Lifeboat Foundation: Safeguarding Humanity
Toggle light / dark theme

Blog

Category: computing

LeWorldModel: Stable End-to-End Joint-Embedding

Joint Embedding Predictive Architectures (JEPAs) offer a compelling framework for learning world models in compact latent spaces, yet existing methods remain fragile, relying on complex multi-term losses, exponential moving averages, pre-trained encoders, or auxiliary supervision to avoid representation collapse. In this work, we introduce LeWorldModel (LeWM), the first JEPA that trains stably end-to-end from raw pixels using only two loss terms: a next-embedding prediction loss and a regularizer enforcing Gaussian-distributed latent embeddings. This reduces tunable loss hyperparameters from six to one compared to the only existing end-to-end alternative. With ~15M parameters trainable on a single GPU in a few hours, LeWM plans up to 48× faster than foundation-model-based world models while remaining competitive across diverse 2D and 3D control tasks. Beyond control, we show that LeWM’s latent space encodes meaningful physical structure through probing of physical quantities. Surprise evaluation confirms that the model reliably detects physically implausible events.

TL;DR: LeWM is a JEPA-based world model that avoids representation collapse using a simple Gaussian regularizer (SIGReg), trains end-to-end from pixels with only two loss terms, and achieves competitive control performance at a fraction of the compute cost.

Catch-bond engineering “turbocharge” T cells to attack prostate cancer

T cells are a powerful weapon in the fight against cancer, forming the basis of treatments such as CAR-T cell therapy and checkpoint inhibitors. This research centers on another type of immunotherapy approach called T cell receptor (TCR) therapy, which engineers T cells to recognize specific proteins on cancer cells, allowing for highly targeted attacks.

Many of these proteins, however, are “self-antigens,” or molecules normally found in the body. To prevent these T cells from attacking healthy tissue, the immune system naturally eliminates the strongest cancer-fighting T cells during development. This leaves behind weaker T cell receptors that may struggle to recognize and destroy tumors, particularly those that have learned to evade immune defenses.

To overcome this challenge, researchers focused on fine-tuning naturally occurring T cell receptors to strengthen their ability to recognize a common prostate cancer protein called prostatic acid phosphatase (PAP), which is commonly expressed on prostate tissue and prostate tumors. The team identified a naturally weak TCR, known as TCR156, that could detect PAP but was not strong enough to effectively kill cancer cells.

Using a novel technique called catch bond engineering, a concept developed by the Lab, the researchers “turbocharged” the T cells. In the body, T cells form brief, mechanical bonds with their targets, known as catch bonds, which help them sense and respond to threats. By altering just one or two amino acids in the T cell receptor, the scientists were able to strengthen these bonds while preserving the T cells’ natural ability to recognize their specific target.

Multiple engineered versions of TCR156 were created and tested. Two candidates proved to be the most effective. These engineered T cells were analyzed for their ability to recognize tumors, release cancer-killing molecules, proliferate, and resist exhaustion. Advanced imaging, single-cell RNA sequencing, and structural analyses were used to confirm that the modifications improved T cell function while maintaining precision and avoiding off-target effects.

Structural and computer modeling studies showed that the catch bond mutations did not change the overall TCR shape but primed it to form a new interaction with PAP when the T cell engaged the tumor, explaining how the engineered T cells could remain highly specific while dramatically boosting their cancer-killing ability.

The researchers found that a single amino acid change created a catch bond hotspot that significantly enhanced T cell function. This change did not directly contact the cancer protein until the T cell engaged dynamically, demonstrating that a tiny modification can have a major effect. Most importantly, the modifications did not make the cells attack healthy tissue.

Continue reading “Catch-bond engineering “turbocharge” T cells to attack prostate cancer” | >

A new entanglement-enhanced quantum sensing scheme

Over the past decades, quantum scientists have introduced various technologies that operate leveraging quantum mechanical effects, including quantum sensors, computers and memory devices. Most of these technologies leverage entanglement, a quantum phenomenon via which two or more particles become intrinsically linked and share a unified quantum state, irrespective of the distance between them.

New X-ray vision for electronics lets scientists monitor working chips remotely

A team of international researchers have developed a breakthrough way to observe what is happening inside electronic chips while they are operating—without touching them, taking them apart, or switching them off. The new technique uses terahertz waves, a safe and non-ionizing form of electromagnetic radiation, to detect tiny movements of electrical charge inside fully packaged semiconductor devices. For the first time, this allows scientists and engineers to monitor electronic components as they function in the real world.

The study, published in the IEEE Journal of Microwaves, involves researchers from Adelaide University in Australia, US technology company Virginia Diodes Inc, the Hasso Plattner Institute and the University of Potsdam, Germany.

Adelaide University Group Leader of the Terahertz Engineering Laboratory (TEL), Professor Withawat Withayachumnankul, said that semiconductors underpin almost every modern technology, from smartphones and medical devices to vehicles, power grids and defense systems.

Frontiers: Information storage and transfer in the brain require a high computational power

Neuronal network display various local or global mechanisms to allow information storage and transfer in the brain. From synaptic to intrinsic plasticity, the rules of input–output function modulation have been well characterized in neurons. In the past years, astrocytes have been suggested to increase the computational power of the brain and we are only just starting to uncover their role in information processing. Astrocytes maintain a close bidirectional communication with neurons to modify neuronal network excitability, transmission, axonal conduction, and plasticity through various mechanisms including the release of gliotransmitters or local ion homeostasis. Astrocytes have been significantly studied in the context of long-term or short-term synaptic plasticity, but this is not the only mechanism involved in memory formation. Plasticity of intrinsic neuronal excitability also participates in memory storage through regulation of voltage-gated ion channels or axonal morphological changes. Yet, the contribution of astrocytes to these other forms of non-synaptic plasticity remains to be investigated. In this review, we summarized the recent advances on the role of astrocytes in different forms of plasticity and discuss new directions and ideas to be explored regarding astrocytes-neuronal communication and regulation of plasticity.

The rules governing changes in synaptic and intrinsic plasticity are diverse and complex, sometimes synergistic and sometimes not (Debanne et al., 2019). Most studies have been neuro-centric, despite growing evidence that astrocytes can intervene or interact to modify or modulate synaptic transmission (Araque et al., 1998; Jourdain et al., 2007; Bonansco et al., 2011), input integration, neuronal excitability (Tan et al., 2017), spike waveform or axonal conductivity (Sasaki et al., 2011; Lezmy et al., 2021). Astrocytes can detect neuronal activity, and depending on the firing rate of action potentials (APs), they can not only release gliotransmitters such as adenosine or glutamate (Hamilton et al., 2008; Lezmy et al., 2021), but also trigger intracellular calcium ([Ca2+]i) oscillations at different frequencies (Pasti et al., 1997).

No exotic physics needed: A new formation mechanism of skyrmions inside magnets

Skyrmions, in which electron spins inside a magnet are arranged like vortices, are a key structure in next-generation spintronics technology. KAIST researchers have shown that skyrmions can form using only the fundamental physical interactions within magnets, without requiring special physical conditions.

This finding, published in the journal Physical Review Letters, expands the possibility of realizing skyrmions in a wide range of magnetic materials and suggests new potential for developing next-generation ultra-low-power information devices with data storage densities tens to hundreds of times higher than current technologies.

A research team led by Professor Se Kwon Kim from the Department of Physics has proposed a new theoretical framework showing that vortex-like magnetic structures can naturally emerge solely through magnetoelastic coupling —the interaction between magnetism and lattice structure.

‘Mini earthquakes’ turn tiny chips into radio signal powerhouses

From GPS satellites to mobile networks, modern technology relies on ultra-precise radio signals. Engineers have long tried to generate them on chips using interactions between light and sound, but the effect was too weak. University of Twente researchers now show in a paper published in Nature Photonics that a thin glass layer creates “mini-earthquake” surface acoustic waves, which make the effect more than 200 times stronger. This enables ultra-pure signals and record-sharp filters on a device thousands of times smaller.

Every time you make a phone call, your signal is filtered out of a crowded radio spectrum using radio frequency filters. These components let through only the frequencies you want and block everything else. The sharper the filter, the cleaner the call. The same principle applies in radar, satellite navigation and future wireless networks like 6G.

Topology helps build more robust photonic networks

Penn-led researchers have shown for the first time that multiple, information-carrying light signals can be safely guided through chip-based, reconfigurable networks using topology, the esoteric branch of mathematics that says donuts and mugs are identical. Because topological properties remain stable even when objects are deformed—hence the field equating mugs and donuts, since both have one opening—the advance could help make light-based technologies for computing and communications more powerful and reliable.

“We already knew how to guide light using topology,” says Liang Feng, Professor in Materials Science and Engineering (MSE) with a secondary appointment in Electrical and Systems Engineering (ESE) within Penn Engineering and senior author of a study in Nature Physics describing the result. “But we had never been able to guide multiple, concurrent signals before.”

That opens the door to building networks of chips that communicate using light while taking advantage of the robustness topology provides. “Signals guided by these principles can be extremely reliable,” says Feng. “It’s like building a highway for light where even large potholes have no effect on traffic—it’s as if the defects simply aren’t there.”

/* */