供资
42AI 为行为世界模式提供6250万美元种子资金
在 Google 上的首选来源中添加 Unite.AI
TechCrunch于2026年11月2日报道，以色列开办的42AI已经以6,250万美元的种子资金和在金融经纪公司运营的行为世界模式退出了隐形。 创始人告诉TechCrunch，系统估计了个别客户在不断变化的情况下如何行动的概率。
首席执行官兼联合创始人Amit Gal-Or告诉TechCrunch，资金来自创始人和无穷集团在三年内，并非融资完全来自外国投资者。 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。 Ehud Ezra是首席技术官和联合创始人。
Gal-Or 称 42AI与10家大型经纪公司合作，登上更多的客户，产生7个以上的收入数字，尽管他没有提供准确的数字。 客户支付固定费用或平台帮助他们获得的价值的百分比。 He 称 1经纪公司通过动态套期保值获得了约3000万美元的额外价值。
行为模式如何运作
Gal-Or和Ezra告诉TechCrunch，行为是概率性的，而不是决定性的。 称 42AI试图找出潜在的动机，而不是简单地推断先前的行动，模式为不同情况下可能作出的决定分配概率。
对于经纪公司客户，Ezra 称的投入可以包括了解客户和登机程序、支持和销售互动、汇款和贸易。 经审查的信号包括情绪、声音、呼叫频率、购买和销售的时间以及同时存在的市场条件。
TechCrunch报告说，这些观测结果被转换成特征的数学表示，并输入一个基于变压器的世界模型。 一个单独的语言模型生成预测的解释。 Ezra 称训练奖励旨在在模拟中保持基本趋势：通常避免风险的人不应该突然被模拟为风险赌手。
培训数据和报告结果
Gal-Or 称 42AI花了3年时间开发这一方法，将心理学研究与神经科学和遗传学相结合，并接受心理学和创伤后精神紧张症专家及化学家丹·谢奇特曼的建议。 团队最初通过心理评估和行为假想研究了数千人，然后增加了政府和生物库数据集，覆盖200多万人。 He 称客户数据集随后扩展至近3000万人。
TechCrunch报告说，42AI的市场贸易预测准确度为98%。 Gal-Or澄清说，它衡量的是贸易流量是否属于预测范围，而不是每个单项贸易的预测是否正确。 他还报告了F1对交易方向和量的0.8到0.85的分数。 F1测量一个分类模型如何平衡对假警报和错失的事件的预测进行校正。
创始人承认有一个准确性上限，因为该系统无法接触一个人经历的一切。 Ezra 称精度在重要情况背景缺失时也会下降，例如朋友或同事建议改变交易决定。
由Ben Yellin, Ezra, Mark Foreman 和 Shula Grinapol 推出于2026年2月19日向arXiv 提交的预印件，对大型行为模型进行了微调，以从结构化的高维心理特征剖面来预测个人的战略选择。 作者介绍了关于将稳定处置、动机状态和情况限制与观察到的选择联系起来的专有数据集的培训。
在搁置的情景评价中，作者报告说，与未适应的Llama-3.1-8B-指令骨干相比，行为预测有所改善，在以五大特征为条件时，其性能与前沿基线具有可比性。 它们还报告说，它们的模型继续受益于密度更大的特征图谱，而以触发为基础的基线则显示出一个复杂性的上限。
中间商用途和时效
在42AI的网站上，公司描述了一个行为堆栈，将原始数据转换成行为驱动器，预测随着时间的推移对特定事件和行为的反应，模拟变化的变量和提供解释。 它列出世界模型的反应时间为1-10毫秒，大行为模型的反应时间为2-12毫秒。
对于经纪公司，该服务站点称可以按仪器、方向和数量预测每个客户的交易时间，最多24小时。 它还列出了其经纪用途中的暴露、动态套期保值、尾端风险管理、以及风险下降和价值减少等预先可见度。 Gal-Or告诉TechCrunch，客户使用该系统来管理流动性，差值要求和套期风险。
Gal-Or 称部署的系统不需要个人识别信息，模型永远不会知道这个人是谁。 TechCrunch报告说，42AI仍然通过关联的客户活动创建了个人行为简介。
当TechCrunch 融资对客户利用对自己客户的预测的可能性时，以斯拉称其为合理的担忧。 创始人提到条例、监测和内部道德委员会，而Gal-Or 称技术可以帮助应对这些风险。 TechCrunch报告说，他们没有解释阻止客户利用预测取得比被剖析人员优势的具体机制。
创始人告诉TechCrunch,42AI正在探索游戏，预防欺诈和网络安全，最终希望机器人和AI agents使用模型来理解人。 Ezra 称模型必须就每个新域的规则和背景进行培训。 Gal-Or以及称公司计划筹集更多资金。
除了有文件证明的经纪部署之外，还需要进行所描述的特定领域培训，而42AI的筹资计划是筹集更多的资金。
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。
Funding
42AI Secures $62.5M in Seed Funding for Behavioral World Model
Add Unite.AI to your preferred sources on Google
TechCrunch reported on November 2, 2026, that Israeli startup 42AI had exited stealth with $62.5 million in seed funding and a behavioral world model operating at financial brokerages. The founders told TechCrunch that the system estimates probabilities for how individual clients may act under changing circumstances.
Chief executive and co-founder Amit Gal-Or told TechCrunch that the funding came from the founders and Infinity Group over three years and was not raised entirely from external investors. Infinity was founded by his father, Amir Gal-Or, who is also 42AI’s chairman and a co-founder. Ehud Ezra is chief technology officer and a co-founder.
Gal-Or said 42AI works with 10 large brokerages, is onboarding more customers and generates more than seven figures in revenue, though he did not provide a precise figure. Customers pay either a flat fee or a percentage of the value the platform helps them gain. He said one brokerage generated approximately $30 million in additional value through dynamic hedging.
How the Behavioral Model Works
Gal-Or and Ezra told TechCrunch that behavior is probabilistic rather than deterministic. They said 42AI tries to identify underlying motivations instead of simply extrapolating previous actions, with the model assigning probabilities to possible decisions under different circumstances.
For brokerage customers, Ezra said inputs can include know-your-customer and onboarding processes, support and sales interactions, money transfers and trades. The examined signals include sentiment, voice, call frequency, the timing of purchases and sales, and concurrent market conditions.
TechCrunch reported that these observations are converted into mathematical representations of traits and fed into a transformer-based world model. A separate language model generates explanations for predictions. Ezra said the training rewards are designed to preserve underlying tendencies during simulations: a normally risk-averse person should not suddenly be modeled as a risk-taker.
Training Data and Reported Results
Gal-Or said 42AI spent three years developing the approach, combining psychological research with neuroscience and genetics and receiving advice from psychology and PTSD experts and chemist Dan Shechtman. The team initially studied thousands of people through psychological assessments and behavioral scenarios, then added government and biobank datasets covering more than 2 million people. He said customer datasets subsequently expanded to cover almost 30 million people.
TechCrunch reported 42AI’s stated figure of 98% accuracy for market-trade prediction. Gal-Or clarified that it measures whether trading flows fall within a predicted range, rather than whether each individual trade is predicted correctly. He also reported F1 scores of 0.8 to 0.85 for trading direction and volume. F1 measures how a classification model balances correct predictions against false alarms and missed events.
The founders acknowledged an accuracy ceiling because the system lacks access to everything a person experiences. Ezra said accuracy also degrades when important situational context is missing, such as advice from a friend or colleague that changes a trading decision.
A preprint submitted to arXiv on February 19, 2026 by Ben Yellin, Ezra, Mark Foreman and Shula Grinapol introduced a Large Behavioral Model fine-tuned to predict individual strategic choices from structured, high-dimensional psychometric trait profiles. The authors described training on a proprietary dataset linking stable dispositions, motivational states and situational constraints to observed choices.
In a held-out scenario evaluation, the authors reported improved behavioral prediction relative to the unadapted Llama-3.1-8B-Instruct backbone and comparable performance to frontier baselines when conditioned on Big Five traits. They also reported that their model continued to benefit from denser trait profiles, while prompting-based baselines exhibited a complexity ceiling.
Brokerage Uses and Stated Limits
On 42AI’s website, the company describes a behavioral stack that converts raw data into behavioral drivers, predicts responses to specific events and behavior over time, simulates changing variables and supplies explanations. It lists response times of 1–10 milliseconds for the world model and 2–12 milliseconds for the large behavior model.
For brokerages, the site says the offering can predict each client’s trades up to 24 hours ahead by instrument, direction and volume. It also lists advance visibility into exposure, dynamic hedging, tail-risk management, and reductions in slippage and value at risk among its brokerage uses. Gal-Or told TechCrunch that customers use the system to manage liquidity, margin requirements and hedging exposure.
Gal-Or said the deployed system does not need personally identifiable information and that the model never knows who the person is. TechCrunch reported that 42AI nevertheless creates individual behavioral profiles from linked customer activity.
When TechCrunch raised the possibility of customers exploiting predictions about their own clients, Ezra called it a valid concern. The founders pointed to regulations, monitoring and an internal ethics committee, while Gal-Or said the technology could help address such risks. TechCrunch reported that they did not explain a specific mechanism that would stop customers from using predictions to gain an advantage over the people being profiled.
The founders told TechCrunch that 42AI is exploring gaming, fraud prevention and cybersecurity and eventually wants robots and AI agents to use the model to understand people. Ezra said the model must be trained on the rules and context of each new domain. Gal-Or also said the company plans to raise additional money.
Expansion beyond the documented brokerage deployments requires the domain-specific training Ezra described, while 42AI’s stated financing plan is to raise additional money.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.