Topic
All digests tagged Large Language Models (LLMs)
This guide details loading pre-trained LLMs (like Qwen3) and optimizing text generation through tokenization, KV caching, and torch.compile. Learn how to achieve high tokens/second throughput for reasoning models.
Read article
MiniMax M3 combines a functional one-million-token context window with native multimodality, achieved through the Minimax Sparse Attention (MSA) architecture. This enables advanced agentic capabilities.
Read article
Jean-Denis Greze argues that multi-agent systems should be reframed as search problems. The core challenge is privacy, which limits the LLM's access to all world data.
Read article
Koray Kavukcuoglu discusses the shift from coding to full software engineering, detailing how agentic workflows are key to building AGI. He highlights Gemini 4 as an ambitious pre-training run.
Read article
Learn how to build reasoning models from scratch, tracing the evolution from conventional LLMs to agent harnesses. The video covers setup using 'uv' and PyTorch best practices.
Read article
NVIDIA Dynamo is a distributed serving layer that scales LLM inference across multi-GPU/multi-node clusters. It coordinates workers, manages routing, ensures fault tolerance, and optimizes KV cache reuse, complementing engines like vLLM.
Read article
IBM unveiled a new dual processor combining z/OS and Arm for mainframes. NVIDIA's strategy involves acquiring Hugging Face to dominate open AI, while Z.ai revealed GLM-5.3-Flash.
Read article
Two winning agents—MasteryWrite and 8kEdu—demonstrate autonomous applications for EdTech. They use advanced frameworks like NemoClaw to build self-improving systems that analyze video lectures and grade student essays.
Read article
Multi-GPU AI workloads are bottlenecked by communication links. The speaker introduces ParallelKittens, primitives that simplify high-performance kernel writing. A benchmark tests if LLMs can generate optimized CUDA kernels for complex multi-GPU patterns.
Read article
Gemini 3.5 Transcribe is an LLM-based transcription model available on the Interactions and Live APIs. It excels at transcribing complex data like email addresses, phone numbers, and multilingual speech.
Read article
Optimizing open source libraries for LLMs requires shifting focus from human SEO to agent primitives. Key strategies include manually curating LLMs.txt and bundling markdown into node_modules.
Read article
DeepWiki, a codebase documentation product for agents like Devin, evolved from orchestration-led v1 to an agentic v2. Learn about context engineering principles, scaling codebases, and building robust knowledge graphs.
Read article
Faraday-27B was trained to replicate scientific research from redacted papers, demonstrating a novel approach to automating AI R&D using CodeX as a tool.
Read article
To train Krea 2, we built a system that handles crashes at scale and uses advanced metrics like tensor core utilization and InfiniBand monitoring to ensure stability.
Read article
This deep dive builds a distributed training framework from first principles using PyTorch. Topics include RoPE, YaRN, MoE, and combining multiple forms of parallelism to train modern LLMs.
Read article
The AI landscape is rapidly evolving, marked by IBM's B300 cluster partnership, Meta's local Muse Glimmer model, and OpenAI's advanced Astra capabilities. We explore the shift to edge computing and industrial-scale infrastructure.
Read article
See how Gemini 3.7 Flash is used to build a playable animated sprite game in Google Antigravity from a single prompt, showcasing rapid prototyping capabilities.
Read article
Scaling Compute on Context addresses how to give LLMs deep expertise from private data when public scaling fails. It reviews methods like KV compaction and distillation to achieve self-improvement.
Read article
This paper proposes Direct On-Policy Distillation, a cheap way to transfer RL benefits from small models to large ones by using policy shifts as dense reward signals.
Read article
To scale RL post-training cross datacenters, the synchronization unit must shift from massive full checkpoints (500 GB) to small 'lossless patches' representing only visible weight changes. This enables elastic rollout fleets.
Read article