Google will let companies run Gemini models in their own data centers (cnbc.com)

Article summary

Google's cloud unit will allow companies to run its Gemini artificial intelligence models in their own data centers, starting in the third quarter. This move is targeted at clients who want to use Google's cloud technology while retaining control of their data. The service will be available through Google Distributed Cloud, and Nvidia will bring Gemini models to its Blackwell graphics processing units. This move may attract new customers, including those in the government sector, who require high security and control over their data.

Main themes

• AI models
• Cloud computing
• Data center security
• Google Gemini
• Nvidia GPUs
• Cloud infrastructure

What commenters say

• Running AI models in physical data centers may increase the risk of models being leaked or pirated, especially if companies are hacked or have poor security measures in place.
• The advantage of running AI models in-house may be short-lived, as models improve rapidly and the benefit of proprietary training data may not last long.
• Some commenters believe that filtering out low-quality data generated by large language models is a significant challenge, while others argue that it is no different from filtering out human-generated garbage.
• There are concerns that large language models trained on other large language models' output may be subject to model collapse, making them less reliable.
• The use of watermarked model weights and liability may help prevent models from being leaked, but it is unclear how effective these measures will be.
• Some commenters think that the risk of models being leaked is low, as only large enterprises will be able to run them, and these companies will be less likely to take risks with proprietary technology.