Modern AI has demonstrated such amazing capabilities, leaving many people fearful that humans will be pushed out of their jobs and replaced with ChatGPT and friends. Indeed, big tech has been laying off a large percentage of their workforce, claiming that the use of AI will be able to replace, and maybe even surpass, the results generated by the intellects of their former employees. Is this the end? Should we acknowledge our inferiority to the machines?
No, I don't think so. AI still has so many failure modes (and will have them for the foreseeable future) that it is supremely unwise for a company to replace a large chunk of its employees with calls to ChatGPT. I will catalog some of the failures that I have encountered recently. After reading them, I hope you will agree with me that humanity isn't done just yet.
Today, let's consider a problem that is very straightforward but involves data-fetching and some tedious analysis. Each state in the U.S. elects to Congress two senators and a number of House representatives that depends on the state's population. The most populous state, California, has 52 representatives. Several states with low populations have one representative each. Suppose we wish to know how many states have four representatives or fewer. This seems like a good task for an LLM, right? Let's ask Google.
Sounds good. But let's ask again to make sure.
Huh? Uh, let's try a few more times...
Hmmmmmm. Thanks, Google, that was very helpful. Google performed the addition correct in each case, but it is getting bad information from somewhere. I was surprised that Google's AI had so much trouble with this, since it is pretty easy to find a list of the states and their corresponding numbers of representatives.
For the record, the correct answer (if I did my math right) is 21. If you calculate it yourself, let me know if I actually did it right. 😁 Surprisingly, Google didn't even come up with this answer by accident.
So take pride in the fact that your brain isn't obsolete, for the time being.
