I just got back from Google I/O 2023 and it's clear the company is betting big on AI, with a keynote that was the most AI-dense I've ever seen from them. They announced PaLM 2, improvements to Bard, Duet AI, a new Pixel phone with on-device AI, and more, all in response to Microsoft's OpenAI push

PaLM 2 is the foundation model powering Google's AI products, and it's a significant upgrade from PaLM 1. It's a better coder, multilingual, and powers everything from Bard to Google Cloud's Vertex AI and on-device features in Pixel products. Google's taking a family approach with different size variants, similar to OpenAI's GPT-4 and GPT-3.5

One of the key challenges in building a large language model like PaLM 2 is the need for massive amounts of high-quality training data. Google has been working on improving its data curation process, using techniques like data distillation and filtering to ensure that the model is trained on the most accurate and relevant information. For example, PaLM 2's training dataset includes a large corpus of text from the web, books, and user-generated content, which is then filtered and ranked using a combination of human evaluation and automated metrics.

Bard's launch in February 2023 was rocky, with a factual error in its demo video that cost Alphabet $100 billion in a day. But by I/O, it had improved substantially, connecting to Google Search for current info, working with images, and showing better reasoning on coding tasks. The technical problems were easier to fix than the reputation damage, though

Duet AI is Google's answer to Microsoft 365 Copilot, integrating AI assistance into Google Workspace. You'll get writing assistance in Docs, formula generation in Sheets, slide design suggestions, and meeting summaries, all powered by LLMs. Both products launched in 2023 and represent a major shift in how we use productivity software

In terms of the technical architecture, Duet AI relies on a combination of Google's PaLM 2 model and its proprietary Workspace APIs to provide a seamless integration with Google's productivity suite. For example, when you're working on a document in Google Docs, Duet AI can use the PaLM 2 model to suggest alternative phrases or sentences, and then use the Workspace APIs to insert those suggestions directly into the document.

The big question is how Google will integrate generative AI into Search, and what that means for the search advertising model. If users get answers directly, the ad click-through model is under pressure. Google's solution is to include links in AI answers and show ads alongside them, but it's unclear how users will respond to this format

Google's approach to AI is all about integration, not just bolting it on as a feature. PaLM 2, Bard, and Duet AI are all part of a larger strategy to make AI a core part of every product, from Search to Workspace. It's a bold move, and one that will be closely watched through 2024

The on-device AI capabilities in the new Pixel phone are also worth noting. With PaLM 2 powering features like image recognition and language processing, it's clear Google is serious about making AI a core part of the mobile experience. This is a direct challenge to Apple's approach to AI, which has been more focused on privacy and security

As I look at the landscape of AI in 2023, it's clear that Google is making a major push to be a leader in this space. With PaLM 2, Bard, and Duet AI, they're covering all the bases, from Search to productivity to mobile. The question is, will it be enough to take on Microsoft and OpenAI

The economics of the search advertising model are going to be a major challenge for Google as they integrate AI into Search. If users don't click through to websites, the ad revenue model is under pressure. Google's solution is to include links in AI answers and show ads alongside them, but it's unclear how this will play out

One thing is certain, though: the integration of AI into Google's products is a major shift, and one that will have far-reaching consequences. As we head into 2024, it will be interesting to see how users respond to these changes, and how Google adapts to the feedback