ChatGPT Weekly News: E5

Bringing you the latest news on ChatGPT things and AI this week

After a few weeks of “clam”, the two giants in the field of generative AI, Google and Microsoft (with OpenAI), have intensified their efforts this week, initiating a new round of competition.

Google

PaLM 2

In Google I/O 2023 which ends this week, AI has undoubtedly been the sole protagonist. At the foundation of Google AI lies the PaLM2 model.

PaLM 2 is a new state-of-the-art language model from Google AI that is more powerful, efficient, and versatile than previous language models.

Comparing reports PaLM2 Technical Report and GPT-4 Technical Report, we can observe the following differences:

PaLM 2

• Model size: 540 billion parameters
• Training data: 1.56T words of text and code
• Training objectives: UL2, masked language modeling, code generation, and text-to-code translation
• Downstream tasks: Text generation, translation, summarization, question answering, code generation, and text-to-code translation
• Results: State-of-the-art results on all downstream tasks
• Efficiency: Faster and more efficient inference than PaLM

GPT-4

• Model size: 175 billion parameters
• Training data: 1.56T words of text
• Training objectives: Masked language modeling
• Downstream tasks: Text generation, translation, summarization, question answering, and code generation
• Results: State-of-the-art results on some downstream tasks
• Efficiency: Not as efficient as PaLM 2

It can be said that PaLM 2 serves as the AI engine for Google's entire product lines.