Imagine a scenario from science fiction: an AI capable of forecasting your actions before you make them, even in novel circumstances.
Israeli startup 42AI claims to be creating a “behavioral world model” with this exact capability, having recently emerged from stealth mode after securing $62.5 million in seed funding. Initially, the company is deploying its predictive system with several financial brokerages, asserting that the model can foresee how individual traders will react during fluctuating market conditions.
“Behavior is not set; it’s always probabilistic,” 42AI CEO and co-founder Amit Gal-Or explained to TechCrunch when pressed to elaborate on his company’s ambitious assertions. He contends that individuals possess fundamental psychological inclinations that shape their choices, irrespective of changing situations, and the startup aims to quantify these tendencies, transforming them into data consumable by the model.
“While you may evolve as a person in a week, your inherent tendencies will consistently manifest within specific domains with a given probability or distribution,” co-founder and CTO Ehud Ezra further noted. This, the company asserts, is the model’s function: to illustrate the likelihood of an individual making a particular choice.
42AI states that its model seeks to decipher individuals’ intrinsic motivations rather than merely inferring patterns from past behaviors. To achieve this, 42AI analyzes how people respond to events that appear unconnected. For instance, an investor or trader’s reactions to tasks like completing KYC, onboarding, support interactions, transferring funds, or executing stock trades could be considered.
“During the initial phase, when individuals interact with support or sales, we analyze sentiment, voice, and numerous other engagements, such as call frequency. All these data points are fed into the model. Subsequently, we observe interactions like buying and selling activities, their timing, and simultaneously consider the market conditions,” added CTO and co-founder Ehud Ezra.
He concluded, “Therefore, our aim is to understand their perception of the world. Within a world model, each individual essentially represents their own world.”
These observations are subsequently transformed into mathematical trait representations, which are then input into a transformer-based world model. This model then outlines the probabilities of various decisions in diverse situations. (A distinct language model is employed by the company to provide explanations for its predictions).
Ezra stated that the model is designed to maintain an individual’s inherent tendencies throughout simulations. He clarified that if a typically risk-averse person is suddenly portrayed as a risk-taker by the model, “the prediction is inaccurate,” detailing the incentive structure of the model’s training.
Over three years, 42AI cultivated this method, integrating psychological research with neuroscience and genetics. This development also incorporated insights from psychology and PTSD specialists, alongside Dan Shechtman, the recipient of the 2011 Nobel Prize in Chemistry.
Gal-Or explained that the team initially examined thousands of individuals through psychological assessments and behavioral situations. Subsequently, they incorporated government and biobank datasets encompassing over 2 million people. He added that their customer datasets have since grown to include nearly 30 million individuals.
During live deployments, the model processes existing customer records, such as transaction histories and support interactions. Gal-Or highlighted that the system does not require personally identifiable information, asserting that “the model never knows the individual’s identity.” Nevertheless, it appears the company still constructs distinct behavioral profiles by linking customer activities.
42AI asserts that its model achieves 98% accuracy in forecasting market trades. However, Gal-Or clarified that this metric primarily indicates whether trading volumes align with a predicted range, rather than perfectly predicting each individual trade.
He further mentioned that the model consistently reaches an F1 score between 0.8 and 0.85 when predicting trading direction and volume. (The F1 score is a machine learning metric evaluating a classification model’s effectiveness in balancing true positive predictions against false positives and negatives).
Given its nature as a confined system, the model cannot be universally applied. Gal-Or stated, “There’s an inherent limit because we truly lack access to every experience a person undergoes.” Ezra elaborated that the system’s precision diminishes when crucial situational context is overlooked, such as when an individual alters trading decisions based on input from a friend or coworker.
Gal-Or reported that 42AI is collaborating with ten major brokerages, which utilize its technology for managing liquidity, margin stipulations, and hedging risks. He noted that the company is attracting more clients and generating revenue exceeding “seven figures,” although he did not disclose an exact amount.
Clients compensate the startup either with a fixed fee or a portion of the value the platform helps them realize. He asserted that a single brokerage achieved an estimated $30 million in extra value by employing dynamic hedging.
Gal-Or clarified that the $62.5 million in seed funding, accumulated over three years, originated from the founders and Infinity Group, and was not exclusively sourced from external investors. His father, Amir Gal-Or, who also serves as 42AI’s chairman and a co-founder, established Infinity. 42AI intends to seek further investment.
Beyond the financial sector, 42AI is currently investigating potential uses for its model in gaming, fraud detection, and cybersecurity. Ultimately, the company aims to apply it to enable robots and AI agents to comprehend human behavior. However, this goal might be distant, as Ezra admitted that the model requires specific training on rules and context for every new application domain.
As with all AI technologies, 42AI’s innovation presents a potentially concerning aspect: if a brokerage can foresee customer actions, what prevents the exploitation of this insight for self-gain? Ezra conceded that this was “a valid concern” when questioned. The founders referenced regulations, oversight, and an internal ethics committee, yet failed to elaborate on what measures, if any, would prevent their clients from leveraging this technology to benefit at the expense of the profiled individuals.
Nonetheless, Gal-Or expressed confidence that the technology could mitigate such risks. He stated, “We are exceptionally proud to be part of the solution; in fact, we are the solution to this problem.”