As AI has grown from a multitude of research projects to a handful of titanic, industry-leading models like GPT-3, the sector needs to evolve – says Dario Amodei, former vice president of research at OpenAI, who went freelance a few months ago to find a new one Start a business. Anthropic, as it is called, was founded with his sister Daniela and has set itself the goal of “creating large-scale AI systems that are controllable, interpretable and robust”.
The challenge that the Amodei siblings face is simply that while these AI models are incredibly powerful, they are not well understood. GPT-3 they worked on is an amazingly versatile speech system that can produce extremely compelling text in virtually any style and on any subject.
But let’s say you had rhyming couplets with Shakespeare and Pope as examples. How does it do it? What is it “think”? Which knob would you turn, which dial would you turn to make it more melancholy, less romantic, or to restrict its diction and lexicon in a certain way? Of course there are parameters to be changed here and there, but nobody really knows exactly how this extremely convincing linguistic sausage is made.
It’s one thing not to know when an AI model is generating poetry and another when the model is monitoring a department store for suspicious behavior or setting legal precedents for a judge about to deliver a verdict. The general rule today is that the more powerful the system, the more difficult it is to explain its actions. That’s not exactly a good trend.
“Today’s large, general-purpose systems can have significant advantages, but they can also be unpredictable, unreliable, and opaque. Our goal is to make progress on these issues, ”says the company’s self-description. “At the moment we are mainly focused on researching these targets. In the future we see many opportunities for our work to create value commercially and for the public benefit. “
The goal seems to be to incorporate security principles into the existing priority system of AI development, which generally promotes efficiency and performance. As in any other industry, it is easier and more effective to integrate something from the beginning than to screw it in at the end. Trying to take apart and understand some of the greatest models out there may be more work than building them at all. Anthropic seems to be starting over.
“Anthropic’s goal is to make the fundamental advances in research that will allow us to build more powerful, generalized, and reliable AI systems and then deploy those systems in a way that benefits people,” said Dario Amodei, CEO of the new company shortly after the company was announced and its $ 124 million funding.
Incidentally, this funding is as high-quality as you might expect. It was led by Skype co-founder Jaan Tallinn and included James McClave, Dustin Moskovitz, Eric Schmidt and the Center for Emerging Risk Research, among others.
The company is a not-for-profit, and for now, as the limited information on the site suggests, the plan is to explore these fundamental questions of how to make large models more manageable and interpretable. We may expect more information later this year as the mission and team grow together and produce initial results.
Incidentally, the name borders on anthropocentric and relates to the relevance for human experience or existence. Perhaps it derives from the “anthropic principle,” the notion that intelligent life in the universe is possible because … well, we are here. If intelligence is inevitable under the right conditions, all that is required is for the company to create those conditions.