Komprise published results from a September 2026 survey of 500 technical decision-makers. While 68% said their organizations were maturing or mature in AI, 28% said they had quantified AI return on investment. Only 7% reported that at least 70% of their data was cleaned and classified for AI; 57% said half or less was ready.
MLCommons was selected to develop the MLPerf Agents benchmark under a two-year, $3.5 million AI Efficiency Challenge supported by the U.S. National Science Foundation. The planned benchmark will measure complete workflows—not just model token serving—including the agent harness, model choice, retrieval, tool execution, orchestration, state management, failure recovery and validation.
Databricks made synchronous REST APIs for document parsing, classification and field extraction generally available. ai_parse_document converts tables and figures into structured text; ai_extract returns requested fields with optional source citations and confidence scores. ai_classify labels documents or extracted entities against a business taxonomy.
Databricks announced its Predictive AI Functions in beta for forecasting, anomaly detection, driver analysis and prediction through SQL. The set includes an updated ai_forecast plus ai_top_drivers, ai_detect_anomalies, ai_predict_value and ai_predict_class. Functions can be called in SQL; some are also exposed through the Genie interface.
OpenAI Economic Research published an analysis of usage across ChatGPT Business and Enterprise accounts in the European Union. The company reports that Codex and ChatGPT Work generated more than half of output tokens in its EU enterprise sample. Weekly active users of these tools at EU small and medium businesses grew more than eightfold from May 2026.
An Arm technical post dated October 13 argues that local AI requires choosing models, frameworks and processors across the whole workflow. It cites event demonstrations of a translation pipeline using more than 30 models and speech synthesis running on one Arm CPU core.
An AWS reference architecture dated October 12 processes public sequencing reads from NCBI SRA with HealthOmics and makes resistance-gene results queryable in natural language. It is a demonstration architecture, not new surveillance data or a clinical system.
An AWS technical guide dated October 12 explains using Bedrock global inference profiles from source Regions in Singapore, Jakarta, Malaysia, Thailand and Taipei. GPT-6 Astra is shown only from Singapore; the global profile can route processing to other supported destination Regions.
An AWS guide dated October 12 demonstrates a Ray LLM container setup on EKS that splits one LLM across two GPU nodes, running prefill and token generation separately. The post focuses on deployment and configuration.
An IBM Research explainer dated October 12 outlines methods in the llm-d framework, including cache-aware routing, KV-cache offloading, flow control and phase disaggregation. Performance figures are attributed to llm-d demonstrations cited in the post.
Alteryx and Google Cloud published an architecture example combining Live Query and BigQuery for unstructured-document workflows. Invoice PDFs remain in Cloud Storage; Alteryx authors the flow in a browser, while SQL pushdown runs transformations and reconciliation in BigQuery. Gemini AI.GENERATE extracts invoice fields in the worked example.
An AWS technical post dated October 9 describes how Postman’s Agent Mode scopes tools by task, uses schema-based reads and shapes context. Postman describes its developer community as 40 million.