AI / ML

Amazon OpenSearch Service has extended automatic semantic enrichment to VPC-enabled domains, previously restricted to public endpoints. This update allows organizations to perform context-aware AI search within private networks without exposing data externally. The service handles all machine learning model management and integration overhead automatically.

• VPC domains can now use AI-powered semantic search without public exposure.
• Eliminates need to self-manage ML models for semantic enrichment.
• Improves query relevance by understanding meaning, not just keywords.
• Previously limited to non-VPC domains; now available for private setups.

KernelArc is a multi-agent framework that autonomously optimizes GPU kernels for heterogeneous workloads. Strategy-specialized agents coordinate via shared memory and a deterministic benchmark guard to generate high-performance implementations. The system successfully produced custom BF16 GEMM, fused MoE backward, and NVFP4 attention kernels evaluated on NVIDIA H100 and B200 hardware.

• Agents use conclusions-only shared memory to coordinate without complex communication overhead.
• A deterministic benchmark guard ensures optimization stability across heterogeneous workloads.
• Results include custom BF16 GEMM and native NVFP4 grouped-query attention kernels.
• Validated on SOL-ExecBench workloads across NVIDIA H100 and B200 GPUs.
• Framework supports complex fusions like shape-gated decoder-layer and paged prefill attention.

PTXBench evaluates LLMs on generating architecture-specific PTX for GPU kernel optimization across H100 and B200 hardware. The benchmark reveals that while models can produce functional code, they often fail to achieve competitive speedups over frontier libraries, particularly in complex attention backward workloads. Even after supervised fine-tuning with Qwen3.6-27B, no model consistently outperformed existing high-performance libraries.

• LLMs generate functional PTX but rarely match frontier library performance on GEMM and attention tasks.
• Success rates drop significantly for complex attention backward workloads on H100 and B200 GPUs.
• Executing target instructions does not guarantee competitive runtime speedup over optimized libraries.
• Supervised fine-tuning on Qwen3.6-27B improves capability but does not close the gap with hand-tuned kernels.

Polaris addresses the misalignment between fluent and effective table descriptions in NL2SQL retrieval tasks. It fine-tunes LLMs using Direct Preference Optimization on preference pairs generated by ranking BM25 effectiveness of candidate descriptions. This method leverages existing relevance judgments in retrieval benchmarks to optimize for search utility rather than natural language quality.

• Shifts table description optimization from fluency to retrieval effectiveness.
• Uses BM25 ranking of candidates to generate preference pairs for DPO.
• Leverages existing query-table relevance data as free supervision signal.
• Improves table-centric NLP tasks like NL2SQL by enhancing initial retrieval.

WhatsApp is beta-testing a Scam Alert feature that leverages on-device machine learning to identify potential scam messages from non-contacts. The architecture ensures message content remains local while employing confidential computing, Oblivious HTTP, and differential privacy to measure performance. This approach allows Meta to monitor model effectiveness and deliver updates without compromising user data privacy.

• On-device ML processes scam detection locally, keeping raw message content off servers.
• Confidential computing and Oblivious HTTP secure model delivery and performance metrics.
• Differential privacy is used to aggregate usage data without exposing individual user behavior.
• Model transparency features help users understand why specific messages are flagged as scams.