How AI agents use sleep-time compute to organize long-term memory
Anthropic explains how Claude uses dreaming and salience judgment to consolidate memory, correcting errors and building knowledge structures during idle time.
RAG, evaluations, inference, fine-tuning and shipping LLM features to production.
Anthropic explains how Claude uses dreaming and salience judgment to consolidate memory, correcting errors and building knowledge structures during idle time.
AWS demonstrates how to scale AI agent security using Bedrock AgentCore Gateway, Cedar policies, and Okta Cross App Access for centralized authorization.
A new wrapper tool called llama-sleep reduces model switching latency on single-GPU setups by parking CUDA state in host RAM instead of reloading weights.
Ramp data reveals AI token volumes are rising while spending flatlines, driven by aggressive price cuts from OpenAI and Anthropic and improved cost controls by enterprises.
NVIDIA details how TensorRT LLM VisualGen accelerates FLUX 3 Image Klein generation using NVFP4 quantization, sparse attention, and multi-GPU scaling on Blackwell hardware.
Apple researchers use small neural networks to approximate support functions, speeding up maximum inner product search for retrieval-augmented generation.
Nubank researchers detail a simulation workflow that screens AI customer experience agents using synthetic users, improving metrics without exposing live customers to errors.
AWS releases Web Answer, a managed tool that returns synthesized, source-cited answers to natural language questions, eliminating the need for custom RAG pipelines.
An empirical study reveals that standard success metrics miss critical infrastructure faults in AI agents, while tool-call volume and scope drift offer better visibility.
AWS Deep Insight demonstrates how to manage LLM context windows using architecture, prompts, file externalization, and verification to prevent token overflow in multi-agent systems.
Anthropic and academic partners release FlashProve, an open-source system using zkSNARKs to prove LLM responses come from specific model weights without revealing them.
Ricursive Intelligence releases Ripple-65B Preview, an open-weight model that achieves state-of-the-art agentic coding scores at one-tenth the cost of previous leaders.
AWS Bedrock AgentCore Evaluations now supports custom code-based evaluators using decision models like Jev and Strands Decider to provide calibrated, low-cost scoring for AI agents.
Amazon Bedrock now routes Kimi K3 inference exclusively to Mumbai and Hyderabad regions, enabling compliant processing for Indian health and financial data.
A new paper demonstrates that large language models can read and modify the raw weights of smaller networks, enabling direct weight space programming without writing training code.