Taming Llama2 with RLHF on Google Cloud

2023-09-01

Google announced at the recent Google Cloud Next event that they will offer Meta's Llama 2 and Code Llama, as well as Falcon LLM, on Google Cloud's Vertex AI. This is surprising because Google is the only cloud service provider that has not collaborated with competing organizations to host Llama 2 or other open-source LLM models.

This seems to be a decision by Google considering Vertex AI's focus on enterprise customers and their desire for more options. Llama 2 is the most popular large language model after GPT-4, considering its open-source and commercial availability. For Llama 2, Google claims to be the only cloud provider offering adapter tuning and RLHF.

Although Meta and Google are advertising competitors, they put aside their competition in the large language model space. Meta does not want to compete with any LLM business and is willing to offer Llama 2 to everyone. Now that Google Cloud has Llama 2, Meta has occupied every possible field. In fact, Meta has just broken the saying, "If you're good at something, never do it for free."

However, Google's acceptance of Llama 2 raises a question: Is PaLM 2 not powerful enough? Google Bard currently uses PaLM 2, which seems to be less favored by enterprises compared to ChatGPT because it cannot be customized according to their requirements and has lower answer quality.

This tech giant claims to have over 100 models in its Model Garden, including enterprise-ready base model APIs, open-source models, and task-specific models from Google and third parties. Google should understand that in the LLM space, it's not about quantity but quality.

Recently, OpenAI also took inspiration from Meta's approach and is now working on providing customization options for GPT-4 and GPT-3.5 while avoiding open-sourcing. To achieve this goal, the creator of ChatGPT recently launched the GPT-3.5 Turbo API for fine-tuning. Additionally, it is partnering with Scale to fine-tune GPT-3.5 to attract enterprises.

The Battle of LLM Cloud Begins

Google may be late to the game, but there is still hope to follow in the footsteps of AWS. Google has realized the importance of hosting multiple LLMs, just like Amazon's Bedrock platform. Currently, Bedrock hosts models from AI21, Cohere, Anthropic Claude 2, and Stability AI SDXL 1.0.

Microsoft is currently actively exploring different LLMs. Microsoft Azure currently includes all OpenAI services through APIs, including Azure OpenAI services. This allows enterprises and developers to create applications using GPT, DALL·E, and Codex.

When Meta launched Llama 2, Azure was announced as the preferred partner. Microsoft seems to not stop there as they further plan to sell a new version of the Databricks software on Azure that can help customers create AI applications for their businesses from scratch or reuse open-source models as an alternative to licensing proprietary models from OpenAI.

In the latest quarter, Azure achieved a 26% revenue growth, thanks to the push of Azure OpenAI services, making it the winner. However, now that all three clouds have Llama 2, it will be fascinating to see who will lead the competition in the LLM cloud in the future quarters.

What about Gemini? With Vertex AI now hosting Llama 2 and shifting focus to smaller models similar to Microsoft and Amazon, the feasibility of creating Gemini to counter GPT-5 is raised. This consideration is particularly important as OpenAI is also shifting focus to serving enterprises.

Let's not forget that Google's Codey has a new competitor called Code Llama, and only time will tell who codes better and to what extent it is adopted by enterprise customers and developers.

RECENT AI NEWS

RECENT AI TOOLS