New technology will always have its disciples and its detractors. Artificial Intelligence (AI) is no different. As technologies go AI has the most vocal disciples and detractors. Depending on who you ask, AI will lead us into a utopian new age or a cyberpunk post-apocalypse nightmare. Somewhere in between those diametrically opposed views exists the current reality of AI.
If one was looking for an accurate yardstick of where AI is today, then OpenAI’s GPT-3 is that yardstick. Initially proposed in an academic paper, it generated buzz around the AI community, but more was to come. When it was released as a closed beta in the second half of 2020 it was done with arguably too much hype and fanfare. Now at the start of 2021, we have a much better understanding of the capabilities and limitations of the technology. It is still ground-breaking, but its limitations are still evident.
What is it?
Starting at the beginning GPT-3 stands for Generative Pre-trained Transformer 3. The tree symbolizes that it is in its third iteration. The technology’s name is fairly nondescript in terms of what it actually does. In short, GPT-3 can generate text using learned algorithms. To do this, 570 GB of text data were fed into its neural network. The data was gained from crawling the Internet and uploading popular data repositories like Wikipedia. By using semantic analytics and a technique known as unsupervised learning GPT-3 generates text.
In practical terms, this means that you can ask GPT-3 a question and you will receive an answer. This is not new but how this is done, and the accuracy GPT-3 achieves is new. Not only will it answer questions, but it can write summaries of long texts, write essays and poems, and can write code. One of the reasons GPT-3 generated so much hype is that it is the largest neural network to date. As to how the technology works in greater depth an excellent article was published on Forbes and is a must-read.
GPT-3 even wrote an article for The Guardian in an attempt to try and convince us, humans, it means us no harm. Whether the article achieved its desired effect is up for debate. A writing critic may even slam the text for being too robotic. The irony of that sentence is not lost on the writer or the AI in all probability. The question of if the article is a genuine assurance or an attempt by the future AI overlords to subvert our attention only time will tell.
Real-world Applications
If 2020 is hailed as the year that AI took a massive leap forward, 2021 will be remembered for how humans looked to reap the rewards. Currently, several start-ups and apps are looking to make full use of GPT-3 to perform several functions. One of the most popular, and practical uses of GPT-3 is for email authoring. Other apps are looking to make use of the technology for better trivia apps and tweeting the masses.
At GMCOLAB we are lucky enough to be part of the closed beta. Of particular interest to us is the email authoring capabilities that can be unlocked with the technology. It is no secret that emails are the bane of many employees’ workday. Any technology that not only automates replies with high levels of accuracy and does not sound like a chatbot will be beneficial. With that said we have encountered some limitations with the technology.
Developers, your Job is still Safe
While GPT-3 can generate text, it still lacks a true understanding of the meaning of words. This also applies to contextual changes of meaning when the word is used differently. The Internet is full of the hilarious side of this with meaningless replies to specific questions. The replies can also be interpreted as exhibiting bias at best, at worst replies generated are downright racist or sexist. This is due in part to the prevalence of such content on the Internet and the AI’s inability to understand the meaning. These limitations have been acknowledged by OpenAI, with the third version never meant to correct all the AI’s failings. Rather, being the third in a long line of substantive improvements.
Other limitations include the closed nature of the API. Those with access do not have unrestricted access to the code. Use of the code is referred to as being done via a black-box API with the source code remaining incredibly private. Another limitation is the vast amount of computing power required to run operations. This results in increased costs. This may place the technology far beyond the scope of smaller organizations to use effectively. This is expected to improve sooner rather than later as hardware prices continue to fall.
For the project manager who is looking to get rid of their developers and hire an AI; best not to schedule a meeting with human resources anytime soon.