(6) Techniques to improve generative AI model output

A limitation of foundation models that performance relies heavily on data. Biased or incomplete data will affect their outputs.

Data Dependency.

A limitation of foundation models that AI models are trained up to a specific knowledge cutoff date, meaning they might lack information about events after that point

Knowledge cutoff

A limitation of foundation models that LLMs learn from large datasets which may contain untrue prejudice or favoritism.

Bias

A limitation of foundation models that assessing the fairness of generative AI models is a key aspect of responsible development.

Fairness

A limitation of foundation models that when AI models produce outputs that aren't accurate or based on real information.

Hallucinations

A limitation of foundation models that rare and atypical scenarios can expose a model's weaknesses, leading to unexpected results

Edge cases

A process where human input and feedback are directly integrated into ML workflows.

Humans in the loop (HITL)

Allows you to keep track of different versions of the model with Model Registry.

Versioning

Allows you to review the model metrics to check the model's performance.

Performance tracking

Allows you to watch for changes in the model's accuracy over time with Model Monitoring.

Drift Monitoring

Allows you to use Agent Platform Feature Store to manage the data features the model uses.

Data management

Allows you to use Model Garden to store and organize the models in one place.

Storage

Allows you to use Agent Platform Pipelines to automate your machine learning tasks.

Automate

The amount of text the model can consider

Context Window

A technique used to enhance a pre-trained or foundation models’ performance for specific tasks or domains.

Fine-tuning