Google Cloud offers access to more than 100 LLMs, but also lets customers customize LLMs. Google has updated its existing LLMs and added models from various parties to its offerings. The models can now also be seamlessly integrated into organizations’ applications with Vertex AI Extensions.
It’s been six months since Google unveiled Vertex AI, and now a huge expansion is coming on top of that. According to Google, the success of Vertex AI is huge, with many enterprise organizations using Vertex AI and 15 times more organisations joining every month.
With Vertex AI, customers can build new machine learning models, as well as customize existing AI models with their own enterprise data. Models can also be integrated with a variety of APIs that allow actions to be performed automatically. A model can be seamlessly integrated and deployed within an organization’s existing applications, while taking into account privacy and security and the use of responsible AI.
Probably the most essential part of Vertex AI is the Model Garden. Here customers can get started with foundational AI models. They can use APIs to access more than 100 LLMs. Whether it’s generating a response to a conversation like ChatGPT, or generating code, images, text summaries, and others.
New and updated models
Google has added new models that make Vertex AI even more powerful. These include the addition of Meta’s Llama 2 model and the Technology Innovative Institute’s Falcon model. Anthropic’s Claude 2 model has been announced but will be available soon.
Proprietary LLMs even better
Google Cloud has additionally provided updates to its own models. This makes the expertise of the Google Deepmind team also available to customers. PaLM 2 is now available in 38 languages and can shape responses with the customer’s own enterprise data. Longer responses and summaries of very long documents, such as research papers, are also possible. PaLM 2 supports 32,000 characters, enough to feed an 85-page document as input.
For generating programming code, Google has the Codey model, this model has received an update and should now be 25% more accurate, according to Google. Finally, Imagen has received an update to generate images even better, in addition it is possible to edit images and add watermarks.
Customization of existing AI models
Probably a much more important new feature is the customization of existing models. As an organization, you can now use your own company data to enhance the PaLM 2 model for text. As a result, the model better supports the writing style within your organization, as well as the context of documents but of course also use the data. Something that will be very valuable in certain industries. The company data will not be added to the PaLM 2 model so any company secrets stay safe..
It’s also possible to apply reinforcement learning via a prompt. Users can tell the model what is correct and incorrect. This is what Google calls Reinforcement Learning with Human Feedback (RLHF). Finally, the Imagen model can be customized so that all images adhere to the organization’s brand guidelines. This can be done by adding up to 10 images that adhere the brand guidelines.
Customizing AI model by using Vertex AI Extensions
The foundation models made available by Google Cloud are extremely powerful but at the same time outdated. This is because after the models are made available, they are no longer updated with new information that becomes available. This is inconvenient for some use cases; in some cases you need real-time data to achieve a good result. Or you want to use your own enterprise data to improve the response of the AI model.
To make this happen you need to tailor the model, which would become extremely expensive if you have to re-train the model for every customer. Google therefore works with extensions. The Vertex AI Extensions provide a set of developer tools that allow models to be connected to real-time data through APIs, or webhooks to perform actions. This allows developers to build very powerful applications, such as digital assistants, search engines and automated workflows.
The possibilities are enormous; for example, organizations can add an HR database and create a chatbot that performs HR tasks for employees. Developers can add their internal codebases and have the code checked for bugs and vulnerabilities. This can also be repeated automatically as new vulnerabilities emerge.
Adding a CRM or Service environment can allow customers to solve their problems faster through a chatbot, or schedule an appointment faster where diagnostics have already been performed and the organization knows what maintenance is needed.
Both for developers and data analysts, the Model Garden is an ideal playground, according to Google.