Google Lays Out Generative AI Challenge to OpenAI and Microsoft
The PaLM model, MakerSuite developer tools, and new features for Docs and Gmail
OpenAI announced GPT-3 in May 2020 and officially introduced its developer beta program two months later. This week, Google announced the PaLM API and the new MakerSuite. The latter is “a tool that lets developers start prototyping quickly and easily,” according to a blog post by Google vice president Scott Huffman and senior director Josh Woodward. PaLM is the acronym for Google’s Pathways Language Model. PaLM’s architecture differs from the more well-known LaMDA model, which is expected to power the Bard search assistant.
Like OpenAI in 2020, Google is announcing PaLM ahead of broad beta availability but is working with “select developers through a Private Preview.” Huffman and Woodward said, “stay tuned for our waitlist soon.”
Google also announced that new generative AI features would soon arrive in Google Workspace, the collection of office productivity apps previously known as G Suite.
More than 3 billion people already benefit from AI-powered features in Google Workspace, whether it’s using Smart Compose in Gmail or auto-generated summaries in Google Docs. Now, we’re excited to take the next step and bring a limited set of trusted testers a new set of features that makes the process of writing even easier. In Gmail and Google Docs, you can simply type in a topic you’d like to write about, and a draft will be instantly generated for you.
Workspace Takes on 365 Copilot
Google announced the new Workspace generative AI features two days before the Microsoft 365 Copilot launch. Google’s new features include:
draft, reply, summarize, and prioritize your Gmail
brainstorm, proofread, write, and rewrite in Docs
bring your creative vision to life with auto-generated images, audio, and video in Slides
go from raw data to insights and analysis via auto completion, formula generation, and contextual categorization in Sheets
generate new backgrounds and capture notes in Meet
enable workflows for getting things done in Chat
This list seems far less ambitious than 365 Copilot, which was presented as an integrated system that enabled natural language inputs to control and augment individual applications as well as execute tasks across applications. By contrast, Google’s announcement seems to point to discrete feature additions to two of its most popular productivity applications. With that said, enabling “workflows for getting things done in Chat” could mean a lot of things. There may be a more sophisticated orchestration idea behind that comment.
These features will be welcomed by Google users regardless of how they compare to the 365 Copilot system. That “system” still has not been independently reviewed in the wild, and some of its more interesting features could still fail to live up to the marketing hype.
Another consideration is that business users are unlikely to switch overnight to 365 from Workspace or vice versa due to a few new generative AI features. So, bringing to market some high-impact features right away could address user needs and increase the perceived value of these applications despite Microsoft’s ambitions.
The bigger impact may be for AI writing assistants such as Jasper AI and Copy AI. Companies in this segment are attempting to carve out a new role in business productivity applications but may seem unnecessary for some Google docs users if similar features are already available in the applications they already use.
Will PaLM Be Competitive with GPT-4?
Google I/O, the company’s annual developer conference, is just two months away. So, it seems logical that the waitlist will appear no later than I/O and likely before. PaLM was first announced in April 2022. That means PaLM in 2023 may be more mature than GPT-3 in 2020, and some data about PaLM in 2022 suggested it was already better for some tasks a year ago.
However, it is logical to assume the solution is behind OpenAI’s generative AI model portfolio, which has nearly three years of experience working with third-party developers on its GPT-3 foundation model. OpenAI has also delivered two significant upgrades this year to that model and last week introduced GPT-4. OpenAI also called out GPT-4 as improved in critical reasoning, one area where PaLM may have shown an advantage over GPT-3 in the past.
What Model is This?
PaLM is known for having about three times more parameters in its model than GPT-3. It also used significantly more data tokens in training. With that said, it seems clear that the largest PaLM model with 540 billion parameters is not what Google will initially offer. Thomas Kurian, CEO of Google Cloud, said:
“For developers who are experimenting with AI, we’re introducing the PaLM API, an easy and safe way to build on top of our best language models. Today, we’re making an efficient model available, in terms of size and capabilities, and we’ll add other sizes soon.”
“An efficient model … in terms of size and capabilities” sounds like a smaller model. Google’s research paper describing PaLM said it compared performance data 8B, 62B, and 540B. So, Kurian may be referring to the 62B parameter model, or there could be a new model size between these.
One additional point to note. Google researchers suggested the improved performance between the 62B and 540B models was significantly higher than the rate of improvement shown between 8B and 62B. Unless there is some point of diminishing returns, that could mean that a model smaller than 540B could be significantly less robust than the references to PaLM’s performance last year.
To better understand the scaling behavior, we present results at three different parameter scales: 8B, 62B, and 540B. Typically, scaling from 62B to 540B results in similar performance as scaling from 8B to 62B, which is consistent with the “power law” rule of thumb often observed in neural network scaling (Kaplan et al., 2020). However, for certain tasks, we observe discontinuous improvements, where scaling from 62B to 540B results in a drastic jump in accuracy compared to scaling from 8B to 62B.
How Will This Impact the Market?
The introduction of PaLM and MakerSuite seems more immediately important than the generative AI additions to Workspace. There are competitors to OpenAI ranging from AI21 Labs and Anthropic to Cohere and Hugging Face. However, Google’s entry into this market is significant. It knows how to support a developer community, and it has cloud hosting services that it can optimize for generative AI model training and inference use cases.
On the developer support front, MakerSuite could be the key to Google’s near-term strategy. MakerSuite includes tools for model tuning, generating synthetic data for model training, generating embeddings, scaling, and adjusting trust and safety settings. If MakerSuite turns out to makePaLM-based generative AI applications easier to develop and scale, it could be Google’s most important differentiator.
There are a lot of unanswered questions everywhere when it comes to generative AI. OpenAI has answered the most questions because it has a lot of users and a lot of developers employing its AI models every day. That makes everyone else a challenger in this market.
Google has all the tools and technical assets to compete. How well Google will perform remains to be seen, but the company is finally providing some visibility into where it’s headed. There is still a big gap between introducing a model and developer tools and turning that into a lot of developers, applications, and end users. Google has come from behind before. Will generative AI be a repeat performance of Android in mobile, or will it be a market where the company waited too long?