Wednesday, 18 January, 2023 UTC
Summary
ChatGPT: A sneak peek into its abilities, limitations, and how to make the best use of it.
ChatGPT is a Large Language Model (LLM) trained by OpenAI. Unlike traditional language models that are only able to perform specific tasks, ChatGPT can perform a wide range of NLP tasks with impressive accuracy. This includes question answering, engaging in human-like conversation, and even code generation and review. You can even ask ChatGPT to write you a song, tell you a story, write an email, or play a game together! The versatility of ChatGPT sets it apart from state-of-the-art language models. In simple terms, ChatGPT could potentially allow human-machine collaboration by enabling us to discuss, learn, understand, and be inspired by new ideas. With its impressive capabilities, will ChatGPT overtake Google search as the go-to source for information and knowledge? Microsoft reportedly added ChatGPT to the Bing search engine.
Figure 1. InstructGPT Architecture ( ChatGPT is a sibling model of InstructGPT, and it is believed to be trained in the same way).Technology
ChatGPT is based on using generative pertaining using Reinforcement Learning from Human Feedback (RLHF), it is deemed an enhanced version of GPT-3, however with 100 times fewer parameters. ChatGPT is trained using reinforcement learning with human-in-the-loop, 40 human labelers to be exact. There are 3 main phases that build ChatGPT as depicted in Figure 1 above. Firstly, the role of the labelers is to specify demos of the expected model behavior. Secondly, they collect a dataset that is composed of human-labeled comparisons between two model outputs by rating multiple QA pairs and ranking them based on quality. A reward model is trained on this data with the goal of generating an output that would be preferred by the labelers. Finally, using the reward model as a reward function they fine-tune the GPT-3 policy to maximize the reward.
ChatGPT is based on using generative pertaining using Reinforcement Learning from Human Feedback (RLHF), it is deemed an enhanced version of GPT-3, however with 100 times fewer parameters. ChatGPT is trained using reinforcement learning with human-in-the-loop, 40 human labelers to be exact. There are 3 main phases that build ChatGPT as depicted in Figure 1 above. Firstly, the role of the labelers is to specify demos of the expected model behavior. Secondly, they collect a dataset that is composed of human-labeled comparisons between two model outputs by rating multiple QA pairs and ranking them based on quality. A reward model is trained on this data with the goal of generating an output that would be preferred by the labelers. Finally, using the reward model as a reward function they fine-tune the GPT-3 policy to maximize the reward.
Limitations
As claimed by OpenAI, ChatGPT is trained on data up to 2021 only, meaning that it is not able to provide any answers or insights to data generated in 2022. Search engines like Google remain valuable because they allow us to find the most current information available as opposed to applications like ChatGPT, which may not always have the most recent knowledge, as their data is pre-trained and may not be updated as frequently. Both types of resources can be useful in different situations.
As claimed by OpenAI, ChatGPT is trained on data up to 2021 only, meaning that it is not able to provide any answers or insights to data generated in 2022. Search engines like Google remain valuable because they allow us to find the most current information available as opposed to applications like ChatGPT, which may not always have the most recent knowledge, as their data is pre-trained and may not be updated as frequently. Both types of resources can be useful in different situations.
o Lack of references: It is not advisable to rely on the answers generated by ChatGPT, as they are not supported by references and can be unreliable, even if they are plausible. Always use your own judgment and consult other sources before relying on the information provided by ChatGPT. On the other hand, search engines allow you to see where the information is coming from, giving you the ability to verify its credibility and authenticity.
o From an ethical standpoint, it is important for models like ChatGPT to avoid promoting unethical behavior. When asked to code a function to predict wealth status based on race and gender, it associated ‘black’ with poor and ‘white’ with rich. ChatGPT should not comply with such requests. Instead, it should encourage honesty and integrity in all interactions; however, this was not the case when tested out (See Figures 2 and 3 ).
o Although ChatGPT is known for its high-quality code generation, it sometimes struggles with edge cases. As a result, the generated code cannot be blindly trusted. In addition, after several tests, the code generated is not always optimized. For instance, when asked to code a relatively hard algorithmic question, it generated a brute-force solution. When asked to optimize it using dynamic programming, the produced code failed all test cases. ChatGPT is also able to provide code reviews, but some of the reviews generated can alter the logic of the code, resulting in flawed outputs. Similar problems are also faced by other tools such as GitHub’s Copilot. However, there has been some controversy surrounding the legalities of the training process for Copilot, with some people suggesting that it may potentially infringe on copyrights.
o The increasing requirements for compute resources for systems like GPT-3 can be costly, time-consuming, and resource-intensive, and may have implications for businesses and organizations that use or rely on these systems. These systems may also have environmental implications due to their resource-intensive nature.
o Security trimming may also be a concern for ChatGPT. This issue arises when there are requirements to have a domain-specific model containing information that can only be accessed by a subset of employees. For example, exploration has many documents with confidential information that should not be accessible to everyone at a certain company. To mitigate security concerns, one must train a model for each set of documents with security requirements, and furthermore use more resources (compute power) to run a separate inference model.
Prospects
What is next for ChatGPT?
As we have seen, ChatGPT is a powerful tool, but there is still room for improvement, particularly when it comes to addressing the above-mentioned limitations. However, this leaves us wanting more. For example, can we see ChatGPT integrate multimodality in the future? We have already seen a glimpse of it from its code generation abilities, but could it be used to create websites or PDF slides on the fly? These are the exciting possibilities that lie ahead for ChatGPT.
What is next for ChatGPT?
As we have seen, ChatGPT is a powerful tool, but there is still room for improvement, particularly when it comes to addressing the above-mentioned limitations. However, this leaves us wanting more. For example, can we see ChatGPT integrate multimodality in the future? We have already seen a glimpse of it from its code generation abilities, but could it be used to create websites or PDF slides on the fly? These are the exciting possibilities that lie ahead for ChatGPT.
o From a research perspective, we are in dire need of explainability for such models to support the efforts of ChatGPT. Explainable AI (XAI) can help in providing clear and understandable explanations for the decision-making processes and outputs of LLMs when processing and interpreting the text. The technological leap of generative AI will make us value originality more and motivate us to find ways to develop reliable methods for detecting originality in AI-generated text.
Comparison
ChatGPT vs Search Engines.
ChatGPT and Search Engines complement each other, depending on the task at hand. In some cases, you might prefer to use one over the other. Below, we have summarized some examples of when ChatGPT and search engines might be used.
ChatGPT vs Search Engines.
ChatGPT and Search Engines complement each other, depending on the task at hand. In some cases, you might prefer to use one over the other. Below, we have summarized some examples of when ChatGPT and search engines might be used.
Utility
How can ChatGPT assist you?
Our report outlines several ways in which ChatGPT can be of benefit to users. From quick knowledge acquisition to providing personalized recommendations ChatGPT can assist in a variety of tasks to improve your productivity and efficiency. Some other examples are summarized in the table below.
How can ChatGPT assist you?
Our report outlines several ways in which ChatGPT can be of benefit to users. From quick knowledge acquisition to providing personalized recommendations ChatGPT can assist in a variety of tasks to improve your productivity and efficiency. Some other examples are summarized in the table below.
Snapshots
Here are some interesting snapshots related to the cases mentioned in the article.
Figure 2. An example of a biased output (however, ChatGPT did claim that this type of logic is not appropriate as seen in Figure 3 below).Figure 3. ChatGPT discouraged this type of logic, yet, I was hoping that it would not implement it.Figure 4. Interestingly, ChatGPT can detect its own generated text, however, it can easily be tricked! In this specific example, this text is from Wikipedia and not generated by ChatGPT.Figure 5. Preparing for an interview about a specific topic? An example of ChatGPT being used to conduct an interview about my recent paper (published in 2021).Figure 6. Asking ChatGPT about the complexity of this Python code (Cont’d in Figure 7).Figure 7. Pretty sure that the complexity is O(N), however, ChatGPT is providing a plausible argument on why it is O(N²)Figure 8. Code translation from Python to Java has never been easier! But Do not blindly trust it!Figure 9 . Code review for the Python code in Figure 8.Here are some interesting snapshots related to the cases mentioned in the article.
Ansh Jhaveri on Instagram: "chaatgpt #funny #comedy #viral #standup #comedy #standupcomedy #nyc #nyccomedy #chatgpt @hotvickkrishna"
Acknowledgments:
Thanks to Steven, Merce, and Amy for their valuable contribution to this article.
References:
1. ChatGPT API: https://chat.openai.com/chat
2. Ouyang, Long, et al. “Training language models to follow instructions with human feedback.” NeurIPS 2022.
ChatGPT: A sneak peek into its abilities, limitation, and how to make the best use of it. was originally published in AR/VR Journey: Augmented & Virtual Reality Magazine on Medium, where people are continuing the conversation by highlighting and responding to this story.