What is ChatGPT and how it compares to Bard and Claude

ChatGPT is a new artificial intelligence chatbot that has taken the world by storm. Developed by OpenAI and launched in November 2022, ChatGPT can engage in conversational dialogues and perform a variety of tasks like answering questions, explaining concepts, summarizing text passages, and even generating new content like articles, poems and computer code.

But what exactly is ChatGPT and how does this impressive new AI work? Let’s break it down.

All the images of this post have been generated using Midjourney.

The History and Development of ChatGPT

ChatGPT is based on a large language AI model called GPT-3, which was created by OpenAI in 2020. GPT-3 uses a specialized deep learning technique called transformers to analyze massive amounts of text data and understand the patterns and structure of human language. Check this article if you want to know more about large language models.

Building on the capabilities of GPT-3, OpenAI trained ChatGPT specifically to have natural conversations and respond to a wide range of prompts and instructions. After over a year of development and training on vast datasets, ChatGPT was unveiled to the public in a beta version in November 2022, immediately capturing people’s attention and imagination.

Other tech companies like Anthropic and Google have also launched their own AI chatbots, such as Claude and Bard, to rival ChatGPT. Each model possesses its unique strengths, and a detailed comparison will be provided later in the article.

How Does ChatGPT Actually Work?

When you give ChatGPT a written prompt or question, it analyzes the text to understand the intent and context. It then generates a response by predicting the most likely next words in a sequence, based on the patterns it learned during its training.

ChatGPT progressively generates its response word-by-word, while continuously referring back to the original prompt for context. This allows it to have coherent, multi-turn conversations that stay on topic.

The key to ChatGPT’s conversational abilities lies in its training methodology. During training, ChatGPT was shown many real examples of human conversations, questions and answers from diverse sources on the internet. It uses these examples to learn the nuances of natural dialogue.

ChatGPT was also trained using a technique called reinforcement learning from human feedback. Essentially, it was rewarded when it generated responses that humans rated as satisfactory, and penalized for inadequate responses. This allowed it to keep improving the quality and relevance of its replies.

The Impressive, Yet Limited, Capabilities of ChatGPT

ChatGPT offers a wide range of useful applications and capabilities that showcase the technology’s current strengths and potential. However, as with any AI system, it also comes with significant limitations that warrant consideration.

First, we will explore some of the key ways ChatGPT can provide value and assistance. While not an exhaustive list, these applications demonstrate ChatGPT’s diverse conversational skills.

ChatGPT’s Current Skills and Use Cases

• ChatGPT has extensive knowledge that allows users to define concepts, ask questions, and find information that may be difficult to search for directly. For example, it can identify a song based on a description when you don’t know the exact lyrics.
• It’s now multimodal (for paid users). This means you can include an image in your prompt and ask something about the image, like which plant is that, or to plan something to eat with what you have on your fridge.
• It excels at planning and boosting productivity by mapping out schedules and steps to achieve goals like learning new skills. For instance, it can outline an effective learning path when starting to program in a new language.
• Generating engaging stories and ideas for kids is another aplication. Parents can use ChatGPT to create personalized bedtime stories based on a child’s interests, especially when feeling too tired to come up with them.
• Explain complex concepts to everyone. You can ask it to explain something for a 10 year old, or 5, or 3.
• For teachers, ChatGPT can develop interactive lessons and educational games covering course material. It can also assist with structuring and writing presentations on various topics.
• As a language partner, ChatGPT provides customized explanations of grammar rules and gives practice exercises when learning a new language. For example, it can ask you questions and correct your answers, or quiz you some vocabulary.
• It acts as an adept writing assistant by drafting, revising, and refining documents like emails, reports, and articles. A brief objective statement is sufficient for it to generate a draft with the desired tone.
• ChatGPT makes for an insightful thinking partner, providing opinions and feedback when asked. It can even emulate famous individuals to offer their perspectives on ideas.
• And finally, with various plugins and the code interpreter, its capabilities expand beyond text to include integrating with platforms like Canva and Worldfram Alpha, or analyzing data, generating graphs or running code.

These examples offer just a sample of ChatGPT’s expanding capabilities. Its conversational nature enables a wide variety of functions through natural prompts and instructions. However, ChatGPT is far from flawless. Next we will discuss key current limitations to understand its deficiencies alongside its utilities.

Limitations to Consider

The breadth of tasks ChatGPT can perform is impressive. It can hold conversations, explain concepts, summarize texts and generate new content in various styles. However, as an AI system still in development, ChatGPT has significant limitations to keep in mind.

• ChatGPT-4’s knowledge only extends to September 2021, as that’s when its training data cuts off. Anything past 2021 remains unknown to the model, creating a major blindspot.
• Only paid users have a ChatGPT option to access the internet. And only if activated. In general, it’s not connected to the Internet.
• ChatGPT also generates sometimes plausible-sounding but incorrect or nonsensical text due to its lack of reasoning skills and grounding in common sense. These shortcomings are known as hallucinations in AI.
• One major limitation of ChatGPT is its struggle with mathematical and logical reasoning (except if using the code interpreter). Since its training data from the internet features small numbers far more frequently, the model is inclined to bias its numerical responses toward lower values rather than provide accurate calculations. ChatGPT does not truly comprehend or reason about mathematical concepts and relationships. Without the capability to understand numbers, it often defaults to guessing probabilities based on what values it saw most during training.
• ChatGPT also faces contextual limitations within prolonged conversations. While it can process thousands of tokens (words) per prompt, this context resets after each response. ChatGPT does not maintain persistent memory and awareness indefinitely. As conversations continue over many prompts, ChatGPT tends to lose focus on earlier statements and facts. This can result in contradictory or irrelevant statements if an exchange exceeds the model’s contextual capacity within a given chat.
• ChatGPT also lacks true long-term memory. Its conversations exist in isolation, with the model forgetting previous interactions once a chat session ends. Asking if it remembers past questions or specific users is ineffective, as the model has no genuine recall ability. This limitation also means any factual errors or contradictions generated by the model tend to compound within a conversation as context is lost.
• ChatGPT was trained by analyzing enormous datasets of actual online content, and as a result, its responses aim to emulate the most common things a human might say in a given context based on that statistical training.
• As an AI trained on internet data, ChatGPT inherently reflects some of the biases and imperfections from its training sources. So problematic responses based on inaccurate stereotypes or toxic viewpoints do occur on occasion.
• Despite efforts by OpenAI, it remains possible for determined users to coax harmful, unethical or dangerous information out of ChatGPT. While OpenAI implemented filters to block certain topics, techniques like carefully crafted prompts can sometimes circumvent these protections. This risk of misuse continues to be an area of concern with conversational models like ChatGPT.
• A key limitation is that ChatGPT provides no sources or citations for the information it provides. As an AI, it generates text based on patterns learned during training, not by referencing documents. All facts, quotes and data are produced from its statistical model, meaning ChatGPT cannot offer credible attribution for its content.
• ChatGPT suffers from compounding errors during conversations. If a factual mistake or contradiction occurs, the model lacks the ability to self-correct. With no long-term memory, any initial errors persist and potentially grow worse as ChatGPT continues generating flawed text devoid of context.

Therefore, while remarkably capable, ChatGPT cannot fully replace human intellect just yet. Relying solely on it for critical advice would be unwise. That said, various plugins and integrations are expanding ChatGPT’s capabilities beyond just text generation. And for less high-stakes applications like drafting emails or summarizing concepts, ChatGPT proves very useful.

While often accurate, ChatGPT cannot be relied upon for 100% factual correctness. Any high-stakes use should involve human verification of its outputs.

While these limitations highlight areas for improvement, they do not negate ChatGPT’s impressive capabilities. Understanding its boundaries allows for responsible, informed use – avoiding overreliance or unrealistic expectations. ChatGPT was built to generate likely text continuations, not possess encyclopedic knowledge. By recognizing its strengths as a conversational model and anticipating its weaknesses, we can have productive interactions without frustration.

Moving forward, it will be insightful to explore how alternative AI systems attempt to address ChatGPT’s limitations. In the next section, we will compare ChatGPT to its two major competitors – Claude and Bard – to see how they stack up in terms of capabilities, limitations, and overall utility. Examining the contrasts and similarities with these other natural language models will provide useful perspective on the current state and trajectory of this rapidly evolving technology.

How ChatGPT Stacks Up Against Claude and Bard

ChatGPT made a big splash as the first widely available conversational AI, but it now faces competition from Claude created by Anthropic, and Bard from Google. How do these alternative natural language models compare?

Claude: A Focus on Safety

Claude is an AI assistant developed by startup Anthropic to prioritize safety and integrity. Like ChatGPT, it uses a transformer neural network architecture for natural language processing. However, Anthropic employed a technique called Constitutional AI to align Claude’s goals and values with human preferences.

As a result, Claude is said to exhibit less harmful bias and is more cautious about generating dangerous or unethical content compared to ChatGPT. It also incorporates self-supervised learning to better handle novel situations outside its training data.

It has a bigger context than ChatGPT or Bard, of about 100k tokens. This gives it the ability to read longer articles, even short books, and retain the information within a chat for much longer.

On the other hand, Claude’s capabilities lag behind ChatGPT and Bard when it comes to conversational breadth and creativity. Its responses tend to be simpler and more constrained. Though possibly safer, Claude lacks some of the expressiveness that makes ChatGPT engaging.

Bard: Google’s Alternative

Google’s Bard aims to rival ChatGPT as an AI assistant that can explain concepts, generate content, and converse naturally. It is based on Google’s LaMDA model rather than GPT, featuring deep neural networks and trillions of parameters.

In initial tests, Bard appears significantly more grounded in facts compared to ChatGPT, with Google’s extensive Knowledge Graph likely providing useful context. However, Bard’s responses are often shorter and it struggles with more complex conversational prompts.

However, Bard also has access to the internet. This means that it can keep its knowledge up-to-date by reading the latest news articles, blog posts, and social media posts. This gives Bard a significant advantage over ChatGPT, which in general, it’s not able to access the internet (only paid users and with the browsing on).

For example, if you ask Bard about the latest news, it will be able to tell you about the most recent events. It will also be able to provide you with links to articles and other sources of information so that you can learn more.

What do I use

In general, I use

• Bard when I need some factual information or recent news
• Claude when I need help to write an article for my blog like this one, as I usually need larger context.
• ChatGPT for creative ideas, brainstorming, bedtime stories for my kids

But my recommendation is that you should try the different tools and discover which one works better for your specific needs. I sometimes ask the same question to two different chats, to compare, and even discover new functionalities. You can have a browser window specifically for your AI tools like Bard, ChatGPT or Claude.

And you, which of them do you use in your daily live? Leave a comment with your specific use cases. I’d like to know!

Follow me to stay tuned to the latest news regarding AI. You can subscribe to my blog here, or to my weekly AI newsletter at marinamele.substack.com.