Lina Irawan
The technological landscape is currently dominated by Large Language Models (LLMs), yet emerging research and industry insights suggest these models face significant hurdles when applied to tasks outside the realm of linguistic synthesis, particularly in predicting human intent and future behavior. Frank Portman, Chief Technology Officer at Yobi, recently highlighted these limitations during an in-depth technical discussion, arguing that while LLMs excel at generating text and synthesizing information within a specific context, they lack the "inductive bias" necessary for robust decision-making under uncertainty. As a behavioral AI company, Yobi is pioneering a different path by building foundation models specifically designed to forecast future actions in sectors such as advertising technology, marketing, and fraud detection. This shift marks a critical juncture in the evolution of artificial intelligence, moving from models that understand what we say to models that anticipate what we will do.
The Limitations of Generative AI in Intent Forecasting
The meteoric rise of generative AI has led many organizations to attempt to use LLMs as "next-token predictors" for non-linguistic data, such as consumer behavior sequences or steps in a business process. However, the fundamental architecture of an LLM is optimized for language. Portman observes that the inductive bias—the set of assumptions a learner uses to predict outputs for given inputs—inherent in LLMs is derived from the world’s collective text and video. This makes them phenomenal at writing code or mimicking Shakespearean prose, but less effective at the nuanced task of forecasting a consumer’s next purchase or an advertiser’s optimal bid in a real-time auction.
Decision-making under uncertainty requires a base representation that is "broadly predictive" of future behavior rather than just being "pleasant or correct in conversation." While LLMs are being integrated into "agentic" workflows—where AI agents take actions on behalf of users—the underlying decision layer often requires a specialized model. Portman argues that the scale of behaviors in the global digital economy is three orders of magnitude larger than the hundreds of thousands of tokens typically found in a language model’s embedding table. This disparity in scale necessitates a specialized foundation model for behavior that can handle the high cardinality of discrete actions, such as website visits, product interactions, and hashed user identifiers.
Chronology of Technical Development and the Shift to Behavioral Models
The transition toward specialized behavioral models is rooted in the history of machine learning and the academic backgrounds of its current leaders. Portman’s own trajectory—from a student of pure mathematics to a CTO in the AI space—mirrors the industry’s shift from theoretical modeling to applied engineering. In the early 2010s, as data science and machine learning became industry buzzwords, the focus was on matrix multiplication and statistical foundations. The publication of the seminal paper "Attention Is All You Need" in 2017 provided the transformer architecture that now powers both LLMs and Yobi’s behavioral models.
However, the application of these architectures has diverged. While the "Frontier Labs" like OpenAI and Google DeepMind focused on general-purpose linguistic intelligence, companies like Yobi began focusing on proprietary, sensitive data that is often identifiable to anonymous IDs or browser sessions. Over the last four years, the industry has seen a move away from "walls of if-statements" and manual heuristics toward automated, scalable models that can fine-tune on per-outcome bases, such as specific marketing campaigns or fraud risk profiles. This chronology suggests that the "second-mover advantage" in AI belongs to those who can apply transformer-based architectures to specialized, high-value data sets that general-purpose LLMs cannot access or interpret effectively.
Infrastructure and Engineering for High-Frequency Decisioning
Building a foundation model for behavior presents unique engineering challenges, particularly regarding inference latency and data throughput. Yobi’s current stack operates at a scale of millions of queries per second (QPS), a necessity in the high-speed environment of ad servers and real-time bidding. To maintain this performance, engineers must balance architectural efficiency with the need for rapid deployment.
Two primary strategies have emerged to handle this scale:
- Pre-computation and Embedding Lookups: Rather than generating a response on the fly—as a chatbot does—behavioral models often rely on pre-computed feature representations. By trading memory for latency, these systems can look up complex embeddings in a fraction of a millisecond, allowing for real-time personalization.
- Advanced Batching and Queuing: Processing a million requests individually is computationally expensive. Modern behavioral AI stacks utilize sophisticated batching mechanisms that allow GPUs to process hundreds of requests simultaneously without a significant increase in latency, maximizing the margin and efficiency of the software business model.
Furthermore, the choice between "inductive" and "transductive" models is a critical architectural decision. A transductive model is limited to the state it was trained with, whereas an inductive model can generalize to new nodes or users it has never seen before. In the fast-moving world of digital behavior, where new websites are created and user paths diverge daily, building inductive architectures is essential for long-term viability.
The Privacy Mandate: Differential Privacy and Homomorphic Encryption
As AI moves deeper into behavioral prediction, the sensitivity of the data involved raises significant privacy concerns. Unlike general-purpose LLMs that train on public internet data, behavioral models often handle proprietary consumer data. This has led to the emergence of "privacy-preserving machine learning" as a core pillar of the industry.
Portman identifies two key academic concepts currently making their way into industrial applications:
- Differential Privacy: This mathematical framework ensures that the output of an algorithm does not reveal whether a specific individual’s data was included in the training set. A common metric is "K-anonymity," which ensures an individual is indistinguishable from a group of "K" other people, typically ranging from 5 to 25. While widely used in data analysis, applying differential privacy to the actual training of deep learning models remains a frontier challenge.
- Homomorphic Encryption: Often described as the "holy grail" of privacy, this allows for "homomorphic machine learning," where models can be trained on encrypted data without ever "seeing" the raw information. While currently computationally expensive and largely academic, its integration into behavioral AI could allow companies to build predictive models while maintaining total data sovereignty.
Broader Impact and Market Implications
The implications of behavioral AI extend far beyond the $600 billion global advertising market. While ad tech provides the economic engine for initial development, the technology is rapidly expanding into "MarTech" (Marketing Technology). This shift focuses on deepening relationships with existing customers through personalized product recommendations, SMS, and email engagement—methods that consumers often find less intrusive than traditional third-party advertising.
Beyond commerce, behavioral AI is poised to become the "decision layer" for the next generation of AI agents. As the industry moves toward "agentic" systems—AI that can schedule meetings, write code, or manage budgets—the need for a model that can forecast the outcome of a specific action becomes paramount. An LLM might be able to write the code for a marketing automation tool, but a behavioral foundation model is required to decide how much should be spent and which audience should be targeted based on expected value.
The rise of these specialized models also serves as a check on the "offloading" of human decision-making. As Portman notes, there is an internal industry debate regarding whether the use of agentic tools for deep-thought tasks, like coding, might "dull" human cognitive abilities over time. By focusing on the "unsent problem"—solving for the correct prediction rather than just the most likely next word—behavioral AI aims to provide a more rigorous, mathematically grounded approach to the future of human-machine interaction.
In conclusion, the path forward for artificial intelligence lies in a dual-track development: general-purpose LLMs for communication and synthesis, and specialized behavioral foundation models for prediction and decisioning. As the industry matures, the integration of these two paths, supported by robust privacy-preserving technologies, will likely define the next decade of technological innovation. The focus is no longer just on making AI smarter, but on making it more predictive and more respectful of the sensitive data that drives the modern economy.
