AI highlights of the week: Claude 2 released, GPT-4 ushered in a strong enemy
In the past week, ChatGPT's GPT-4 is no longer the most powerful language model that has swept all directions. Its old opponent in the GPT-3.5 era-Claude has been powerfully updated, and the LLM war has officially entered the 2.0 era.
Posted by Claude 2
The GPT-4 rival is here, free and powerful
Claude, known as "ChatGPT's strongest competitor", officially released the second generation version.
Claude 2 is trained with new data up to the beginning of this year. Currently, it can support the input of up to 100K tokens and the output of 4000 tokens. In conversion, it can output a coherent article of about 3000 words at a time.
After testing, Claude 2 has greatly improved its coding ability, and its logical ability also has a good performance. Of course, compared to the advantages of these parameters, the biggest feature of Claude 2 is the " free " strategy as always. Now users can try Claude 2 for free on the newly launched Claude Chat web version (currently only for UK and US IP).
ChatGPT online code interpreter
The "strongest function" since the launch of GPT-4
Last weekend, ChatGPT officially launched the Code Interpreter function to Plus subscribers .
The code interpreter can bring users a better interactive programming experience and data visualization support, and is known as the "strongest function" since the launch of GPT-4.
Some Twitter netizens once cited the things that can be done with the code interpreter, such as simple animation generation, analysis and interpretation of data, and output visualization results . Made a game , amazing code interpreter magic.
GPT-4 Technical Content Leaked
Contains 1.8 trillion parameters, using mixed expert models
On July 11, a breaking news article about the technical details of GPT-4 was published on SemiAnalysis. This document disclosed the specific parameters of GPT-4 model architecture, parameter quantity, data set composition, token number, training cost, etc. And the content, because the author of the document has previously revealed that the information is accurate, so the credibility is very high.
The article revealed that GPT-4 contains a total of 1.8 trillion parameters in 120 layers, which is about 10 times more than GPT-3, and uses a mixed expert (MoE) model to maintain a reasonable cost.
In the article, the author also bluntly stated that the closedness of the GPT-4 architecture is due to the fact that the things built by OpenAI are reproducible , and it is estimated that Google, Meta, Anthropic (Claude developer), Tencent, Baidu and other giants will also be in the short term. Has a model as powerful as GPT-4.
Google Bard ushered in a big update
Chinese communication is now supported
Google Bard has recently launched a major update, which not only adds communication support for more than 40 languages such as Chinese , Arabic, and German, but also adds functions such as speech-to-text and image analysis.
In addition, users can now choose the tone and style of Google Bard's responses. There are currently five expectations: simple, long, short, professional, and casual. Currently, this feature only supports English, and more languages will be supported in the future.
Stability AI launched Stable Doodle
Let AI add billions of details to your sketches
On July 14th, Stability AI launched Stable Doodle, a tool that transforms simple sketches into exquisite images , allowing your "few strokes" to become "drawing flowers".
Currently, users can use Stable Doodle on the website and application of Clipdrop by Stability AI. Just use the mouse to draw a basic sketch, choose one of 14 art styles, click "generate", and AI will add to your sketch. On the details of "billion points", let AI help you become a master of art.