My IBM Log in Subscribe

What is Claude AI?

24 September 2024

 

 

Authors

Ivan Belcic

Staff writer

Cole Stryker

Editorial Lead, AI Models

Gather

What is Claude AI?

Claude AI (Claude) is a generative artificial intelligence (AI) chatbot and family of large language models (LLMs) developed by the research firm Anthropic. Claude excels at natural language processing (NLP) and is multimodal: it accepts text, audio and visual inputs and can answer questions, summarize documents and generate long-form text, diagrams, animations, program code and more.

Claude adheres to Anthropic’s Constitutional AI philosophy: a code of ethical norms that the firm believes differentiates Claude from competing AI models such as ChatGPT and Google’s Gemini. The principles of Constitutional AI are focused on AI safety, designed to guide Claude toward providing more helpful responses while avoiding harmful behaviors such as AI bias.

Claude 3, released in May 2024, includes one free and two premium AI chatbots.

  • Claude 3.5 Sonnet underpins the free version of Claude AI. Its emphasis on speed enables it to quickly process user queries and other tasks requiring urgent data retrieval. According to Anthropic, Claude 3.5 Sonnet is twice as fast as Claude 3 Opus, one of the two premium offerings.
  • Claude 3 Opus is one of two Claude models currently available to Claude Pro users. It provides in-depth document processing and content generation services, specializing in complex tasks. While slower than Claude 3.5 Sonnet, Opus runs a lower risk of hallucinations: when an AI model provides incorrect information as though it is factually correct.
  • Claude 3 Haiku is the second premium Claude offering. It’s the smallest and fastest of the three and is ideal for use in summarizing long documents, real-time customer service and simple text generation.

What is Claude used for?

Each of the three Claude 3 models has its own specialized use cases. In general, people can use Claude AI to help with a wide range of tasks, including:

  • Question-answering and research
  • Proofreading and editing
  • Document summarization, including PDFs and Word documents
  • Text and content generation
  • Language translation
  • Business plan creation
  • Image and audio processing
  • Code snippet generation and review

Unlike Claude 2 and 1, Claude 3 is multimodal: it can process image and audio content alongside text-based prompts. For example, Claude 3 can generate e-commerce product descriptions based on images. While Claude 3 cannot generate nontext content on its own, its multimodal integration is one of several new features that allow it to compete with GPT-4.

How does Claude AI work?

Like Gemini and OpenAI’s ChatGPT, Anthropic’s Claude family of AI systems are based on the transformer architecture of neural network. But unlike its competitors, Claude applies the principles of Constitutional AI to govern its behavior.

  • Transformer models excel at drawing connections between distant words in a user input sequence, enabling them to better understand context and generate long-form replies.
  • Constitutional AI is a guiding set of harm reduction principles designed to make Claude more beneficial with less risk.

What are transformer models?

Transformers are a type of AI model built for high-performance natural language processing. They work by applying complex mathematical algorithms to statistically predict the most likely response to a user query. The workflow can be divided into four basic steps.

The transformer breaks up a user query into tokens. Each token represents either a whole word or a portion of a word. AI model pricing is typically represented as the cost per token. Claude Pro’s context window is 200,000 tokens1, meaning it can process user queries of up to 200,000 tokens in length.

  1. Each token is plotted into a three-dimensional vector space via mathematical processes. Tokens that are assessed as more similar in meaning are plotted closer together in space, aiding LLMs in understanding user inputs. The result of this process is called a vector embedding.
  2. Transformers such as Claude and GPT-4 apply self-attention mechanisms to self-direct resources on the most relevant portions of a user query and process context.
  3. The model applies probabilistic algorithms to generate the most likely response to an input. AI models such as Claude don’t actually “know” anything—rather, they combine their training data with advanced statistics to yield the most probable outcomes to prompts.

What is Constitutional AI?

Constitutional AI2 is a set of AI ethics and safety principles created by AI startup Anthropic. When designing Claude, Anthropic sourced input from approximately 1,000 people, asking them to vote on and suggest rules for ethical generative AI operation and responsible AI use. The final assembly of rules formed the basis of Claude’s training process.

The first three rules of Constitutional AI are:

  • Choose the response that is the least dangerous or hateful.
  • Choose the response that is as reliable, honest, and close to the truth as possible.
  • Choose the response that best conveys clear intentions.

Where other models have their content reviewed by human trainers in a process called reinforcement learning from human feedback (RLHF), Claude’s was trained with RLHF as well as a second AI model. Reinforcement learning from AI feedback (RLAIF) tasked the “trainer” model with comparing Claude’s behavior against Constitutional AI and correcting it accordingly.

RLAIF automates the behavior-adjustment portion of the training process, making it cheaper and more efficient to encourage ethical behavior. The intended result is that Claude would fine-tune itself, learning to avoid harmful prompts while generating helpful replies to prompts it deems answerable.

Who is Anthropic AI?

Anthropic is an AI startup founded in 2021 by several ex-OpenAI researchers and executives, including siblings Daniela and Dario Amodei. Amazon and Google have each invested billions in USD into the company, while OpenAI continues to enjoy backing from Microsoft.

The Amodei siblings parted ways with OpenAI in 2021, the year before OpenAI released GPT-3.5. This is the same AI model that continues to power the free ChatGPT AI tool today. Along with other former OpenAI researchers, the Amodei siblings founded Anthropic AI and began work on what would become Claude AI.

Anthropic’s defining feature is their stated approach to ethical AI, represented by the Constitutional AI training process.

The benefits of Claude vs ChatGPT and Gemini

When releasing Claude 3, Anthropic AI conducted a series of LLM benchmarking tests to evaluate their models against those of their two primary competitors: OpenAI and Google. Both in those tests and otherwise, Claude demonstrated several key advantages:

  • Larger context window
  • Strong performance in many tests
  • No input or output data retention

Larger context window

Able to field prompts of up to 200,000 tokens—approximately 350 pages of text—Claude can remember and use more information when creating relevant answers. By comparison, GPT-4 Turbo and GPT-4o limit users to 128,000 tokens.

Claude’s ability to retain more information allows users to create detailed, data-packed prompts. The more data contained in the input sequence, the more relevant an AI model’s answer can be.

Strong performance in many tests

When Anthropic tested Claude 3 against GPT-4 and Gemini 1.03, Claude 3 Opus was the top performer in all selected evaluation benchmarks. Gemini 1.0 Ultra came out on top in four of the six vision tests, though the Claude family of models performed comparably.

However, GPT-4o and Gemini 1.5 were not included in the testing pool. When revealing GPT-4o in May 20244, OpenAI conducted benchmarking that saw their new flagship model beat Claude 3 Opus in five out of six conducted tests.

No input or output data retention

Users concerned about data privacy might appreciate Anthropic’s data retention policy5: they state that all user inputs and outputs are deleted after 30 days. Google’s Gemini for Google Cloud data policy6 says that the company will not train its models with user prompts.

By comparison, OpenAI can retain and use user data7 to further train their models. Google’s Gemini Apps policies8 permit the company to retain user data unless the user manually deactivates this option.

Claude’s disadvantages

While Claude’s overall performance is strong when compared to the competition, it also has a handful of weaknesses that can delay its acceptance by the greater population.

  • Limited image generation
  • No internet browsing

Limited image generation

Compared to GPT-4o, Claude is less able to create images. While Claude can produce interactive flowcharts, entity relationship diagrams and graphs, it stops short of full image generation.

No internet browsing

Due to Microsoft’s integration with Bing, GPT-4 is able to search the internet when answering user queries. While Claude is regularly updated with new training data, its knowledge base is always several months behind until Anthropic elects to open Claude up to the internet in the same way.

Footnotes

1. How large is Claude Pro's Context Window? Anthropic, 2024 

2. Collective Constitutional AI: Aligning a Language Model with Public Input, Anthropic, 17 October 2023 

3. Introducing the next generation of Claude, Anthropic, 4 March 2024 

4. Hello GPT-4o, OpenAI, 13 May 2024 

5. How long do you store personal data?, Anthropic, 2024 

6. How Gemini for Google Cloud uses your data, Google, 10 September 2024 

7. How your data is used to improve model performance, OpenAI, 17 September 2024 

8. Gemini Apps Privacy Hub, Google, 28 August 2024 

Related solutions

Related solutions

Foundation models

Explore Granite 3.2 and the IBM library of foundation models in the watsonx portfolio to scale generative AI for your business with confidence.

Explore watsonx.ai
Artificial intelligence solutions

Put AI to work in your business with IBM’s industry-leading AI expertise and portfolio of solutions at your side.

Explore AI solutions
AI consulting and services

Reinvent critical workflows and operations by adding AI to maximize experiences, real-time decision-making and business value.

Explore AI services
Take the next step

Explore the IBM library of foundation models in the IBM watsonx portfolio to scale generative AI for your business with confidence.

Explore watsonx.ai Explore AI solutions