Your LLM works harder after a short nap
A new paper proposes letting LLMs 'sleep' between processing chunks, consolidating context into fast weights and clearing the cache. The results on multi-hop reasoning are 3x better than baseline.
NESTFRONTIER / ARCHIVE
All entries
Technical analysis, model releases, research, and deployment notes.
PAGE_20
294 PUBLISHED ENTRIESVision models finally broke free from the token-by-token cage
NVIDIA's LocateAnything uses Parallel Box Decoding to make vision-language models 10x faster at visual grounding while improving accuracy on LVIS and COCO. The 3B model hits 12.7 boxes per second and handles dense detection, GUI, and document grounding from a single checkpoint.
Picking the right AI model is now a $1.3 billion problem
OpenRouter raised $113M at a $1.3B valuation as the multi-model AI shift accelerates. Behind the numbers: 25 trillion weekly tokens, 8 million users, and the team making it happen with only 50 people.
Your AI future just got a 42,300-word warning from the Vatican
Pope Leo XIV released Magnifica Humanitas, a 42,300-word encyclical on AI, alongside Anthropic co-founder Chris Olah. It calls for an international ban on autonomous weapons, warns AI labs cannot self-regulate, and frames AI as the defining moral question of the decade.
Chinese open-weight models are eating Silicon Valley's lunch
Chinese open-weight models now handle over 60% of tokens on OpenRouter, up from 2% a year ago. Xiaomi, Alibaba, DeepSeek, and others are winning on price-performance at 10-20x lower cost for comparable quality. Here is how it happened and what it means for developers.
Your AI coding agent is burning money on every cache miss
Reasonix is an open-source terminal agent built specifically around DeepSeek's prefix cache. 99.82% hit rate, $0.05 per turn, and it proves that single-provider agents beat multi-provider ones on cost.
Your AI coding agent wastes tokens on files it doesn't need
Claude Context is an open-source MCP server from Zilliz that gives AI coding agents semantic code search, cutting token usage by ~40%. Hit #1 on GitHub Trending with 10K stars in one week.
Long Context Still Breaks Memory. Gated DeltaNet-2 Fixes It
Gated DeltaNet-2 gives linear attention separate erase and write gates, targeting the memory interference that breaks long-context retrieval.
Your AI coding assistant is not yours
Microsoft just canceled thousands of internal Claude Code licenses and forced developers onto Copilot CLI. The reason has nothing to do with quality. It is about who gets to own your development loop.
Scaling Laws Need Fewer Questions Than Anyone Thought
Item Response Scaling Laws estimate model scaling with about 50 questions per benchmark after calibration. That cuts evaluation cost, not model-training cost.