Here's a question that seems to come out of a science fiction novel: What if artificial intelligence could predict what you're going to do next before you even make a decision, even in a completely unfamiliar situation?

Israeli startup 42AI claims to be developing a “Behavioral World Model” that can achieve this, and has just completed a stealth launch, securing $62.5 million in seed funding. The company is initially deploying its predictive system with a handful of financial brokers, stating that the model can help predict how individual traders will react during times of significant market volatility.

"Behavior is not deterministic; it is always probabilistic," said Amit Gal - Or, CEO and co-founder of 42AI, to TechCrunch when I asked him to explain the company's bold statement. He believes that people have certain underlying psychological tendencies that influence their decisions even as the environment changes, and what startups try to do is to measure these tendencies and convert them into data that models can understand.

"Looking one week from now, you might be a different person, but your tendencies will always manifest in certain areas with some probability or distribution," added co-founder and chief technology officer Ehud Ezra. The company believes that this is precisely the purpose of the model: to show the probability of a person making a certain decision.

42AI indicates that the model attempts to understand people's underlying motivations, rather than inferring patterns based on their past behavior. To this end, 42AI studies people's reactions to seemingly unrelated events. Taking investors or traders as an example, such events may include completing a KYC process, opening an account, receiving customer service calls, making fund transfers, or trading stocks.

"When this person communicates with customer service or sales during the initial process... we look at their emotions, their voice, many of their interactions, how many times they make calls, and so on," added co-founder and chief technology officer Ehud Ezra. "All of these data points are fed into the model. Afterwards, you will see the subsequent interactions: how and when you buy and sell, and at that stage, we observe the market."

“So what we try to understand is how they perceive this world. In a world model, every individual is this world itself,” he added.

Subsequently, the company will transform these observations into mathematical representations of features, and then input them into a world model based on Transformer to list the probabilities of possible decisions in different scenarios. (The company also uses another language model to generate explanations for its predictions.)

Ezra indicates that during training, the model is required to retain a person's underlying tendencies. Therefore, if someone usually dislikes risk, but the model suddenly classifies them as a risk-taker, 'the goal is then off,' he said, and explained the design of the rewards used in model training.

It took 42AI three years to develop this method, combining psychological research, neuroscience, and genetics, as well as advice from experts in psychology and post-traumatic stress disorder. The opinions of Dan Shechtman, the Nobel Prize winner in Chemistry in 2011, were also taken into consideration.

Gal-Or indicates that the team initially studied thousands of people through psychological assessments and behavioral scenarios, and later incorporated data from government and biological sample databases covering over 2 million people. He said that since then, their customer database has expanded to nearly 30 million people.

In actual deployment, the model will analyze the customer's existing records, including transactions and customer service calls. Gal-Or points out that the system does not require information that can identify individual identities, stating that "the model never knows who this person is." However, from the description, it seems that the company still creates personal behavior profiles based on associated customer activities.

In terms of predicting market transactions, 42AI claims that its model has an accuracy rate of 98%, but Gal-Or clarifies that this indicator is more about whether the transaction flow falls within the predicted range, rather than whether each individual transaction is accurately predicted.

He also stated that in terms of predicting the direction and volume of transactions, the F1 score of this model remains stable between 0.8 and 0.85. (The F1 score is an indicator in machine learning used to measure the balance between correct predictions and false positives or missed detections in classification models.)

However, this model is not completely useless, as it is after all a "machine inside a box." "There is a natural limit to its effectiveness, because you simply cannot be aware of everything that a person experiences," says Gal-Or. Ezra explains that if the system misses important contextual information, the accuracy will also be affected, for example, if someone changes their decision to make a transaction based on the advice of a friend or colleague.

Gal-Or indicates that 42AI is collaborating with 10 major brokers, who use its services to manage liquidity, margin requirements, and hedging exposures. He said that the company continues to acquire more clients, and its revenue has surpassed the "seven-figure" mark, but he refused to disclose the exact figure.

Customers can pay a fixed fee, or they can pay a certain percentage of the value that the platform helps them to obtain. He claims that one of the brokers created an additional value of about 30 million US dollars through dynamic hedging.

Gal-Or said that this $62.5 million in seed funding came from the founders and Infinity Group, to be completed over three years, and did not entirely come from external investors. Infinity was founded by his father Amir Gal - Or, who is also the chairman and co-founder of 42AI. 42AI now plans to continue raising funds.

In addition to finance, 42AI is also exploring the application of this model in gaming, anti-fraud, and network security, and hopes that one day it will be able to help robots and AI agents understand humans. However, this may still take some time: Ezra acknowledges that every time the model enters a new field, it needs to be trained according to the corresponding rules and context.

Like all AI tools, the technology of 42AI can also have its concerning aspects: if a broker is able to predict what customers might do, what prevents others from exploiting this knowledge for their own gain? When I raised this point, Ezra acknowledged it as a “reasonable concern.” The two founders mentioned regulation, monitoring, and an internal ethics committee, but they did not explain what mechanisms are in place to prevent customers from taking advantage of this technology to exploit those whose profiles are being created.

However, Gal-Or is full of confidence that this technology can help to address these risks. “We are very proud because, in fact, we are part of the solution; we are the solution to this problem,” he said.