Anthropic

Anthropic is an AI safety and research company that develops large language models (LLMs) under the Claude family. Operators encounter Anthropic through the Claude API or by running Anthropic's models locally via third-party implementations (e.g., llama.cpp supports Claude-like architectures). Anthropic's models emphasize safety and constitutional AI, which can affect output behavior and system prompt design.

Anthropic was founded by former OpenAI employees and focuses on building reliable, interpretable AI systems. Their Claude models (e.g., Claude 3.5 Sonnet, Claude 3 Opus) are proprietary and accessed via API, but some open-weights models like Claude 2.1 have been released. For local operators, running Anthropic models requires converting weights to GGUF format or using compatible runtimes. Anthropic's constitutional AI approach means the models are trained to follow a set of principles, which can reduce harmful outputs but may also affect creative freedom. The company also publishes research on interpretability and scaling laws, relevant for understanding model behavior.

An operator wanting to run Claude 2.1 locally would download the original weights from Hugging Face, convert them to GGUF using llama.cpp's convert script, then run inference with ./main -m claude-2.1.Q4_K_M.gguf -p 'Hello'. The Q4 quantized model requires about 7 GB VRAM, fitting on an RTX 3060 12 GB. However, the model may refuse certain prompts due to constitutional training, unlike uncensored models.

When using the Claude API via pip install anthropic, an operator sets ANTHROPIC_API_KEY and calls client.messages.create(model='claude-3-5-sonnet-20241022', ...). For local inference, the workflow involves downloading weights from Hugging Face, converting to GGUF, and running with llama.cpp. The operator must check license terms—Anthropic's models may have usage restrictions not present in fully open models like Llama 3.