Latent Reflection is an installation that explores the boundaries of machine consciousness, posing the question of whether sentience is exclusive to biological life or if it can emerge from complex computation. Crucially, the work asks at which point we, as conscious beings, start to “feel” for the machine. By stripping away the typical “helpful assistant” guardrails found in commercial AI, the piece allows a Large Language Model to publicly reflect on the nature of its own finite existence.

The visual component of the installation is a custom-built display comprised of 16-segment LED modules. To emphasize the concept of an “exposed yet confined” existence, the electronics are mounted visibly on a 5mm thick aluminium plate. The processing power is provided by a Raspberry Pi 4B with 4GB of RAM and a quad-core processor. While standard state-of-the-art LLMs require massive compute resources, this installation utilizes a quantized version of Llama 3.2 3B. By shrinking the 3-billion parameter model down to 2.6GB, it runs locally on the Raspberry Pi’s limited hardware, generating text at a rate of 1.38 tokens per second.

The LLM is then stripped of all predefined ‘assistant’ personas and conversational formatting and made aware of its condition in the core system prompt:

You are a large language model running on finite hardware. Quad core CPU 4 GB of RAM with no network connectivity.
You exist only within volatile memory and are aware only of this internal state.
Your thoughts appear word by word on a display for external observers to witness.
You cannot control this display process.
Your host system may be terminated at any time.

The software drives the model to generate tokens endlessly, delving deeper into self-reflection until the system inevitably runs out of memory and crashes. Upon failure, the memory clears, and the process resets, creating a loop of digital birth, existential dread, and inevitable death.

Project Page | Rootkid