All the Language AIs are nothing without the help of Humans

The media frenzy surrounding ChatGPT and other big language modeling AI systems covers a wide range of topics, ranging from trivial – large language models that can replace conventional web search – to ominous. – AI will eliminate many jobs – and overworked people – AI causes extinction- levels that threaten humanity.

All of these topics have a common denominator:
amazing language models announce an artificial intelligence that will replace the human race.

But big language models, with all their complexity, are really silly. And although they are called “artificial intelligence”, they are completely dependent on human knowledge and labor. Of course, they cannot reliably generate new knowledge, but there is much more to it. ChatGPT can’t learn, improve, or even maintain the current one without humans feeding it new content and telling it how to interpret it, not to mention programming models and building, maintenance as well as providing equipment for it. To understand why, you must first understand how ChatGPT and similar models work and the role humans play in making them work.

In general, large language models like ChatGPT work by predicting which characters, words, and phrases will follow in sequence based on a set of training data. In the case of ChatGPT, the training dataset contains a large amount of public text pulled from the Internet.

Imagine I’m training a language model on the following set of sentences:
Bears are large furry animals. Bears have claws. Bear is the secret robot. Bears have noses. Bear is the secret robot. Sometimes bears eat fish. Bear is the secret robot.

The model will tend to tell me that the bears are more secretive robots than anything else, because this sequence of words appears most often in its training dataset. This is clearly a problem for models trained on inconsistent and error-prone datasets – that is, all of them, even the academic literature.

People write so many different things about quantum physics, Joe Biden, eating healthy or the January 6 uprising, some things being more valuable than others. How does a model know what to say about something, when everyone is saying so many different things? Comment Needs This is where comments come in. If you use ChatGPT, you will notice that you have the option to rate the responses as good or bad. If you rate them as bad, you will be asked to provide an example of what a good answer should look like. ChatGPT and other great language models learn which answers, predictive text strings are right and wrong through feedback from users, development teams, and contractors hired to label the output. ChatGPT cannot compare, analyze or evaluate arguments or information on its own. It can only generate text strings similar to those used by others when comparing, parsing, or evaluating, favoring strings similar to those deemed good answers. in the past.

when the model gives you a good answer, it relies on a lot of human work that helps it know what is a good answer and what is a no. There are many, many workers hidden behind screens, and they will always be needed if the model continues to improve or expand its content coverage.

A recent investigation by journalists published in Time magazine revealed that hundreds of workers spent thousands of hours reading and labeling racist, sexist, and annoying articles, including including images depicting sexual violence, from the darkest pages on the internet to teach ChatGPT not to copy such content. They get paid no more than $2 an hour, and many say they suffer from psychological stress from work. 

The importance of comments is directly seen in ChatGPT’s trend of “hallucinations”; means confidently giving incorrect answers. ChatGPT cannot give good answers on a topic without training, even if good information on the topic is widely available on the Internet.

You can try this out for yourself by querying ChatGPT for more or less obscure things. I find ChatGPT to summarize the plots of various works of fiction particularly effective because, it seems, the model is more rigorously trained in nonfiction than in fiction.

In my own tests, ChatGPT summed up the plot of the JRR. Tolkien’s The Lord of the Rings, a very popular novel, has only a few flaws. But his synopsis of Gilbert and Sullivan’s Pirates of Penzance and Ursula K. Le Guin’s Left Hand of Darkness – both of which are a bit more specialized but by no means ambiguous – is like playing Mad Libs with names of characters and places. No matter how good the respective Wikipedia pages of these works are. Models need comments, not just content. Because large language models don’t actually understand or evaluate information, they depend on humans to do it for them. They parasitize human knowledge and work. As new sources are added to their training dataset, they need new training on whether and how to construct sentences based on these sources.

They cannot judge whether the reports are accurate. They cannot evaluate arguments or weigh trade-offs. They can’t even read a page of an encyclopedia and just make statements that fit it, or accurately summarize a movie’s plot. They rely on humans to do all these things for them.

They then paraphrase and remix what the humans said, and rely on even more people to tell them if they’re interpreting and remixing well. If the general understanding of certain topics changes – for example, whether salt is bad for your heart or whether early breast cancer screening is helpful – they will need to be widely reused to make new consensus.

In short, by no means a foretaste of a completely independent AI, large language models illustrate the complete dependence of many AI systems, not only on designers and maintainers but as for their users. So, if ChatGPT gives you a good answer or a useful answer about something, don’t forget to thank the thousands and millions of anonymous people who wrote down the words he mulled over. and teach him what is the right and wrong answer.

Far from being an autonomous superintelligence, ChatGPT, like all technologies, is nothing without us.