לוח שנה משולב

Two major theories compete to explain the origin of aging. The first, proposed by Leo Szilard in 1959, attributes aging to DNA damage. The second, articulated by Robin Holliday in the 1980s, emphasizes epigenetic alterations. While both reveal plausible molecular origins of aging, they leave important puzzles unresolved. First, mutation and epimutation burdens increase linearly with age, whereas aging phenotypes exhibit strongly nonlinear behavior. Second, key aging phenotypes cannot be traced to specific genetic or epigenetic changes; instead, they emerge collectively from their cumulative effects on cellular function.In this talk, I will present a network resilience theory of aging that resolves these puzzles. Network resilience is formalized as the ability of a network to sustain its basic functions under changes in its topology and dynamical variables. Our theory conceptualizes aging as a progressive loss of network resilience as cells approach a novel critical mutation-epigenetic line. We identify two regimes of cellular stability, with young cells remaining resilient while older cells exhibit increased susceptibility. Using GTEx data and numerical simulations, we link transcriptional noise to cellular susceptibility and reproduce delayed immune activation observed in aging. Overall, our theory offers a novel perspective on aging based on resilience and critical phenomena.

Large language models often produce errors—factual inaccuracies, biases, and reasoning failures known as "hallucinations". We show that LLMs' internal representations encode rich information about truthfulness, but in surprising ways. First, truthfulness information concentrates in specific tokens, allowing a dramatic improvement in error detection compared to using other token locations. However, these detectors don't generalize across datasets, revealing that truthfulness encoding is multifaceted rather than universal. Second, internal representations can predict the types of errors a model will make, enabling targeted mitigation strategies. Finally, we uncover a striking discrepancy: LLMs sometimes internally encode correct answers while consistently generating incorrect ones. We'll also discuss follow-up work building on these findings and their implications for developing more reliable language models.

Bio:
Hadas is a Research Fellow at Harvard's Kempner Institute, where she studies the internal mechanics of large AI models to improve their robustness, safety, and reliability. She focuses on problems where scaling compute and data falls short—such as hallucinations, harmful outputs, and biases—with the broader goal of developing controllable AI systems. She completed her PhD in the Technion under the supervision of Yonatan Belinkov.

Satellite instruments, such as TROPOMI, are routinelyused to quantify tropospheric nitrogen dioxide (NO2)based on its narrowband light absorption in the UV/visible spectral range. The key limitation of suchretrievals is that they can only return the „verticalcolumn density“ (VCD), defined as the integral of theNO2 concentration profile. The profile itself, whichdescribes the vertical distribution of NO2, remainsunknown.This presentation showcases „NitroNet“, the first NO2profile retrieval for TROPOMI. NitroNet is a neuralnetwork, which was trained on synthetic NO2 profilesfrom the regional chemistry and transport model WRFChem,operated on a European domain for the month ofMay 2019. The neural network receives NO2 VCDs fromTROPOMI alongside ancillary variables (meteorology,emission data, etc.) as input, from which it estimates NO2concentration profiles.The talk covers:• an introduction to satellite remote sensing of NO2.• the theoretical underpinnings of NitroNet, how themodel was trained, and how it was validated.• practical new applications that NitroNet enables.