This week - GPT-3 vs human brain; turning retired F-16s into drones; AI is also struggling to adjust to 2020; how Excel messes up with geneticist; and more!
In this video, Lex Fridman asks a question, how much would it cost to train OpenAI's GPT-3 model with 175 billion parameters to match the scale of a human brain (100 trillion synapses). Right now, the answer is a lot. But if the efficiency of training AI models continue to double every 16 months, in 12 years we could be able to scale GPT-3 to human brain level for the same cost as training GPT-3.
Even AIs can't understand 2020. The sudden change in our behaviour confuses many AI models which, like humans, were not prepared for all what happened. This article focuses on one example of such confusion - face recognition AIs can't recognise faces of people wearing masks.
Researchers studying spiking neural networks have found that network simulations became unstable after continuous periods of unsupervised learning. When they exposed the networks to states that are analogous to the waves that living brains experience during sleep, stability was restored. So maybe androids will dream of electric sheep after all.
Ben Goertzel shares his thoughts on GPT-3. He states that GPT-3 has "some truly novel and exciting capabilities for language processing and related tasks" but it should not be hailed as artificial general intelligence (AGI) and even goes so far as declaring GPT-3 to be a dead-end path towards AGI.
Here is a picture from National Geographic's Instagram showing the conversion of retired F-16 fighter jets into drones which will be used in training exercises. Interestingly, the article describing this transformation is not about AI, robotics or drones. It's about recycling.
We live in the times where technology is advanced enough for someone in Boston to control a robotic dog in San Francisco. All we need now is a superfast internet connection because the image from Spot was far away from 4K 60fps.
A fearless and loving challenges roboticist took a robot designed to transfer movements for remote needle biopsy and instead of a needle, he attached a straight razor blade and ask a friend to give him a shave.
Scientists studying the human genome faced an unexpected problem - when you put some gene names into Excel, it converts it into a date. It might sound like a silly problem but roughly one-fifth genetic data in papers had been affected by Excel errors. Solution - rename troublesome genes because it’s easier to rewrite genetics than to update Excel.
While looking for the newest greatest cat gif on Imgur, I stumbled upon this post listing some developing technologies that moved or soon will move from science fiction to reality, like nanomedicine, implants, exoskeletons and more. Interesting read.