This is a question from SF. What if AI could predict your behavior before you decide to do it, even in completely new situations?
Israeli startup 42AI says it’s developing a “behavioral world model” that can do just that, and has just emerged from stealth with $62.5 million in seed funding. The company is initially running its predictive system at a small number of financial brokerage firms, and says the model helps predict how individual traders will react to volatile situations.
“Behavior is never deterministic. It’s always probabilistic,” 42AI CEO and co-founder Amit Gal-Or told TechCrunch when I asked him to justify his company’s lofty claims. He argues that people have underlying psychological tendencies that influence their decisions even when circumstances change, which the startup is trying to measure and translate into data that models can understand.
“A week from now, you may be a different person, but your tendencies will always be reflected in the probabilities and distributions in certain areas,” added his co-founder and CTO Ehud Ezra. And the company claims that what the model does is show the probability that a person will make a certain decision.
42AI says the model seeks to uncover people’s inner motivations, rather than inferring patterns based on their past behavior. To do this, 42AI investigates how people react to seemingly unrelated events. For example, for investors and traders, such events may include completing a KYC process, onboarding, support calls, sending money, trading stocks, etc.
“When that person is talking to support or sales in the early process (…), we look at their sentiment, their voice, the number of interactions they have, the number of calls they make, etc.,” added CTO and co-founder Ehud Ezra. “And all those data points are fed into the model. Then there’s the interaction of how to buy and sell, when to do it, etc. We’re looking at the market at that stage.”
“So we’re trying to see how they see the world. In the world model, every individual is the world,” he added.
The company then converts these observations into mathematical representations of properties and feeds them into a transformer-based model of the world to show the probabilities of possible decisions under different circumstances. (The company uses a separate language model to generate prediction explanations.)
Ezra said the model is trained to maintain fundamental human tendencies during simulations. So if someone is normally risk-averse and the model suddenly treats them as a risk-taker, “you’re off target,” he said, explaining how the model’s training rewards are structured.
42AI spent three years developing this approach, combining psychological research with neuroscience and genetics, as well as advice from psychology and PTSD experts and 2011 Nobel Prize in Chemistry winner Dan Shechtman.
Gal-Or said 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. Since then, the company’s customer dataset has expanded to cover about 30 million people, he said.
In a live deployment, the model analyzes the customer’s existing records, such as transactions and support calls. Gal-Or pointed out that the system does not require any personally identifying information, arguing that “the model will never know who the person is.” However, the company still appears to be creating behavioral profiles of individuals from linked customer activity.
While 42AI claims its model achieves 98% accuracy in predicting market trades, Gal-Or clarified that this is a measure of whether trade flow is within the predicted range, not whether each trade was predicted correctly.
He also said the model consistently achieves F1 scores of 0.8 to 0.85 for predicting trade direction and volume. (The F1 score is a machine learning metric that measures how well a classification model balances correct predictions with false alarms or missed events.)
And considering this model is still a boxed machine, it won’t work everywhere. “There are natural limits because you don’t really have access to everything that people experience,” Gal-Or says. Ezra explained that the system’s accuracy also decreases when it misses important situational context, such as when changing trade selections based on advice from friends or colleagues.
Gal-Or said 42AI works with 10 large brokerage firms, which rely on 42AI to manage liquidity, margin requirements and hedging exposures. He said the company is gaining more customers and generating more than “seven figures” in revenue, but he declined to provide numbers.
Customers pay startups a flat fee or a percentage of the value that the platform helps them capture. He claimed that one brokerage firm generated approximately $30 million in additional value through dynamic hedging.
Gal-Or said the $62.5 million in seed funding was raised over a three-year period from the company’s founders and Infinity Group, and not entirely from outside investors. Infinity was founded by his father, Amir Ghaluoğlu, who is also the chairman and co-founder of 42AI. 42AI is currently planning to raise further funds.
Beyond finance, 42AI is currently exploring applications for its model in gaming, fraud prevention and cybersecurity, and hopes to one day use it to help robots and AI agents understand humans. But that may still be a long way off. Ezra acknowledged that models need to be trained based on rules and context for each new domain.
And, like any AI tool, there are potentially troubling aspects to 42AI’s technology. If a brokerage firm can predict the behavior of its customers, how can it prevent someone from using that knowledge for their own benefit? When I raised the issue, Ezra acknowledged that this was a “legitimate concern,” and the founders cited regulation, oversight, and internal ethics committees, but did not explain whether there was anything stopping customers from using the technology to gain an advantage over the very people they were profiling.
However, Gal-Or was confident that this technology could help address such risks. “We’re really proud to be part of the solution. We’re the solution to this problem,” he said.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
