millionwaysai 发布 Thorsten-4：主打识人动机的 Large Psychology Model
一家叫 millionwaysai 的公司发了个不走寻常路的模型，数据不是爬网页，而是心理学家面访了几千人标出来的。做销售、招聘类 AI 智能体的开发者可以看看，它专门补 LLM 读不懂人的短板。
millionwaysai 发布 Thorsten-4，定位为 Large Psychology Model，宣称能以 90% 的验证准确率预测对话对象的动机、决策风格和可能反应。该模型基于 15 年以上心理学研究：50 余名受过训练的心理学家在 4 年内反复访谈 7,000 多人（对象覆盖 8 岁儿童到亿万富翁），人工标注了 3,500 万词的对话文本。
训练数据无法从互联网抓取，团队还用成熟心理学量表和这些受访者数年间的实际行为来校验模型预测。Thorsten-4 接入企业现有的会议、面试、邮件和客服对话数据，对每个人打 50 项行为与心理特质分数，每次判断都可追溯到证据。文中引用的对比数据称，10 个主流 AI 模型在追踪人的想法与情绪随时间变化方面比人类低约 45%，且在被用户反驳时有 46% 的概率推翻自己原本正确的答案。
该产品瞄准的痛点是：当前 AI 智能体在销售电话、面试和谈判中只会读字面意思，听不出客户一句 sounds great 背后的犹豫。它宣称被质疑时能坚持自己的判断，区别于容易被说服的通用 LLM。
Brilliant release and this is such a different kind of AI.
@millionwaysai launched Thorsten-4 today, AI that reads people’s motives
Every AI company trains on the same internet, and the internet doesn’t record why people do what they do.
So the AI agents now running sales calls, job interviews and customer service keep misreading people, and each misread costs a dea.
> Thorsten-4 is built on 15+ years of psychology research that started long before ChatGPT,
> 50+ trained psychologists interviewed 7,000+ people repeatedly over 4 years (from an 8-year-old to billionaires),
> hand-tagged 35M+ words from those conversations, and checked the model’s predictions against what people actually did.
It’s a Large Psychology Model that lets AI read the person behind the words, predicting their motives, decision style and likely reactions with 90% validated accuracy.
AI agents now run sales calls, job interviews, and negotiations, but they miss the hesitation and pushback that decide how those conversations end, like the doubt behind a prospect’s “sounds great”.
LLMs were built to read words, not people, and 10 leading AI models fall about 45% short of humans at tracking how a person’s thoughts and feelings change over time.
They also cave easily, since leading AI models flip a correct answer 46% of the time when a user pushes back.
Thorsten-4 reads the conversations companies already have (meetings, interviews, emails, chats and support calls) and scores each person on 50 behavioral and psychological traits, showing what drives them, how they decide and how they react under stress.
Its training data can’t be scraped off the internet, because 50+ trained psychologists interviewed 7,000+ people again and again over 4 years and hand-tagged 35M+ words from those conversations.
The team then checked the model’s predictions against established psychology tests and against what those people actually did over several years.
Every read is explainable and traces back to the evidence behind it, and Thorsten-4 holds its read when a user pushes back.
Companies plug it into the AI tools they already use, so their agents know when to push, when to pause and how to adapt.
🧵1/n
https://t.co/iNkHHcAcmp
本条情报归属 AITOP「模型」核心追踪栏目。在跨区域产业观察坐标中，该动态直接关联「AI 开发者工具链、Coding Agent 与工程化实践」研究主线，反映出当前技术演进与商业落地的最新风向。涉及核心实体与技术议题包括：Thorsten-4、millionwaysai、Large Psychology Model、智能体，读者可通过下方标签进入对应专题页查阅历史脉络与关联对比。
信源核验说明：本文原始线索采编自「rohanpaul_ai」（@rohanpaul_ai），经 AITOP 多维评分管线结构化提炼。截至目前，全网已有 3 家独立信源对该事件进行交叉印证与跟进报道（含 Ate-a-Pi、kimmonismus 等），属于近期高关注度行业信号。
