Google expands cloud AI offering with Meta, Anthropic and PaLM update

Summary

Google is expanding its cloud AI offerings to include third-party models like Llama 2. The company also announced several updates to its products.

At its Next’23 cloud event in San Francisco, Google announced a more open cloud AI platform: In addition to its own PaLM language models, the company is now offering Meta’s Llama 2 and Anthropic’s Claude 2. In total, more than 100 major AI models for a variety of tasks will be available through Google’s Vertex AI model garden.

Among major cloud providers, Google now joins Microsoft and Amazon, which also offer AI models from multiple providers. Microsoft has exclusive access to OpenAI’s GPT models, while Google has its PaLM models as an exclusive LLM offering.

By offering AI models, cloud providers aim to increase their cloud market share. It makes sense to offer as many models as possible and have an additional exclusive offering to be the most attractive platform for generative AI applications. Until GPT-5 is released, Google may have a strong case for its cloud with the upcoming Gemini multimodal model.

32K upgrade for PaLM 2

Google also announced that PaLM 2 is now available in 38 languages and with a 32K context window. This puts PaLM 2 on par with GPT-4, which is also available in a 32K version for OpenAI’s ChatGPT Enterprise.

A larger context window enables the model to process more text simultaneously, allowing it to analyze larger documents or write longer articles, for example. Anthropic’s Claude 2 currently offers the largest context window at 100K.

Google’s AI code generator, Codey, now offers 25 percent better code generation quality in the most popular programming languages, according to Google.

Imagen, Google’s generative image system, is also getting an update: It’s supposed to produce prettier images now and offers new features like editing, captioning, and the ability to answer visual questions about images. Imagen will also get Deepmind’s experimental watermarking feature SynthID.

Google is also expanding its AI model customization offerings. PaLM 2 now supports fine-tuning, including Reinforcement Learning with Human Feedback (RLHF), familiar from GPT-3.5 and ChatGPT. With Style Tuning, businesses can customize Google Imagen to match their brand’s style with ten or fewer image examples.