Keep pulling the thread on Yashar Delju.
Yashar Delju predicts that the core function of recommender systems is shifting from ranking items to performing complex tasks for the user, driven by the capabilities of large language models and language agents.
A significant danger of large language model hallucinations is that the incorrect information they produce often appears natural and persuasive, posing a high risk in sensitive applications like healthcare.
Since 2023, the emergence of large language models like ChatGPT has significantly impacted every subtopic within artificial intelligence, including recommender systems.
A targeted adversarial attack on a recommender system can involve a company dishonestly altering its product descriptions to penetrate a business market and promote its items for economic gain.
Hallucination is a new risk in large language models where a system might recommend items that do not exist or provide incorrect metadata for existing items.
Context drift is a risk in large language models where the model fails to remember the user's original goal during a conversation, causing it to change the topic.
Risks associated with large language models can be categorized into two types: new emerging risks like hallucination, and exaggerated pre-existing risks like fairness and biases.
Large language models are prone to projecting stereotypes, thereby exaggerating existing fairness and bias risks, because they are trained on vast amounts of unregulated internet data.
Collaborative filtering models for recommender systems have reached a high level of maturity, with few 'mind-blowing' new models being proposed in the last couple of years.
Large language models can augment recommender systems by connecting in-domain data to external world knowledge, enabling the extraction of new information and relationships between entities.
Yashar Delju argues that collaborative filtering models should not be abandoned, as they remain a trustable source that provides an acceptable minimum level of quality by learning from a company's own data.
Early large language models typically underwent three levels of training: pre-training on a large corpus of internet data, instruction tuning on question-answer pairs, and alignment with human feedback to make responses more human-like.