Wissenschaftler warnen, dass KI-Modelle in großen Sprachen nicht für den Einsatz in der realen Welt geeignet sind – selbst geringfügige Änderungen führen dazu, dass ihre Weltmodelle zusammenbrechen

A recent study by researchers from MIT, Harvard, and Cornell reveals significant limitations in large language models (LLMs) like GPT-4 and Claude 3 Opus when applied to real-world scenarios. Although LLMs can produce impressive outputs, the study found that their underlying “world models” — the frameworks they use to infer and generate answers — are often incoherent or inaccurate.

Key findings include:

1. **Inaccurate Representations**: LLMs provided accurate turn-by-turn driving directions in New York City, but the underlying maps they used contained nonexistent streets and routes. When unexpected changes like detours were introduced, their accuracy plummeted, raising concerns about their reliability in dynamic environments.
2. **Challenges with Dynamic Environments**: The study showed that even small changes, such as closing 1% of streets, reduced accuracy drastically (from nearly 100% to 67%), highlighting LLMs’ fragility in adapting to real-world variability.
3. **Testing with Deterministic Finite Automations (DFAs)**: Researchers tested LLMs on problems involving sequences of states, such as game rules (Othello) and city navigation. Metrics assessed whether the models could form coherent world models and compress sequences effectively. While models trained on random data performed slightly better, none succeeded in creating fully accurate or adaptable representations.
4. **Divergent Performance**: Transformers trained on random data showed better adaptability, likely because they encountered a broader range of possibilities. However, neither model type could reliably handle detours or dynamic changes in navigation tasks.

The researchers emphasize the need for new approaches to developing AI systems that can create coherent, accurate world models. They caution against overestimating the understanding and capabilities of LLMs based solely on their ability to generate impressive outputs. This study highlights the importance of critically evaluating AI’s limitations, particularly for real-world applications like autonomous systems.

– real-world useful AI generated summary

“ChatGPT couldn’t change my flat tire, is it stupid?”