Sequoia
·
4 months ago
Scanner, a log search engine designed for cloud storage, enables security teams to search petabytes of logs in seconds by building an inverted index that maps directly to file regions in S3 rather than scanning all data. The company has signed customers including Notion, Ramp, and Benchling, with nearly a third of customers using it with AI agents within weeks of an MCP release, and agents now generating 80% of platform queries. This enables autonomous AI-driven security investigations that previously required hours of manual work, marking a shift toward AI-native security operations.
OpenAI Blog
·
4 months ago
Researchers developed IH-Challenge, a training method that teaches large language models to prioritize instructions from trusted sources over conflicting inputs. The approach improved instruction hierarchy performance by enabling models to better distinguish between legitimate directives and injected prompts. Models trained with this method showed increased resistance to prompt injection attacks and greater safety control without sacrificing general capabilities.
OpenAI Blog
·
4 months ago
ChatGPT now offers interactive visual explanations for math and science topics, allowing students to manipulate formulas and variables during learning sessions. The feature enables real-time exploration of how changes to variables affect outcomes in equations and scientific concepts. Students can work through problems dynamically rather than passively reading static explanations, potentially improving comprehension of abstract concepts.
.png)
Together AI
·
4 months ago
Together AI introduced autoscaling, role-based access control, observability dashboards, and self-healing capabilities to its GPU Clusters platform. The autoscaling feature uses Kubernetes to automatically add or remove GPU nodes based on demand, while health checks and self-repair can restore failed nodes within minutes. These production-grade features enable teams to run large distributed training jobs and inference workloads without manual infrastructure management or losing compute time to hardware failures.
Hugging Face Blog
·
4 months ago
Researchers surveyed 16 open-source reinforcement learning libraries to understand how they handle asynchronous training, where inference and gradient computation run in parallel on separate GPU pools. The generation phase for reasoning models can consume 28 minutes to 3.7 hours per training step on a single GPU, leaving training GPUs idle while waiting for data. The standard solution separates inference and training onto different hardware, connects them with a rollout buffer, and synchronizes model weights asynchronously so neither process blocks the other.
Hugging Face Blog
·
4 months ago
Hugging Face introduced Storage Buckets, mutable object storage on its Hub designed for intermediate ML artifacts like checkpoints, optimizer states, and processed data that don't require version control. Buckets are built on Xet, a chunk-based backend that deduplicates content across related files, reducing bandwidth and storage costs, with pre-warming available to bring frequently-accessed data closer to compute infrastructure in AWS and GCP regions. Users can now manage training pipelines, data processing, and agent traces in a single Hub-native workflow, with a clear separation between mutable working storage and versioned repos for final deliverables.
