OPINION In a post-apocalyptic setting – a zombie outbreak, for example – a small, off-grid language model running on something like a Raspberry Pi sounds like a survival cheat code.

It’s certainly a clever way to get folks to check out your physical AI platform and maybe even splash out 7x the MSRP for an Nvidia Jetson Orin Nano with a battery enclosure, as one startup has done. But the scenario also sets a dangerous precedent, and in a real-world survival situation that local LLM could end up being a bigger liability than a lifesaver.

Roleplaying a zombie apocalypse with a local AI model is harmless fun. Zombies aren’t real and you’re probably not LARPing your way through the backcountry pursued by the undead. 

The problem is there’s a certain group of people who are going to see the concept presented in videos like this and think an off-grid AI is a great idea for a bug-out bag or to bring along on a hiking trip, and it’s not hard to see why.

Bit-for-bit, these models’ ability to retain information is unrivaled. You could download the entire English text of Wikipedia (about 24.7 GB compressed), and still not approach the information density of a four-billion-parameter model taking up 2 GB of memory.

Add in vision capabilities, support for translation, and maybe layer in a few text-to-speech and speech-to-text models and it starts to sound like the perfect survival companion. But for reasons you have probably already experienced over the past four years, we wouldn’t recommend it. 

Even the largest LLMs still hallucinate — that is, they’ll confidently fabricate information and present it as fact. That’s because while models are trained on tens of trillions of words (really, tokens) of data, their ability to retrieve that data is inherently imperfect.

LLMs don’t actually know what words belong together, just that statistically one word or another has the highest probability of coming next. It’s like auto-complete on steroids. Modern chatbots from OpenAI, Anthropic, or Google attempt to mitigate potential hallucinations by pulling in live web data to better inform their responses. The model is no longer answering your query on its own, but rather conducting research and presenting its findings. Even then, there’s no guarantee the information it retrieved is accurate or applicable to your specific context.

Medical professionals have been sounding the alarm on the dangers of relying on LLMs for medical advice for some time now. According to OpenAI, more than 230 million people globally query ChatGPT with health and wellness related questions each week.

The problem, medical professionals at Duke University School of Medicine note, is that these models can’t read between the lines to understand what the patient is really trying to say, and often suffer from sycophantic tendencies that lead them to tell users what they want to hear.

“When a patient comes to us with a question, we read between the lines to understand what they’re really asking,” Ayman Ali, a fourth-year surgical resident at Duke Health, said in a blog post earlier this year. “We’re trained to interrogate the broader context. Large language models just don’t redirect people that way.”

Off grid, an LLM can’t fall back on the web for help and almost certainly doesn’t benefit from the same guardrails that major chatbots use to protect users against potentially dangerous outputs.

As we’ve shown time and time again, getting local LLMs running on inexpensive hardware has never been easier. For a contrived scenario like roleplaying a zombie apocalypse, the risk is minimal. But change things up and ask the model whether the berries pictured are edible, and you could be playing Russian roulette with a probability engine. ®