Google’s Bet on Synthetic Data; AI Copilots Reach Finance; Another Nvidia Blowout

Google’s artificial intelligence group is having a moment.

After getting its flagship Gemini large language model off the ground and roughly on par with OpenAI, Google on Wednesday appeared to take a sudden lead in a parallel race to develop open-source LLMs, which Meta Platforms has been dominating. In the process, Google also gave a shot in the arm to a branch of AI model science that’s been bubbling for the past 12 months: using computer-generated text and imagery to train new models, in addition to using data from the real world.

In launching Gemma, a general purpose, open-source LLM that beats out Meta and Mistral’s LLMs based on Hugging Face’s leaderboards, Google may be accomplishing several goals.

First, Google could theoretically upsell users of Gemma to its proprietary Gemini LLM, as the two are related. Even if it doesn’t tempt them to pay for Gemini, any technical advances or fixes open-source developers make to Gemma might help Google improve Gemini. Google also will appease some of its coveted AI researchers who are big fans of open-source technology and have been frustrated to see Meta jump into the lead. Gemma may also give Google brownie points in the eyes of regulators concerned it has too much power in the field by giving away some of its tech for free.