AI / ML

Amazon SageMaker AI now allows users to submit a prioritized list of instance types for training and processing jobs instead of specifying a single option. This feature automates capacity selection, enabling the service to pick the first available instance from the provided list. It eliminates the need for complex custom retry logic or concurrent job submissions during periods of high GPU demand.

• Submit multiple instance types in a prioritized list to let SageMaker select available capacity automatically.
• Reduces job submission latency during peak demand without requiring custom orchestration scripts.
• Simplifies workflows for workloads that function equivalently across different instance sizes or types.
• Removes the operational overhead of managing concurrent jobs or complex retry mechanisms.

Amazon SageMaker JumpStart now hosts three new foundation models: IBM's granite-speech-4.1-2b, Kakao's kanana-2-30b-a3b-instruct, and OpenFold3. These additions expand AWS's AI portfolio with specialized capabilities for multilingual speech recognition, bilingual agentic workflows, and biomolecular structure prediction. The models are ready for deployment on AWS infrastructure to address specific enterprise AI challenges.

• IBM granite-speech-4.1-2b handles multilingual ASR and translation across six languages.
• Kakao kanana-2-30b-a3b-instruct enables bilingual agentic AI capabilities.
• OpenFold3 provides high-performance biomolecular structure prediction.
• All models are immediately deployable via SageMaker JumpStart.

Researchers propose OrchSLM, a non-interactive orchestration paradigm that combines heterogeneous small language models to handle agentic subtasks. This approach addresses the latency, privacy, and cost constraints of large cloud-based models while mitigating SLM limitations in long-horizon reasoning. The study probes the dynamics of this complementary architecture as an alternative to interactive debate or iterative verification strategies.

• SLMs offer lower latency and better privacy for repetitive agentic subtasks compared to monolithic LLMs.
• OrchSLM uses a non-interactive, heterogeneous model pipeline to overcome SLM context window limits.
• This paradigm avoids the overhead of iterative verification and debate required by smaller models.
• Specialized SLMs may outperform general LLMs on narrowly scoped tasks in agentic workflows.
• The approach aims to reduce computational costs while maintaining reasoning capabilities.

AWS has integrated TwelveLabs Marengo 3.0 into Amazon Bedrock Managed Knowledge Base, enabling direct multimodal embeddings for video, audio, and images. Unlike previous text-only approaches that relied on transcription, Marengo 3.0 encodes visual scenes and speech cues directly into compact 512-dimensional vectors. This allows for natural language search across media assets with state-of-the-art retrieval accuracy and precise segment timing.

• Marengo 3.0 enables direct multimodal embedding of video, audio, and images in Bedrock KB.
• Captures visual and audio semantics beyond what transcription-based text embeddings provide.
• Produces compact 512-dimensional vectors for efficient storage and high-accuracy retrieval.
• Supports natural language search with precise start/end segment results for media assets.