Did you miss a session from the Future of Work Summit? Head over to our Future of Work Summit on-demand library to stream.
There’s a growing need for investment in foundational AI technologies. With deep learning potentially approaching computational limits and subfields like natural language running up against intractable technical barriers, novel AI and machine learning techniques have arguably never been in higher demand.
NYU psychologist Gary Marcus, Google software engineer Francois Chollet, and Facebook head of AI Jerome Pesenti, among others, have argued that the lack of progress isn’t surprising, as researchers face challenges both algorithmic and scientific. Even the most sophisticated AI models can suffer from catastrophic forgetting, or a tendency to abruptly forget previously learned information, in addition to a lack of reproducibility, explainability, stability, and reliability.
That’s why Marek Rosa, a Slovakian entrepreneur and computer programmer, founded GoodAI, a company dedicated to the research and development of general artificial intelligence (AGI). He’s the CEO and founder of Keen Software House, an independent video game design studio with a headquarters in Prague, the capital of the Czech Republic.
Rosa founded GoodAI in 2014 with a $10 million investment, then announcing the company publicly and its first research roadmap in 2015 and 2016, respectively. In 2017, he founded the General AI Challenge, pledging $5 million in prize money to tackle critical research problems in “human-level” AI development.
GoodAI now employs around 20 researchers and engineers. Its newest endeavor is the GoodAI Grants Initiative, which aims to fund efforts in areas like curiosity and continual learning. To date, the GoodAI Grants Initiative has awarded over $650,000 — all from Rosa — to nine projects that GoodAI considers a part of its roadmap to general AI.
“What makes us different [from other grant organizations] is our openness and flexibility and our willingness to work with potential grantees in creating a fitting proposal,” GoodAI PR manager Will Millership told VentureBeat in an email interview. “We really don’t want to be limited in who we work with by bureaucracy and therefore we work with individual scientists, groups of researchers, private companies, and even individual students. We do a lot of work to make sure that all the intellectual property from the projects is shared but this doesn’t necessarily mean completely open. Each agreement in place aims to respect the academic and business interests of both GoodAI and the receivers of the grants.”
GoodAI grant projects
In December 2019, Rosa and the GoodAI team published Badger, a unifying AI architecture defined by a principle GoodAI calls “modular lifelong learning.” Badger, which outlines the direction of GoodAI’s research, seeks to create a system of AI agents capable of adapting to a growing, open-ended range of tasks while remaining able to reuse knowledge acquired in previous tasks.
“Our aim is to develop safe general AI — as fast as possible — to help humanity and understand the universe,” Millership said. “We see the creation of human-level AI as the biggest challenge to mankind and a task far beyond that of an individual researcher or research group. That’s why we believe collaboration — and not competition — is the best way forward.”
Among GoodAI’s grant recipients is Deepak Pathak, an assistant professor at Carnegie Mellon University who’s taking inspiration from developmental psychology and particularly how curiosity drives human’s early developmental learning. Another is Ferran Alet, a Ph.D. student at MIT’s Computer Science and Artificial Intelligence Laboratory, who’s aiming to make an AI model that generalizes to new tasks in new environments from small amounts of data and previous experiences.
GoodAI’s ambition — AGI, or the hypothetical intelligence of a machine with the capacity to understand or learn from any task — has its detractors. Facebook chief AI scientist Yann LeCun believes that it can’t exist, because there’s no such thing as general intelligence. He argues that even human intelligence is very specialized, requiring many different systems to accomplish different individual tasks.
In something of a rebuttal to this, GoodAI recently released its latest research roadmap, which spotlights some of the technical challenges related to creating human-level or general AI. GoodAI asserts that AGI must “learn to learn” and engage in lifelong learning, both continuously and at a gradual cadence. It also believes that AGI should be able to engage in open-ended exploration and self-invent goals as well as generalize “out of distribution” and extrapolate to new problems.
“Each of these features reflects the ways in which humans learn throughout their lifetime and therefore we see them as key to creating AI that’s able to generalize to new problems in different environments, much like humans do,” Millership said. “We [plan to] work closely with the grantees during their projects, offering support if they need it, and [put] on a seminar in the summer, where all grantees can share their ideas and projects. We’re trying to create an international community of researchers crossing the boundaries of academia and industry.”
Despite recent breakthroughs in solving barriers to AGI, it’s clear the road to more humanlike AI will be long and winding. However, efforts like GoodAI, along with nonprofit organizations and open communities like ContinualAI and EleutherAI, look to accelerate progress by tapping into the broader pool of AI and machine learning expertise.
Thanks for reading,
AI Staff Writer
VentureBeat's mission is to be a digital town square for technical decision-makers to gain knowledge about transformative enterprise technology and transact. Learn More