Applied AI Summit

Today’s Google Cloud ‘Applied AI Summit’ primarily focused on introducing Gemini to developers. Here are my key takeaways:

If this were a few months ago, it would be mind-blowing. But, with OpenAI’s substantial lead and innovation, there weren’t any breakthrough surprises. That said, Google is back in the running in the LLM space.

The most common theme was multi-modality within one LLM. One demo showed uploading a video and two images and getting a detailed description of a house’s interior and exterior that required all three images to get the whole picture.

They demoed their essential tools: Duet AI, AI Studio, Vertex AI, and Gemini Models.

Vertex AI is for enterprises, is sandboxed, and doesn’t leak training or usage back to the LLM.

They claim that:
– More than 50% of startups are on Google Cloud
– More than 70% of AI unicorns are built on Google Cloud

As announced last week, there are three LLM models:
– Gemini Nano (for on-device tasks)
– Gemini Pro (best for a wide range of tasks)
– Gemini Ultra [coming early 2024] (most capable)

https://bard.google.com is now using Gemini Pro.

They now have 𝘍𝘶𝘯𝘤𝘵𝘪𝘰𝘯𝘴 (like OpenAI) which can make API calls to other applications.

They’ve made significant strides in their TPU hardware and cloud offerings for training and inference of large AI models.

They have similar features to OpenAI, like chat prompts (like assistants), fine-tuning, and structured and unstructured output.

As a user, try it out at https://lnkd.in/eng3r2Yi

As a developer, get started with AI Studio: https://lnkd.in/eqVrfCwc (the old name, Maker Suite, is still reflected in the URL).

Or, try out Vertex AI Studio: https://lnkd.in/e6fsj_5Q

Or, try Duet AI: https://lnkd.in/ebssCU_A

Quick side note: I’ve been pleasantly surprised and have been testing Gemini Pro (in Bard) for the last few days for my day-to-day usage. My unscientific take is that it’s on par with GPT 3.5, with some wins and some losses.

My summary (based on Gemini Pro):
– Against OpenAI and GPT 3.5, Google Cloud’s Gemini Pro comes in as a fully valid 𝘤𝘩𝘰𝘪𝘤𝘦, but not a new 𝘭𝘦𝘢𝘥𝘦𝘳.
– While Google hasn’t leapfrogged OpenAI with their new release of Gemini, they have caught up in many ways and we should expect intense competition.
– The release of Gemini Ultra in early 2024 will put OpenAI and Gemini neck-in-neck. We should expect 2024 to be an exciting year of regular leapfrogging of innovations between the top LLM providers.
– One important note is the enterprise focus of Vertex AI which is focused on privacy. OpenAI doesn’t do this today, although you have options with Microsoft’s Azure OpenAI.

I’ll be watching Gemini closely, including experimenting and working with their API, and using bard.google.com as a top two choice for my day-to-day usage.