Anthropic presenta Claude 3, con l’obiettivo di superare Gemini e ChatGPT thumbnail

Anthropic presents Claude 3, aiming to surpass Gemini and ChatGPT

It is now well established: among the big tech companies that are working on generative artificial intelligence, the challenge is very open and constantly evolving. The aim is to release increasingly powerful software on the market with increasingly astonishing performance. And since it is an almost newborn technology, in the first period it is reasonable to expect notable improvements in a short time.

Until a few months ago, when deepfakes were easily unmasked due to gross errors (especially in figures not in the foreground), for example, the precision shown by the video clips of Sora, the OpenAI AI that generates images starting from from a text command.

Furthermore, the major companies that are producing generative artificial intelligence software always present updated versions every few months after the last launch. This is the case of Anthropic, which just showed the world Claude 3.

Anthropic presenta Claude 3

We didn’t even have time to read version 2.1, and Anthropic launched Claude 3.

He did so through a long post published on his official website on Tuesday 4 March. Where we can read: “Today we announce the Claude 3 family of models, which sets new industry benchmarks for a wide range of cognitive tasks.

The family includes three cutting-edge models in ascending order of capacity: Claude 3 Haiku, Claude 3 Sonnet and Claude 3 Opus. Each successive model offers increasingly powerful performance, allowing users to select the optimal balance of intelligence, speed and cost for their specific application.

Opus and Sonnet are now available for use in claude.ai and in the Claude API which is now generally available in 159 countries. Haiku will be available soon.”

The novelty of Claude 3

The main novelty of the Claude 3 family is its multimedia. That is, the fact that software is able to understand and process both texts and images.

They are able to provide answers to even complex questions within a few seconds, and compared to previous versions the company has made it known that all three models have a much lower probability of rejecting requests due to a lack of understanding of the context.

Even Haiku, the least powerful of the three systems, is “capable of reading an article full of graphs and tables in less than three seconds”.

The three models were trained on a combination of non-public internal datasets, third-party data, and public data available as of August 2023. Training was done using Amazon Web Services (AWS) hardware and Google Cloud.

Recall that last year Amazon invested 4 billion dollars in Anthropic.

The performances

Secondo Anthropic, Claude 3 models demonstrated graduate-level reasoning abilities superior to those of GPT-4. In a specific test in this sense, they obtained 50.4% against 35.7% of GPT-4 of the more advanced OpenAI software.

Claude 3 would also surpass GPT-4 in writing code, solving mathematical problems and understanding logical reasoning.

In the note, the company explains that “all Claude 3 models demonstrate enhanced capabilities for analysis and prediction, nuanced content creation, code generation, and conversation in languages ​​other than English such as Spanish, Japanese, and French.”

The Claude 3 model family will initially have a context window of 200,000 tokens. However, Anthropic specifies, “all three models are capable of accepting inputs greater than 1 million tokens and we may make it available to selected customers who need greater processing power”.

The security

A large part of the note explains that the Claude 3 family of models is supervised by several teams that monitor various dangers: from misinformation to the spread of child pornography to electoral interference.

Furthermore, “Claude 3 shows less bias than our previous models according to the Bias Benchmark for Question Answering (BBQ). We remain committed to promoting techniques that reduce bias and promote greater neutrality in our models, ensuring they are not biased towards any particular partisan position.”

Costs

Opus, the most powerful model, costs $15 per million tokens for input and $75 per million tokens for output.

For Claude 3 Sonnet the cost is $3 for input and $15 for output (per million tokens).

And Haiku costs $0.25 for input and $1.25 for output, again for every million tokens.

Walker Ronnie is a tech writer who keeps you informed on the latest developments in the world of technology. With a keen interest in all things tech-related, Walker shares insights and updates on new gadgets, innovative advancements, and digital trends. Stay connected with Walker to stay ahead in the ever-evolving world of technology.