Teaching AI How to Forget - The Data Exchange with Ben Lorica Recap

Podcast: The Data Exchange with Ben Lorica

Published: 2026-01-15

Duration: 44 min

Summary

In this episode, Ben Luria discusses the challenges of deploying AI in enterprises, focusing on the concept of 'unlearning' to enhance AI's reliability and trustworthiness. He explains how traditional methods fail to address core behavioral issues in AI models.

What Happened

Ben Luria, CEO and co-founder of Hirondo, joins the podcast to explain the pressing issue of AI reliability in enterprise settings. He emphasizes that while AI and large language models (LLMs) hold significant promise, they also pose inherent risks, such as bias and unexpected behaviors. Many enterprises are eager to adopt AI, but the potential for these risks often holds them back from deployment, particularly for mission-critical tasks. Hirondo's approach addresses the core problem of AI 'memory'—the fact that once AI learns something, it cannot easily forget it, leading to unintended consequences in behavior and decision-making.

Luria and his team, which blends non-technical and technical backgrounds, began their journey by interviewing various data science teams to identify common pain points. They discovered that when issues arise with AI models, it often becomes too late to resolve them effectively. This discovery informed Hirondo's focus on 'unlearning'—a process that allows AI to remove problematic data and behaviors from its models. Luria argues that traditional methods like context engineering and fine-tuning only address symptoms rather than the underlying issues, making them insufficient for ensuring AI models are trustworthy.

A unique metaphor used in the discussion is that of neurosurgery for AI, where the focus is on modifying the model's internal representations rather than just filtering external inputs. Luria's insights highlight a pivotal shift in the AI landscape, where understanding how to make AI forget certain learned behaviors could be the key to its safe and effective deployment in enterprises.

Key Insights

• The need for AI models to be trustworthy and deployment-ready is critical for enterprise adoption.
• Traditional methods like fine-tuning and guardrails often fail to address the core issues in AI behavior.
• Unlearning allows AI to remove problematic data and behaviors, enhancing reliability.
• A mix of technical and non-technical expertise can yield innovative solutions to complex AI challenges.

Key Questions Answered

What are the risks of AI in enterprises?

Ben Luria discusses various risks associated with AI, including its tendency to know and act on information it shouldn't. These risks manifest as biases, hallucinations, and vulnerabilities to attacks, making AI less reliable for mission-critical tasks. Enterprises are cautious in deploying AI due to these inherent risks, as they can lead to undesired and unexpected outcomes.

How does Hirondo define unlearning?

Unlearning, as defined by Luria, is the process of teaching AI to forget certain learned behaviors and problematic data. This goes beyond simple forgetting; it involves altering the model's internal weight structures to remove undesirable influences. The goal is to make the AI more reliable and trustworthy, especially in enterprise applications where stakes are high.

What traditional methods are used to address AI issues?

Luria identifies several common approaches, including context engineering and retrieval augmented generation (RAG), which attempt to mitigate problems by adding context to the AI's inputs. However, he points out that these methods often fail because they don't tackle the internal issues of the model itself. Instead, they merely filter or adjust external inputs without solving the underlying problems.

What is the significance of the neurosurgery metaphor?

The neurosurgery metaphor highlights the deep, internal changes that need to be made within AI models to improve their behavior. Just like neurosurgery involves intricate adjustments to the brain, unlearning requires precise modifications to the model's weights and internal representations. This metaphor underscores the complexity of the challenge and the innovative nature of Hirondo's approach.

What insights did Luria gain from interviewing data science teams?

Through interviews with various data science teams, Luria and his co-founders identified a common frustration: issues with AI models often become apparent too late to address effectively. This revelation guided Hirondo's focus on solving core problems within AI systems rather than applying superficial fixes, emphasizing the need for a deeper understanding of AI's learning and forgetting capabilities.