This week - edible drone; an implant tapping directly into the human speech system; Meta's new protein-predicting AI; and more!
More Than A Human
Edward Chang shares his research into a neuroprosthetic that enables paralysed patients to communicate by thinking about what they want to say. The implant taps directly into the speech system promising quicker and more natural communication (Shoutout to Alexandra for sharing this article with me).
Can you build a prosthetic arm with Lego? As David Aguilar shows, yes, you can.
Machine learning could help develop new types of metals with useful properties, such as resistance to extreme temperatures and rust, according to new research. Instead of a process of trial and error, the new approach uses AI to find promising combinations of metals which then can be tested in a lab by humans.
Researchers from Google Research created an AI that can predict how a chemical will smell like based only on its structure. This opens the possibility of digitising smells and be a foundation for a new breed of AIs that can synthesise any scent.
Computers are good at math until you give them word problems to solve. AI researchers have learned the hard way that training language models to solve math problems is hard. However, a new AI from Google named Minerva has made enormous progress, reaching 50 to 80% accuracy while other AIs got 7-20%.
Not every AI application needs a full data center to crunch numbers. Sometimes you have to fit a neural network on a small device and make it as energy efficient as possible. In this video, Asianometry explains how engineers approach the problem of edge AI, from both a software perspective (like shrinking and optimizing neural nets) and from a hardware perspective.
Swiss engineers present a drone whose wings are made from rice cakes. It might sound ridiculous but the idea here is that in case of emergency, these drones can deliver life-saving supplies and then you can eat them. "The edible wing tastes like a crunchy rice crisp cookie with a little touch of raw gelatin (which worked as an edible glue to hold the rice cookies as a flat plate shape). No artificial flavor has been added yet", said the lead researcher.
There is a new kid on the protein-predicting AIs scene. It comes from Meta and it is named ESMFold. Meta's researchers admitted their AI isn’t quite as accurate as AlphaFold but it is about 60% faster for short sequences than DeepMind's AlphaFold.