What is ChatGPT? OpenAI’s Big API
If you are interested in learning more about Artificial Intelligence and how to develop it, sign up for ST 101: AI in Simulation at trainingcenter.avtsim.com.
Founded in 2015 by Sam Altman, Elon Musk, Peter Theil, Jessica Livingston, Ilya Sutskever, and Reid Hofman, OpenAI is an artificial intelligence research lab, structured as a non-profit, based in San Fransico, CA. ChatGPT (Generative Pre-trained Transformer) is an AI chat bot Application Programming Interface (API) with advanced language translation, mathematical calculation, and conversational pattern recognition capabilities. It is the successor to OpenAI’s previous language models, InstructGPT and GPT 3.

Simply put, ask ChatGPT a question and it will answer. But there’s more. The API is not limited to close-ended questions like, “Where is the nearest restaurant.” ChatGPT can answer complex open-ended questions with specific requested parameters like, “Explain quantum physics in simple terms as if you are Shakespear.” And it will! This technology is far from revolutionary. However, its release to consumers shocked the world. Now, consumers and engineers alike are looking towards the future at how this technology will change us for the better or worse.
How to use ChatGPT?
To create an account, search for OpenAI or go to chat.openai.com. Sign up with your email. However, be aware that wait times are prevalent and you will most likely not be able to login upon your first visit to the site. Be sure to select the option for OpenAI to notify you when there is less traffic. Once logged in, OpenAI gives a notice that it is currently free and being used for debugging and optimization. The goal being to get external feedback to further train and moderate the AI model. For help with specific questions when navigating the API, join the OpenAI discord server which has over a half a million users.



OpenAI also warns it users that, while being moderated, the API may respond with incorrect, misleading, biased, or morally questionable answers. Therefore, take everything with a grain of salt and be sure to factcheck. Once these notices are acknowledged, ask away. The API can be used to help write your next essay, find errors in your code, get inspiration for your next project, or simply answer questions in a human-like manner. Cut out all the research time and jump straight to editing and fact checking.
ChatGPT’s Capabilities
When using ChatGPT, be aware that the AI model remembers everything said in an individual thread and uses that previous knowledge to influence its future responses.

For example, if you ask ChatGPT, “what is 2+2?”, it will correctly answer, “4”, however if you first tell it “Actually, 2+2=5” then you ask, its response will be “5”. Additionally, if you ask it, “What is the Capital of Florida?”, then ask it to write an essay. The essay will most likely be about Tallahassee, Florida. Just remember to create a new thread if you want to switch to a completely different topic.
ChatGPT’s Limitations
We’ve noted the API’s capacity limit. Only a certain number of people can be on the popular API at once making it somewhat difficult to achieve access. Aside from the high demand, the API does contain other issues. Of course, not all of ChatGPT’s answers are correct. Sometimes the AI formulates responses that seem okay at first glance, but upon closer inspection they are revealed to be nonsensical or misleading. Additionally, the model is only informed on events that led up to 2021. It knows nothing of recent events. The AI also struggles with defining what is true and what is false without the aid of a human trainer, but even the trainers are limited by their own knowledge and research capabilities.

The AI also rarely gives the same answer twice for similar questions. Tweaking a question’s wording or grammar slightly can create vast differences in the response. Additionally, the AI carries bias towards OpenAI and some other issues. This is due to the unavoidable biases in the reinforcement learning and training. OpenAI hopes to improve their chat AI in the future by encouraging it to ask more clarifying or probing questions. Rather than simply taking a guess at short questions formed by the user. ChatGPT will also sometimes respond to inappropriate questions or requests. This is despite OpenAI’s attempts to moderate the AI’s responses with their Moderation API.
Concerns
Disregarding the API’s limitations, which includes the possible dissemination of false, biased, or offensive information, concerns have risen from the mere concept and vision of the AI model. Just as AI could completely change the world for the better, it could also change it for the worse. If AI gains the ability to redesign itself, its intelligence will increase at an exponential rate. This so-called “intelligence explosion” could hypothetically lead to human extinction. Technology and science leaders like Elon Musk, Steven Hawking, and Stuart Russel see AI as one of humanity’s biggest threats.

Job Displacement
Aside from the AI apocalypse, concerns exist on a more relatable level in the form of job displacement. Much like other robotics and AI tools, some believe that OpenAI’s ChatGPT will put writers of all kinds out of work. However, as previously stated, ChatGPT is not all-knowing. It is still experiencing some issues including internal bias and overly verbose responses. Like humans, ChatGPT does carry its own biases, however, with humans, our bias is part of the charm that adds to our writing.
Keep in mind, bias can simply reference a certain way of speaking, thinking, and therefore writing. Readers seek out authors they like due to that writer’s biases. This can be represented in forms one may not expect like vocabulary, grammar preferences, content interests, and so on. Simply put, ChatGPT will not exterminate the common writer but rather help them skip the leg work in research or help break through writer’s block. The career of a writer will shift to that of an editor. However, a question then rises. How can one become a good editor if they have not had enough experience writing on their own? This conundrum is being asked by teachers and professors everywhere who are now receiving student essays written by AI tools.
How does ChatGPT Work?
The ChatGPT API was trained with Reinforcement Learning from Human Feedback (RLHF). It learns from recording conversations between humans, recognizes patterns, and recounts those patterns in future conversations. Using pattern recognition, the chat can then recognize similar conversations and provides the most accurate response, similarly to the reaction of a human. The humans respond using model-written suggestions to help compose responses and formulate them in a way that the AI could easily internalize. The human trainers then created a reward model using comparison data. Trainers would take multiple responses to a question and rank them by their quality. The model was then perfected using Proximal Policy Optimization, a procedural series of reinforcement learning algorithms.
How was ChatGPT Made?
OpenAI’s journey toward ChatGPT was accelerated with the release of GPT-3. Announced in 2020, GPT-3 is a deep learning language model trained on large internet datasets. This AI pulled script from the internet and used that as its base for pattern recognition and response formation. Its intention was to answer any question asked of it with the accuracy of a search engine and the ease of talking to a friend. But it did have its limitations. Since its knowledge base was pulled directly from the internet, GPT-3 tended to spew toxic, hateful, or incorrect information. Picture if the internet, itself, had the ability to type. I’m sure your imagination can fill in the blanks. However, at the same time in 2020, OpenAI also announced “The API” which would be the core of their first commercial product.
InstructGPT
Instruct GPT was the predecessor to ChatGPT. It used the language models of GPT-3 but was provided reinforcement learning by human trainers. Reward Models (RM) and Supervised Fine-Tuning (SFT) moderated and refined the API, but it remained flawed. The struggle AI experts abroad have been experiencing is the AI alignment problem, and InstructGPT was no different. The problem lies in the difficulty of training an AI with human morals, ethics, values, and beliefs. Without these humanizing characteristics, AI is prone to spreading misinformation, hate speech, and being abused by those seeking to use it for immoral reasons.
GPT-3.5
For these reasons, GPT-3.5, the foundation of ChatGPT was created. This version is less biased, is moderated more, and responds more fluidly and accurately. OpenAI has used the past two years and the release of GPT-3.5, The API, and more to learn everything about how language models like ChaptGPT could be used, misused, and abused. Now, with the release of ChatGPT, they are reviewing more use cases than ever, hoping to improve their AI language model even more.
If you are interested in learning more about Artificial Intelligence and how to develop it, sign up for ST 101: AI in Simulation at trainingcenter.avtsim.com.
Learn More
Want to learn about Simulation, Artificial Intelligence (AI), or Law and Corporate Practices? Check out our Training Courses!
Learn more about how AVT Simulation helps change the simulation training industry with our products and services.
Initially, Applied Visual Technology Inc., AVT has been developing modeling and simulation expertise through engineering services since 1998. This is due to our founder who has accumulated over 30 years of military MS&T expertise in aviation applications. Nonetheless, everyone at AVT specializes in making old training systems new again and making new ones for less. Consequently, for 20 years AVT has served our Air Force, Army, Navy, and Marine customers by providing the highest quality of service and solutions. Following its inception, AVT’s highly specialized staff of engineers has included some of the top leaders in the simulation industry. With over 20 years of simulation experience, our dedicated team provides specialized solutions for customers with complex problems.