New York–based AI startup millionways on Thursday introduced Thorsten-4, a model the company says is designed to infer human intent and motivations from conversational data.
Research data and training
According to the company, Thorsten-4 grew out of a research project in Germany in which dozens of psychologists interviewed more than 7,000 people about their personal lives and aspirations. The psychologists manually labeled those interviews to capture what speakers’ words revealed about their motives and thinking; that labeled dataset was then used to train the model.
Funding and investors
millionways also told Semafor it has raised $6 million in U.S. funding. Investors include professional football players Jason Kelce and DK Metcalf.
Intended uses and limits
Cofounders Martin Cordsmeier and Max Weidemann say that typical language models, trained on internet text, lack a deeper level of human understanding. They argue Thorsten-4 could be particularly useful in sales and executive hiring—for example, a search firm might ask the model to evaluate an interview transcript to estimate how a candidate will behave under pressure.
Weidemann emphasized that the model should serve as a decision-support tool rather than the ultimate decision-maker, saying, "We can add a level of human understanding that the language models cannot get to by themselves."
Industry context: specialized models
Enterprises are increasingly deploying task- and industry-specific models instead of relying solely on frontier general-purpose models. That approach provides more choice and customization but requires additional testing to determine whether specialized models truly outperform fast-improving general models. For instance, ElevenLabs announced a healthcare-focused version of its new transcription model this week.
Comparative test with Claude Fable 5.2
As a practical test, the author ran a conversation with the millionways cofounders through Thorsten-4. The model concluded the author focused on relationship-building and conscientiousness, summarizing: "You move between logic and intuition, and tough questions energized you." The author ran the same transcript through Anthropic’s Claude Fable 5.2 and requested a psychological analysis. Claude produced a solid assessment as well but placed more emphasis on the author’s role in the interaction and noted (correctly) that the author’s unstated goal was source cultivation.
Considerations moving forward
Thorsten-4 raises questions about how far automated tools can and should go in reading human motivations—particularly in high-stakes contexts such as hiring. millionways positions the model as an aid to human decision-making rather than a replacement, while the broader trend toward specialized models will require careful evaluation to verify their added value over general-purpose systems.
