Lifeboat Foundation: Safeguarding Humanity
Toggle light / dark theme

Blog

Get the latest international news and world events from around the world.

Log in for authorized contributors

Ploutos Browser

Forking: when AI models suddenly memorize too much.

AI models are expected to learn patterns from their training data and apply what they learn to new, unseen examples. But what happens when learning too much from repeated examples actually makes a model worse at handling unfamiliar ones? https://arxiv.org/abs/2610.00394v1

This paper explores a phenomenon called forking, discovered during NanoGPT autoresearch. The researchers found that when a model repeatedly trains on the same data, its performance on the training set can continue to improve while its performance on unseen validation data suddenly deteriorates. When plotted, the two loss curves split apart at certain training boundaries, resembling the branches of a fork.

The researchers investigated this behavior in a controlled NanoGPT setup and reproduced it in a DeepSeek-style model using a memory mechanism called Engram, which stores and retrieves patterns of word sequences. Their findings suggest that repeated training strengthens the model’s predictions for familiar sequences while making it less likely to predict alternative continuations it has not encountered. Rare contexts appear to contribute significantly to this problem.

Interestingly, the team also observed forking in certain short-training scenarios involving supervised fine-tuning and reinforcement learning-like methods. However, the effect was less pronounced when training budgets were larger or when the memory tables were more crowded.

Why does this matter? The discovery introduces another unexpected behavior in deep learning, alongside phenomena such as grokking and double descent. More importantly, it highlights a potential danger in AI-driven research automation: techniques that appear to improve model performance can introduce unintended problems that are easy to overlook.

The key takeaway: Better performance on training data does not always mean a better model. As AI agents increasingly automate machine learning experiments, their discoveries and proposed optimizations still need careful validation to ensure that apparent improvements translate into genuine generalization.

Continue reading “Ploutos Browser” | >

Molecular pendulum shuttles protons to enable rapid hydrogen production in enzymes

Nearly all metabolic processes require protons—positively charged hydrogen atoms. Although these are abundant in the cellular fluid, the biocatalysts involved depend on mechanisms for efficiently capturing protons and transporting them to the site where catalysis occurs. Enzymes from microorganisms can produce hydrogen (H₂) by combining electrons with protons. These so-called hydrogenases are so efficient that one biocatalyst can capture nearly 10,000 protons per second.

A team of researchers from the Photobiotechnology group at Ruhr University Bochum, Germany, has now discovered how they do this. The researchers reported their findings in the Journal of the American Chemical Society on Oct. 17, 2026.

Alcohol in space: Exotic methanol discovered in planet-building zones

An international team of astronomers using the Atacama Large Millimeter/submillimeter Array (ALMA), of which the U.S. National Science Foundation National Radio Astronomy Observatory is a partner (NSF NRAO), has combined expertise from across the globe to open a new window on how the organic molecules that seed planets, and potentially life, form in space.

New results from the ALMA COMPASS Large Program (Complex Organic Molecules in Protostars with ALMA Spectral Surveys) show that young, still-forming stars are surrounded by a surprisingly sophisticated “chemical starter kit,” including organic molecules such as methanol that may one day enrich planets.

“We are getting completely new insights into the complex chemistry occurring around the youngest protostars,” said Jes Jørgensen, professor at the Niels Bohr Institute of the University of Copenhagen and principal investigator of the program. “The interesting question is how these chemical ingredients may influence the conditions on planets and potentially the origin of life there.”

A suite of engineered mice for interrogating psychedelic drug actions

This national research on lsd and ultra lsd called large scale docking where a digital cell is used and then automatically new compounds and discoveries can be made by the millions to find the right amount of scaffolding for the compounds to work in a non hallucinogenic way to essentially even work with schizophrenia patients but his work is not done yet as it is a very complex process.

Chiu et al. have developed engineered HTR2A mouse lines enabling in vivo mapping, targeting and modulation of receptor-expressing neurons and validated their utility across biochemical, electrophysiological and behavioral studies of psychedelic mechanisms.

The past, present and future of de novo protein design

An excellent overview/primer of the current state of de novo computational protein design with AI. The review was written by a panel of illustrious experts including Nobel Laureate David Baker, who is perhaps the most prolific and impactful protein engineer in history.

A Review of de novo protein design highlights key methodological advances and achievements, current challenges and future applications.

Bespoke library docking for 5-HT2A receptor agonists with antidepressant activity

This new compound can lead to better treatment of depression using essentially a non hallucinogenic version of lsd. This lead to essentially many discoveries including the 5-HT2A receptor in the human brain which can help make the new compounds more accurate for the new compound. This would help with more effective treatment of treatment resistant depression.

New agonists of the serotonin 5-HT2A receptor (5-HT2AR) confer high brain permeability and antidepressant activity and—in contrast to classic 5-HT2AR agonists—lack psychedelic activity.

Old Blood Cells Have Been Transformed Into Brain Stem Cells That Get Younger As They Age

Aging is the price paid for the ability of experiencing existence, one might philosophize.

Natural cellular aging begins with differentiation. Once a stem cell becomes a specific cell type, it’s locked in forever: fat cells can never become muscle cells, no matter how hard one wishes, or pokes and prods at them in front of the mirror.

Or so it goes in nature, but scientists can definitely transform cells, as if biologically compensating for all the failed alchemical attempts of the past.

Detecting the Kick of a Single Gas Molecule

An experiment resolves the tiny momentum transferred from an individual gas molecule to a levitated nanoparticle.

For two centuries, collisions between atoms or molecules and small suspended objects have been understood only statistically. In 1827, Robert Brown first described the random motion of pollen in water. Eight decades later, Albert Einstein and Marian Smoluchowski independently explained the jitter as the accumulated effect of countless unseen molecular impacts—convincingly demonstrating the physical reality of atoms and molecules. However, individual kicks were never resolved. Only their collective consequence, Brownian motion, was ever visible. Modern trapped-ion experiments succeeded in watching individual atom–ion collisions perturb a single confined ion’s motion [1]. But there, the “projectile” and the “sensor” were both individual atomic-scale particles of comparable mass. Now David Moore at Yale University and his colleagues have extended single-collision detection to a very different regime [2].