What it says it is building
According to TechCrunch on November 2, Israeli startup 42AI has come out of stealth with $62.5 million in seed funding. What it is building is a "behavioral world model": instead of extrapolating from what a person has done before, it tries to infer their underlying motivations and psychological tendencies, then predict the probability of their different decisions in new situations. The system is initially running at a handful of financial brokerages, predicting how individual traders will react in volatile markets.
CEO Amit Gal-Or puts it as "behavior is not deterministic; it's constantly probabilistic", and CTO Ehud Ezra adds that a person may be different a week from now but their tendencies will keep playing out according to some probability distribution. In practice, it draws signals from seemingly unrelated interactions: KYC and onboarding, the sentiment, voice and number of calls with support or sales, money transfers, and when people buy and sell. These observations are converted into mathematical representations of traits and fed into a transformer-based world model that lays out probabilities for different decisions under different circumstances, with a separate language model generating explanations. The training rewards are designed to preserve a person's underlying tendencies: if the model suddenly treats a usually risk-averse person as a risk-taker, "the target is off."
Where the data comes from, and how accurate it is
The company says it spent three years developing the approach, combining psychology, neuroscience and genetics research, with advice from experts in psychology and PTSD and from Dan Shechtman, winner of the 2011 Nobel Prize in Chemistry. The team first studied thousands of people through psychological assessments and behavioral scenarios, then added government and biobank datasets covering more than 2 million people; customer datasets have since grown to almost 30 million people. Gal-Or stresses that the system doesn't need personally identifiable information and that "the model never knows who the person is", though TechCrunch observes that the company still builds individual behavioral profiles from linked customer activity.
On accuracy, 42AI claims 98% in predicting market trades, but Gal-Or clarified that this measures whether trading flows fall within a predicted range, not whether each trade was predicted correctly; the model's F1 score for predicting trading direction and volume is consistently 0.8 to 0.85. The company also acknowledges a natural ceiling: the model can't see everything a person experiences, and accuracy drops when it misses context such as a trader changing course on advice from a friend or colleague.
What to watch
All of the methods and numbers above are the company's own account, and TechCrunch does not mention independent validation. In a brokerage's hands, a model like this could help manage liquidity and risk exposure, but it could also be used to identify which customers are more likely to take certain actions under pressure, and the report gives no detail on how the company guards against the latter. For readers following financial AI, the news is useful mainly for showing how the term "world model" is being applied to predicting individual behavior, and what a number like "98% accuracy" actually measures.