TheChek News Flow

OpenAI’s success in overtaking Google with an artificial intelligence–powered chatbot has achieved what seemed impossible in the past: It has forced the two AI research teams within Google’s parent, Alphabet, to overcome years of intense rivalry to work together.

Software engineers at Google’s Brain AI group are working with employees at DeepMind, an AI lab that is a sibling company within Alphabet, to develop software to compete with OpenAI, according to two people with knowledge of the project. Known internally as Gemini, the joint effort began in recent weeks, after Google stumbled with Bard, its first attempt to compete with OpenAI’s chatbot.

The release of Bard was also marred internally by the resignation of a prominent Google AI researcher who had raised red flags about its development to Alphabet CEO Sundar Pichai and other executives.

THE TAKEAWAY

• Researchers from DeepMind and Google are working on software to rival OpenAI’s GPT-4

• High costs and computing needs forced researchers from the two units to work together

• Key AI researcher quit in January after raising red flags about Google’s Bard chatbot

DeepMind has operated more like a rival than a collaborator with regard to Google Brain since Google acquired DeepMind in 2014. The two have been competing to improve some of Google’s products and services and to gain global notoriety for research breakthroughs. Now, though, employees at both of Alphabet’s AI labs agree that OpenAI has outflanked them. Plus, the startup has hired away some of their key engineers and researchers, The Information has reported.

OpenAI in November launched ChatGPT, a chatbot that gives humanlike answers and has become one of the fastest-growing apps in history. Google’s own chatbot, Bard, which became available to a limited group of users last week, appears to perform worse on some tasks compared to ChatGPT.

Reflecting Gemini’s importance, Jeff Dean, Brain’s leader and the most senior AI research executive at Google, has taken a technical role in the project, writing code to help make a machine-learning model that would try to match the capabilities of OpenAI’s GPT-4, the model that powers ChatGPT, one of these people said.

The effort is one sign of how Google has shaken up the product road maps at countless teams including search and cloud as a result of competition from ChatGPT and OpenAI’s work to incorporate its tech inside products made by Microsoft—which is funding the startup and paying for its computing needs.

Gemini is a forced marriage. Alphabet’s two AI labs have seldom collaborated or shared computer code with one another. But now, because both wanted to develop their own machine-learning model to compete with OpenAI and needed an inordinate amount of computing power to do so, they had little choice but to work together, said a person with knowledge of the situation.

Red Flag

Bard’s early stumble demonstrates Google’s problems. The chatbot was created in a process so contentious that a prominent Google AI engineer, Jacob Devlin, resigned in January and immediately joined OpenAI, according to a person with direct knowledge and another person briefed about it. Devlin was the lead author of a seminal paper on a method of training machine-learning models to improve their understanding of groups of sentences—an invention OpenAI has incorporated into its language models.

Devlin quit after sharing concerns with Pichai, Dean and other senior managers that the Bard team, which received assistance from Brain employees, was training its machine-learning model using data from OpenAI’s ChatGPT. Specifically, Devlin believed the Bard team appeared to be relying heavily on information from ShareGPT, a website where people publish conversations they’ve had with ChatGPT.

Some Google employees felt the use of such chat logs would have violated OpenAI’s terms of service, which prohibit the use of “output…to develop models that compete with OpenAI,” according to its website. Devlin also told executives he was concerned that by relying on the ChatGPT chat logs from ShareGPT, Bard’s answers would mimic or resemble those of ChatGPT too much.

After Devlin raised the issue, Google stopped using ChatGPT data to train Bard, one person said. The Bard team has been led in part by Sissie Hsiao, a vice president who previously was in charge of Assistant, which is similar to Apple’s Siri voice assistant, said two people with knowledge of the group. A Google spokesperson didn’t comment on Devlin’s resignation and allegation. Spokespeople for OpenAI didn’t immediately respond to a request for comment.

1 Trillion Parameters

The business implications of OpenAI’s early lead over Alphabet aren’t clear. Though it makes plenty of mistakes, millions of people are already using ChatGPT and the technology behind it to quickly generate blog posts and other content, summarize meeting notes and automatically create spreadsheets, among many other tasks, according to a person with knowledge of the figures.

The product has generated relatively little revenue to date, though that could change. Microsoft has attempted to boost its own search engine, Bing, by incorporating a ChatGPT-like feature with OpenAI’s help, but so far that doesn’t appear to have taken meaningful market share away from Google Search. Still, OpenAI appears to have heralded a new era of applications that can understand requests customers make using natural language.

The Gemini project aims to develop a large-language model—a computer program that can understand and generate human-like language—that would have up to 1 trillion parameters, a measure of calculations in a machine-learning model, said one of the people briefed about it. GPT-4 also has around 1 trillion parameters, Semafor reported. Google’s effort requires tens of thousands of tensor processing units. The company’s equivalent to Nvidia’s graphical processing units, TPU microchips are particularly well suited to training large machine-learning models, this person said. Still, the Gemini effort may take months to complete, this person said.

Meanwhile, other teams at Google have been developing their own large-language models as part of a companywide effort to integrate AI into all of its products. Pandu Nayak, a Google executive overseeing search rankings, has been developing a separate model to handle certain search queries, while Google Cloud is creating its own model to sell to cloud customers, similar to the one customers can buy from OpenAI or Microsoft’s Azure cloud unit, this person said.