Keynote, Conference Presentation
Atindriyo Sanyal, CEO of Galileo: Making Enterprise Gen AI “Reliable”
- The majority of Generative AI applications are expected to remain in internal or prototype phases for an extended period due to the difficulty in predicting model behavior.
- LLM technology is predicted to become strictly smarter within the next six to 18 months, yet leadership confidence in shipping applications is dwindling despite this capability growth.
- Future AI applications will evolve from linear workflows to complex, dynamic agentic systems communicating via protocols like MCP, increasing the difficulty of de-risking errors and preventing recurrence.
- In the agentic era, stakes are projected to be very high as systems transition from theoretical operations to executing actions such as financial transactions and insurance decisions.
- Approximately 90% of Retrieval-Augmented Generation (RAG) failures are estimated to stem from data quality issues, with the remaining 10% caused by improper instructions.
- The pressure to deliver precise information and instructions to models is expected to intensify as model intelligence increases and agent paths become more compounded by user queries.
- Tool selection is identified as a significant challenge in modern agentic systems due to semantic overlap within growing tool stores.
- An Anthropic blog from December 2024 suggests that 80% of agents will operate using 10 to 15 traditional design patterns, while the remainder require customized architectures.
- A new metric lifecycle is anticipated to progress from a baseline evaluation yielding seven out of 10 accuracy, to instruction tuning reaching 90–95% accuracy, and finally requiring fine-tuning for last-mile precision.
- Developers utilizing traditional LLM judges for production evaluations are predicted to quickly encounter limits imposed by rate restrictions, context length constraints, high latency, and cost.
- Small language models (Luna) with three to eight billion parameters are expected to resolve the conflict between latency and accuracy, providing guardrails in under 200 milliseconds.
- Reliance on third-party endpoints for LLM judges is viewed as a risk that places developers at the mercy of external service constraints.