Yobi CTO Frank Portman Argues LLMs Fall Short for Intent Prediction

Yobi CTO Frank Portman has highlighted the limitations of current Large Language Models (LLMs) when it comes to predicting human intent, arguing that their core training mechanism of next-token prediction is not an ideal fit for this complex task. In a recent interview on the Stack Overflow podcast, Portman explained that while LLMs excel at language synthesis and generation, their inductive bias is not optimized for forecasting behavior or making decisions under uncertainty.

Portman, whose company Yobi builds “foundation models of behavior,” distinguished their approach from the general-purpose nature of LLMs. He suggested that the massive datasets and training methodologies for LLMs, focused on predicting the next word or token, do not inherently equip them for the nuances of understanding and predicting human actions.

The Inductive Bias Challenge

The core of Portman’s argument lies in the “inductive bias” of machine learning models. This refers to the set of assumptions a model makes to learn from data and generalize to new, unseen data. For LLMs, this bias is geared towards linguistic patterns.

“It’s not clear to me that the inductive bias of let’s gather all the texts in the world… get very good at predicting sort of sometimes very long sequences… would make these LLMs good for… forecasting or prediction or decision making,” Portman stated. He contrasted this with Yobi’s focus on building a “foundation model of behavior,” which requires different data inputs and perhaps different architectural approaches.

While LLMs can perform emergent tasks like writing code or creative text, Portman suggested that decision-making under uncertainty, a key aspect of intent modeling, is a fundamentally different problem. He believes that simply fine-tuning an LLM for conversational tasks does not imbue it with the necessary capabilities for accurate behavioral prediction.

Specialized Models for Behavior

Yobi’s approach involves using transformers and graph neural networks, rather than chat-style LLMs, to create what Portman describes as a “foundation model of behavior.” This specialized model aims to capture the complexities of human actions and predict future behavior, which is crucial for applications in ad tech, marketing, and personalization.

Portman also touched upon the practical challenges of running such systems at scale, noting Yobi’s ability to handle millions of personalization decisions per second while prioritizing consumer data privacy. This suggests that their specialized models are not only theoretically distinct but also practically optimized for real-world performance and ethical considerations.

The Role of LLMs in Agentic Systems

Despite his critique of LLMs for direct intent prediction, Portman acknowledged their potential when combined with other tools. He believes that LLMs can be valuable components within larger “agentic” systems, rather than being the sole solution. This perspective aligns with the growing trend towards multi-tool AI agents that leverage the strengths of different models and technologies for specific tasks.

Portman’s background, rooted in mathematics before transitioning to software and machine learning, provides a strong foundation for his analytical view on model architectures and their suitability for different problems. He sees software problem-solving as a series of puzzles, a mindset that likely informs Yobi’s development of specialized behavioral models.

Key facts

The implications for the AI industry are significant. As companies increasingly seek to understand and predict user behavior for personalized experiences and targeted services, relying solely on general-purpose LLMs may prove insufficient. The development of specialized behavioral AI models, as pioneered by Yobi, could become a critical differentiator. This shift suggests a move away from a “one LLM to rule them all” mentality towards more tailored AI solutions for specific domains, particularly where understanding intent and predicting actions is paramount.

Source: Stack Overflow Blog – Why intent prediction needs more than an LLM – https://stackoverflow.blog/2026/06/30/why-intent-prediction-needs-more-than-an-llm/