Jailbreaking LLMs with ASCII Art

The document presents a technical argument that digital neural networks (DNNs), due to their layered structure, finite precision, and nonlinear activation (e.g., rectifiers), behave like one-way functions that cannot be feasibly inverted, drawing parallels to cryptographic constructs and emphasizing the impracticality of mapping high-dimensional input states to outcomes. It underscores the mathematical and engineering constraints behind reversibility, notes the potential for chaotic behavior when real-world assumptions fail, and briefly touches on the legal and regulatory implications by recalling Australia’s 2017 debate over encryption.