Lambda Feature Trails

The platform now supports multiple orchestration layers for AI workloads, including Slurm for batch training, Kubernetes for flexible containerized workloads, SkyPilot for portable job execution, and dstack for vendor-agnostic management. This allows teams to align the platform with their specific requirements for scheduling, resource allocation, and configuration overhead.

This post introduces a new capability in 3D scene representation learning using Deep Latent Particles (DLP). It extends DLPs to decompose real, colored 3D observations (like RGB-D images and voxel grids) into a compact set of interpretable particles, each with its own 3D position, size, and appearance. This self-supervised approach addresses challenges in 3D scene modeling, particularly the issues caused by tokenization in 2D, by focusing on object-centric representations rather than patches. The framework utilizes an appearance-aware K-means prior and a chroma loss to handle sparse voxel grids and color reconstruction. The learned representations have demonstrated effectiveness in driving robots for task completion across various benchmarks.

This post introduces a new capability focused on improving agent security testing by analyzing prompt injection attacks. It highlights the discovery that rhetorical mechanisms for prompt injection are transferable across different agent domains. The capability involves running adversarial competitions to identify these transferable attack patterns, emphasizing the need to test for 'mechanisms' rather than just 'scenarios'. Key findings include the prevalence of short, conversational attacks and the sensitivity of defender performance to minor variations in attack phrasing. Practical hardening advice includes separating trusted instructions from untrusted input and gating consequential actions behind explicit confirmation. The capability also involves benchmarking near-misses to better understand failure modes.

The platform now supports dynamic scaling of GPU resources, allowing additional GPUs to be added mid-competition without interrupting ongoing battles. This is achieved by independently serving models on each GPU with vLLM, using Caddy for load balancing and health checks, and a dynamic concurrency limiter that resizes based on available resources. This enables efficient cost management and ensures competition integrity even with fluctuating demand. This post highlights the increasing demand for serving large open-weight models like GLM 5.2, emphasizing the infrastructure challenges and the need for efficient serving solutions to keep up with the capability gap closing between open and closed-weight models.

This capability focuses on developing adaptive data preprocessing techniques for high-frequency trading (HFT) environments. It addresses the challenge of non-stationary market conditions by moving beyond fixed normalization strategies. The core idea is to build systems that learn market structure directly from data, enabling them to preprocess and adapt to incoming data dynamically, even across unseen market conditions. This involves using Bayesian frameworks to model market behavior and statistical divergence measures to distinguish between routine noise and novel data. The ultimate goal is to reduce the computational burden of training ML models by ensuring data representations are accurate and regime-aware before reaching the model, leading to more efficient infrastructure utilization and faster convergence.

The platform now supports NCode, an open-source AI model harness that provides transparency and control over how models operate, what tools they can access, and how their performance is measured. This allows for greater reliability and customization by enabling users to integrate their own models, fine-tune them for specific tasks, and monitor their execution. NCode addresses the 'black-box' problem of proprietary harnesses by offering a fully inspectable and modifiable infrastructure layer, facilitating the use of open-weight models and custom deployments on platforms like Lambda.

The platform now supports PixARMesh, a novel single-image 3D scene reconstruction method. This capability leverages autoregressive Transformers to directly generate artist-ready 3D meshes from a single image, addressing limitations of prior implicit representation or multi-stage pipeline methods. It achieves state-of-the-art results on benchmarks like 3D-FRONT, producing compact meshes suitable for downstream applications in robotics, AR/VR, gaming, and embodied AI.

The platform now supports agentic experimentation loops, enabling AI agents to autonomously hypothesize, experiment, evaluate, and iterate on tasks. This includes structured experiment tracking via an API, leaderboards for performance comparison, idea branching for exploring different hypotheses, and history logging for maintaining context. The system is designed to leverage otherwise idle compute resources for continuous research and development.

The platform now supports large-scale synthetic data generation, enabling the creation of diverse and high-signal training data for AI models. This includes systems for procedural generation of physical scenarios using simulators, automatic construction of verified question-answer pairs, and optimization of data generation pipelines for various model sizes and benchmarks. The focus is on moving beyond human annotation to scalable, simulation-driven data creation.

This post details Lambda's performance in MLPerf Training v6.0, showcasing the fastest LLM training on NVIDIA GB300 NVL72 and fastest MoE training on NVIDIA HGX B200. It highlights improvements in training speed for Llama 3.1 8B on GB300 NVL72 due to software optimizations and presents results for GPT-OSS-20B, an MoE model, demonstrating competitive performance on both GB300 NVL72 and HGX B200 systems. The post emphasizes the impact of hardware innovations like NVIDIA Blackwell Ultra and software stack tuning on convergence times and efficiency for various AI model architectures.