Why Chat GPT is a Big Deal
ChatGPT is a breakthrough in generative AI–artificial intelligence that can produce its own unique content. It’s a large language model (LLM) that has trained on trillions of data points to gain massive predictive power when it comes to text generation. It’s an incredibly efficient information organizer that can create written content, code, and other formulas that are accessible based on its knowledge base.
This kind of big-brained, lofty generative AI has been attempted before, but not with astonishing success. Without enough parameters, training, and feedback they’ve often flopped comically.
Take Meta AI’s Galactica, released last November as an interactive database for scientific knowledge. The LLM trained on 48 million scientific papers, and its answers sounded legitimate. There was just one problem: They were drivel. Information was usually wrong and sources were bunk. Meta AI suspended the tool after two days.
ChatGPT, however–an extension of OpenAI’s GPT-3–is trained on 570 gigabytes of text and has over 175 billion parameters. Its responses have been fine-tuned over the course of around 300 years of training completed over the course of a few months. Needless to say, content creators aren’t laughing any more.
How Does Chat GPT Learn?
GPT-3 and GPT-3.5 were trained with a combination of techniques (and one particular innovation) that gave them unprecedented predictive capabilities. First, they learned on massive labeled datasets–like really massive. For comparison, the Pile is an open-source language modeling data set that’s 825 GiB. That’s just a small fraction of the data needed to prepare an LLM like ChatGPT for the functionality it’s trying to attain.
GPT-3 trained on these datasets, then was fine-tuned during a phase of data comparison and reward model implementation. Labelers rank several data outputs based on their accuracy, then use that to train a reward model. Where ChatGPT breaks the training mold, however, is through its use of Reinforcement Learning with Human Feedback (RLHF).
Not sure what RLHF is? Just ask ChatGPt and it’ll tell you itself (as long as you ask the right way):
ChatGPT uses a variant of reinforcement learning called Reinforcement Learning with Human Feedback (RLHF) in addition to the pre-training and fine-tuning process.
RLHF is a method of training a language model using reinforcement learning where the model receives rewards or penalties based on feedback from human evaluators. This allows the model to learn from feedback that is more in line with human preferences and expectations, rather than relying solely on the likelihood of the training data.
ChatGPT is the first Large Language Model to train via RLHF. Basically, GPT-3 was fine-tuned with lots and lots of data that reinforced human preferences. The result was ChatGPT, which can craft outputs that account for sophisticated levels of communication nuance and human intent. Thus, reinforcement learning enables it to produce the most conversationally authentic, AI-generated answers in the history of artificial intelligence.
ChatGPT is a remarkable innovation in generative AI, but that doesn’t mean it’s perfect. At its core it’s still just a really really good word and content predictor. It isn’t sentient, it isn’t alive. Because of this, it has a few drawbacks.
It struggles to distinguish between truth and falsehood
GPT-3 has been trained on thousands of web pages, language conventions, and content pieces, but it can’t always evaluate or distinguish between the quality of information (it even gave me false info when I asked it to describe how it was trained). Sometimes ChatGPT will go too far in its content delivery to the point of even making up and misattributing quotes. So if you’re looking for help on writing a research paper or your latest piece of journalism, tread with caution on its sourcing.
The latest information is from 2021
The dataset pre-training takes months. Over the course of that time things have a tendency to change. If you’re looking for the latest information, search engines are still the way to go, and ChatGPT will lack the context of the most up-to-date content.
Also, while it pulls and organizes information based on an incredibly intricate system, it can’t create new, accurate information. Therefore, it struggles to provide anything substantive on recent topics, like a current event or a new product
Stilted, generic voice and communication style
ChatGPT is essentially a huge formula so it makes sense that the writing would be a little formulaic. The algorithm does an amazing job replicating many nuances of human speech and communication. But for anything that requires a distinct voice and originality, ChatGPT may be more of a rhetorical starting point than an endpoint. It can create some fun, rhyme-schemed diddies, but struggles when its artistic conventions are overstretched. Art, therefore, still seems to be the mostly exclusive playground of the human soul.
So How Can I use LLMs like ChatGPT in My Own Work?
When used responsibly, ChatGPT and other natural language models are an incredibly effective content generation, programming, and research tool. It can help program apps, research, structure content, write content, brainstorm–the possibilities are limited, but somehow endless.
Halda, a company that specializes in creating personalized content platforms for website visitors, leverages a large language model for efficiency and effectiveness. Natural language models help clients build custom forms that prompt visitors to share information about themselves. Then that information becomes the foundation for a customized, personalized content experience based on that visitor’s interest.
Large language models make this process more effortless and efficient. It’s just one of the ways that LLMs like ChatGPT help companies streamline their workflow. And that’s the ultimate goal: To leverage AI and its amazing capabilities to have a palpable effect on our daily productivity. To turn the miraculous into modern practice.
AI and Higher Ed: An Increasingly Common Combination
Artificial intelligence functions in almost all the layers of the education system. That’s no surprise considering that much of...
Will Chat GPT and AI Take My Job in Higher Ed? Absolutely Not (If you know how to use AI)
To academics and those associated with higher education, artificial intelligence can seem a little redundant. After all, who...
How does AI work? And What Are the Different Types of AI?
At this point at the dawn of 2023 you’ve probably heard it said, “AI might not replace people, but the people who don’t learn...