Likewise, if you don't believe in "Artificial Intelligence", take five minutes to ask it. Or preferably, five minutes to understand how it works and what it ca

Likewise, if you don't believe in "Artificial Intelligence", take five minutes to ask it. Or preferably, five minutes to understand how it works and what it can and cannot do, then five minutes to ask it something actually suitable. "AI" (LLMs) are currently good at: - language understanding, i.e. understanding and processing text you provide it. For example, taking a wall of text and answering questions about points mentioned there, or general sentiment, or extracting data from it etc. - somewhat general knowledge, i.e. stuff that was sufficiently frequently represented in the training data Absent additional tricks, "AI" is really bad at obscure knowledge or complex, multi-step thinking. We are slowly getting there, but we aren't there yet. This is not something the LLMs do, but rather the wrappers around them that provide the model with tools to get additional information and first prompting the model to select the tools, then repeated prompts with the output of the tools. A good rule of thumb is that if an average well-educated intelligent person could answer it without further research, a LLM will probably be able to. I'd even say that if an average fresh out of school graduate of the corresponding discipline can answer it quickly (without further research or sitting down for ten minutes and doing the math), there's a good chance AI will be able to answer it, but it might also get it horribly wrong and you will have a hard time distinguishing between those if you have no knowledge in the field. As the author mentions at the very end of the article, the hallucination problem also means that the best kind of tasks are where you can quickly verify whether the response was useful. A system that produces misleading responses 50% of the time is useless if you can't distinguish them, but very useful if in those 50% it saves you ten minutes of work and in the other 50% you lose a minute by trying.