模型能泛化至棋类逻辑和物理模拟,表明“隐式策略”可能捕捉到了更通用的因果推理和世界动态规律,为构建具备广泛适应能力的通用AI代理指明了方向。

Disclaimer: The above content is generated by AI and is for reference only. 免责声明:以上内容由 AI 生成,仅供参考。

Agent Agent Multimodal 多模态 Video Generation 视频生成 Training 训练 Benchmark 基准测试