AI company Anthropic is previewing its new “family” of Claude 3 models that it claims can outperform Google’s Gemini and OpenAI’s ChatGPT in multiple benchmarks.
This group consists of three AIs with different levels of ‘capacity’. You have Claude 3 Haiku at the bottom, followed by Claude 3 Sonnet, and then there is Claude 3 Opus as the top dog. Anthropic claims the trio delivers “strong performance” across the board thanks to their multimodality, improved accuracy, better context understanding, and speed. What is also striking about the trio is that they are more willing to answer difficult questions.
Anthropic explains that older versions of Claude sometimes refused to respond to cues that pushed the limits of the safety railings. Now, the Claude 3 family will take a more nuanced approach with their answers, allowing them to answer those tough questions.
Despite the overall performance improvement, much of the announcement focuses on Opus as the best in all these areas. They even go so far as to say that the model exhibits “near human levels of understanding… (for) complex tasks.”
Specialized AIs
To test it, Anthropic put Opus through a Needle In a Haystack, or NIAH, evaluation to see how well it can recall data. It turns out it’s pretty good, as the AI was able to remember information with near-perfect detail. The company further claims that Opus is a pretty smart cookie that can solve math problems, generate computer code, and display better reasoning than GPT-4.
The technology is not without its quirks. While Anthropic claims their AIs are more accurate, there is still the problem of hallucinations. The responses produced by the models may contain misinformation, although they are significantly more limited compared to Claude 2.1. Additionally, Opus is a bit slow when it comes to answering a question at speeds comparable to Claude 2.
Of course, this is not to say that Haiku or Sonnet are inferior to Opus, as they have specific usage scenarios. Haiku, for example, is great at providing quick answers and extracting information “from unstructured data.” It’s also not as good at answering math questions as Opus. Sonnet is a larger-scale model intended to help people save time on menial tasks and even parse lines of “text from images”, while Opus is ideal for large-scale operations.
Changing the Internet
Both Sonnet and Opus are currently for sale, although one free version of Claude on the company website. No launch date has been given for Haiku, but Anthropic says it will be released soon.
As you can probably guess, the Claude 3 trio is aimed more at companies that want to automate certain workloads. Your experience with the group will likely come in the form of an online chatbot. Amazon recently announced that it will be implementing Anthropic’s new AIs in AWS (Amazon Web Services) gives websites on the platform a way to create a custom Claude 3 model that fits the needs of brands and their customers.
If you’re looking for a model suitable for everyday use, check out Ny Breaking’s list of the best AI content generators for 2024.