AI has exploded with the development of Large Language Models. Trained on massive amounts of input, these models are the algorithmic basis for GenAI tools, which are rapidly improving. Tech giants are currently racing to create the best AI chatbot. The competitors are: ChatGPT, Bard, Bing (Copilot) and Claude. Check out our comparison of these 4 tools.
ChatGPT by OpenAI is the only generative AI product available with two different LLMs, GPT-3 and GPT-4. Its launch in November 2022 broke all records for popularity in consumer software and web applications.
It is one of the most advanced AI chatbots that can imitate human conversations, providing users with an easy and personalised user experience. Its ability to provide long, good-quality answers and generate specialised texts on various topics set a high bar for both the LLM itself and the competition.
OpenAI continues to develop its product, regularly adding new features. Currently (November 2023), using ChatGPT, you can search the internet, generate images or have voice conversations with chat.
More about ChatGPT in our other article: ChatGPT and Generative AI – it’s time to understand them.
- a large selection of plug-ins for various applications;
- ability to enter Custom instructions for better personalisation of results;
- web search function (Browse with Bing) in the paid version;
- ability to generate images using the built-in DALL-E 3 model (note: we generated the first 4 images in this article in ChatGPT);
- voice conversations with chat in the mobile app (also in Polish);
- the context window (in GPT-4) supports up to 128,000 tokens! (for comparison: Bard – 1000).
- training data is limited to April 2023, which, according to OpenAI, is to be gradually improved (the Browse with Bing function partially compensates it);
- occasional wrong answers and susceptibility to the so-called hallucinations;
- user interface;
- the free version does not support many features.
When OpenAI made ChatGPT publicly available, Google rushed to release a version of its text-based generative AI product, Bard, in February 2023. It was also a response to the integration of ChatGPT with Bing from Microsoft.
The Bard chatbot was designed to simulate conversations with a human and generate realistic answers to questions. To provide even more relevant search recommendations, Bard continuously searches the entire internet in real-time. It has been its flagship feature from the very beginning.
- friendly user interface;
- cites sources and gives the most up-to-date answers;
- has a button to back up your statements via search results or to explain when there are no sources;
- generates text on a given topic and offers continuation options for completed entries – you can switch between alternative versions and choose the optimal one for you;
- generates and debugs code in 20 programming languages and translates code from one language to another.
- does not generate or read documents such as PDF files, images, etc.;
- generating results can be slow.
3. Bing (now Copilot)
Seeing the massive success of ChatGPT, Microsoft, in March 2023, integrated similar AI features into its Bing web search engine (one of the biggest search engines in the world, but with only 3% of searches). Maintaining close technical and financial cooperation with ChatGPT, it created Bing AI Chat.
It is important to mention the recent renaming of this tool. Bing Chat was rebranded as Copilot during the Microsoft Ignite 2023 event (14-17 November), which is part of Microsoft’s strategy to strengthen the tool’s branding.
Copilot (previously Bing) uses the LLM but has additional features to make it more experimental and accurate. It is based not only on training data but also on indexing functions and the so-called web scraping, which involves the automated downloading of data from the internet. Thanks to this, it can generate answers based on current data.
- a clear and pleasant to use tool;
- it has different user interfaces for various purposes;
- demonstrates depth and accuracy as well as an exciting and entertaining conversational style;
- searches the internet by default;
- cites his sources;
- generates good quality images using the DALL-E 3 engine.
- generally only works in Microsoft Edge;
- still prone to errors in results – they can be hit or miss;
- not very good at generating or displaying code.
Claude is an AI chatbot developed by the research company Anthropic (founded by former OpenAI members). It is designed to be user-friendly and more accessible to use than, for example, ChatGPT. It excels in creative writing and programming using Python and emphasises security, control, and ethics.
Claude is intended to be used not only to write human-like content but also to imitate some non-cognitive behavioural characteristics. Historical data do not limit it and provide current results.
It was trained using the so-called constitutional artificial intelligence.
It means it focuses on honest conversation to produce “less harmful” results, generating responses that are socially responsible and ethical. So far, the results have been mixed, with reviewers emphasising that it continues to create content that may be biassed, false or potentially harmful.
The current versions of Claude 1, Claude 2 and Claude Instant have recently become available to a broader audience, a big shift from their earlier versions, which were solely for businesses.
- sharp practical thinking abilities, but avoids problematic answers;
- designed to recognise best whether a given piece of information is accurate;
- function of attaching files and analysing them (asking questions about them);
- the context window supports up to 100,000 tokens (for comparison: Bard – 1000).
- lack many multimedia and internet functions (compared to ChatGPT), e.g. image generation.
Architecture and technology of AI chatbots
It is based on a deep learning architecture called Generative Pre-trained Transformer (GPT). It allows you to generate entirely accurate and highly personalised answers to user queries.
The GPT architecture uses a technique called unsupervised learning – it can learn from both labelled and unlabelled data to make sure it understands questions. Labelled data allows the chatbot to accurately identify the intent behind user queries and generate appropriate responses.
It uses LaMDA technology and neural network architecture to answer complex user questions in real-time. It takes into account the context, topic and previous statements. Thanks to advanced machine learning techniques, LaMDA is able to learn new concepts and be used in many different fields.
Google’s Transformer, the basis for other AI generation tools such as the GPT-3.5 language model, is also used in ChatGPT. They tend to imitate answers, which can lead to errors and mistakes. The scientists who created the LaMDA model tried to solve this problem by increasing the size of the model. However, they found that this approach did not work in areas and industries where facts were constantly changing.
Bing (now Copilot)
It is based on Microsoft’s proprietary Prometheus technology. This AI model combines a comprehensive search index, rankings and results from Bing searches with creative inference of the most advanced OpenAI GPT models. It then combines the output generated by Bing and GPT to develop internal queries using a component called Bing Orchestrator. The combined response is presented to the user in the form of human-like text.
Internal queries and Bing searches are critical elements of the Prometheus architecture. Thanks to them, it provides current information to the primary language model, enabling answers to present questions and reducing inaccuracies – this method is called grounding.
Claude’s work was based on an innovative approach to AI called “cognitive architecture”. It aims to replicate how the human brain processes information, enabling AI systems to learn and reason in a more human-like way.
By combining cognitive architecture with cutting-edge machine learning techniques, Claude aims to achieve a level of intelligence and adaptability that is likely to exceed existing AI systems.
- While Bard and Claude are still in the improvement stages, ChatGPT and Bing are already demonstrating advanced intelligence and usability – feedback generally seems more consistent with expectations and responses more refined.
- ChatGPT (despite the new image generation function) and Claude specialise mainly in text-based interactions, while Bing and Bard aim to build a semi-automatic internet search engine. Early demonstrations suggest that Bard has some accuracy issues with other chats, but Google’s vast resources could give Bard an advantage in fact-based conversations.
- ChatGPT and Claude clearly outperform the other two chatbots in terms of context window size. ChatGPT (GPT-4 Turbo) is currently capable of processing 128,000 tokens! (equivalent to 300 pages of text), and Claude – 100,000 tokens. This allows them to better read and generate responses in context, which brings great potential for tasks requiring understanding and precision.
Where are these AI chatbots useful?
- Possibility to conduct multi-round conversations. Compared to other chatbots operating in a question-answer format, ChatGPT can maintain the context of the conversation and generate answers that are relevant to previous rounds.
- ChatGPT’s NLP engine includes a sentiment analysis component to understand the emotions behind user queries. This feature is helpful for companies that want to provide empathetic customer service and provide them with personalised product recommendations based on their preferences and purchase history.
- It can help companies across industries automate customer service operations, reducing staff workload and response times.
- Individuals can use it as a virtual assistant to improve their language skills – you can naturally practise not only writing but also speaking with this assistant.
- Because it constantly searches the entire internet and collects the latest data from around the world, its answers are always up-to-date and are additionally modified.
- It has a new feature of uploading images using Google Lens and receiving replies in the form of images from Google Search.
- You can use it for everyday life tasks, such as creating an itinerary document or meeting, preparing an online ad or writing a blog article.
Bing (now Copilot)
- Developed primarily for search and optimised for a better understanding of natural language processing than, for example, ChatGPT. And although Bing is less flexible and creative, it has several unique functions that result from direct integration with the search engine.
- The image generation function allows you to maintain a smooth “conversation”, providing comments on an ongoing basis, without having to enter subsequent image requirements each time.
- What is innovative about it is its learning from many different data sources – text, images, and human behaviour. It allows the system to develop a rich understanding of the world and its complexities; this could have significant implications for industries such as customer service to provide them with more engaging support.
- The ability to quickly and accurately identify objects and patterns in images can prove invaluable in specific fields, such as medical imaging – AI systems can help doctors more accurately diagnose and treat patients.
Cost and availability
OpenAI offers three plans to use:
- basic – free, with access to the GPT-3.5 model;
- ChatGPT Plus – USD 20 per month, with access to the GPT-4 model, plug-ins, web search and image generation;
- ChatGPT Enterprise – a dedicated solution for companies, with prices set individually by the sales department.
It is entirely free and is said to remain so, at least for some time. Google plans to integrate this tool with its other products.
Initially, it was available in the USA and Great Britain, and from July 2023, it is accessible throughout Europe and in over 40 languages, including Polish.
Bing (now Copilot)
Until recently, it was only available in Microsoft Edge; at the end of August 2023, Bing AI also came to the Google Chrome browser and is slowly adapting.
Microsoft is working to bring its features to other browsers on desktop and mobile devices. Bing became a free-access phenomenon from the very beginning. Current innovation maintains this – users from all over the world have access to both versions of advanced tools: Bing AI and Bing Image Creator. You can generate 50 free images daily in Bing (with a Microsoft account). The chatbot understands commands in both English and Polish.
Until recently, Claude 2 only allowed users from the US and UK and offered a priority paid service (with five times the number of messages compared to the free version).
This fall, the company has opened access to 95 countries worldwide – in Europe, Asia, Africa, South America and Australia.
Information about the new subscription plan for each version, i.e. Claude 1, Claude 2 and Claude Instant can be found on the Anthropic website. For now, it is known that different price levels allow you to adjust access to your needs.
AI chatbots and the speed at which they are growing, show their enormous potential. Despite of the fact that these LLMs still require caution due to the potential misinformation, they provide an exciting glimpse into the future of AI.
Changes in the chatbot and GenAI market are happening so dynamically that it takes a lot of effort and ability to assess the situation to stay up to date. If you are looking for a reliable and up-to-date source of knowledge, check out our YouTube channel, Beyond AI. We regularly publish about the latest solutions and practical use of AI there.