Fast company logo
|
advertisement

TECH

Google, not OpenAI, has the most to gain from generative AI

While OpenAI has captured the public’s imagination with ChatGPT, ultimately the technology may not change the balance of power among the tech giants.

Google, not OpenAI, has the most to gain from generative AI

[Photo:
cottonbro studio
/Pexels]

BY Mark Sullivan10 minute read

After spending billions on artificial intelligence R&D and acquisitions, Google finds itself ceding the AI limelight to OpenAI, an upstart that has captured the popular imagination with the public beta of its startlingly conversant chatbot, ChatGPT. Now Google reportedly fears the ChatGPT AI could reinvent search, its cornerstone business.

But Google, which declared itself an “AI-first” company in 2017, may yet regain its place in the sun. Its AI investments, which date back to the 2000s, may pay off, and could even power the company’s next quarter century of growth (Google turns 25 this year). Here’s why.

OpenAI supercharged natural language processing (NLP) models by making them larger and feeding them massively larger amounts of training data, culled from free e-books, Wiki pages, discussion boards, and fan fiction all over the internet. But OpenAI didn’t invent the NLP model that powers ChatGPT. GPT stands for “generative pretrained transformer,” and it was Google that invented the transformer language model in 2018 with BERT (bidirectional encoder representations from transformers), which Google now uses to enhance its search and translation capabilities.

But Google didn’t stop working on NLP after BERT. In fact, Google claims that it has an NLP chatbot called LaMDA that’s more powerful than ChatGPT (one Google engineer swears the thing is “sentient”). Alphabet subsidiary DeepMind is also considering releasing an AI chatbot called Sparrow later this year, and Google’s AI image generator Imagen is said to rival OpenAI’s DALL-E 2.

So far Google has been more cautious about making these tools available to the public. That’s not entirely surprising: Google is a far larger company than OpenAI, with much more to lose from any AI missteps. Google considers generative AI to be an immature technology, its risks still not properly understood. The company reportedly fears it could face liability if one of its AI tools violates someone’s privacy, or violates a copyright, or creates some form of monopoly that might invite an antitrust lawsuit. 

But Google has a lot to gain from AI, too, and has set itself up to offer new and novel AI-powered features. The company’s approach has been to use its foundational AI models to run in the background and make its apps and services work better.

Reinventing search

Search is a good example. Google has already been using BERT language models to interpret the intent behind the keywords and phrases that users type into the search bar. BERT was designed to infer meaning from text, but more advanced language models—generative models—can be used to compose search results in the way that ChatGPT generates textual answers to user questions. Instead of just returning a long list of decreasingly relevant links (accompanied by ads), Google Search might generate a narrative—supplemented with images, video, statistics, and links—that answers the searcher’s questions directly. 

In fact, Google has been working on this generative approach to search for a while now. Its multitask unified model (MUM) is designed to help users with complex, or many-faceted, search requests. For example, a user might say or type “How do I prepare for a trip to Tibet?” and MUM will infer all the questions inherent in the query, then generate a customized, multimedia package of information that addresses all of them and more.

Google also has the advantage of its long experience crawling and cataloging the web and its vast contents. The company may be able to access more and better-quality training data for its models, which could result in impressive AI apps and services that don’t generate incorrect or toxic content.

“I do think that large language models like ChatGPT can be disruptive to the web search business,” says Landing AI founder and CEO Andrew Ng. But Ng is quick to add that making search a back-and-forth conversation with an AI model is no easy feat: How do you actually train an NLP model, even a large one, on the petabytes and petabytes of content on the web? How do you develop the training data? Then there’s the issue of monetization.

“The core dynamics of Google’s business model is to return links and place ad links next to the results,” says Ng, formerly a founding lead of the Google Brain team. “So [if] Google is now just returning the answer directly [from the model], how does that affect the ad business model? I think these are all really complicated questions.”

It may be a somewhat simpler matter for Google to integrate generative AI models into its productivity apps. The tech could, for example, be used to help Gmail users draft messages, or help Docs users compose documents or presentations, or offer Meet users summaries of video chats. 

OpenAI and Microsoft

Until recently, the OpenAI-Google rivalry looked like a David and Goliath story: a small, agile player menacing a long-entrenched tech behemoth with a novel, disruptive technology. OpenAI may have decided to open ChatGPT to the public in hopes of perpetuating this impression, raising its public profile, and perhaps to attract more investment money.

That strategy seems to have worked. OpenAI already had a distribution partnership with Microsoft, but this week Microsoft announced an expansion of the partnership as part of an additional $10 billion investment in the company. Microsoft now owns a significant share in OpenAI and intends to build the technology into its Bing search engine, as well as into its productivity apps.

Because of that, Index Ventures partner Mike Volpi points out, we’re not talking about a David and Goliath situation anymore.

“OpenAI is not a small player; it’s basically a subsidiary of Microsoft,” says Volpi, whose firm has been investing in AI startups since 2016. “A majority of the funding comes from Microsoft, so I think it’s fair to think of it as an extension of Microsoft; and in that context, a lot of the other things become very clear. Bing’s market share [in the search advertising business] is small relative to Google’s, so Microsoft definitely has a desire to disrupt that flow.”

Cloud matters

Microsoft sees OpenAI as a way to offer advanced generative AI services through its Azure cloud, which is second only to Amazon AWS in the cloud market. Microsoft may also be using its association with OpenAI to lend credibility to its claim that the Azure cloud is preferable to other cloud services for hosting advanced AI models. 

“Azure delivers best-in-class performance and scale for the most compute-intensive AI training and inference workloads,” wrote Azure AI Platform VP Eric Boyd in a blog post this month. “It’s the reason the world’s leading AI companies—including OpenAI, Meta, Hugging Face, and others—continue to choose Azure to advance their AI innovation.” 

Google is working hard to promote its own cloud service (and its cloud business is headed for profitability) but it still trails Azure and Amazon’s AWS by a significant margin. Azure’s addition of OpenAI services could make Google’s effort to catch up harder.

advertisement

An AI arms race?

It can be tough to get a read on Google management’s take on OpenAI. In mid-December Google’s AI chief, Jeff Dean, said his company had more to lose from launching a generative AI tool that made mistakes or spewed toxic content. But the company is taking OpenAI’s bid for gen AI leadership seriously. The New York Times reports that Google brought in founders Larry Page and Sergey Brin, who retired from day-to-day management of the company in 2019, to discuss ways of countering the upstart’s current mojo.

In a December strategy meeting, The Times reports, Google said it will “recalibrate” the level of risk it’s willing to take when releasing new public-facing generative AI tools.

In other words, ChatGPT may have jump-started an AI arms race. It now seems very likely that Google will turn out a number of new AI features and products this year. Some of them will be aimed at developers: Google will likely develop tools like GitHub’s Copilot, which, trained on millions or billions of lines of code, can intuit the needs of the developer and suggest lines or chunks of new code.

Index’s Volpi believes it’s very likely that Google will release a natural language chatbot that looks and acts a lot like ChatGPT. DeepMind may also release its Sparrow chatbot. Google Search may offer a chat-style experience for some types of searches.

A year from now OpenAI may look like a first mover in generative AI, but its models may not look quite as novel and distruptive as they look now. Ultimately, the OpenAI story might not significantly change the balance of power among the Big Tech players.

“I’m not sure that it changes the power hierarchy,” Volpi says. “Google is going to have a [NLP chat] product, Amazon’s going to have a product for sure. I can’t imagine that Apple is not going to improve Siri to do about the same things. So in the end . . . my guess is you’re going to see a similar shape to the game.”

We civilians can sit back and watch the companies fight it out, and enjoy the results. But there’s an element of danger here, too. The pace of AI development (within the well-funded companies that can afford to do it) is likely to accelerate in the wake of ChatGPT. As competition grows, so might secrecy. As Big Tech companies invest more in AI R&D they may be less willing to divulge the details of their models and how they work in the research papers they publish.

The normal collegiality and scientific openness of the AI research community has already diminished somewhat over the past year, says Percy Liang, director of the Center for Research on Foundation Models at Stanford’s Institute for Human-Centered AI. Liang says it’s important that tech companies divulge the details of their models and training methods and policies so that other researchers can re-create the models and their outputs, thereby validating the research. OpenAI does publish papers on the new models it creates, but Liang says the company hasn’t always provided enough detail about its methods.

“The latest OpenAI models, we just have no idea what’s behind them,” he says. “And until recently we don’t even know what size they are, what data they’re trained on, whether it’s the same model [described in the research paper] that we are hitting using the API [to access and test the model]. They’ve since released a small blog post that explains some of this, which is, you know, one small step in the right direction, but there’s still a lack of transparency.”

Liang says the final output of an AI model is greatly influenced by the downstream fine-tuning and filtering imposed by human beings. Developers must do this to prevent the model from outputting text or imagery that’s incorrect, or toxic, or biased in favor of certain user types and against others. For instance, if you asked the early version of OpenAI’s GPT-3 model to write a brief story about a doctor, the doctor character would always be male. By the time ChatGPT was released, this bias was gone, likely fixed with additional training. But that’s one example. There are very likely hundreds of these issues to address.

“OpenAI likes to talk about value alignment, so these models are going to be aligned to human values . . . so they don’t do bad stuff, but whose values are we talking about? What are these values?” Liang asks. “There’s nothing [published] about what these values are, and it’s some sort of opaque decision process.”

OpenAI willingly fixed some of these problems on the fly while the ChatGPT beta was open to the public. It even has a “bias bounty” program wherein users can report incorrect or biased output in exchange for a chance to win $500 in API credits. Google has not been willing to expose AI products and features with such shortcomings. But the increasing pace of AI development may put pressure on Google, and all the Big Tech companies, to launch new models before all the kinks are worked out.

And here the losers could be us—the users who might see generative AI tools begin adding even more false, toxic, and biased content to the growing tide we already face in online life.

PluggedIn Newsletter logo
Sign up for our weekly tech digest.
This site is protected by reCAPTCHA and the Google Privacy Policy and Terms of Service apply.
Privacy Policy

ABOUT THE AUTHOR

Mark Sullivan is a senior writer at Fast Company, covering emerging tech, AI, and tech policy. Before coming to Fast Company in January 2016, Sullivan wrote for VentureBeat, Light Reading, CNET, Wired, and PCWorld More


Explore Topics