NVIDIA
ideas
Here are some ideas and potential areas of research for Tensort:
-
Model analysis and interpretability: Develop new techniques for analyzing and understanding what large language models have learned from their training. This could help ensure they behave as intended.
-
Constitutional AI techniques: Apply concepts from Stability AI's Constitutional AI research like value specification languages, capability restrictions, and alignment incentives to improve Tensort's safety.
-
Interactive debriefing tool: Create an interactive interface that allows users to have detailed conversational debriefing sessions with Tensort to understand its capabilities and limitations.
-
Self-supervised pretraining methods: Research how to utilize self-supervised pretraining approaches like contrastive learning to implicitly teach helpfulness, harmlessness and honesty without labeled data.
-
Adversarial robustness: Make Tensort more robust to adversarial examples, ambiguous queries, and attempts to deviate it from its intended purpose.
-
Helpfulness evaluation: Develop better automatic and human-based evaluation methods to quantitatively measure how helpful Tensort is across different contexts and demographics.
-
Stakeholder values: Conduct values research to better understand the priorities and concerns of different potential stakeholder groups interacting with an AI system like Tensort.
-
Lifelong learning: Research how to continuously update Tensort's knowledge and abilities over its lifetime through reinforced self-supervised learning from interactions.
-
Explainability of reasoning: Improve Tensort's ability to transparently explain to users the reasoning behind its responses and recommendations.
Tensorrt is just a infer engine, according to your ideas, tensorrt should have an embedded llm to drive itself.
As said, TensorRT is mainly for inference optimizations. For training related issues please refer to each frameworks' approach. Thanks!
