Google DeepMind is developing Gemini, an AI system intended to combine the language capabilities of large models with techniques associated with AlphaGo. The project’s stated ambitions include problem-solving and planning, though the source report says Google DeepMind had not said when Gemini would be released.
Bringing language models and AlphaGo techniques together
DeepMind CEO Demis Hassabis described Gemini as combining strengths from AlphaGo-type systems with the language abilities of large models such as GPT-4, the technology behind OpenAI’s ChatGPT. He also referred to new innovations, without detailing them in the report.
AlphaGo is associated here with reinforcement learning and tree search. The report says Gemini aims to use techniques such as these to support capabilities including problem-solving and planning. In practical terms, that points to a system designed to do more than produce language: it is also intended to work through problems and plan.
Hassabis said the company’s reinforcement learning expertise could give Gemini distinctive features. That describes the project’s direction, rather than confirming how well those features will work or what users will be able to do with them.
A system still being developed
Gemini was officially unveiled in May, and Hassabis said development and training would continue for several more months. The report did not give a release date. Google DeepMind had not said when the system would become available, leaving its timing unresolved.
The model was also described as multimodal, with capabilities not seen in previous models, and as being designed to integrate tools and APIs efficiently. The report does not specify which kinds of inputs or tools Gemini would support. Those descriptions therefore indicate intended areas of capability, rather than a complete account of the finished system.
Gemini is expected to come in multiple sizes. It is also being designed with future innovations such as memory and planning in mind. Taken together, these details describe a project intended to develop in stages and accommodate different versions, while leaving the exact functions of each size unspecified.
Scale, cost and uncertainty
The report says development could cost tens or hundreds of millions of dollars, and that the project could play a critical role in Google’s response to ChatGPT and other generative AI technologies. The wording is prospective: it describes possible costs and significance, not confirmed final figures or outcomes.
Rumors in March said Gemini would have a trillion parameters, like GPT-4 reportedly does. The report also says the project was said to use tens of thousands of Google’s TPU AI chips for training. These points are presented as rumors or reported claims, not as specifications confirmed by Google DeepMind.
That distinction matters when assessing claims about a system still in development. The account gives a broad picture of Gemini’s intended combination of language, reinforcement learning, planning and tool use, but it does not establish the model’s final performance, confirmed scale or release schedule.
What the report says about the AI race
Gemini is framed as part of Google’s response to ChatGPT and the wider development of generative AI. The report also notes that OpenAI CEO Sam Altman said GPT-5 was still a long way from launch and would not begin training for at least six months; it described a launch in 2024 as seeming likely.
Those remarks offer context for the competition described in the article, but they do not settle when either company’s next system would arrive. For Gemini, the clearest points are its proposed blend of large language model capabilities and AlphaGo techniques, its planned range of sizes, and the absence of an announced release date.