My fifteenth installment of interesting research papers that I have read over the past few weeks and would like to share with my community.
Category: futurism
Next-Latent Prediction Transformers Learn Compact World Models
View recent discussion. Abstract: Transformers replace recurrence with a memory that grows with sequence length and self-attention that enables ad-hoc lookups over past tokens. Consequently, they lack an inherent incentive to compress history into compact latent states with consistent transition rules. This often leads to learning solutions that generalize poorly. We introduce Next-Latent Prediction (NextLat), which extends standard next-token training with self-supervised predictions in the latent space. Specifically, NextLat trains a transformer to learn latent representations that are predictive of its next latent state given the next token. Theoretically, we show that these latents provably converge towards belief states, compressed information about the history necessary to predict the future.
How oxygen sneaks into a corked wine bottle long before the first pour
The main reason for sealing wine bottles with a cork is to protect the liquid from oxygen. However, it is not an impermeable barrier, and a small amount of air leaks in, which is not always entirely bad news. The gas helps the wine mature and develop a more complex flavor.
In a paper published in the journal Science Advances, researchers highlight several mechanisms that control how oxygen enters and behaves inside the bottle.
Oxygen is, in fact, a crucial consideration. Too much, and the wine can oxidize and spoil, while too little can stunt its development and lead to unpleasant aromas. So winemakers have a delicate balancing act to get it right.
Light-controlled microgripper bridges the gap between precision and force
For some time, researchers have used optical tweezers to manipulate tiny objects with incredible precision, using carefully controlled beams of laser light. So far, however, this technique has always come with strict limits on how much force it can exert.
Through new research published in Nature, a team led by Dong Wu at Anhui University, China, has unveiled an improved design: a miniature mechanical gripper controlled by light signals through an optical fiber.
Combining the precision of light-based tools with the gripping strength of mechanical devices, the device could make it far easier for researchers to manipulate and assemble objects at the microscale.
Bio-Computing: Making Computers with Human Neurons (Interview with FinalSpark)
Are human brain cells the computer chips of the future? Three companies are competing to make biocomputers out of human neurons, and using them to play Doom, detect explosive molecules, and more.
This is a conversation with Dr. Fred Jordan and Dr. Ewelina Kurtys of FinalSpark. We talk about the differences between AI’s \.