Anthropic released Claude 3 Haiku, a large language model, like ChatGPT, with vision capabilities. Anthropic promotes its LLMs for advanced reasoning, that can perform complex cognitive tasks. But also for vision analysis where LLM can transcribe and analyze static images, including handwritten notes, graphs and photographs. Additionally, Claude can write code, e.g. HTML and CSS and turn images into structured JSON data. It also is capable of multilingual procesing, including content creation and translating between languages in real-time. Anthropic emphasizes Haiku’s speed, such as when analyzing quickly large datasets and generating timely output, for example in customer support. Claude 3 Haiku is said to process 21K tokens (~30 pages) per second for prompts under 32K tokens.
Anthropic prices place Claude at half the cost of OpenAI. Anthropic also says it’s meant for enterprise workloads which often involve longer prompts where it can quickly analyze large volumes of documents. For example, Claude 3 Haiku processes and analyzes 400 Supreme Court cases or 2,500 images for 1 USD.
Knowing the corporate clients concern for data security, Anthropic promises enterprise-grade security and robustness to reduce the likelihood of harmful outputs and jailbreaks. Layers of defense include continuous systems monitoring, endpoint hardening, secure coding practices, strong data encryption protocols, and stringent access controls to protect sensitive data. They also conduct regular security audits and work with penetration testers to identify and address vulnerabilities.
API access is available, allowing integration into exisiting pipelines. Claude 3 Haiku is also available on Amazon Bedrock and soon at Google Cloud Vertex AI.
Anthropic envisions the users will adapt Claude family LLMs to a wide range of enterprise solutions. Besides Haiku, the Claude family includes also Sonnet and Opus models. Up to 200K context window is available. Anthropic sees it as a GPT-3.5 competitor and not GPT-4. It’s not gonna be as smart as competitors’ top tier models, but for customer service use cases, it will respond fast. Hence, combining it with retrieval augmented generation (RAG) approaches, a customer chatbot could provide answers fast and with satisfactory results.
Read more
https://www.anthropic.com/claude
Claude can be used at
https://claude.ai
(requires a phone nr; EU phone nr not yet accepted; Google Voice nr-s do not work)
References
https://www.anthropic.com/news/claude-3-haiku
https://www-cdn.anthropic.com/de8ba9b01c9ab7cbabf5c33b80b7bbc618857627/Model_Card_Claude_3.pdf