以下是科幻题：如果AI能预测你决定之前要做什么，即使在全新的情况下呢？
以色列开办的42AI 称公司正在开发一个能够做到这一点的“行为世界模式”，它刚刚以6 250万美元的种子资金从隐形处出来。 该公司最初在少数金融经纪公司运行其预测系统，该模型可以帮助预测个体贸易商如何应对不稳定的条件。
42AI的首席执行官兼共同创始人Amit Gal-Or在我要求TechCrunch为其公司高额索赔辩护时，他提出，人们有根本的心理倾向，即使在他们的情况发生变化时也会影响他们的决定，这就是启动者试图衡量和转换成模型能够理解的数据。
“一个星期后，你可能是一个不同的人，但你的趋势总是在某些领域表现出来，有概率或分布”，他的共同创始人和CTO Ehud Ezra补充说。 公司认为，模型就是这么做的：显示一个人做出某种决定的概率。
42AI 称模式试图找出人们的内部动机，为此，42AI研究人们如何将react与看似无关的事件联系起来。 例如，对投资者或交易商而言，此类活动可包括完成KYC程序、登机、呼叫支助、汇款或交易库存。
“当此人在初始过程中谈论支持或销售时[......]我们审视情绪，“所有这些数据点都进入模型。 后来，你们有了互动：你们如何买卖，何时，我们在那个阶段审视市场。
“所以我们想看看他们是如何看待世界的。 他补充说，在一个世界模式中，每个人都是世界。
公司然后将这些观测结果转换成特征的数学表示，然后输入一个基于变压器的世界模型，为在不同情况下可能作出的决定规定概率。 (公司使用单独的语言模型来生成对其预测的解释).
Ezra 称模型在模拟中被训练为保存一个人的基本倾向。 因此，如果某人通常不愿冒险，但模型却突然把他们当作风险标牌，“目标关闭了，”他称解释了模型的训练奖励是如何安排的。
42AI花了三年时间开发这种方法，将心理学研究与神经科学和遗传学相结合，以及心理学和PTSD专家的建议，丹·谢奇特曼也获得了2011年诺贝尔化学奖。
Gal-Or 称团队首先通过心理评估和行为假想研究了数千人，然后加入了涵盖200多万人的政府和生物库数据集。 其客户数据集从此扩大至近3000万人，他称。
在现场部署中，模型分析客户的现有记录，包括交易和支持呼叫。 Gal-Or指出这个系统不需要个人识别信息，然而，似乎该公司仍然通过关联客户活动创建个人行为简介。
在预测市场贸易时，42AI 声称它的模型达到了98%的准确度，但Gal-Or澄清说，这更是衡量贸易流量是否属于预测范围，而不是如果对每项贸易的预测正确的话。
他同时也是称，该型号在预测交易方向和量量时始终达到0.8比0.85的F1分。 (F1分是机器学习的衡量标准，衡量分类模型能够在多大程度上在错误警报和错失事件之间平衡正确的预测).
也不像是模特儿能到处工作，”Gal-Or 称。 Ezra解释说，当系统错过重要的情况背景时，其准确性也会受到影响，比如有人根据朋友或同事的建议选择改变交易选择。
Gal-Or 称 42AI正与10家大型经纪公司合作，这些经纪公司正利用它来管理流动性，差值要求和套期保值风险。 公司在客户、他称上船
客户向启动支付平价费，或平台价值的一定百分比帮助他们增益。 他声称，一个经纪公司通过动态套期保值获得了约3 000万美元的额外价值。
6250万美元的种子资金来自创始人和无穷集团三年来，并非完全来自外部投资者Gal-Or 称。 Infinity由他的父亲Amir Gal-Or创立，他也是42AI的董事长和共同创始人。 42AI现在计划筹集更多资金。
除了金融，42AI现在还在探索其模式在游戏，预防欺诈和网络安全方面的应用，有一天想用它来帮助机器人和AI agents了解人。 这或许是一条出路：以斯拉承认，模型需要针对每个新领域进行规则和背景培训。
像所有人工智能工具一样，42AI技术也有潜在令人不安的一面：如果经纪公司可以预见其客户能做什么，那么阻止某人利用这种知识为自己谋取利益又有什么办法？I融资的创始人指出条例、监控和内部道德委员会，虽然他们没有解释为何客户会阻止他们使用技术
然而，Gal-Or相信该技术有助于应对这些风险。 “我们感到非常自豪的是，实际上我们是解决办法的一部分；我们是这一问题的解决办法，”
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Here’s a question straight out of sci-fi: What if AI could predict what you were going to do before you decided to, even in a completely new situation?
Israeli startup 42AI says it is developing a “behavioral world model” that can do just that, and it has just come out of stealth with $62.5 million in seed funding. The company is initially running its predictive system at a handful of financial brokerages, where it says the model can help predict how individual traders will respond to volatile conditions.
“Behavior is not deterministic; it’s constantly probabilistic,” 42AI’s CEO and co-founder Amit Gal-Or told TechCrunch when I asked him to justify his company’s tall claim. He argues that people have underlying psychological tendencies that influence their decisions even when their circumstances change, and that’s what the startup is trying to measure and convert into data that the model can understand.
“A week from now, you might be a different person, but your tendencies will always play out in certain areas to a probability or distribution,” his co-founder and CTO Ehud Ezra added. And that, the company argues, is what the model does: show the probability of a person making a certain decision.
42AI says the model tries to figure out people’s internal motivations instead of extrapolating patterns based on their previous actions. To do that, 42AI examines how people react to seemingly unrelated events. For an investor or trader, for example, such events might include completing KYC processes, onboarding, support calls, money transfers, or trading stocks.
“When the person is talking to support or sales in the initial process […] we’re looking at the sentiment, we’re looking at the voice, we’re looking at a lot of interactions that they’re doing, how many times do they call, and so forth,” added CTO and co-founder Ehud Ezra. “And all of those data points are going into the model. Later on, you have the interactions: how you buy and sell, and when, and we’re looking at the market at that stage.”
“So we’re trying to look at how they’re looking at the world. In a world model, every individual is the world,” he added.
The company then converts those observations into mathematical representations of traits, which are then fed into a transformer-based world model that lays out probabilities for possible decisions under different circumstances. (The company uses a separate language model to generate explanations for its predictions).
Ezra said the model is trained to preserve a person’s underlying tendencies during simulations. So if someone is usually risk-averse, but the model suddenly treats them as a risk-taker, “the target is off,” 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 experts in psychology and PTSD, and Dan Shechtman, who was awarded the 2011 Nobel Prize in Chemistry.
Gal-Or said the team first studied thousands of people via psych assessments and behavioral scenarios, before adding government and biobank datasets covering more than 2 million people. Its customer datasets have since expanded to cover almost 30 million people, he said.
In live deployments, the model analyzes customers’ existing records, including transactions and support calls. Gal-Or pointed out the system doesn’t need personally identifiable information, arguing that “the model never knows who the person is.” However, it seems the company still creates individual behavioral profiles from linked customer activity.
In predicting market trades, 42AI claims its model achieves 98% accuracy, but Gal-Or clarified that this is more a measure of whether trading flows fall within a predicted range, not if each trade was predicted correctly.
He also said the model consistently achieves an F1 score of 0.8 to 0.85 for predicting trading direction and volume. (F1 score is a metric in machine learning that measures how well a classification model can balance correct predictions against false alarms and missed events).
And it’s not like the model can work everywhere, given it’s still a machine in a box. “There’s a natural ceiling because you really don’t have access to everything that a person experiences,” Gal-Or said. Ezra explained that the system’s accuracy also suffers when it misses important situational context, like if someone chooses to change their trading choices based on advice from a friend or a colleague.
Gal-Or said 42AI is working with 10 large brokerages, who are using it to manage liquidity, margin requirements, and hedging exposure. The company is onboarding more customers, he said, and is generating more than “seven figures” in revenue, though he wouldn’t specify a number.
Customers pay the startup a flat fee, or a percentage of the value the platform helps them gain. He claimed one brokerage generated approximately $30 million in additional value through dynamic hedging.
The $62.5 million seed funding came from the founders and Infinity Group over three years, and wasn’t entirely raised from external investors, Gal-Or said. Infinity was founded by his father, Amir Gal-Or, who is also 42AI’s chairman and a co-founder. 42AI now plans to raise more money.
Beyond finance, 42AI is now exploring applications for its model in gaming, fraud prevention and cybersecurity, and one day wants to use it to help robots and AI agents understand people. But that may be a ways off yet: Ezra acknowledged that the model needs to be trained on rules and context for each new domain.
And like with all AI tools, there’s a potentially troubling side to 42AI’s technology: if a brokerage can anticipate what its customers may do, what’s to stop someone from using that knowledge to their own benefit? Ezra admitted that this was “a valid concern” when I raised it, and the founders pointed to regulations, monitoring and an internal ethics committee, though they didn’t explain what, if anything, would stop their customers from using the technology to gain an advantage over the very people it’s profiling.
Gal-Or, however, was confident that the technology could help address such risks. “We are very proud that actually we are part of the solution; we are the solution to this problem,” he said.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.