LLMs automate formal verification for open-source libraries at scale
A new collaboration demonstrates that large language models can generate formal proofs for over 100 open-source libraries, proving security properties at the assembly level.
RAG, evaluations, inference, fine-tuning and shipping LLM features to production.
A new collaboration demonstrates that large language models can generate formal proofs for over 100 open-source libraries, proving security properties at the assembly level.
NVIDIA's new Global Performance Tuner in TensorRT 11.1 uses over 200 compiler knobs to find faster inference paths while maintaining accuracy constraints.
Microsoft and Hugging Face improved Qwen3.8-27B reliability by using executable database checks as rewards, tripling consistent task success.
OpenAI has made its Decisions API widely available, offering fast, low-cost inference by reconfiguring ChatGPT Luna to compete in the emerging decision model market.
NVIDIA now supports eight-node DGX Spark clusters, offering 1 TB of unified memory and new tools for efficient local inference and training.
New benchmarks show DeepSeek’s updated open-weight model beating Anthropic’s Haiku 5.5 on accuracy for two of three tasks while costing significantly less.
Learn to use torch.profiler and Neuron Explorer to identify performance bottlenecks in ML inference on AWS Trainium hardware.
Researchers demonstrate how floating-point precision quirks allow arbitrary code execution inside PyTorch's gradient descent, even when loss metrics report zero.
Nvidia launched Agent Knowledge Foundation, an open-source toolkit combining retrieval, ontology, and prediction models to help AI agents understand enterprise data without moving it.
A technical guide details three methods for training Qwen3 models on AWS Trainium using torchtitan, including Hugging Face integration and custom NKI kernels.
Hugging Face releases FairStress, a new benchmark showing that large language models apply different levels of scrutiny to decisions based on candidate demographics.
A developer reduced hallucinations in a 0.6B-parameter model by refining training data style rather than increasing model size, achieving better refusal rates in browser-based search.
DeepSeek has open-sourced DeepInfer, a high-performance inference framework optimized for DeepSeek-V4.1-Flash models on NVIDIA SM100 and Huawei Ascend 950 hardware.
i14 and AWS trained the Laplace 3B linear attention model on Trainium2, achieving high throughput by replacing GPU kernels with standard PyTorch operators.
AWS demonstrates how bi-directional key-value cache transfer in vLLM reduces time to first token by up to 3.3x for long conversational contexts.