Computing power to develop artificial intelligence doesn't come cheap. While simple AI chatbots are almost free to build, training fine-tuned models on large datasets can cost millions of dollars.
To reduce the costs associated with training AI, the UK's Advanced Research and Inventions Agency (ARIA) has launched a program called Scaling Compute.
The initiative is committing £42m to finding new, more economical alternatives to the energy-intensive hardware currently used to support the explosion in computing demand due to the arrival of generative AI. Masu. We also encourage funding applicants to look to natural processes and biomaterial alternatives to silicone for inspiration.
Suraj Bramhavar, director of the program, said that while the evolution of the digital world has traditionally relied on the fact that computers have become progressively cheaper and faster over time, this is no longer the case. I am.
“This trend is facing physical limits,” Blumhaber said. in a statementadded that the current cost of training AI has far-reaching social and geopolitical implications.
“The remarkable thing about AI is that all these amazing capabilities that we've seen so far are actually dependent on a very narrow set of algorithms and an even narrower set of underlying hardware. That is,” he continued. “And they worked phenomenally well and scaled phenomenally well.”
However, Bramhavar believes that various alternatives that can accomplish the same thing remain unexplored. He said the program will take inspiration from nature and the human brain to rethink how computing and memory are separated and how chips interconnect.
What AI can be accessed globally at a fraction of the cost?
It will also look at hardware materials, potentially finding other biological alternatives to silicon semiconductors to run AI.
The institute expects this will create additional avenues for expanding computing infrastructure at lower economic and environmental costs. The goal is to reduce the cost of training AI by a factor of 1,000.
TNW Conference 2024 – Group Ticket Offer
Save up to 40% with our group offers and join Europe's leading technology festival in June.
“If successful, this program will go far beyond the limits of current computational power and efficiency… paving the way for universally accessible, secure, and innovative AI.” Managing Director Ilan Gha said. financial times newspaper.
Indeed, one of the concerns about generative AI from an equity perspective is that most of the benefits of promised efficiency and productivity gains will go to those who can afford to pay for their own underlying models. . This means many parts of the world may be left behind in the next technological leap.
ARIA is currently recruiting for: Submission of concept paper The submission deadline is March 27th. The deadline for full proposals is May 7th. However, applicants may submit a complete proposal without having participated in the concept paper round.
The Institute encourages applications from scientists and engineers in a wide range of fields, including biology, chemistry, and neuroscience. Start-ups and existing companies are also eligible for funding.
ARIA was created by an Act of Parliament and is sponsored by the Department for Science, Innovation and Technology, with total funding of £800m. It is tasked with funding early-stage, potentially breakthrough innovations through open competitions and calls. We do not retain any intellectual property rights to the work we fund, nor do we acquire any equity in the resulting spin-out.