Google X trends in generative AI: teaching computers to learn

Google X trends in generative AI: teaching computers to learn

Market news |
By Nick Flaherty

Cette publication existe aussi en Français

Generative AI such as ChatGPT and Bard have seen a huge interest in AI, but it is the real world application and the scaffolding around the AI that is more important, says David Andre, chief scientific officer at Google X.

“The role of AI is evolving incredibly quickly and its easy to feel overwhelmed. It might cause us to rethink our place in the world, The question is now to do this and avoid the pitfalls,” said Andre at the Leti Innovation Days this week in Grenoble, France.

Google X is the moonshot division of Alphabet, developing ideas with a horizon of decades rather than years. “We create seed crystals for moonshots so that when you apply more resources to create sustainable businesses or technologies Google.” He points to driverless car company Waymo and drone delivery company Wing as examples of where Google X thinking has led to real world developments.

Project Mineral is working on sustainable agriculture. “We use machine vision algorithms in recognising the state of health of each of the plants.”

He points to the Chorus moonshot that is developing new sensor technology, software and machine learning tools to radically improve our real-time understanding of where physical goods are, what state they are in, and how they are used. This used reinforcement learning to support the delivery of 11m vaccines in New Zealand during Covid with a sensor in each freezer box.

“Another question we asked was what if a computer could teach itself? This team moved to Google became Google Brain, and developed Google Translate, the transformers that are used in most AI systems today and in Google’s Bard AI.”

Generative AI for EDA

“Generative AI has to be coupled with the real world and this is what we are really good at at Google X. The challenge is figuring out where the AI and ML goes in the real world applications,” he said. “All of this scaffolding is often more valuable than the AI particularly for large language models.”

“One of the key requirements is very powerful simulation,” he said. “Waymo got to where it is was by using 20bn miles of simulation and that allows them to drive in three cities with over 1m miles of journeys without a driver in the front seat.”

Teaching code to write itself is key to this.

“If you can write code faster you can write simulations faster,” he said, and that has all kinds of benefits. For example it can be used for synthesis to shrink chip designs, combining neural net relaxation with simulated annealing.

“We at X remain incredibly interested in cases that combine AI with real world scaffolding. That’s where we think the future is.”





Linked Articles