Anthropic Unveils Claude 2: An Enhanced AI with Advanced Capabilities
On Tuesday, Anthropic launched Claude 2, an advanced language model akin to ChatGPT with the ability to craft code, analyze extensive text, and compose writings. Distinct from its predecessor released in March, Claude 2 can be trialed for free via its beta site and is commercially accessible for developers as an API.
Anthropic’s vision for Claude revolves around simulating interactions with a supportive colleague or personal assistant. Responding to user feedback from the previous version, they remark on Claude’s improved conversational ease, clear thought processes, diminished harmful outputs, and an extended memory.
Claude 2’s improvements manifest in three domains: coding, mathematical tasks, and logical reasoning. The new model boasts an impressive 76.5% score in the Bar exam’s multiple-choice segment, a significant increase from Claude 1.3’s 73.0%. Furthermore, it surpasses the 90th percentile of graduate school aspirants in the GRE’s reading and writing tests and matches the median score in quantitative reasoning.
A standout feature of Claude 2 is its amplified input and output capacity. The model now processes text fragments, or tokens, up to 100,000, enabling analysis of lengthier documents, from technical manuals to entire books. Coding proficiency has also surged, evidenced by its raised score from 56% to 71.2% in the Codex HumanEval Python test. Additionally, it exhibited improvement in grade-school math tests.
Anthropic’s commitment to producing less controversial or offensive outputs shines in Claude 2, with internal evaluations indicating it’s twice as safe in its responses compared to its prior version.
Currently available in the US and UK, Claude 2 has been integrated into business operations by companies like Jasper and Sourcegraph. Yet, while Claude 2 can process intricate texts, Anthropic cautions against blind reliance, suggesting users validate its outputs, especially for familiar subjects. They advocate for its use in common tasks like summarizing but advise against its application in health and wellness scenarios.
Source: Edwards, B. (n.d.). New ChatGPT rival, Claude 2, launches for open beta testing. Ars Technica.