Yashar Delju

AI expert and author specializing in generative AI, language agents, and recommender systems.

Mentions

Appeared in

Discussed in

Key positions and views

The primary function of recommender systems is evolving from item ranking to complex task execution, orchestrated by AI agents.

While LLMs offer powerful new capabilities, they introduce dangerous risks like persuasive hallucinations and exaggerate existing problems like bias, requiring new evaluation dimensions.

AI agents are superior to standard LLMs for complex tasks because they can leverage external tools and maintain multi-layered memory.

Mature technologies like collaborative filtering should not be discarded; they provide a reliable, high-quality baseline that learns from a company's proprietary data.

LLMs can be applied in recommender systems through three main paradigms: complete replacement of old systems, augmentation of existing systems (e.g., for data generation), or for simulation and evaluation purposes.

Podcast consensus on Delju

Points of consensus

▶Large Language Models (LLMs) introduce novel risks like hallucination and context drift while simultaneously amplifying pre-existing issues such as fairness and bias, necessitating more complex evaluation methods.May 2026

▶The fundamental purpose of recommender systems is undergoing a paradigm shift, moving away from simple item ranking towards executing complex, multi-step tasks for users, driven by AI agents.May 2026

▶AI agents represent an evolution of standard LLMs, distinguished by their key capabilities to leverage external tools and services and to maintain various forms of memory (working, episodic, semantic).May 2026

▶While LLMs bring powerful new capabilities by connecting internal data to external world knowledge, older, mature models like collaborative filtering remain valuable as a trusted baseline for quality.May 2026

Points of debate

▶Delju champions the transformative potential of LLMs and agentic systems, yet simultaneously cautions against abandoning mature, reliable technologies like collaborative filtering, suggesting a hybrid future rather than a complete replacement.

▶He highlights the power of LLMs to augment systems by drawing on vast external knowledge, while also warning that this same reliance on unregulated internet data is the source of exaggerated stereotypes and biases.May 2026

▶Delju describes a field in rapid flux, noting the massive impact of LLMs since 2023, while also observing that core areas like collaborative filtering have reached a plateau with few recent 'mind-blowing' innovations.May 2026

Key themes

▶The Agentic Transformation of Recommender SystemsMay 2026

This theme covers the shift in recommender system functionality from ranking items to performing complex tasks. Delju posits that AI agents, defined by their use of tools and memory, will orchestrate this change, operating under paradigms of system replacement, augmentation, or simulation.

Investors should look for companies building platforms that support this task-oriented, agentic approach, as it represents a fundamental expansion of the market from simple recommendations to automated personal assistance.

▶The Double-Edged Sword of Large Language ModelsMay 2026

Delju details both the benefits and dangers of integrating LLMs. While they can connect in-domain data to external world knowledge, they also introduce new risks like hallucination and context drift, and amplify existing ones like bias due to their training on unregulated internet data.

Analysts must heavily discount the value of AI products that lack robust, transparent mechanisms for mitigating hallucination and bias, as these risks pose significant reputational and operational threats, especially in sensitive sectors.

▶A Pragmatic Approach to AI IntegrationMay 2026

This theme reflects Delju's balanced perspective, arguing against the complete abandonment of established models. He views mature techniques like collaborative filtering as a 'trustable source' that provides a baseline level of quality, suggesting a hybrid approach where new LLM capabilities augment, rather than entirely replace, proven systems.

Companies that can successfully integrate cutting-edge LLMs with their existing, mature AI models will likely have a competitive advantage, as they can leverage new features without sacrificing the reliability of their core systems.

▶Evolving AI Architectures and EvaluationMay 2026

Delju outlines the technical underpinnings of modern AI, from the three-stage training process of early LLMs to the multi-component memory systems of AI agents. Consequently, he argues that evaluation must also evolve beyond simple accuracy to encompass new dimensions like hallucination, latency, and fairness.

The increasing complexity of AI systems and their evaluation metrics creates a market opportunity for specialized AI auditing and testing services that can validate model trustworthiness across these new dimensions.

Source episodes

Sentiment over time

Not enough data for timeline

Changes over time

Pre-2023

Describes a period where collaborative filtering models for recommender systems had reached a high level of maturity, with few 'mind-blowing' new models being proposed in the preceding years.

2023

Identifies 2023 as a pivotal year when the emergence of large language models like ChatGPT began to significantly impact every subfield of AI, including recommender systems.

Present

Discusses the current era of integrating LLMs and agents into systems, focusing on new capabilities like data augmentation and new challenges such as managing hallucination, context drift, and exaggerated biases.

Future

Predicts the core function of recommender systems will transform from ranking items to performing complex tasks, driven by the advanced capabilities of language agents.

2026

States the expected publication date for his educational book on generative AI and language agents, which will include Python exercises.

Suggested prompts

How does Yashar Delju's advocacy for retaining collaborative filtering models contrast with his vision for a future dominated by agentic systems? &nearr;What are the practical steps a company should take to mitigate the 'new' and 'exaggerated' risks Delju identifies when integrating LLMs into their recommender systems? &nearr;According to Delju's framework, how would the use of an AI agent for simulation-based evaluation change the traditional A/B testing lifecycle for recommender systems? &nearr;How does Delju's distinction between AI agents and standard LLMs (tools and memory) inform the architectural choices for building next-generation recommender systems? &nearr;

Key concepts

Large Language Models (LLMs) 1 ep Recommender Systems 1 ep AI Agents / Language Agents 1 ep AI Risks & Trustworthiness 1 ep Hallucination 1 ep Collaborative Filtering 1 ep AI Evaluation 1 ep Adversarial Attacks 1 ep Fairness & Bias 1 ep

Notable quotes

“Now, instead of do this ranking for me, we can say with large language models and language agents where we are heading to, and it's being transformed to, please do these tasks for me. So from rank, you're going to tasks.”

Yashar Delju · The Future is Agentic in Recommender Systems

“one of the bad things about this hallucination is that they look natural. So if they are used in task-sensitive applications like medical domain, you may see a system producing drug or medicine suggestions, which is wrong, but they don't seem wrong. And this is something that's scary because they look very persuasive.”

Yashar Delju · The Future is Agentic in Recommender Systems

“Since 2023, the emergence of large language models like ChatGPT has significantly impacted every subtopic within artificial intelligence, including recommender systems.”

Yashar Delju · The Future is Agentic in Recommender Systems

“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.”

Yashar Delju · The Future is Agentic in Recommender Systems

Report last updated: May 5, 2026

Get started free

Back to Entities Intelligence Report