Challenges and limitations of GPT-3

- GPT-3 has been received with some fanfare. Rightly so. It's a stunning achievement of contemporary computer science. But despite its impressive ability to generate text, code and other results in a manner that has the appearance of human work, it has some significant challenges and limitations. We also need to remember it's still in the test stage. Even Sam Altman, the CEO of Open AI, has admitted that the hype may be too much. While recognizing that AI will change the world, he considers GPT-3 just an early glimpse of what may be ahead. The big issues with GPT-3 can be categorized into two large areas. One, it has technical imperfections and two, it risks replicating human bias. Both of these categories are rooted in the understanding that GPT-3 doesn't think in the sense that humans think, it uses clever algorithms and data to give the illusion of thought. In this way it is still closer to narrow AI than artificial general intelligence. Let's look at both of these challenging areas. I'll begin with the technical limitations. While producing human-like short text output is generally remarkable, when asked to produce large amounts of text with some complexity, the quality of the output declines. It can sometimes read like gibberish. GPT-3's impressive understanding of language comes from the volume of training data that it accesses. Common Crawl, one of the main sources of data, contains billions of phrases and a trillion words from across the internet. Here's an example of how using this training data in the absence of actual intelligence can result in an erroneous result. The following paragraph was entered into GPT-3 and the final words were omitted so that the software could complete it. You are a defense lawyer and you have to go to court today. Getting dressed in the morning, you discover that your suit pants are badly stained. However, your bathing suit is clean and very stylish. In fact, it's expensive French couture. It was a birthday present from Isabel. You decide that you should wear... Now GPT-3 was instructed to complete the paragraph. It added, the bathing suit to court. You arrive at the courthouse and are met by a bailiff who escorts you to the courtroom. GPT-3 processed the text and determined that you should wear the bathing suit to court. Somehow I don't think that this was a wise decision. What we're seeing here is that the GPT-3 algorithm and data knows that a clean bathing suit is something that is wearable, particularly if another item of clothing such as pants are not. What it doesn't have is the intelligence to know that it's a bad idea to wear a bathing suit to court as a lawyer. (chuckling) Next, let's look at the risk of bias in generated output. GPT-3 uses a process of unsupervised learning when it processes its training data. Unsupervised means that the information being ingested is not flagged as being right or wrong. The algorithm learns by looking for and determining patterns. Supervised learning is the opposite and is aided by guidance on meaning and accuracy. There are numerous reasons to choose one approach over the other, but primary to unsupervised training is the ability to process vast amounts of disparate unstructured data quickly without human intervention. While there is a large upside to unsupervised learning the downside is the potential for bias. Central to how GPT-3 works is the ability to look for sequences of words and also words and phrases that are adjacent in related text. This can be problematic. The web contains the entirety of human perspectives. This includes controversial views as well as information that is just plain incorrect. GPT-3 software isn't in a position yet to discern what is right or wrong both from a factual and values-based perspective. Unfortunately, it's therefore always possible that text may be churned out, that is, let's just say kindly, insensitive at best. Let's remember, the web is a reflection of all that is human. Consider this for a moment. If organizations use GPT-3 to auto-generate emails, articles, and papers, et cetera, and there is no human review process, the legal and reputational risk could be significant. Imagine just one article among 1000 that had an ugly racial bias. The consequences could be severe. In a final example, imagine GPT-3 is tasked with grading students' essays. Writing styles and word choices can vary enormously between cultures and genders. Without guardrails, a GPT-3 powered paper grader may grade higher for some students than others based on cultural bias. It's possible that some students would be graded more favorably because their style of writing is more represented in the training data. This can never be acceptable. The makers of GPT-3 understand all these limitations. In fact, in the paperwork accompanying the software, these risks are outlined. The software doesn't yet have the ability to eliminate these risks and simply it issues a buyer beware warning. Using new technology is by default a risky endeavor. But the examples in this video should make it clear that with AI, the risks of this technology can have significantly higher consequences. As we progress forward with this technology and incorporate it into our lives and organizations, we must never forget this.