Sep 23, 2022
Andy and Dave discuss the latest in AI news and research,
starting with a publication from the UK’s National Cyber Security
Centre, providing a set of security principles for developers
implementing machine learning models. Gartner publishes the 2022
update to its “AI Hype Cycle,” which qualitatively plots the
position of various AI efforts along the “hype cycle.” PromptBase
opens its doors, promising to provide users with better “prompts”
for text-to-image generators (such as DALL-E) to generate “optimal
images.” Researchers explore the properties of vanadium dioxide
(VO2), which demonstrates volatile memory-like behavior under
certain conditions. MetaAI announces a nascent ability to decode
speech from a person’s brain activity, without surgery (using EEG
and MEG). Unitree Robotics, a Chinese tech company, is producing
its Aliengo robotic dog, which can carry up to 11 pounds and
perform other actions. Researchers at the University of Geneva
demonstrate that transformers can build world models with fewer
samples, for example, able to generate “pixel perfect” predictions
of Pong after 120 games of training. DeepMind AI demonstrates the
ability to teach a team of agents to play soccer by controlling at
the level of joint torques and combine it with longer-term
goal-directed behavior, where the agents demonstrate jostling for
the ball and other behaviors. Researchers at Urbana-Champaign and
MIT demonstrate a Composable Diffusion model to tweak and improve
the output of text-to-image transformers. Google Research publishes
results on AudioLM, which generates “natural and coherent
continuations” given short prompts. And Michael Cohen, Marcus
Hutter, and Michael Osborne published a paper in AI Magazine,
arguing that dire predictions about the threat of advanced AI may
not have gone far enough in their warnings, offering a series of
assumptions on which their arguments depend.
https://www.cna.org/our-media/podcasts/ai-with-ai