RISCai

One day, we woke up and had a massive revelation - if AI can do anything, and computers can do anything, then surely AI can take over the role of computers? So, for our project, we set out to emulate a CPU and then replace parts with AI models:

1. The ALU --> now performs calculations through a neural network.
2. The CU --> handles control based operations via a separate neural network.
3. The Memory --> Through prompt engineering, we've mastered the art of retrieving data by asking the LLM where the element is in an array. And if the LLM retrieves the wrong data? Just ask again.

We built it by first emulating a computer in python, with our own fully fledged assembler and custom ISA (RISCai). Then, using the data generated from the emulator, we can replace the parts of the emulation with ai models.

Surprisingly, trying to fit AI into components that need to be 100% reliable to work produced a few issues:

1. The AI may produce unexpected/incorrect results sometimes.
2. See point 1

Besides the AI, emulating the computer and coming up with a suitable instruction set and assembler was time consuming.

As our first AI project, we're proud that we've managed to successfully create something, even if the final result doesn't work all the time.

Built With

• keras
• langchain
• python
• riscai-assembler
• riscai-isa
• tensorflow