On March 4th, Anthropic unveiled their new model family Claude 3. The Claude 3 model family consists of three different models:
* Claude 3 Haiku
* Claude 3 Sonnet
* Claude 3 Opus
Each subsequent model is more powerful than the previous one, giving users the ability to find the optimal balance between intelligence, speed, and price according to their specific needs.
Opus and Sonnet are now available on the claude.ai platform and through the Claude API in 159 countries. Haiku will also become available soon.
Claude 3 Opus is the most intelligent model in the family, surpassing the benchmarks of most artificial intelligence systems to date, including bachelor's level domain knowledge (MMLU), master's level reasoning ability (GPQA), elementary mathematics (GSM8K), and much more. It demonstrates near-human comprehension and expression abilities for complex tasks, pushing the boundaries of general intelligence.
All Claude 3 models are more capable in terms of analytical and predictive ability, nuanced content creation, code generation, and communication in Spanish, Japanese, French, and other languages.
Haiku is the fastest and most cost-effective model in its intelligence category on the market. It can process a research paper on the arXiv database (~10,000 characters) with diagrams and graphs in less than three seconds.
Sonnet is twice as fast as Claude 2 and Claude 2.1 for most workloads while offering higher intelligence. Opus offers similar speed to Claude 2 and 2.1 but at a significantly higher level of intelligence.
The Claude 3 models also have advanced visual analysis capabilities comparable to other leading models. They can process various visual formats, including photos, graphs, tables, and technical drawings.
Of these, the benchmarks for the most powerful one, Opus, indicate that for the first time, a large foundational model more powerful than any currently available OpenAI model has been made accessible to regular users.
Initially, we can only rely on Anthropic's own measurements, as independent verification is still pending.
Data: Anthropic, March 2024
However, our initial tests show that there is no longer much difference whether the respondent is GPT-4, Claude 3, Gemini 1.0 Ultra, or even Mistral Large.
You can read more about Claude 3 here: https://buff.ly/49CvbJq
What Happened to OpenAI?
But how did Claude 3 suddenly become the world's best large model, and more importantly: where is OpenAI's GPT-5?
Never before has it happened that OpenAI has been left "without a response". As a rule, whenever Google or Anthropic released something new, OpenAI would immediately release a new version and show that they are still far ahead. But not this time.
It is speculated that OpenAI's hands are currently tied from releasing new models due to lawyers and lawsuits. The latest case with Elon Musk seems to be the most curious and pointless.
What is Elon Musk's Case About?
Elon Musk sued OpenAI, the creators of ChatGPT, claiming that the company has violated the principles he helped establish when it was founded in 2015. The lawsuit is also filed against OpenAI's CEO Sam Altman. Musk claims that the company has departed from its original non-profit and open-source mission. Instead of trying to "benefit humanity" as it was originally set out to do, OpenAI is now focused on "maximizing profits" for its major investor Microsoft. It doesn't seem like the most defensible position, but here's a "but"...
This "but" lies in the fact that OpenAI's deal with Microsoft only covers AI technologies up to AGI. With the completion of AGI, things should change, meaning Microsoft should not get AGI. However, it is currently unknown whether OpenAI's secret model Q* is ready and whether it is an AGI.
Musk is likely trying to drag OpenAI to court to get information on where OpenAI really is in its development and what this Q* model is.
In court, however, OpenAI will likely have to show that they do not have any AGI yet and that they can continue working with Microsoft. Therefore, OpenAI (at least for now) cannot release models more capable than GPT-4, even if they already exist.
This situation likely made it possible for Claude 3 to currently be the best large multimodal model, and Sam Altman cannot do anything about it. It would be nice, of course, if I'm wrong and we get to try GPT-5 at the beginning of next week ;)
You can read more about Elon Musk's case here:
https://buff.ly/49wjlAD
PS. The illustrative image was created using the new image AI model/app Ideogram 1.0, which can currently combine text and images better than other models on the market.
PPS. I improved the Ideogram-made image with Magnific.ai, which significantly enhances image detail and improves quality.
