Lifeboat Foundation: Safeguarding Humanity
Toggle light / dark theme

Blog

Category: information science

Paul Vitányi

Consider teaching a computer how to read by giving it billions of books. You don’t teach it grammar rules or logic; you simply ask it to play a game: “Look at these words, and guess what word comes next.” To win this game at a world-class level, the computer can’t just memorize phrases. It has to start figuring out how the world works. If it’s reading a mystery novel, it needs to deduce who the killer is to guess the final sentence. If it’s reading a math textbook, it has to understand addition to predict the answer to a problem. This is the core idea explored in a recent scientific paper titled “Algorithmic Compression via Pretrained Neural Networks.”*The researchers look under the hood of today’s Large Language Models (LLMs)—like the AI assistants we use every day—to explain a fascinating mystery: Why does a machine trained merely to predict the next word end up looking like it can think, reason, and solve complex problems? Think about how a ZIP file works on your computer. If you have a massive text file filled with the word “apple” repeated a million times, a compression program won’t save all million words. It will compress it into a short rule: “Repeat ‘apple’ 1,000,000 times.” It turns a massive mountain of data into a tiny, elegant recipe. (learning how to learn). Because the AI is fed a massive, diverse diet of information, it can’t just memorize everything. Instead, it is forced to find the underlying “recipes” or rules behind the data it sees. When you type a prompt into an AI, it doesn’t just look up an answer in a database. It looks at your text, infers the “generative algorithm” (the underlying pattern or logic of what you are asking), and uses that pattern to compress the problem and generate the correct response. In essence, it deduces the hidden rules of the game on the fly. * Discover Complex Logic: When given a sequence of chess moves, the AI doesn’t just guess random moves; it actually reconstructs the abstract rules and evaluations of a chessboard in its digital “mind.” While this framework helps explain why AI is getting so smart, it also opens up big new questions. We know these models are compressing data and finding rules, but we still don’t fully understand the absolute limits of this approach. How close can a practical AI get to that theoretical “perfect” intelligence? What happens when the AI runs out of human data to learn from?

Vitányi was appointed professor of computer science at the University of Amsterdam, and researcher at the National Research Institute for Mathematics and Computer Science in the Netherlands (CWI, initially Mathematical Centre [MC]) where he is currently a CWI Fellow. He was guest professor at the University of Copenhagen in 1978; research associate at the Massachusetts Institute of Technology in 1985/1986; Gaikoku-Jin Kenkyuin (councilor professor) at INCOCSAT at the Tokyo Institute of Technology in 1998; visiting professor at Boston University in 2004, at Monash University in 1996 and at the National ICT of Australia NICTA at University of New South Wales in 2004/2005; visiting professor at and adjunct professor of computer science at the University of Waterloo from 2005.

Scientists discover strange “narwhal” waves that trap light beyond known limits

Physicists at Peking University have uncovered a new way to confine light far beyond conventional limits — without relying on metals and their inherent energy dissipation. By formulating the singular dispersion equation, the team discovered narwhal-shaped wavefunctions that trap light at deep-subwavelength volumes in purely dielectric materials. The advance, dubbed singulonics, could pave the way for ultra-efficient photonic chips, new quantum technologies, and imaging tools with unprecedented resolution.

Smartphones may soon be able to track hidden objects using LiDAR

Modern smartphones are packed with incredible technology, from high-resolution cameras and advanced graphics chips to AI processors. In premium models, this hardware includes LiDAR (light detection and ranging), which helps power augmented reality features and improve depth sensing.

And that capability could soon be in for a seriously impressive upgrade. Researchers at the Massachusetts Institute of Technology (MIT) have developed an algorithm that lets a phone’s LiDAR sensor detect objects hidden around corners. Details are in a paper published in the journal Nature.

Typically, this type of non-line-of-sight (NLOS) capability is found in labs and relies on bulky, expensive research-grade hardware. But the team’s breakthrough makes it possible for consumer LiDAR sensors to peek behind obstacles.

AI-powered stretchable computing patch can run algorithms directly on the body

A new skin-like computing patch developed at the University of Chicago Pritzker School of Molecular Engineering (UChicago PME) can analyze health data using artificial intelligence in an unprecedented way. Unlike today’s wearable devices, it carries out its AI computations directly on the body, in mere milliseconds, without relying on a wireless connection.

While your current smartwatch may be able to track your heart rate or movements, it doesn’t analyze what it finds. The analysis happens elsewhere, after it shuttles data to an external server. In some situations—detecting ventricular fibrillation in the heart, for instance—that few-seconds lag to communicate with the server is too long.

The new device, designed and tested in collaboration with researchers at Argonne National Laboratory, was made possible by the development of manufacturing processes that allow organic electrochemical transistors to be printed onto flexible surfaces.

AI atlas reveals hidden whole-body-damage caused by obesity

Obesity affects far more than metabolism and fat storage. It alters immune activity, nerve structure, and tissue organization across multiple organ systems, increasing the risk of diseases including type 2 diabetes, cardiovascular disease, stroke, neuropathy and cancer. Yet despite these systemic effects, researchers have lacked tools capable of studying disease-associated changes across the entire body in intact organisms and at high resolution.

A team led by Prof. Ali Ertürk, Director of the Institute for Biological Intelligence (iBIO) at Helmholtz Munich and Professor at the LMU, has now developed MouseMapper, a suite of foundation-model-based deep-learning algorithms designed to analyze whole-body biological imaging data. The framework automatically segments 31 organs and tissue types while quantitatively mapping nerves and immune cells throughout the body, enabling comprehensive multi-system analysis in intact mice.

“MouseMapper is built on a foundation model, which means it generalizes far beyond the data it was originally trained on,” says Ying Chen, co-first author of the study published in Nature.

Adverse impact of acute Toxoplasma gondii infection on human spermatozoa

Ultimately, QIML proves that we don’t need a fully fault-tolerant quantum computer to see results. By using quantum processors to learn the complex “rules” of chaos, we can give classical computers the boost they need to make reliable, long-term predictions about the most turbulent environments in the natural world.

Modeling high-dimensional dynamical systems remains one of the most persistent challenges in computational science. Partial differential equations (PDEs) provide the mathematical backbone for describing a wide range of nonlinear, spatiotemporal processes across scientific and engineering domains (1–3). However, high-dimensional systems are notoriously sensitive to initial conditions and the floating-point numbers used to compute them (4–7), making it highly challenging to extract stable, predictive models from data. Modern machine learning (ML) techniques often struggle in this regime: While they may fit short-term trajectories, they fail to learn the invariant statistical properties that govern long-term system behavior. These challenges are compounded in high-dimensional settings, where data are highly nonlinear and contain complex multiscale spatiotemporal correlations.

ML has seen transformative success in domains such as large language models (8, 9), computer vision (10, 11), and weather forecasting (12–15), and it is increasingly being adopted in scientific disciplines under the umbrella of scientific ML (16). In fluid mechanics, in particular, ML has been used to model complex flow phenomena, including wall modeling (17, 18), subgrid-scale turbulence (19, 20), and direct flow field generation (21, 22). Physics-informed neural networks (23, 24) attempt to inject domain knowledge into the learning process, yet even these models struggle with the long-term stability and generalization issues that high-dimensional dynamical systems demand. To address this, generative models such as generative adversarial networks (25) and operator-learning architectures such as DeepONet (26) and Fourier neural operators (FNO) (27) have been proposed. While neural operators offer discretization invariance and strong representational power for PDE-based systems, they still suffer from error accumulation and prediction divergence over long horizons, particularly in turbulent and other chaotic regimes (28, 29). Recent work, such as DySLIM (30), enhances stability by leveraging invariant statistical measures. However, these methods depend on estimating such measures from trajectory samples, which can be computationally intensive and inaccurate in all forms of chaotic systems, especially in high-dimensional cases. These limitations have prompted exploration into alternative computational paradigms. Quantum machine learning (QML) has emerged as a possible candidate due to its ability to represent and manipulate high-dimensional probability distributions in Hilbert space (31). Quantum circuits can exploit entanglement and interference to express rich, nonlocal statistical dependencies using fewer parameters than their promising counterparts, which makes them well suited for capturing invariant measures in high-dimensional dynamical systems, where long-range correlations and multimodal distributions frequently arise (32). QML and quantum-inspired ML have already demonstrated potential in fields such as quantum chemistry (33, 34), combinatorial optimization (35, 36), and generative modeling (37, 38). However, the field is constrained on two fronts: Fully quantum approaches are limited by noisy intermediate-scale quantum (NISQ) hardware noise and scalability (39), while quantum-inspired algorithms, being classical simulations, cannot natively leverage crucial quantum effects such as entanglement to efficiently represent the complex, nonlocal correlations found in such systems. These challenges limit the standalone utility of QML in scientific applications today. Instead, hybrid quantum-classical models provide a promising compromise, where quantum submodules work together with classical learning pipelines to improve expressivity, data efficiency, and physical fidelity. In quantum chemistry, this hybrid paradigm has proven feasible, notably through quantum mechanical/molecular mechanical coupling (40, 41), where classical force fields are augmented with quantum corrections. Within such frameworks, techniques such as quantum-selected configuration interaction (42) have been used to enhance accuracy while keeping the quantum resource requirements tractable. In the broader landscape of quantum computational fluid dynamics, progress has been made toward developing full quantum solvers for nonlinear PDEs. Recent works by Liu et al. (43) and Sanavio et al. (44, 45) have successfully applied Carleman linearization to the lattice Boltzmann equation, offering a promising pathway for simulating fluid flows at moderate Reynolds numbers. These approaches, typically using algorithms such as Harrow-Hassidim-Lloyd (HHL) (46), promise exponential speedups but generally necessitate deep circuits and fault-tolerant hardware.

Quantum-enhanced machine learning (QEML) combines the representational richness of quantum models with the scalability of classical learning. By leveraging uniquely quantum properties such as superposition and entanglement, QEML can explore richer feature spaces and capture complex correlations that are challenging for purely classical models. Recent successes in quantum-enhanced drug discovery (37), where hybrid quantum-classical generative models have produced experimentally validated candidates rivaling state-of-the-art classical methods, demonstrate the practical potential of QEML even before full quantum advantage is achieved. Despite these strengths, practical barriers remain. QEML pipelines require repeated quantum-classical communication during training and rely on costly quantum data-embedding and measurement steps, which slow computation and limit accessibility across research institutions.

Quantum-centric supercomputing simulates 12,635-atom protein

The scale of chemistry simulations with quantum computing has increased dramatically in just the last few months. In the latest milestone for the field, researchers from Cleveland Clinic, RIKEN, and IBM used a quantum-centric supercomputing (QCSC) framework to calculate the electronic structure of a pair of large protein-ligand complexes, reaching a scale of 12,635 atoms in the largest simulation.

The molecules were T4-Lysozyme, a protein from a family of proteins involved in the immune system degradation of peptidoglycans in bacterial membranes, and Trypsin, produced in the pancreas and used in digestion. The team simulated these proteins binding to molecules they interact with in nature and immersed in a liquid water solution, at scales of 11,608 atoms and 12,635 atoms respectively. Bringing together an international team of researchers from across the United States and Japan made it possible to develop the necessary algorithm and workflow enhancements to reach this milestone.

The researchers achieved this scale just four months after modeling the 303-atom miniprotein Trp-cage using quantum computing for the first time. Today’s new result not only demonstrates a 40-fold increase in system size compared to the Trp-cage result, it represents a 210-times improvement in accuracy from previous state-of-the-art QCSC approaches in a specific step of the workflow.

Generalization Dynamics of LM Pre-training

An AI has a limited amount of “capacity” (brainpower). Early in training, it develops quick, shallow circuits to memorize data because that’s the easiest way to get the right answer. Later, it develops complex circuits for actual reasoning. Because space is limited, these two internal systems are constantly competing for control. Whichever type of data the AI happens to be reading in a specific moment determines which circuit wins the battle.

People typically assume that LMs stably mature from pattern-matching parrots to generalizable intelligence during pre-training. We build a toy eval suite and show this mental model is wrong: throughout pre-training, LMs frequently and suddenly hop between parrot-like and intelligence-like modes, i.e. distinct algorithms implemented by distinct circuits. We call this mode-hopping. Across our suite, LMs can suddenly latch onto memorized or in-context patterns instead of in-context learning, use System 1 instead of System 2 thinking, pick up what sounds true instead of what is true, fail at multi-hop persona QA, out-of-context reasoning, and emergent misalignment — then just as suddenly revert and generalize. Mode-hopping is not explained by standard optimization dynamics: it is locally stable and can not be fixed by checkpoint averaging. We instead think of it as a capacity allocation problem: in a capacity-bounded model, generalizable circuits must compete with the shallow ones learned early in training, and the data in each pre-training window decides which circuits win. Our suite provides a cheap set of pre-training monitors and a new lens on generalization. Building upon our insights, we demonstrate three applications: (i) select intermediate pre-training checkpoints that strongly generalize reasoning and alignment, better than the final pre-or mid-training checkpoints, (ii) select pre-training data that controls and stabilizes generalization dynamics, and (iii) test prior generalization predictors, falsifying the monolithic belief that “simpler solutions generalize better”

Building general AI without generalization is doable but meh. We want an intelligence that learns deep, transferable structure, not a parrot that matches shallow patterns. Real generalization would unblock many today’s key open problems: data-efficient (online) learning, shortcut learning, transfer capabilities from verifiable domains (math, coding) to broader non-verifiable yet economically valuable domains, and maintain a coherent character that truly aligns with human values.

The distinction between parrots and intelligence is computational. Parrots repeat in-context patterns; intelligence infers in-context functions. Parrots encode a persona as bags of disconnected facts and traits; intelligence learns a shared persona representation that connects all. Parrots memorize reasoning steps; intelligence forms general reasoning circuits for entity tracking, backtracking, or even for highly abstract concepts like truth.

/* */