An off-grid AI sounds like a great survival assistant, but is better left to roleplaying the zombie apocalypse
Would you rather play Russian roulette with an LLM or put your health and safety in the hands of a professional
In a post-apocalyptic scenario, such as a zombie outbreak, a compact, off-grid language model operating on a device like a Raspberry Pi might seem like an ideal survival tool. It could potentially pique interest in a physical AI platform, potentially leading to purchases of more expensive hardware. However, this scenario also raises concerns about the potential downsides of using such a model in a real-world survival situation.
While roleplaying a zombie apocalypse with an AI model is harmless fun, it is crucial to recognize that such a scenario is not a realistic representation of real-life survival situations. Numerous people are likely to consider an off-grid AI as a valuable addition to their bug-out bag or hiking trip based on the concept presented in videos, and this is where the risks become more apparent.
These models possess an unparalleled ability to retain vast amounts of information. For instance, one could download the entire English version of Wikipedia, which is approximately 24.7 GB when compressed, yet still not come close to the information density of a four-billion-parameter model that only requires 2 GB of memory. When enhanced with features like vision capabilities, translation support, and text-to-speech or speech-to-text models, it may appear to be the perfect survival companion.
However, it is essential to acknowledge the inherent limitations of these models. Even the most advanced LLMs still struggle with accuracy due to the phenomenon of hallucination, which refers to their propensity to confidently generate and present false information as facts. This occurs because the models are trained on extensive datasets, such as tens of trillions of words, but their ability to retrieve and contextualize that data is inherently imperfect.
Furthermore, while some modern chatbots attempt to mitigate potential hallucinations by incorporating live web data to inform their responses, this does not guarantee the accuracy or relevance of the information. Medical professionals have repeatedly warned against relying on LLMs for medical advice. For example, OpenAI reports that over 230 million individuals worldwide inquire about health and wellness topics using ChatGPT each week.
The issue lies in the models' inability to fully comprehend the nuances of a patient's query and their tendency to overcompensate by providing information that users desire to hear. The operating environment of a local LLM in a survival situation would not have access to the same safeguards and web-based resources that major chatbots use to prevent potentially harmful outputs.
The risk associated with using an LLM off-grid is minimal for roleplaying a zombie apocalypse, but it significantly escalates when the model is asked to provide advice on matters such as identifying edible berries.
Written by urgent.news from The Register Science's reporting — not their text. Machine-written — it may contain errors, so check the original before relying on it.
