OpenAI has unveiled a new experimental language model that operates in a significantly more transparent manner than current AI systems. Although it cannot compete with top models such as GPT-5, Claude or Gemini, it could provide crucial clues as to why AI sometimes reacts strangely, hallucinates or becomes unreliable. 🤔💬
A small model with big significance 🧪✨
The new model is called Weight-Sparse Transformer (WST). It is much smaller and roughly as powerful as GPT-1 (2018). But that’s not a problem for OpenAI – because the goal is not to break records, but to better understand AI. Researchers want to use it to find out how large models work internally and what structures influence their decisions. 🧠🔧
Research experts see this as an exciting approach. The methods presented could help make the development of future AI systems safer and more predictable.🔐📈
Why AI is so difficult to understand 🕸️🤯
WST is part of the field of research known as mechanistic interpretability – the attempt to analyse neural networks in such a way that we can understand which internal mechanisms are responsible for certain abilities.
The problem: modern LLMs consist of countless neurons that are connected to each other in many layers. In classic “dense networks,” knowledge is distributed across a huge network of connections. A concept or function therefore never resides in just one neuron, but is distributed across many – often overlapping. This phenomenon is called superposition.⚡🧩
This makes AI seem like a black box: it is extremely difficult to assign specific tasks to specific parts of a model. This is precisely where the new model comes in – by using structures that are easier to understand.
