A UK research and development unit today unveiled a futuristic vision of “quantitative safety guarantees” for AI.
The Advanced Research and Invention Agency (ARIA) compares the guarantees to the high safety standards in nuclear energy and passenger aviation. In machine learning, the standards include a probabilistic guarantee that no harm will result from a given action.
At the heart of the ARIA plan is a “gatekeeper” AI. This digital guard takes care of that Other AI agents only work within the guidelines set for a specific application.
ARIA will provide £59m for the scheme. By the end of the program, the agency intends to demonstrate a scalable proof-of-concept in one area. include suggestions Power grid balancing and supply chain management.
If effective, the project could protect sophisticated AI applications, such as improving critical infrastructure or optimizing clinical trials.
The <3 of EU technology
The latest rumors from the EU tech scene, a story from our wise old founder Boris and questionable AI art. It's free in your inbox every week. Join Now!
The program is the brainchild of David “Davidad” Dalrymple, who co-invented the popular cryptocurrency Filecoin.
Dalrymple has also studied AI technical security extensively, which sparked his interest in the gatekeeper approach. As program director of ARIA, he can now put his theory into practice.
The Gatekeeper Guarantee
ARIA's gatekeeper will rely on scientific models of the world and mathematical evidence. Dalrymple said the concept combines commercial and academic concepts.
“The approaches being explored by major AI companies are based on finite samples and provide no guarantees about how AI systems will behave when deployed,” he told TNW via email.
“Meanwhile, if we focus too heavily on academic approaches such as formal logic, we risk effectively trying to build AI capabilities from the ground up.”
“The gatekeeper approach gives us the best of both worlds, matching the limit capabilities as an engine for fast driving, but on the rails of mathematical reasoning.”
This fusion requires deep interdisciplinarity Collaboration – this is where ARIA comes into play.
The British DARPA?
ARIA was founded last year and funds “high-risk, high-reward” research. The strategy has drawn comparisons to DARPA The Pentagon’s “Mad Science” unit.
Dalrymple drew another parallel with DARPA. He compares ARIA's new project with DARPA's HACMS program, which created an unhackable quadcopter. The project has proven that formal verification can create bug-free software.
“Vulnerabilities can be eliminated, but only with assumptions about the scope and speed of intrusion an attacker can make into the physical embodiment of a system,” Dalrymple said.
His plan is built on an approach supported by Yoshua Bengio, a renowned computer scientist. Bengio has also called for a Turing Award winner “quantitative security guarantees.” But he is disappointed with the progress made so far.
“Unlike methods for building bridges, medicines, or nuclear power plants, current approaches to training frontier AI systems – the most powerful AI systems currently available – do not allow us any quantitative safety guarantees,” Bengio wrote in a blog post last year.
Dalrymple has the chance to change that. This would also be a huge boost for ARIA, which is being considered by politicians.
Some lawmakers have questioned ARIA's budget. The organization has attracted £800 million in funding over five years – a significant sum but a fraction of others state research institutions.
ARIA can also indicate potential savings. A program launched last month aims to train AI systems at 0.1% of current costs.
One of the themes of this year's TNW conference is Ren-AI-ssance: The AI-Powered Rebirth. If you want to delve deeper into all things artificial intelligence or just want to attend the event (and say hello to our editorial team), we have something special for our loyal readers. Enter code TNWXMEDIA at checkout to receive 30% off your product Business passport, Investor pass or starter packages (Bootstrap & Enlarge).
Comments are closed.