Just when we were still shocked by the actual effect of OpenAI's Sora and marveled that the future has come. In just one month, OpenAI's "home" was stolen, and in the early hours of last night Beijing time, Anthropic announced the official release of Claude-3, with three high-performance model forms, and announced that it surpassed ChatGPT-4 in AI logic benchmarks.
Maybe few domestic partners know about Claude, which is an AI model that has attracted much attention overseas and has a large user base, which has been in love with GPT since its birth, and is chasing after it on the AI performance rankings. It is also a large language model based on GPT (Generative Pre-trained Transformer) technology, which has learned the ability to understand and generate natural language through pre-training on large-scale text data.
Amazon today announced the use of Claude3 to optimize its business logic
The series consists of three distinctive models: the Claude 3 Haiku, the Claude 3 Sonnet and the Claude 3 Opus, each with its own focus on performance and functionality, designed to meet the needs of a wide range of applications.
claude 3 haiku
The Haiku model stands out for its extremely fast processing power and cost-effective price. It is capable of reading a data-dense ARXIV study** with approximately 10,000 markers, including the understanding of charts and graphs, in just three seconds. Haiku is particularly suitable for platforms that require extremely high processing speed but have a relatively light performance load.
claude 3 sonnet
The Sonnet model is faster at most workloads than the Claude 2 as well as the Claude 21, and the model performance is consistent with performance, and provides a higher level of intelligent analysis capabilities. It is particularly good at handling tasks that require quick responses, such as knowledge retrieval and sales process automation, among others.
claude 3 opus
The Claude 3 Opus is the high-end model in the series, with sophisticated visual processing capabilities on par with other top models on the market, capable of handling a wide range of visual formats, including charts, graphs, and technology. Compared to Claude 21. OPUS doubles the accuracy of solving open-ended questions, while also significantly reducing the proportion of incorrect answers.
Other highlights:
The full range of models can handle more than 1 million labeled inputs, providing reliable support for customers who need more processing power. The Claude 3 Series excels at executing complex, multi-step instructions, especially when it comes to following brand tone and response guidelines, creating a customer experience that users can trust. In addition, these models are also adept at generating popular structured output formats such as JSON.
Now that OPUS and Sonnet are available through APIs, developers can now sign up and start experiencing the power of these cutting-edge models. For example, PoE already supports the Claude-3-Opus model, which can be experienced after purchasing a PoE "monthly card".
Beyond GPT-4; Take into account the visual function
According to Anthropic, the Claude 3 OPUS surpassed GPT-4 in 10 AI standard tests, including MMLU (undergraduate-level knowledge), GSM8K (elementary math), Humaneval (programming), and hellaswag (general knowledge).
Some of these victories had a very small margin, such as Opus, who scored 86 out of five attempts in MMLU8 score, while GPT-4 gets 864. Some gaps are larger, such as 90 on humaneval opus7 while GPT-4 scored 670％。This could mean that Claude 3 is more friendly to novice coders.
Compared to its predecessor, the Claude 3 series shows improvements in analytics, content creation, generation, and multilingual dialogue. The models also reportedly possess enhanced visual capabilities, allowing the models to work with visual formats such as graphs, diagrams, and graphs, similar to GPT-4V and Google's Gemini
In the actual test, Claude 3 is faster than GPT-4V for PDF interpretation, and the logic and optimization of Chinese output are significantly better than the previous generation, which also reaches the level of GPT-4 replacement.