In this AI tools guide, we'll provide a detailed assessment of the viral chatbot that took AI mainstream: ChatGPT, including it's key capabilities, limitations, competitors, and more.
About ChatGPT
Developed by the renowned AI research firm OpenAI, ChatGPT is a conversation AI assistant that has taken the world by storm and become the fastest growing app of all time. Launched on November 30, 2022, ChatGPT crossed 1 million users in its first week:
ChatGPT launched on wednesday. today it crossed 1 million users!
— Sam Altman (@sama) December 5, 2022
For context, here's how long it took other major tech companies to crack the 1 million user mark:
It took Netflix 3.5 yrs to hit 1 million users.
— Contrary Research (@Contrary_Res) December 8, 2022
It took Twitter and GPT-3 24 months.
It took Pinterest 20 months
It took Facebook 10 months
It took Copilot 6 months
It took Spotify 5 months.
It took Dall-E and Instagram 2.5 months.
And it took ChatGPT 5 days.
h/t @alliekmiller
What is ChatGPT?
ChatGPT is an AI chatbot with a knack for generating human-like text using natural language processing. As OpenAI writes:
ChatGPT is a sibling model to InstructGPT, which is trained to follow an instruction in a prompt and provide a detailed response.
ChatGPT is a highly versatile AI assistant that can provide answers to complex questions, generate code, and much more. ChatGPT offers a conversational dialogue format that allows it to engage in follow-up questions, admit errors, challenge faulty premises, and decline inappropriate requests.
Capabilities of ChatGPT
ChatGPT's ability to generate content, information, and interactivity is undoubtedly phenomenal (in our opinion). It can be a valuable assistant in a wide variety of tasks, including programming, content creation, and creative writing—making it far more than just an AI chat program. Its strength comes from its breadth of knowledge (albeit limited to September 2021) that enables better learning as well as giving substantive and quality content in different fields of expertise.
Another key component of ChatGPT interface is its evolving ecosystem with ChatGPT plugins, web browsing, code interpreter, and (upcoming) multimodal inputs.
ChatGPT Plugins
As highlighted in OpenAI's ChatGPT Plugin post:
Plugins are tools designed specifically for language models with safety as a core principle, and help ChatGPT access up-to-date information, run computations, or use third-party services.
The demand for plugins has been clear ever since ChatGPT’s inception with many developers exploring similar concepts. The inclusion of plugins unlocks an incredibly wide range of potential applications.
This roll out includes plugin developers, users of ChatGPT Premium, and following an alpha testing period API users who might be interested in integrating plugins into one’s products.
Internet Browsing (Alpha)
One key plugin that OpenAI developed interanlly is web browsing. Integrating browsing capabilities into language models greatly widens the scope of topics ChatGPT can accomplish. With browsing, the LLM model is enabled to pull in up-to-date information from the web, augmenting their training data with current facts and trending issues.
Code Interpreter (Beta)
Code Interpeter is another plugin OpenAI developed internally that provides the model with a working Python interpreter safely stored inside their isolated environment. When you use the interpreter plugin to run code, it’s evaluated in a session that persists as long as your chat conversation is saved, though there are time limits by default. So it’s kind of like building a code tower where every call forms a new block stacked upon the others. Plus, you can load files directly into your conversation workspace and then download the results when you’re ready.
Limitations of ChatGPT
Just like all other LLM-enabled chatbots, ChatGPT has its limitations and it's important to understand what ChatGPT can and cannot do.
1. Knowledge Cutoff Date
One key limitation is the knowledge cutoff date of September 2021 at the time of writing. This means it doesn’t know about events that have happened after this point in time. However, the aforementioned Web Browsing plugin is an attempt to bridge this gap by enabling ChatGPT to access and provide more recent information from the web for use in conversations. There's also a good chance OpenAI will update the knowledge cutoff date, especially with competitors like Bard much more recent context windows.
2. Hallucination
Another key limitation of ChatGPT and many other AI assistants is "hallucination", which means ChatGPT will generate information that seems as if it could be true but actually isn’t or is entirely made up. This occurs because often times the AI just filling gaps in its knowledge based off of the data depending on how far back it had been trained on. Often times it gets the details wrong, so any sort of factual information gathering should always be human-verified
3. Bias
ChatGPT has also been shown to mirror that which exists within the data, which includes human biases. OpenAI continually works on reducing these biases and improving how the model responds to different inputs, although this is an activate area of research & development.
How was ChatGPT trained?
As highlighted in OpenAI's ChatGPT blog post, ChatGPT was developed through a process known as Reinforcement Learning from Human Feedback (RLHF). This process shares some similarities with how InstructGPT developed. However, there are some differences in the way data is gathered for this model.
The initial model was fine-tuned in a supervised learning environment in which human AI trainers provided conversations assuming both roles—that of the user and the virtual assistant. To facilitate responses, the trainers could opt to make use of suggestions generated by the model itself. This curated dialogue dataset was then woven into the InstructGPT dataset, which had been reorganized into a dialogue-friendly format.
One element in reinforcement learning involves developing a reward model. Since this required comparative data, essentially two or more model-generated responses evaluated and ranked according to quality, OpenAI used conversations between the AI trainers and the chatbot. In these dialogues, a message crafted by the model was chosen at random, followed by the generation of several alternative endings. These alternatives were then ranked by the AI trainers.
Using an approach called Proximal Policy Optimization, the chatbot was fine-tuned and iterated over this process many times to improve ChatGPT’s performance.
GPT 3.5 vs. GPT-4
At the time of writing, GPT-4 is the latest OpenAI model achievement and has demonstrated to be a considerable improvement performance. As OpenAI writes:
GPT-4 is a large multimodal model (accepting image and text inputs, emitting text outputs) that, while less capable than humans in many real-world scenarios, exhibits human-level performance on various professional and academic benchmarks.
For example, GPT-4 scored within the top 10% on the simulated bar exam—a dramatic improvement from GPT-3.5’s score which was at the bottom 10%. This standpoint advanced through six months of iterative alignment using insights gained from OpenAI’s adversarial testing program and from ChatGPT to yield impressive improvements such as factuality and steerability.
The difference in these two models may be immediately noticeable at first glance, although for certain tasks such as programming it's remarkedly better than GPT 3.5.
While OpenAI hasn't rolled out multimoel input just yet, they have announced plans to support uploading visual inputs such as images. If you're interested in learning more about GPT-4, check out the developer livestream below:
ChatGPT Pricing
As of June 2023, ChatGPT has two offerings: the free version and ChatGPT Plus, a subscription-based service priced at $20/month. The paid version promises faster response times, priority access even during peak usage times, and first access to new features and enhancements.
ChatGPT also has an API, enabling developers to integrate ChatGPT models into their apps and products.
Learning Curve and Usage
Mastering the use of ChatGPT might take a bit more practice, especially if you are just starting out on this AI journey. Much like becoming more efficient & proficient as a writer, learning how to pose effective prompts to the AI is something that can only be learned over time, education, and experience. While it might seem daunting at first, this learning journey is very worthwhile in terms of exploring the potential of AI and large language models in general.
ChatGPT Competitors
Along with the AI gold rush comes many ChatGPT competitors that are trying to steal its rapidly adopted users. Key ChatGPT competitors include Bard from Google (Alphabet) and LLaMA, developed by Meta.
Also, while not a competitor in the traditional sense, MicroSoft is integrating ChatGPT into Bing Chat (as they call it, the new Bing). There are many other smaller competitors, and surely new larger players will enter the market in this the expanding landscape of conversational AI.
Summary: ChatGPT Review
ChatGPT provides an incredibly impressive set of features, which continuously evolves and expands to provide improved performance.
If you know how to use it effectively, ChatGPT serves as one the most useful companion tool, and many users (myself included) find it well worth the $20 monthly investment, as well as time to commitment to educate yourself and practice how to effectively. That said, once you get the hang of it, ChatGPT can drasticall improve your business and personal productivity in many ways.
Related ChatGPT Tutorials
Here are a few more tutorials that will help you start using or master ChatGPT:
Peter Foy
Peter FoyPeter FoyPeter Foy
