Sparsifying Full Attention in Just 100 Steps — RTPurbo Changes the Conventional Wisdom of LLM Inference
Long-context inference in large language models (LLMs) has long been bottlenecked by the quadratic cost of full attention. Existing efficiency methods rely on either sparse training or heuristic token pruning, forcing a trade-off among efficiency, training cost, and accuracy.
The newly announced RTPurbo is a new approach built on the discovery that full-attention LLMs are already inherently sparse. The research team reports three key observations.
First, only a small subset of attention heads genuinely require long-context processing. Second, long-range retrieval is governed within a low-dimensional subspace, so a 16-dimensional indexer can efficiently retrieve relevant tokens. Third, because the budget of useful tokens depends heavily on the query, dynamic top-p selection is better suited than fixed top-k.
Building on these findings, RTPurbo keeps the full KV cache only for retrieval heads and introduces a lightweight token indexer for the remaining heads. Remarkably, just 100 training steps can convert a full-attention model into a highly accurate sparse model. This looks like a practical breakthrough for cutting the real cost of long-context inference.
Generating a Short Drama from a Single Sentence — A Multi-Agent Video Production System
"One Sentence, One Drama," featured on Hugging Face Daily Papers, is a multi-agent framework that generates a fully produced short drama from just a single sentence of input.
It adopts a hierarchical multi-agent architecture made up of three main components. First, story generation via multi-agent debate, which ensures the pacing and narrative consistency that short dramas demand. Next, 3D scene generation for cinematic visual expression. And finally, a personalization feature that delivers works tailored to each user's preferences.
AI-driven video production has seen remarkable tool-level advances, but few attempts have simultaneously delivered both narrative consistency and visual quality. This research shows that generative AI is moving not only into the downstream production stages of the creative industry but also into upstream planning.
AI Infrastructure Unicorns Keep Coming — Exa, Modal, TurboPuffer
In the AI infrastructure space, new unicorns are being born one after another. According to Latent Space, Exa with its search API, Modal with its serverless GPU platform, and TurboPuffer with its vector database have all reached valuations above $1 billion in quick succession.
Exa provides a high-precision search API for AI applications and is used as a knowledge base for LLMs. Modal simplifies GPU computation in the cloud, accelerating AI model development and deployment. TurboPuffer is a database specialized for AI-native vector search.
As LLM applications move into genuine widespread adoption, demand is surging for the surrounding infrastructure rather than the models themselves. While competition in the model layer intensifies, it is noteworthy that companies in the infrastructure layer are establishing stable business models.
AI May Become an Inflationary Force Before It Boosts Productivity
An Axios analysis offers a perspective worth heeding on AI's economic impact: before AI delivers productivity gains, it may first act as inflationary pressure.
At this stage, corporate AI investment is pushing up demand while real productivity gains remain limited. Data center construction, GPU demand, and rising power consumption are driving up costs, which then spill over into consumer prices. AI's productivity effects are not expected to arrive in earnest for several years, and the analysis argues that this gap is what generates inflationary pressure.
The debate over the economic returns of AI investment is growing in importance not just as a matter of technical interest, but in the macroeconomic context as well.
Average ¥34 Million Bonuses for Samsung's Semiconductor Division — Rewards of the AI Boom
Samsung has decided to pay bonuses of up to $26.6 billion (roughly ¥4 trillion) to employees in its semiconductor division. The average payout per employee is expected to reach about $340,000 (around ¥34 million).
This exceptional bonus is underpinned by the explosive growth in demand for AI semiconductors. Samsung's semiconductor division has stayed strong on the back of HBM (high-bandwidth memory) production for AI chips, and a final agreement with the labor union sealed the payout.
The benefits of the AI boom extend beyond model developers to the entire supply chain. Led by demand for NVIDIA GPUs, related industries — memory, packaging, cooling equipment, and more — continue to grow rapidly, and the competition for talent is heating up. Samsung's massive bonus is also seen partly as a move to retain talent in this competitive environment.