AI 模型和平台
42AI 发布预测人类决定的世界行为模式
在 Google 上的首选来源中添加 Unite.AI
大部分AI系统是用来解释信息或生成答案的。 42AI在追求另一个问题：在情况变化时，特定的人会如何回应？公司以所谓的"行为世界模型"(Beavioral World Model)从隐蔽中脱颖而出，将人类行为定位为企业AI的一个显著的预测问题。
特拉维夫公司称公司至今拥有融资6250万美元，并开始进入金融市场。 它的发布在企业面前提出了一个雄心勃勃的建议：了解一系列交易背后的人可能与分析交易本身同样重要。
以人为重点的世界模式
在机器人学中，一个世界模型典型地代表着一个环境随着物体移动和行动展开而发生的变化。 42AI将世界模范思想应用到人身上，结合了个人特征，变化中的情况，以及之前的行为等信息来预测未来的行动。
该公司描述了一种包含心理、行为和基因组数据的专有方法。 其公告没有解释这些数据集的来源或同意安排，也没有解释在商业部署中是否需要基因组投入。 这一区别很重要：描述研究投入并不能确定客户必须提供遗传信息。
在其网站上，42AI将平台呈现为行为风险智能。 拟议的工作流程从原始数据转移到对行为驱动器的理解，然后到对事件和行为随时间推移的反应的预测。 其上市应用包括曝光管理，欺诈检测，网络安全，投资研究，以及个性化AI.
实际吸引力在于背景。 面对相同信息的两个人可能做出不同决定，而同一人可能随着激励、约束或经历的变化而做出不同的反应。 一个有用的行为模式必须代表这些差异，而不是将每个用户视为人口组中可互换的成员。
行为预测背后的研究
2026年2月由42AI链接的"解码人的因素：战略前瞻的高清晰度行为预测"(High Fidelity Behavior President for Strategy Foresight)的研究预印，提供了对其研究方向的更具体的看法。 它描述了由微调 Llama 3.1-8B-指令所构建的大型行为模型，其结构剖面从心理测量评估中抽取。
研究人员将处置、动机状态和情况限制与观察到的选择联系起来。 他们报告说，经过调整的模型改进了在被搁置的情景中未经修改的骨干结构，而较丰富的特征描述不仅能产生广泛的人格类别，还能带来好处。
这说明了要求通用聊天员想象某人的行为和专门培训使用结构化行为证据的模型之间的技术区别。 然而，预印本并没有独立地确定42AI的宣布每件商业产品的业绩，其情景评价不应与公司的交易预测相混淆。
金融市场提供第一次商业试验
42AI 称的最初应用预测个体贸易商的行动，包括工具、方向、时间和贸易量，在1、4和24小时内。 公司报告，数百万贸易商的准确度超过98%。
这个数字是公司索赔。 公告没有提供关于准确性计算、比较基线或验证数据的足够细节，无法独立评估。 预测商人的行为也不同于预测市场价格或显示有利可图的投资策略。
对于经纪人来说，即时使用的案例是预测风险：客户的集体行动如何改变其风险状况。 这使行为预测成为一种操作目的，即使它不被用来推荐投资。
从行为预测到企业决定
根据其领导网页，42AI由CEO Amit Gal-Or和CTO Ehud Ezra领导，Amir Gal-Or为联合创始人和主席。 它的研究领导包括认知AI专家Roi Reichart教授和行为研究负责人Eyal Fruchter教授。 诺贝尔奖得主丹·谢奇特曼教授担任高级顾问。
更大的机会超越金融。 一个AI助手，预计有人会如何解释一项建议，可以更有益地调整其互动。 同样的能力还可能影响对客户或雇员作出相应的决定，作出同意、适当使用以及对部署的核心预测提出质疑的能力。
42AI的发布推出是解决一个困难AI问题的方法：模拟人类的变异性，而不是简单地产生流利的反应。 它的下一次测试表明，这些预测仍然是有用和可靠的外部可控评价——企业可以不把概率预测作为一个人的固定描述来对待就采取行动。
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。
AI Models & Platforms
42AI Launches a Behavioral World Model to Predict Human Decisions
Add Unite.AI to your preferred sources on Google
Most AI systems are built to interpret information or generate an answer. 42AI is pursuing a different question: how will a particular person respond when circumstances change? The company has emerged from stealth with what it calls a Behavioral World Model, positioning human behavior as a distinct prediction problem for enterprise AI.
The Tel Aviv company says it has raised $62.5 million to date and is starting in financial markets. Its launch puts an ambitious proposition in front of businesses: understanding the people behind a stream of transactions could be as important as analyzing the transactions themselves.
A World Model Focused on People
In robotics, a world model typically represents how an environment changes as objects move and actions unfold. 42AI applies the world-model idea to people, combining information about individual characteristics, changing circumstances, and previous behavior to anticipate future actions.
The company describes a proprietary approach incorporating psychological, behavioral, and genomic data. Its announcement does not explain the provenance or consent arrangements for those datasets, or whether genomic inputs are needed in commercial deployments. That distinction matters: describing the research inputs does not establish that customers must supply genetic information.
On its website, 42AI presents the platform as behavioral risk intelligence. The proposed workflow moves from raw data to an understanding of behavioral drivers, then to predictions of reactions to events and behavior over time. Its listed applications include exposure management, fraud detection, cybersecurity, investment research, and personalized AI.
The practical attraction is context. Two people confronted with the same information may make different decisions, while the same person may respond differently as incentives, constraints, or experience change. A useful behavioral model would have to represent those differences rather than treat every user as an interchangeable member of a demographic group.
The Research Behind Behavioral Prediction
A February 2026 research preprint linked by 42AI, Decoding the Human Factor: High Fidelity Behavioral Prediction for Strategic Foresight, offers a more concrete view of its research direction. It describes a Large Behavioral Model built by fine-tuning Llama 3.1-8B-Instruct with structured profiles drawn from psychometric assessments.
The researchers connect dispositions, motivational states, and situational constraints with observed choices. They report that the adapted model improves on its unmodified backbone in held-out scenarios, and that richer trait profiles can provide benefits beyond prompting with broad personality categories.
This illustrates a technical distinction between asking a general-purpose chatbot to imagine someone’s behavior and training a model specifically to use structured behavioral evidence. The preprint does not independently establish the performance of every commercial product announced by 42AI, however, and its scenario evaluations should not be conflated with the company’s trading forecasts.
Financial Markets Provide the First Commercial Test
42AI says its initial application predicts individual traders’ actions, including the instrument, direction, timing, and volume of trades, at horizons of one, four, and 24 hours. The company reports accuracy above 98% across millions of traders.
That figure is a company claim. The announcement does not provide enough detail about the accuracy calculation, comparison baselines, or validation data to assess it independently. Predicting a trader’s behavior also differs from predicting market prices or demonstrating a profitable investment strategy.
For a broker, the immediate use case is anticipating exposure: how the collective actions of customers might change its risk position. That gives behavioral prediction an operational purpose even when it is not used to recommend an investment.
From Behavioral Forecasts to Enterprise Decisions
According to its leadership page, 42AI is led by CEO Amit Gal-Or and CTO Ehud Ezra, with Amir Gal-Or as co-founder and chairman. Its research leadership includes cognitive AI specialist Professor Roi Reichart and behavioral research head Professor Eyal Fruchter. Nobel laureate Professor Dan Shechtman serves as a senior advisor.
The wider opportunity extends beyond finance. An AI assistant that anticipates how someone will interpret a recommendation could adapt its interaction more usefully. The same capability could also influence consequential decisions about customers or employees, making consent, appropriate use, and the ability to challenge a prediction central to deployment.
42AI’s launch introduces an approach to a difficult AI problem: modeling human variability rather than simply producing fluent responses. Its next test is showing that those predictions remain useful and reliable outside controlled evaluations—and that enterprises can act on them without treating a probabilistic forecast as a fixed description of a person.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.