Google Brain and DeepMind are combining their efforts on Gemini, a project described as a competitor to OpenAI’s GPT-4. The collaboration follows concerns that Google’s Bard had not kept pace with ChatGPT, and the project is reported to draw on substantial computing resources and senior technical involvement.
Two teams, one model effort
According to The Information, Google Brain is working with DeepMind on an alternative to GPT-4. Work on Gemini is said to have begun in recent weeks. The report describes a large language model, though it does not establish whether the system will handle more than text.
The teams had each wanted to build their own GPT-4 competitors. Pursuing separate models would have required enormous amounts of computing power, and a person familiar with the situation told The Information the groups had “little choice but to work together.” The merger of effort makes Gemini a shared undertaking rather than two separate attempts to catch up.
Ambition comes with a long build
The planned model is said to have a trillion parameters, like GPT-4. Training is reportedly expected to use tens of thousands of Google’s TPU AI chips. Those details point to a project designed at a very large scale, but they do not mean the model is ready: completion could take months.
Several important details remain unknown. The report does not say whether Gemini will be multimodal, meaning able to work across different kinds of input or output. It also offers no confirmed release date or account of what the finished model will be able to do. For now, the reported scale and schedule describe the ambition and effort, not a demonstrated result.
A response to competitive pressure
The project reflects pressure on Google and DeepMind from OpenAI. The Information reports that employees at both companies believed “OpenAI has outflanked them.” The startup had also recruited important researchers from Google and DeepMind, adding another reason for the companies to coordinate their work.
DeepMind, based in London, otherwise operates almost independently of Google, according to the report. Gemini therefore represents a notable joint effort between organizations that have not typically worked as one closely integrated unit. The account also says Jeff Dean, head of Google Brain and Google’s most senior AI research executive, is taking a technical role and writing code for the project.
Gemini sits alongside Google’s other AI plans
Google is continuing to develop Bard while Gemini is underway. A separate report says Bard is expected to merge with Google Assistant. The company also plans to bring generative AI to Gmail, Docs, Slides, and other products.
Those efforts serve different parts of Google’s AI strategy as described in the source: Gemini is the reported bid to build a GPT-4 competitor, while Bard and Google Assistant are connected to the company’s products. Google is also making its own AI models available to developers through the cloud.
The picture is one of several initiatives moving forward at once, with Gemini’s training still expected to take months. Whether the model will be multimodal remains unresolved, and the report does not establish how it will perform. The immediate development is the collaboration itself: Google Brain and DeepMind are pooling their work on a project with significant computing demands and senior attention.