What’s actually inside a “task”?
Building and scaling RL environments for LLM training.
Category: robotics/AI – Page 51
There Is No Formula: Why AI Cannot Solve What Matters Most
There is no formula for love. No formula for meaning. No formula for great art, for grief, for living a life that matters.
But we keep looking for one anyway. Increasingly, we look for it in AI.
In my new essay, I argue that this is a category error with a real cost. Some problems lend themselves to calculation: fusion, protein folding, and route optimization. With enough compute, they yield. Other problems do not bend at all. They cannot be solved. They can only be lived.
When we mistake the second category for the first, we bring what I call the Hammer of AI to questions that ask for wisdom, presence, and judgment.
Then we are surprised when the hammer keeps breaking the very thing we were trying to mend.
The piece draws on Tolkien, Vaclav Havel, Carlos Castaneda, and the Japanese art of Kyudo to argue that what we actually need in the age of AI is not another formula. It is the wisdom to know when there is no formula at all.
When complexity arrives in your life, do you reach for the hammer or for something else?
Efficacy and Safety of Amifampridine in Myasthenia GravisA Randomized, Double-Blind, Placebo-Controlled Crossover Trial
Class I evidence that in patients with AChRAb+ myasthenia gravis, the addition of amifampridine to pyridostigmine was not superior to treatment with pyridostigmine alone and was associated with a higher incidence of adverse events.
This website uses a security service to protect against malicious bots. This page is displayed while the website verifies you are not a bot.
Cellular and subcellular specialization enables biology-constrained deep learning
Galloni et al. introduce “dendritic target propagation”: a Dale’s law-compliant learning algorithm for cortical microcircuits with soma-and dendrite-targeting inhibition and realistic connectivity constraints. By combining experimentally derived BTSP and Hebbian rules, dendrites compute local error proxies via E/I mismatch, supporting gradient-based deep learning during simultaneous bottom-up and top-down signaling.
Silicon oscillators solve computer problems that would take thousands of years using semiconductors
In the era of big data and artificial intelligence, a new approach has emerged for solving combinatorial optimization problems, which involves finding the most efficient solution among many possible options and can otherwise take thousands of years to compute.
A KAIST research team has developed computational hardware that can be implemented entirely using existing silicon processes, enabling deployment on existing fabrication lines without additional facilities. This is expected to enable faster and more accurate decision-making across various industries, including logistics, finance, and semiconductor design.
The research is published in Science Advances.
Scientists program materials just by spinning them
There is something universally appealing about the slap bracelet, and the way a simple tap causes it to switch between a straight shape and a curled one. What you probably didn’t know is that a slap bracelet’s satisfying snap is the same principle behind bistable structures. These can toggle between two stable positions (one representing 0 and the other 1) to store data directly within their physical forms as mechanical bits (m-bits).
Because of their exciting potential for efficient control of robotic and other mechanical systems, researchers have been engineering special materials with programmable structures (programmable metamaterials) for years. But until now, actual programming of such systems has been a major challenge: mechanical bits must typically be controlled individually, which is extremely cumbersome and time-consuming.
Now, researchers in the Flexible Structures Laboratory (fleXLab) in EPFL’s School of Engineering, the Dutch research institute AMOLF, and Leiden University have found a way to program metamaterials globally with a surprisingly simple solution: rotation. By tuning a spinning platform’s speed, direction, and acceleration, the researchers can harness forces arising in a rotating system—such as centrifugal and Euler forces—to make elastic beams snap back and forth, creating a simple new way to “write” multiple mechanical bits at once.
Zuckerberg Trying to Simulate Human Biology at the Cellular Level
Mark Zuckerberg is following a path paved by fellow billionaires Bill Gates and Warren Buffet: laundering his untold billions through a health research prestige project.
Called the Chan Zuckerberg Biohub — his wife Priscilla Chan, a pediatrician, is also involved — the foundation’s stated long-term mission is to “cure and prevent all disease through AI-powered biology, frontier research, and state-of-the-art technology.”
True to those enormous goals, the Biohub recently announced a $500 million investment into AI models of human cells, specifically, in order to “accelerate the cure and prevention of all diseases,” Euronews reported.
Temporal superposition and feature geometry of RNNs under memory demands
Abstract: Understanding how populations of neurons represent information is a central challenge across machine learning and neuroscience. Recent work in both fields has begun to characterize the representational geometry and functionality underlying complex distributed activity. For example, artificial neural networks trained on data with more features than neurons compress data by representing features non-orthogonally in so-called *superposition*. However, the effect of time (or memory), an additional capacity-constraining pressure, on underlying representational geometry in recurrent models is not well understood. Here, we study how memory demands affect representational geometry in recurrent neural networks (RNNs), introducing the concept of temporal superposition. We develop a theoretical framework in RNNs with linear recurrence trained on a delayed serial recall task to better understand how properties of the data, task demands, and network dimensionality lead to different representational strategies, and show that these insights generalize to nonlinear RNNs. Through this, we identify an effectively linear, dense regime and a sparse regime where RNNs utilize an interference-free space, characterized by a phase transition in the angular distribution of features and decrease in spectral radius. Finally, we analyze the interaction of spatial and temporal superposition to observe how RNNs mediate different representational tradeoffs. Overall, our work offers a mechanistic, geometric explanation of representational strategies RNNs learn, how they depend on capacity and task demands, and why.
Supplementary Material: zip
Primary Area: interpretability and explainable AI.
Artificial intelligence accelerates discovery of next-generation disinfectants
Chemists and computer scientists tapped AI to find new disinfectants to combat the growing threat of dangerous “superbugs.”
The Journal of Chemical Information and Modeling published their computational-experimental framework for developing quaternary ammonium compounds, or QACs, to kill bacteria.
The method yielded 11 new QACs that show activity against antimicrobial-resistant bacteria.