Anthropic Launches Claude 3.5 Models with New Interactive Skills
Anthropic has announced major improvements to its AI models, including the enhanced Claude 3.5 Sonnet and Claude 3.5 Haiku models, with a new feature that allows these models to interact directly with computer interfaces.
The Claude 3.5 Sonnet model has seen significant improvements in programming tasks, and the company claims that its performance outperforms all currently available models, including specialized programming systems.
One of the key features being announced is the ability for the Claude 3.5 model to interact with computer interfaces. Rather than developing specific tools for each task, the company took a more holistic approach by teaching the model general computer skills: the system can move a cursor, click on objects, and enter information using a virtual keyboard.
In the OSWorld test, which evaluates the ability of AI models to use computers in a human-like way, Sonnet 3.5 scored far above other systems.
However, Anthropic recognizes that Claude's ability to interact with computers is still not perfect, with some challenges in simple tasks, such as scrolling, dragging, and zooming, and the company recommends that developers start relying on this feature for low-risk tasks.
The new models showed improvements in logical reasoning, mathematical problem solving, and programming tasks. Although the new Sonnet 3.5 model made a slight improvement on the MMLU test of language comprehension compared to the old version, overall performance was better than before.
The company also introduced the new Claude 3.5 Haiku model, which outperforms the previous Claude 3 Opus model in many criteria, while maintaining the same speed and cost as the previous model. However, the company has not announced any plans to launch a new version of the Opus model anytime soon.
The new Haiku model is expected to launch later this month, and the company promises to shake up the competition for AI systems that can interact with computers.