Runway Claims GPU Count for 57B Video Model Cut From Eight to Four
Runway says it cut the GPUs needed to run a 57-billion-parameter video model at 24 fps from eight to four. The claim reached us through a repost by...

Created by Yifeng Peng
Daily curated AI papers with reproducible code, benchmarks, and practical industry takeaways
Explore the latest content tracked by AI Research Pulse
Runway says it cut the GPUs needed to run a 57-billion-parameter video model at 24 fps from eight to four. The claim reached us through a repost by...
The article introduces Standard Model Templates (SMT), an approach that standardizes model design into a curated set of reusable templates.
The...
This paper systematically reviews ML applications in four core areas: credit scoring, fraud detection, algorithmic trading, and a customer-related...
The authors argue that current video MLLMs reduce video to static multi-image reasoning. Whole-frame pooling can dilute small moving targets, and...
Turning document vectors into visibility graphs is an unusual approach to text classification. The preprint treats an averaged word-embedding vector...
The authors treat patent-office search reports as expert supervision, distilling 38,397 JPO reports into executable Boolean strategies.
Highlight detection has to handle two timescales at once: brief impact events and longer narrative arcs, while also producing semantic descriptions...
The Multi-Branch Transformer proposes parallel branches of full blocks as a complement to making Transformers wider or deeper. The authors' reported...
Transferring agent skills with on-policy distillation (OPD) can change interaction habits well beyond the training domain. In this study,...
Transformers were chosen because they train well with gradient descent and run fast on GPUs. This paper argues those reasons no longer apply when an...
Interactive video generation depends on latency, and the paper says existing autoregressive models are orders of magnitude too slow for fluid...
Most world-action models supervise complete future observations. But observations often change little over an action chunk, so the loss mostly copies...
The paper argues that sum-pooling acts as a mean-field (Hartree) model. That suits nearsighted quantities, but a close approach between two particles...
The paper argues that control-based alignment rests on a misunderstanding of digital agency. Its author contends that current reasoning models meeting...
The article argues AI should speed up electronics design without taking over the decisions that matter.
Anthropic argues that models are already strong at refusing blunt requests, but real misuse is often gradual, with users keeping the conversation...
The author expected to trade accuracy for speed, but this small, informal test in progress didn't show that pattern. GPT-6.1 Terra (high) scored 96...
The source reports that Google released a paper on Self-Improvement Memory (SIM), which it says improves continual self-improvement for AI agent...
This approach efficiently transfers pretrained weights from a single template network to candidate architectures, which the source says accelerates...
This article frames a path-based view as recurring across reinforcement learning and planning and control.
In planning and control, the paths are...