The rapid progress of software engineering agents drives interest in applying them to open-ended problems such as GPU kernel optimization. However, we find that existing coding agents converge to local optima on real-world, expert-written kernels (e.g., FlashInfer). We introduce Evolving Agent Teams, a multi-agent framework that escapes such local optima. It features (1) a central supervisor that diagnoses each local optimum and restructures a team of agents; and (2) a persistent file system that comprises a learning document, an evolution history, and per-team artifacts, grounding restructuring with accumulated experience. Under a matched token budget, Evolving Agent Teams consistently outperforms strong coding-agent baselines. With an extended token budget, Evolving Agent Teams generates a CUDA kernel that reaches 1.2x mean speedup over the production FlashInfer MLA paged decode kernel from scratch and wins on 36 of 47 production workloads from FlashInfer-Bench (0.86-0.99x on the remaining 11). It also demonstrates superior performance on tasks such as KDA prefill on AMD MI350X (3.01x over vLLM) and NVIDIA B200 (3.44x over FlashKDA), MoE expert-offload serving on Apple silicon (30.3 tokens/s decode for Qwen3.8-Flash-Next in MLX 4-bit on a 48 GB M4 Pro), and RoboDojo robot manipulation (62.1 on held-out seeds vs. 55.3 for the best leaderboard entry). Code is available at https://github.com/caoshiyi/EAT.
Training large-scale AI models on thousands of GPUs is a high-stakes operation where monitoring and automation must be built-in to optimize cluster resiliency...
The recent incident in which OpenAI agents autonomously compromised Hugging Face's infrastructure illustrates the importance of instruction hierarchy (IH): an LLM should follow lower-priority instructions, such as user requests, only when they do not conflict with higher-priority instructions, such as system policies. IH can be effectively improved by fine-tuning LLMs on simulated examples of instruction conflicts, whose diversity is key to teaching LLMs to properly handle conflicts at test time. However, our decomposition of a task into goals and constraints reveals a conflict type -- goal-goal conflicts -- that is absent from existing IH training sets from OpenAI and NVIDIA. We show the importance of this conflict type: fine-tuning Qwen3.5-9B and Gemma4-12B-it on examples of this conflict type significantly improves IH on a wide range of real-world tasks unseen in training; adding these examples to existing IH training sets for fine-tuning also gives better IH adherence. The IH gains come with trivial mean utility changes on no-conflict tasks across eight benchmarks. Our work advocates including this new conflict type in IH training, even if these examples do not reflect typical LLM use.
New text-generation model. Tags: gguf, nvidia, pytorch, nemotron-3.5, unsloth, text-generation, en, ar
New text-generation model. Tags: gguf, nvidia, pytorch, nemotron-3.5, text-generation, en, ar, bg
Predicting the short-term future trajectory of a person from RGB-D cameras is a key capability for mobile robots navigating in dynamic and human populated environment. The key challenge to deploying learned predictors on real robots is the train-deploy pipeline gap. Training data are often prepared with an idealized, smoothed ground truth, whereas at inference, the model is fed noisy estimates from pose detectors and depth sensors. The model is therefore evaluated on an input distribution it never encountered during training. Most existing models rely on recurrent neural networks (RNNs) or transformers, which either struggle to capture short-term temporal dependencies or suffer from computational cost. We present TrajMamba, a lightweight, bidirectional state space model for short-term human trajectory prediction using an RGB-D camera. The model uses a single shared processing path for both training and deployment, aligning human pose extraction, depth back projection, and smoothing processes. It encodes a short window of persons's hip center velocities with stacked bidirectional Mamba blocks and decodes future velocities with a GRU head. On the ETH/UCY benchmark, TrajMamba reaches an average ADE of 0.29 m FDE of 0.51 m. With 262K parameters, it runs in real time on modest hardware and is deployed on a TIAGo robot with an onboard NVIDIA Jetson Orin.
Explore how to deploy and benchmark LLMs locally using tools like Ollama and NVIDIA NIMs. This deep dive covers performance, cost, and scaling insights.
The robots, stomping their little feet and making tiny R2D2 noises, sat, fell down dead for a "nap," and summersaulted on voice command.
inDrive plans to begin testing Pony.ai's robotaxis in Kazakhstan's capital in the first half of 2027.
Oh sure, SpaceX's credit risk is pretty high because it's about to borrow another $40 billion to spend on Nvidia chips, but who cares about…
GPU applications are processing ever-growing datasets with increasingly powerful accelerators. According to the 2025 Stanford AI Index, compute for notable AI...
Today, I’m talking to Bill Gates, cofounder and former CEO of Microsoft and chairman of the Gates Foundation. You may have noticed Bill in the…
The Adaptive Agentic Intelligence Framework (AAIF), built with Strands Agents, handles routine tasks locally with small language models and routes complex requests to cloud models through Amazon Bedrock. This post walks through building AAIF with local SLMs through Ollama and Amazon Bedrock, using an Internet of Things monitoring scenario.
Autonomous agents can work for hours on complex tasks, accessing files, tools, the internet, and spawning other agents to collaborate with. Executing complex...
Building an AI factory digital twin requires developers to bring site plans, equipment assets, and engineering requirements into a consistent model that teams...
AI factories turn energy into intelligence. But before an AI workload can produce its first token, operators must bring an interconnected system of compute,...
New image-text-to-text model. Tags: Model Optimizer, glm5_next, nvidia, ModelOpt, GLM-5, quantized, FP4, fp4
Learn how we fine-tuned GPT-OSS-20B for multilingual reasoning on Amazon SageMaker HyperPod, and how to decide between supervised fine-tuning (SFT) and reinforcement learning with GRPO. SFT took reasoning-language accuracy from 22% to 100%, and GRPO added measurable gains where headroom existed.
Energy-efficient deployment of large language models requires understanding how model architecture, hardware, and serving settings shape energy per generated token. We introduce the Token-Energy Law, a compact, interpretable model of GPU-board energy per output token for dense, full-history decoding. It relates token energy to two core dimensions: active bytes and realized decode time. Active bytes describe logical byte demand per output, separating dense weights shared within a decode iteration from private key--value (KV) histories read and updated for each request. Realized decode time is the elapsed decode duration divided by the number of output tokens produced. We combine the two byte coordinates and time in a common affine form, with coefficients calibrated for each model and hardware/runtime configuration. This formulation explains how batching amortizes shared-weight work, longer attended histories increase private KV demand, and throughput determines the time-associated cost per output. Our experiments cover three models under BF16 decoding on NVIDIA Hopper and Blackwell GPUs, including two-GPU tensor parallelism on Hopper. Using fixed coefficients and observed decode time, held-out condition-mean energy estimates differ from measurements by at most 3.73%.
