OpenX

Tag: Nvidia

  • Beyond LLMs: The Three Architectural Approaches Teaching AI to Understand Physics

    Beyond LLMs: The Three Architectural Approaches Teaching AI to Understand Physics

    Large language models excel at writing poetry and debugging code, but ask them to predict what happens when you drop a ball and you’ll quickly discover their limitations. Despite mastering chess, generating art, and passing bar exams, today’s most powerful AI systems fundamentally don’t understand physics.

    This gap is becoming increasingly apparent as companies try to deploy AI in robotics, autonomous vehicles, and manufacturing. The solution? World models鈥攊nternal simulators that let AI systems safely test hypotheses before taking physical action. And investors are paying attention: AMI Labs raised a billion-dollar seed round, while World Labs secured funding from backers including Nvidia and AMD.

    The Problem with Next-Token Prediction

    LLMs work by predicting the next token in a sequence. This approach has been remarkably successful for text, but it has a critical flaw when applied to physical tasks. These models cannot reliably predict the physical consequences of real-world actions, according to AI researchers.

    Turing Award recipient Richard Sutton warned that LLMs just mimic what people say instead of modeling the world, which limits their capacity to learn from experience. DeepMind CEO Demis Hassabis calls this jagged intelligence鈥擜I that can solve complex math olympiads but fails at basic physics.

    The industry is responding with three distinct architectural approaches, each with different tradeoffs.

    1. JEPA: Learning Abstract Representations

    The Joint Embedding Predictive Architecture, endorsed by AMI Labs and pioneered by Yann LeCun, takes a fundamentally different approach. Instead of trying to predict what the next video frame will look like at the pixel level, JEPA models learn a smaller set of abstract, or latent, features.

    Think about how humans actually observe the world. When you watch a car driving down a street, you track its trajectory and speed鈥攜ou don’t calculate the exact reflection of light on every leaf in the background. JEPA models reproduce this cognitive shortcut.

    The benefits are substantial: JEPA models are highly compute and memory efficient, require fewer training examples, and run with significantly lower latency. These characteristics make it suitable for applications where real-time inference is non-negotiable鈥攔obotics, self-driving cars, high-stakes enterprise workflows.

    Healthcare company Nabla is already using this architecture to simulate operational complexity in fast-paced medical settings, reducing cognitive load for healthcare workers.

    2. Gaussian Splats: Building Spatial Worlds

    The second approach, adopted by World Labs led by AI pioneer Fei-Fei Li, uses generative models to build complete 3D spatial environments. The process takes an initial prompt鈥攅ither an image or textual description鈥攁nd uses a generative model to create a 3D Gaussian splat.

    A Gaussian splat represents 3D scenes using millions of tiny mathematical particles that define geometry and lighting. Unlike flat video generation, these 3D representations can be imported directly into standard physics and 3D engines like Unreal Engine, where users and AI agents can freely navigate and interact from any angle.

    World Labs founder Fei-Fei Li describes LLMs as wordsmiths in the dark鈥攑ossessing flowery language but lacking spatial intelligence and physical experience. The company’s Marble model aims to give AI that missing spatial awareness.

    Industrial design giant Autodesk has backed World Labs heavily, planning to integrate these models into their design applications. The approach has massive potential for spatial computing, interactive entertainment, and building training environments for robotics.

    3. End-to-End Generation: Physics Native

    The third approach uses an end-to-end generative model that continuously generates the scene, physical dynamics, and reactions on the fly. Rather than exporting to an external physics engine, the model itself acts as the engine.

    DeepMind’s Genie 3 and Nvidia’s Cosmos fall into this category. These models ingest an initial prompt alongside continuous user actions and generate subsequent environment frames in real-time, calculating physics, lighting, and object reactions natively.

    The compute cost is substantial鈥攃ontinuously rendering physics and pixels simultaneously requires significant resources. But the investment enables synthetic data factories that can generate infinite interactive experiences and massive volumes of synthetic training data.

    Nvidia Cosmos uses this architecture to scale synthetic data and physical AI reasoning. Waymo built its world model on Genie 3 for training self-driving cars, synthesizing rare, dangerous edge-case conditions without the cost or risk of physical testing.

    The Hybrid Future

    LLMs will continue serving as the reasoning and communication interface, but world models are positioning themselves as foundational infrastructure for physical and spatial data pipelines. We’re already seeing hybrid architectures emerge.

    Cybersecurity startup DeepTempo recently developed LogLM, integrating LLMs with JEPA elements to detect anomalies and cyber threats from security logs. The boundary between AI that thinks and AI that understands the physical world is beginning to dissolve.

    As world models mature, expect AI systems that can not only tell you how to change a tire, but actually understand what happens when you apply torque to a rusted bolt. The physical world is finally coming into focus for artificial intelligence.

  • Nvidia’s Nemotron-Cascade 2: How a 3B Parameter Model Wins Gold Medals in Math and Coding

    Nvidia’s Nemotron-Cascade 2: How a 3B Parameter Model Wins Gold Medals in Math and Coding

    The prevailing assumption in AI development has been straightforward: larger models trained on more data produce better results. Nvidia’s latest release directly challenges that orthodoxy鈥攁nd the training recipe behind it may matter more to enterprise AI teams than the model itself.

    Nemotron-Cascade 2 is an open-weight 30B Mixture-of-Experts model that activates only 3B parameters at inference time. Despite this compact footprint, it achieved gold medal-level performance on three of the world’s most demanding competitions: the 2025 International Mathematical Olympiad, the International Olympiad in Informatics, and the ICPC World Finals. It is only the second open model to reach this tier, after DeepSeek-V3.2-Speciale鈥攁 model with 20 times more parameters.

    Nvidia Nemotron-Cascade 2 Performance

    The Post-Training Revolution

    Pre-training a large language model from scratch is enormously expensive鈥攐n the order of tens to possibly hundreds of millions of dollars for frontier models. Nemotron-Cascade 2 starts from the same base model as Nvidia’s existing Nemotron-3-Nano鈥攜et it outperforms that model on nearly every benchmark, often surpassing Nvidia’s own Nemotron-3-Super, a model with four times the active parameters.

    The difference is entirely in the post-training recipe. This is the strategic insight for enterprise teams: you don’t necessarily need a bigger or more expensive base model. You may need a better training pipeline on top of the one you already have.

    Cascade RL: Sequential Domain Training

    Reinforcement learning has become the dominant technique for teaching LLMs to reason. The challenge is that training a model on multiple domains simultaneously鈥攎ath, code, instruction-following, agentic tasks鈥攐ften causes interference. Improving performance in one domain degrades it in another, a phenomenon known as catastrophic forgetting.

    Cascade RL addresses this by training RL stages sequentially, one domain at a time, rather than mixing everything together. Nemotron-Cascade 2 follows a specific ordering: first instruction-following RL, then multi-domain RL, then on-policy distillation, then RLHF for human preference alignment, then long-context RL, then code RL, and finally software engineering RL.

    MOPD: Reusing Your Own Training Checkpoints

    Even with careful sequential ordering, some performance drift is inevitable as the model passes through many RL stages. Nvidia’s solution is Multi-Domain On-Policy Distillation鈥攁 technique that selects the best intermediate checkpoint for each domain and uses it as a “teacher” to distill knowledge back into the student model.

    Critically, these teachers come from the same training run, sharing the same tokenizer and architecture. This eliminates distribution mismatch problems that arise when distilling from a completely different model family. According to Nvidia’s technical report, MOPD recovered teacher-level performance within 30 optimization steps on the AIME 2025 math benchmark, while standard GRPO required more steps to achieve a lower score.

    What Enterprise Teams Can Apply

    Several design patterns from this work are directly applicable to enterprise post-training efforts. The sequential domain ordering in Cascade RL means teams can add new capabilities without rebuilding the entire pipeline鈥攁 critical property for organizations that need to iterate quickly. MOPD’s approach of using intermediate checkpoints as domain-specific teachers eliminates the need for expensive external teacher models.

    Nemotron-Cascade 2 is part of a broader trend toward “intelligence density”鈥攅xtracting maximum capability per active parameter. For enterprise deployment, this matters enormously. A model with 3B active parameters can be served at a fraction of the cost and latency of a dense 70B model. Nvidia’s results suggest that post-training techniques can close the performance gap on targeted domains, giving organizations a path to deploy strong reasoning capabilities without frontier-level infrastructure costs.

    For teams building systems that need deep reasoning on structured problems鈥攆inancial modeling, scientific computing, software engineering, compliance analysis鈥擭vidia’s technical report offers one of the more detailed post-training methodologies published to date. The model and its training recipe are now available for download, giving enterprise AI teams a concrete foundation for building domain-specific reasoning systems without starting from scratch.