Andrej Karpathy, who recently departed from OpenAI to launch his own AI education company, continues to focus on advancing AI applications.

He emphasizes that interactions with AI systems are fundamentally based on averaged responses from human data labelers, rather than being driven by a magical AI system.

Karpathy explains that the fine-tuning of large language models (LLMs) involves training on extensive datasets and conversations, with human annotators guiding the responses to enhance user experience.

This fine-tuning process has made interactions feel more personable and less mechanical, contributing significantly to the success of models like ChatGPT.

For specialized inquiries, companies often employ expert data labelers, such as physicians for medical questions and mathematicians for math-related queries.

While LLMs can produce useful responses, they may not always guarantee expert-level answers due to limitations in their underlying knowledge and reasoning capabilities.

Karpathy advises against using AI for complex policy issues, likening it to asking a random individual to research the answer within constraints.

He illustrates this point by explaining that when users inquire about topics like 'top 10 sights in Amsterdam,' the AI generates responses based on previous human responses to similar questions.

For queries not found in the training data, the AI creates answers by mimicking human patterns learned from earlier data.

Responses to controversial topics often include neutral phrases, a result of guidelines given to human labelers during the fine-tuning process.

Karpathy critiques the method of reinforcement learning from human feedback (RLHF) as a temporary solution, noting its lack of objective success criteria compared to more definitive systems like DeepMind's AlphaGo.