Goldman Sachs Safeguards Apprenticeship Culture in AI Era
Goldman Sachs Marquee head Chris Churchman warns that overreliance on AI tools risks undermining institutional knowledge transfer.
PYMNTS AI·24 Aug 2026
Use this view to inspect the underlying evidence corpus. For ranked developments, decision posture and interpretation, use Signals.
Raw feed or Signals?
Raw feed is chronological evidence. Signals ranks and interprets material change.
Goldman Sachs Marquee head Chris Churchman warns that overreliance on AI tools risks undermining institutional knowledge transfer.
Synchrony announced an enterprise partnership with OpenAI to integrate card loyalty benefits into conversational AI purchasing workflows.
Chinese threat actors are integrating DeepSeek and open-source AI models into offensive cyber operations against foreign targets.
AWS introduced the Agent Registry and open Agentic Resource Discovery (ARD) standard for cross-environment agent governance.
Google Cloud released its State of AI Infrastructure report highlighting agentic security and governance challenges for enterprises.
AWS details patterns for using LLM agents and human-in-the-loop workflows to automate enterprise metadata harmonization.
NVIDIA claims its Blackwell and upcoming Vera Rubin GPU architectures significantly reduce power consumption for agentic inference workflows.
Starling Bank has launched an agentic AI assistant available to all of its business banking customers.
Hugging Face is reportedly exploring a sale that could value the open-source AI platform at $13 billion.
Agentic AI deployments introduce unpredictable API and compute consumption patterns, complicating enterprise financial planning for the next fiscal year.
Financial APIs are increasingly queried by autonomous AI agents, disrupting legacy assumptions about human-driven API interactions.
Researchers introduced aiXamine, a black-box framework evaluating trade-offs across safety, security, and privacy in LLMs.
Research demonstrates that clustered calibration data inflates effective sample sizes, undermining safety filter and conformal thresholds.
GroupSegment-SHAP introduces a Shapley-based framework for multivariate time-series models to explain spatial and temporal interaction signals.
Researchers introduced a task-exchangeability framework for statistical inference using synthetic data and LLM-as-a-judge evaluators.
Research shows LLMs retain internal representational biases even when passing standard behavioral bias evaluations.
Research on 132,000 prompt variants reveals token-level scaling laws governing prompt brittleness and lexical sensitivity in LLMs.
Research proposes OneModel, replacing multi-component agent pipelines with a single model that internalizes complex business workflows.
Researchers introduce AgenticRAG-FP, a benchmark for causal failure attribution and root-cause error tracking in multi-hop agentic RAG.
Researchers introduced Quantization-Aware Healing, a training recipe that recovers reasoning and math performance in 4-bit compressed LLMs.
Research proposes a method to preserve model uncertainty and abstention behavior during LLM quantization using target-aware calibration data.
Research introduces scenario-level out-of-distribution benchmarks to prevent false confidence in SMS and voice phishing detection models.
ArXiv research demonstrates that user emotional context amplifies sycophantic behavior and confirmation bias in LLM evaluations.
Research identifies a memory failure in agentic models where contextual prerequisite blocks are evicted before retrieval occurs.
Researchers tested eight open-weight models and found zero empirical ability for models to accurately report on their internal computation state.
Researchers introduced an ontology-driven RAG framework designed to ensure source traceability and auditability in enterprise FP&A workflows.
Researchers proposed an Evaluation Agent middleware to detect knowledge poisoning and factual inaccuracies in RAG retrieval outputs.
arXiv research reveals LLM judges show bias when evaluating functionally correct code that has superficial formatting or style variations.
New research analyzes LLM library hallucinations in code generation, highlighting software supply chain risks like slopsquatting.
Researchers benchmarked LLM agent-generated code security across 186 real-world software engineering tasks using the SUSVIBES benchmark.
© 2026 OneBench: AI Insights. All rights reserved.
Evidence before opinion