In an era where AI models typically demand hundreds of gigabytes (GB) of video random - access memory (VRAM) and boast trillion - parameter scales, developer HarryR has achieved a remarkable feat. He has managed to make the 8 - bit Z80 processor, which was introduced back in 1976 with a mere 64 kilobytes (KB) of memory and lacking floating - point capabilities, run conversational AI through the Z80 - μLM project. This project even enables the processor to support a 20 - question puzzle - solving game.

Several key optimizations have made this possible. Firstly, the inference engine, model weights, and interaction interface have been compressed into a 40KB file. Secondly, instead of relying on floating - point calculations, 16 - bit integer operations are employed. Thirdly, 2 - bit weight vectorization is applied. In this method, each weight is compressed to one of the values in the set {-2, -1, 0, +1}, and 4 weights are stored per byte.

The project presents two examples to showcase its functionality. The Tinychat bot responds to greetings in a minimalist manner. For instance, it simply says “OK” to confirm something and “WHY?” when questioning. On the other hand, Guess guides users through a process of solving hidden AI puzzles by asking 20 questions.

HarryR openly admits that this system is far from passing the Turing test. However, its true value lies in exploring the lower limits of AI size. By designing ambiguous responses, the system forces humans to infer the context or use closed - ended questions to gauge the AI's level of understanding.