9 misconceptions about Large Language Models

There's lots of noise behind any new technology but even more so with something as disruptive as Large Language Models (LLMs). In every workshop the same misunderstandings come up, so here are the nine I encounter most often.

LLMs can be bias-free. LLM bias shows up in three ways: mimicking human falsehoods, mimicking human inaccuracies, and being limited by outdated or incomplete data. Since LLMs are designed to simulate human language it's hard to see how those biases can ever be entirely eliminated. We need tools to reduce and mitigate them to avoid causing harm.

LLMs are information machines. LLMs are limited to the data they were trained on. Sam Altman is on record as saying we should consider them as 'reasoning machines'. They can paraphrase but can't quote. Misunderstanding this leads to unexpected results. If you ask for a set of URLs about a subject an LLM doesn't have data on it will hallucinate, confidently giving you non-existent URLs.

LLMs are creative. LLMs do a good job of mimicking humans but they have no creativity. We've started talking about 20:60:20 where the first and last 20% requires human creativity and the middle 60% is LLM graft. They do have emergent properties but they're not creative.

LLMs are always learning. They're not. LLMs are limited to their original data-set. To "learn" they need to be retrained. The model you're interacting with at any given time is static. The only thing you have control over is the prompt.

LLMs are all seeing. Most LLMs aren't connected to the internet. GPT4 and GPT3.5 can't "surf" the web and even those that can, like Bing, are using URLs as part of their prompt rather than as training. They can't see everything.

LLMs are good in any language. LLMs are very good at English, and computer languages written in English, but they tail off dramatically in languages with fewer speakers. Given the volume of English-language data on the internet it's likely English will continue to be lingua franca for these models. If you work with communities whose first language isn't English, that matters.

LLMs are only for generation. This one frustrates me. LLMs predict the next token in a conversation. That makes them very good at tasks beyond generation: summarisation, question-answering, reformatting, categorising and analysing. Treating them as generation-only tools is selling them short.

LLMs are sentient. The machine doesn't have feelings but it can do a good impression of having them because it's mimicking human discourse. There's a strange dynamic here: using 'please', 'thank you' and generally minding your manners can produce better results because the machine is simulating how a helpful person would behave.

LLMs are deterministic. I've left the most important one for last. We're used to machines taking 2 plus 2 and giving us 4. Since computers appeared in the 1950s they've always been deterministic, following rules where the same input receives the same output. But LLMs aren't deterministic. LLMs are stochastic. They're random, just like humans, which means a deterministic approach to 'talking' to them won't work.

These misconceptions matter because they shape how people approach the tools. If you assume an LLM is deterministic you'll be baffled when the same prompt gives you different results on Tuesday. If you assume it's always learning you'll wonder why it keeps making the same mistakes. Getting the mental model right is the first step to getting useful results.