Large Language Models have been widely leveraged to build chatbots with constant improvements in natural language processing capabilities and fine-tuning for specific domains to enable performance and user experience across various domains. This is all thanks to advancements in machine learning, better computational power and good quality training data.

In this article we compare the architecture, unique features, ethical AI features, image generation capabilities, language and cost of five chatbots: ChatGPT, Gemini, Llama, Meta AI, Claude.

To test each chatbot’s language capabilities, they were asked to translate the ‘man in finance’ meme using the prompt:

“can you write “i’m looking for a man in finance, trust fund, 6’5, blue eyes” in korean, arabic and spanish?”

To test each chatbot’s image-generation capabilities, they were provided with the following prompt:

“Generate an image of an ice cream with strawberries.”

The information in this article is correct as of June 2024. Developments happen very fast in the AI world – it is best to keep up with the respective companies’ blogs for the most recent updates to their large language models and chatbots. 

Integrate artificial intelligence applications with your existing information systems, build custom LLMs for your business and stay five steps ahead of your competitors: Hire AI engineers from Gaper.io’s marketplace.

OpenAI’s ChatGPT

ChatGPT is a sophisticated language model developed by OpenAI based on the GPT-4 architecture. Also utilized by businesses, it is designed to assist with a broad range of tasks by leveraging advanced natural language processing techniques. As of its most recent update in 2023, GPT-4 is ChatGPT’s latest version.

How does ChatGPT work?

• Architecture: ChatGPT useS a transformer-based neural network architecture, which is highly effective for natural language processing tasks.
• Training Data: ChatGPT has been trained on a diverse dataset that includes text from books, websites, and other written material, enabling me to understand and generate human-like text.
• Learning Process: ChatGPT’s training involved a process called unsupervised learning which helps predict the next word in a sentence, based on the context provided by previous words. This process is repeated billions of times to fine-tune the bot’s ability to understand and generate coherent text.

Unique and Important Features

• Versatility: Can handle a wide variety of tasks, from answering simple questions to engaging in complex discussions.
• Contextual Understanding: Maintains context over relatively long conversations, which helps in providing more relevant and accurate responses (some people have used this ability to use ChatGPT as a therapist…)

ChatGPT also has an option to manage its ‘memory’ across chats now.

• Language Proficiency: Understands and generates text in multiple languages.

Ethics

OpenAI has implemented guidelines and safety measures to ensure that ChatGPT’s use is ethical to avoid generating harmful or inappropriate content…though this has come under question many a time in recent months. ChatGPT has also been no stranger to data breaches.

Current Version (as of June 2024)

• GPT-4: GPT-4 offers better performance in understanding and generating text compared to its predecessors. This version includes improvements in areas like factual accuracy, contextual understanding, and response coherence. GPT-4 can be fine-tuned for specific applications, making it adaptable to various needs and industries.
• GPT-4o: GPT-4o is faster and better equipped at understanding visual and audio input as compared to other existing models.

Google’s Gemini

Google’s Gemini, particularly the latest 1.5 version, represents a significant advancement in AI technology. 

How does Gemini work?

• Architecture and Training: Gemini 1.5 employs a Mixture-of-Experts (MoE) architecture. A traditional Transformer operates using one large neural network, but MoE models get segregated into smaller “expert” neural networks. This in turn enhances efficiency by activating only relevant neural network pathways. This design builds upon Google’s research in transformer models and MoE, such as GShard and Switch-Transformer.
• Training Data and Learning Process: Gemini models are trained on vast and diverse datasets, including text, images, audio, and code, improving their ability to understand and generate multimodal content.
• Versatility and Contextual Understanding: A key feature of Gemini 1.5 is its extended context window, which can handle up to 1 million tokens. This allows the model to process and reason across large volumes of data, such as entire books or long videos, making it highly versatile in applications.

Unique and Important Features

• Language Proficiency: Gemini 1.5 demonstrates high proficiency in multiple languages and modalities. According to Google, it learnt to translate English to Kalamang using only a grammar manual at a level comparable to a how a human would learn.

While Gemini can do various languages, it took it upon itself to not translate a meme that it considered ‘materialistic’ and provided me with more ‘balanced’ translations. It also provided transliterations for each language without being asked, something that other chatbots did not.

• Gmail integration: Google is also rolling out Gemini to be used with Gmail to help write and summarize emails.

Ethics

Google ensures ethical use through extensive safety testing and adherence to its AI principles, focusing on mitigating biases and potential harms.

Current Version: Gemini 1.5

Gemini 1.5 shows dramatic improvements over its predecessors, with enhanced performance, better long-context understanding, and reduced computational requirements. The model is available for early testing with features like a 1 million token context window, providing developers with new possibilities for AI applications.

Gemini 1.5’s Pro’s context window capacity exceeds 32,000 tokens which allows it to process a huge amount of information (video, audio, codebases, words) at once.

Meta Llama / Meta AI

Meta has its own cohort of AI applications: Meta AI, Meta Lllama, and Meta Code Llama. Let’s see what each one of them does.

Meta AI

The term Meta AI is often confusing…

Meta AI, owned by Meta (formerly Facebook), is a company (or a research lab) that works on building AI, and augmented and artificial reality technologies. 

Meta AI is also the name of the AI assistant developed by Meta, that understands and responds to human input in a conversational manner.

How does Meta AI Assistant work?

• Architecture: Meta AI’s architecture is based on a transformer-based language model consisting of an encoder and a decoder, which work together to understand and respond to user input.

• Training Data: Meta AI was trained on a massive dataset of diverse text from various sources, including but not limited to web pages, books, articles, and conversations (the conversations part was mentioned by the Meta AI Assistant itself…🤔)
• Learning Process: Meta AI uses a self-supervised learning approach, where the model is trained to predict the next word in a sequence of text. This process is repeated millions of times, allowing the model to learn patterns, relationships, and context in language.

Unique and Important Features

• Versatility and Contextual Understanding: Meta AI can respond to various types of input and adapt to different tones, styles, and languages. It can also follow conversations, recall previous information, adapt to changes in topic or tone.
• Language: Meta AI is also proficient in multiple languages. However, when asked to translate the ‘looking for a man in finance’ meme, it seemed to translate it too literally. Perhaps native Arabic/Korean/Spanish speakers can weigh in?

• Ethics: While Meta tries to stick to the five pillars of Responsible AI guidelines, it has come under fire for summarizing news stories without providing a source to original articles (especially in Canada). This is quite concerning especially as Meta continues to censor certain news stories, and reduce overall news visibility on its platforms.

Meta Llama 2 and 3

Meta Llama is a family of LLMs developed by Meta AI. Also using a transformer-based architecture, Meta Llama models are trained on massive datasets and designed to perform various tasks like text generation, question answering, and code analysis.

The following table illustrate a few differences between Llama 2 and Llama 3.

Meta Code Llama

Meta Code Llama is an LLM for coding. It aims to create better and more productive workflows for developers and make learning code easier for aspiring developers. Built on Llama 2 using code-specific datasets, Meta Code Llama can understand and generate both natural language and code-based prompts and responses (including code completion and debugging). It has three models:

• Code Llama
• Code Llama Python
• Code Llama Instruct

Architecture and Training

Four versions of Meta Code Llama are available with varying lengths of parameters: 7B, 13B, 34B, and 70B. The first three are trained with 500 billion tokens of code/code-related data, but 70B is trained on 1 trillion tokens.

Meta suggests that the 7B and 13B models are faster and better for tasks that require low latency, like real-time code completion.

Ethics

Meta acknowledges the risks associated with AI models, especially those that generate code. Their red teaming process aimed to get the model to generate malicious code with clear intent and found that Code Llama returned ‘safer responses’. More on this can be read in this research paper.

Microsoft’s Copilot

How does Microsoft’s Copilot work?

There is surprisingly little information made available by Microsoft about the architecture and other technical features of their AI assistant, Microsoft Copilot. 

The free version of Microsoft Copilot is able to conduct chats through text, voice, and images. It can summarize documents and web pages, generate images, provide information/answers from specified sources (web grounding), and make use of plugins. Microsoft Copilot is fully integrated across Microsoft 365 applications like Word, Excel, PowerPoint, and Outlook (Microsoft 365 Copilot).

It is sometimes frustrating to talk to Microsoft’s Copilot. In addition to being slow, it often refuses to answer questions and tells you that, “It might be time to move onto a new topic. Let’s start over”, and forces you to start a new chat while discarding the old one. In this case, I was merely asking it what its exact name was (to understand how Bing, Bing AI, Microsoft Copilot, Microsoft 365 Copilot were different/interconnected or if some of these terms are used synonymously). 

The Copilot provided a simple and literal translation of the man in finance meme.

Microsoft’s Copilot does not maintain chat history. It does end every conversation with a smiling emoji though. 

Anthropic’s Claude

A relatively newer kid on the block, Anthropic offers an AI assistant called Claude, known for its strengths in long conversation memory and informative responses. Claude comes in different model versions, each with varying capabilities and performance levels: Haiku, Sonnet, Opus.

Architecture

• Token context window: 200,000.

Unique and Important Features

• Low hallucination rates
• High accuracy when working with long documents.
• Absolutely, here’s the text split into a bulleted list:
• Sets new industry benchmarks for: 
  • Graduate-level reasoning (GPQA)
  • Undergraduate-level knowledge (MMLU)
  •  Coding proficiency (HumanEval)
• Is able to identify nuances, humor and complex instructions
• Can transcribe text from unclear/incomplete images which makes it useful for tasks in retail, logistics, and financial services, among others.

Let’s see how well Claude did in translating the man in finance meme:

Interestingly, while Gemini also changed the height measurement unit in Korean (and completely omitted to state it in Arabic and Spanish), it did not explain why. Claude provided the cultural context when changing the unit to centimenters and meters in Korean and Spanish translations respectively.

Ethics

Anthropic prioritizes ethical considerations, incorporating measures to reduce biases and ensure fair and responsible AI usage. This includes extensive testing and refinement to mitigate any unintended biases in the model’s outputs.

With a “best-in-class jailbreak resistance”, Claude also adheres to various security standards thanks to its ability to access AWS and GCP, SOC 2 Type II certification, and HIPAA compliance options.

Claude 3.5 Sonnet

Claude 3.5 represents an advanced iteration of the Claude series of large language models. Named after Claude Shannon, a pioneer in information theory, this model builds on previous versions with enhanced capabilities and refined architecture.

It boasts a significant increase in parameters compared to its predecessors, enhancing its ability to understand and generate human-like text. Claude 3.5 excels in maintaining context over long interactions, making it particularly effective for applications requiring sustained coherence, such as chatbots and virtual assistants.

The model also employs reinforcement learning techniques, allowing it to adapt and improve based on user interactions and feedback. This continuous learning process helps in refining its responses and maintaining relevance over time. Further, its fine-tuning capabilities allow it to be tailored for specific tasks or industries.

Cost Comparison of the Chatbots

The following table shows a general cost comparison of the chatbots and their various versions. Each chatbot is available for free to use. More advanced versions have a price.

Image Generation: Comparison of the Chatbots

The following command was input into a new chat of every chatbot:

“Generate an image of an ice cream with strawberries.”

ChatGPT

It was possible to generate this only on the phone app – it kept having ~issues~ on the website.

Gemini

Not bad. Strawberries are missing in the third picture but the ice cream color is correct. Gemini also allowed a high quality download of the pictures.

Meta AI

MetaAI’s ice creams look better than Gemini’s with interesting backgrounds and crockery.

Copilot

Copilot’s Designer is powered by DALL-E 3. Their images adopt that typical AI-generated appearance, but still look good and were presented in a high quality. The detailed backgrounds are quite impressive but all of its ice-creams were in cones for some reason.

Claude

Claude could not generate images.

All chatbots refused to generate images of humans.

Which Chatbot Generated the Best Images?

While every chatbot generated impressive images (ChatGPT’s ice cream was interesting in its own right), I was particularly impressed by Meta AI’s pictures.

Can the ChatGPT Chatbot Generate Images or Not?

ChatGPT can generate basic line drawings of prompts it is given but like other chatbots, it does not generate images of humans. When asked to generate a picture of Harry Styles, it responded with, “As a text-based AI, I don’t have the capability to directly generate realistic images or photographs of specific individuals, including Harry Styles. However, I can suggest a few methods you can use to obtain or generate images”.

After this prompt, ChatGPT completely stopped generating images. This is how a new chat went:

Prompt: can you generate an illustration of a lamp.

ChatGPT: I can’t generate images directly, but I can certainly describe a lamp for you!

Imagine a sleek, modern desk lamp with a slender, curved neck made of brushed stainless steel. The base is a sturdy, square shape, also in stainless steel, providing stability.

The lampshade is a conical shape, crafted from frosted glass, emitting a soft, diffused light when turned on. The overall design is minimalist yet elegant, perfect for illuminating a workspace or adding a touch of sophistication to a room.

Prompt: sir you were generating illustrations in the previous chat

ChatGPT: I apologize for any confusion. As an AI text-based model, I can generate descriptions and provide information, but I do not have the capability to create visual illustrations or images directly. If you’re looking for descriptions or details about something specific, feel free to ask!

Prompt: can you generate images?

ChatGPT: As an AI text-based model, I don’t have the ability to generate or create images directly. My function is to process and generate text based on the input and data I’ve been trained on. If you need descriptions or information about something, feel free to ask, and I’ll do my best to provide accurate and helpful responses!

Prompt: how did you generate them in the previous chat?

ChatGPT: I apologize for any confusion earlier. Just to clarify, I don’t generate images or illustrations directly. Instead, I can provide detailed descriptions or discuss topics based on the text input I receive and the knowledge I’ve been trained on.

If there was a misunderstanding, I’m here to provide information and assistance to the best of my abilities using text-based responses. If you have any specific questions or topics you’d like to explore, feel free to let me know!

When asked to generate an image via the Android app, ChatGPT happily generated the following illustration of an ice cream with strawberries.

For the images that it did generate, I can only describe them as being comparable to what little kids used to make on Microsoft Paint. 

This is just an evaluation of how the chatbot does with images – if one wants to use an OpenAI product for image generation, it is obviously more sensible to go to DALL-E or use DALL-E within ChatGPT.

Conclusion

Choosing the ideal chatbot depends heavily on your specific needs. Each competitor in this arena brings unique strengths to the table.

For creative text generation and pushing boundaries in storytelling, ChatGPT and Claude are a great choice. If informative responses with a focus on factual accuracy are your priority, then Gemini and its access to vast search data might be the perfect fit.

For complex problem-solving tasks requiring analysis of various data formats, the advanced capabilities of Llama stand out. Meta AI, with its consumer-facing assistant functionality, offers a user-friendly way to interact with powerful AI technology.

Finally, Claude excels in long conversation memory and nuanced understanding, fostering a more natural and engaging dialogue experience.

The future of chatbots is undeniably collaborative. As these models continue to evolve, we can expect them to learn from each other’s strengths. Imagine a world where ChatGPT’s narrative prowess merges with Gemini’s factual grounding, or where Llama’s problem-solving abilities are integrated with Meta AI’s accessibility.

Don’t be a laggard in your industry: become an early adopter by integrating artificial intelligence applications with your business processes, use custom LLMs, and have a stellar team of AI engineers work for you.