Granite 4.1 models adopt LLMON to block prompt injections
Hugging Face releases Granite 4.1 3B and 8B preview models trained on LLMON, a markup language that separates instructions from data to prevent indirect prompt injections.
RAG, evaluations, inference, fine-tuning and shipping LLM features to production.
Hugging Face releases Granite 4.1 3B and 8B preview models trained on LLMON, a markup language that separates instructions from data to prevent indirect prompt injections.
Amazon details a method to detect silent data corruption in SageMaker HyperPod by adding three checks to PyTorch training loops, preventing wasted compute on faulty hardware.
Anthropic released Fable 5.2, a model that matches Opus 5.5 on coding benchmarks at half the task cost, yet advises developers to keep using Opus for daily work.
MIT CSAIL's weight-space programming research shows that a simple linear program can match an AI agent's complex weight edits, achieving similar accuracy gains with far fewer changes.
A new training method teaches multi-vector encoders to preserve retrieval signals after compressing document vectors, reducing storage by 75% with minimal quality loss.
vLLM improves LLM inference throughput by replacing static batching with continuous batching and using PagedAttention to manage memory efficiently.
Amazon Bedrock now hosts Kimi K3.1 Pro Highspeed, offering faster output speeds and cost-saving tiers for agentic workflows.
Anthropic releases Claude Fable 5.2, surpassing Gemini 4 Carbon on key benchmarks through extensive self-verification and subagent usage, despite higher per-task costs.
Meta released Triton Low-Level Extensions (TLX), a new programming model that gives engineers explicit control over warp-group orchestration on modern GPUs without losing Triton's high-level productiv
Google released TPU Sync, an open-source system that cuts time to first token by up to 90% for large models on Cloud TPUs.
Amazon Bedrock Managed Knowledge Base now supports single-shot RAG and retrieval-only requests via the AgenticRetrieveStream API, allowing developers to balance latency and accuracy.
Anthropic released Claude Fable 5.2 with a 20% price cut, faster token generation, and improved coding benchmarks for complex technical tasks.
Mihail Eric updates Stanford’s CS146S course to reflect agent-native workflows, emphasizing open-source contribution and custom tooling over basic code generation.
Video Rebirth releases HyperFlow H3 Flash, a compressed MiniMax-H3 model that generates joint video and audio on consumer hardware using INT8 quantization and GGUF formats.
A new pipeline improves speech-to-text accuracy by reading on-screen accessibility trees and biasing a frozen Parakeet model, avoiding heavy LLMs or fine-tuning.