Google’s DeepMind has a long-term goal of artificial general intelligence

When DeepMind, a subsidiary of Alphabet, started over a decade ago, solving some of the most pressing research questions and problems with AI wasn’t at the top of the company’s mind.

Instead, the company started AI research with computer games. Every score and every win was a measuring stick of success, and a sign that DeepMind’s AI was heading in the right direction.

Colin Murdoch

“Five years ago, we conquered the game of Go. It was a great moment,” Colin Murdoch, the business manager, said during a fireside chat Tuesday at the AI Hardware Summit held in Santa Clara, California.

– Advertising –

Many years later, the gaming experience has now evolved into a much more ambitious AI effort to solve massive problems in areas such as protein folding, nuclear fusion, and quantum chemistry.

The most notable DeepMind research project is AlphaFold, which can predict the 3D structures of over 200 million known proteins. Protein folding is fundamental to the drug discovery process, and DeepMind’s AI has been used in Covid-19 vaccine and drug research.

“That means we’ve gone from years to…reporting.” We can now fold a protein in minutes,” Murdoch told the audience.

Murdoch also talked about DeepMind’s AI in nuclear fusion reactors. When building such reactors, the plasma – which is very hot and volatile – must be controlled. Researchers have worked for a number of years to control plasma, but DeepMind has been able to use its AI research to control nuclear fusion and plasma.

In its quest to solve complex scientific problems, DeepMind has not forgotten everyday problems. Murdoch said DeepMind researchers have developed technology to optimize the battery life of Android smartphones, which is a feature requested by smartphone users.

DeepMind’s priority is much higher than fixing smartphone or data center problems, however – the organization has a long-term goal of creating an “artificial general intelligence” system, which is more like a computer system. general-purpose AI capable of performing routine human tasks. For example, robots with an AGI will be able to perform routine tasks performed by humans.

“With artificial general intelligence, you can play chess, tic tac toe. He could write an essay. You can answer questions. He can do more things than we can do as humans,” Murdoch said.

Murdoch acknowledged that the company was trying to recreate the functions of a human brain, but clarified that the idea was not to create sentient AI, which has been a controversial topic lately. A Google engineer earlier this year claimed that an AI chatbot had gained sensitivity.

DeepMind’s efforts in areas such as protein folding and nuclear fusion solve specific problems and fall outside the general problem-solving functionality of the brain. But Murdoch said natural language processing — where you can get coherent answers by talking to computers in full or partial sentences — was more challenging for general AI intelligence.

Large language models are able to do a better job completing things like filling out emails, summarizing transcripts and writing code, which are everyday human tasks, Murdoch said.

Google TPU v4 OK datacenter May
Part of one of Google’s Cloud TPU v4 Pods (courtesy Google)

“What we’ve found is that these models are often able to do things that we haven’t heard of yet,” Murdoch said.

Natural language AI models are becoming complex, approaching nearly 1 trillion parameters and expected to increase further. While this makes training AI models more accurate, it also requires more computational resources.

DeepMind is creating what Murdoch called a “model factory” of AI, where large models of AI can be turned into bite-sized versions depending on the task and computing needs. Some of these derived AI models may be part of DeepMind’s long-term AI plans.

Murdoch also said Google resources such as Google Cloud and TPU chips have been critical to DeepMind’s ability to run research projects.

“We have a diverse line of hardware systems to match this wide range of research programs,” Murdoch said.

Google’s TPUs – which are application-specific integrated circuits – have been particularly useful for DeepMind in training large-scale language models. DeepMind uses a variety of processors and GPUs, but Google’s hardware roadmap helps DeepMind shape its search roadmap, Murdoch said.

We would like to give thanks to the author of this article for this remarkable material

Google’s DeepMind has a long-term goal of artificial general intelligence

Find here our social media accounts as well as other pages that are related to them.