This week - DeepMind's AI learns simple physics like a baby; autonomous cars jammed San Francisco; the coming AI chip boom; and more!
More Than A Human
After over 30 years of not being able to feed himself, a paralysed man, with help from mind-controlled robotic arms, was able to eat a cake. What is interesting about these arms is that they are semi-autonomous which makes controlling them much easier.
Last week, I shared with you an article arguing why BLOOM Is the Most Important AI Model of the Decade. This article from MIT Technology Review focuses more on the differences between BLOOM and other LLMs, and its potential impact on AI research. “One new large language model is not going to change the course of history,” says Teven Le Scao, a researcher at Hugging Face who co-led BLOOM's training. “But having one good open language model that people can actually do research on has a strong long-term impact.”
Researchers from DeepMind have created an AI that learns simple physics just like a baby. They trained the AI using hours of videos showing rolling down a slope or two balls bouncing off each other to predict how these objects should behave. The AI also can be "surprised" when something impossible happens.
In this video, Asianometry explains how in pursuit of more powerful and efficient AI systems, companies are looking at using AI accelerators - specialised chips designed just to do AI computations. The AI accelerator hardware market is estimated to be worth over $35 billion. Venture capitalists poured out nearly $2 billion for AI chip startups in 2021 and TSMC considers AI accelerator hardware as one of their top secular drivers in revenue for the near future.
AI researchers asked themselves a question: is machine learning better equipped than humans to create a society that divides resources more equitably? The answer, according to researchers at Google’s DeepMind, seems to be yes.
A year ago, autonomous racing drones beat the best human pilots. The drone pilots then requested a fair rematch, meaning all the vision processing and decision-making has to be done on the drone without any fancy motion-capture systems. The result - drones won again but not as decisively as last time.
Nearly early 60 vehicles equipped with Cruise autonomous driving system suddenly stopped working in San Francisco because the cars could not reach Cruise's servers. As many as 20 cars, some of them halted in crosswalks, created a jam in the city’s downtown.
In a step toward robots that can learn on the fly like humans do, a new approach expands training data sets for robots that work with soft objects like ropes and fabrics, or in cluttered environments. It could cut learning time for new materials and environments down to a few hours rather than a week or two. In simulations, the expanded training data set improved the success rate of a robot looping a rope around an engine block by more than 40% and nearly doubled the successes of a physical robot for a similar task.
Researchers from Japan have found a way to clone rats from their somatic cells (the cells that make up our bodies) rather than sperm or eggs. Although the success rate of this new method is roughly 5% at most, this new cloning method can make biopreservation easier and potentially revive genetic variations of near-extinct species.
The University of Copenhagen researchers utilized a mouse model to discover an alternate path that some cells follow to build organs and used that information to exploit a new kind of stem cells as a possible supply of organs in a dish.