Učitavanje video playera...
00:51:50
KulenDayz 2024 - Enhancing LLM Capabilities: Advanced Techniques and Strategies
KulenDayz 2024
Enhancing LLM Capabilities: Advanced Techniques and Strategies
Antonio Krizmanić
The saying "garbage in - garbage out" has never before manifested as vividly and concretely as it did in the world of generative AI. Although the hype surrounding LLM's is about-a-year in the making, it already became exhausting to keep track with the best practices in achieving the state-of-the-art performance of the AI agents that businesses are motivated to build and deploy.
[BEYOND THE RAG] Prompt engineering techniques and semantic similarity searches became the cornerstone of AI agent development, currently being offered and/or used by every AI enthusiast on the market. However, the saturation of the offer is often not justified - PE + SSS is not enough in most cases, especially if LLM's are used for internal process automatization. Advanced RAG methods will be presented, together with how they integrate into prompt pipelines, with real-world examples and science to back them up.
[IT TAKES A VILLAGE] Similar to a deep singular tree v. shallow random forest algos, however big the token limit gets (current standard is above 100k), combining multiple 'experts' will result in a better performance than giving all of the instructions to a singular agent. Learn about 'mixture of experts', 'task delegations' and 'agent swarms' and their ability to improve performance, how to build them, where are they implemented and how they can save you money.
[IN AND OUT OF THE PIPELINE] Get some practical ways to leverage multi-agent systems and boost your performance, both inside the pipeline (on-line processing of the query) and outside of the pipeline (ML model training for future usage). Learn about planners, troubleshooters, ensemblers, primers, self-informers and synthetic data generators.
Enhancing LLM Capabilities: Advanced Techniques and Strategies
Antonio Krizmanić
The saying "garbage in - garbage out" has never before manifested as vividly and concretely as it did in the world of generative AI. Although the hype surrounding LLM's is about-a-year in the making, it already became exhausting to keep track with the best practices in achieving the state-of-the-art performance of the AI agents that businesses are motivated to build and deploy.
[BEYOND THE RAG] Prompt engineering techniques and semantic similarity searches became the cornerstone of AI agent development, currently being offered and/or used by every AI enthusiast on the market. However, the saturation of the offer is often not justified - PE + SSS is not enough in most cases, especially if LLM's are used for internal process automatization. Advanced RAG methods will be presented, together with how they integrate into prompt pipelines, with real-world examples and science to back them up.
[IT TAKES A VILLAGE] Similar to a deep singular tree v. shallow random forest algos, however big the token limit gets (current standard is above 100k), combining multiple 'experts' will result in a better performance than giving all of the instructions to a singular agent. Learn about 'mixture of experts', 'task delegations' and 'agent swarms' and their ability to improve performance, how to build them, where are they implemented and how they can save you money.
[IN AND OUT OF THE PIPELINE] Get some practical ways to leverage multi-agent systems and boost your performance, both inside the pipeline (on-line processing of the query) and outside of the pipeline (ML model training for future usage). Learn about planners, troubleshooters, ensemblers, primers, self-informers and synthetic data generators.
