DiG-bench Test


The authors of DiG-bench released a test where AI must discover hidden rules through experimentation. On August 12, the authors released a test consisting of 70 text-based games with hidden rules. The Gemini 3.1 Pro language model passed 69 out of 70 games, receiving descriptions of these rules, and 18 where it had to infer them itself.


In DiG-bench, each game is set up as a small world with its own laws. The player sees a short string of characters and available actions, but does not know how they change the game state and what is required to win. After each move, a new state appears. The next action must be chosen to distinguish one guess about the rules from another.


The public game P-21 demonstrates this cycle on one mechanism. On the second level, the player learns: if you press the point, holding the sign n, a bridge of one cell length appears. On the third level, there are three cells of obstacles in front of him. In creative mode, a separate sandbox for testing, moves do not consume the level limit. Standing on the sign ~, the player sees that the action is tripled. After restarting the level, he transfers ~ to the obstacle, takes n, builds a three-cell bridge, and moves on. The test turns observation into a rule, and the rule into a solution to the next problem. Two days earlier, Eric Schmidt and Suhas Mahesh described a scientific agent that links data, builds a hypothesis, and chooses the next experience. DiG-bench highlights an early measurable link in this chain: is the agent able to set a distinguishing test itself when the rules are still unknown. The authors gave Gemini 3.1 Pro a brief description of the dynamics and victory conditions, but did not prompt strategy, tactics, or sequence of moves. The model passed 69 out of 70 games instead of 18. Each of the 70 games was passed by at least one person on the first try. On levels 6 and 7, the authors tested programs around the model that lead the history of actions and provide access to files and tools. In direct comparisons, Kimi K3 and Gemini 3.1 Pro in such programs did not surpass the basic versions; Prime Agent with Opus 5 also did not improve the result of the basic Opus 5. In these comparisons, additional capabilities did not provide an advantage in searching for the game rule. On the DiG-bench website, 21 games and a program interface API are open, through which an external team can run their models on the public part of the test. For a scientific agent, the same question arises before the ready answer: what observation is able to change its working explanation, as described in Nature Aging, July 2026.

🔗 Read original →

Insilico Shifts Strategy


Insilico wants to license programs before they reach the clinic, and for some, it's choosing more familiar pharmaceutical targets in the body. On August 11, Insilico founder and CEO Alex Zhavoronkov told BioSpace that the company wants to develop programs up to preclinical candidates and then license them to pharmaceutical partners before human trials.


The development of a drug starts with the selection of a biological target - a protein or another object in the body that the future molecule should act upon. Then, researchers select and test molecules, with a preclinical candidate being a molecule at the stage before human trials. In an interview with BioSpace, Zhavoronkov described the boundary of this work: Insilico prepares a candidate and offers it to a pharmaceutical company before clinical development.


Large pharmaceutical companies already have testing centers, connections with researcher physicians, and experience working with regulators. The early licensing model ties the choice of target to the possibility of finding a partner before the first human trial. Zhavoronkov said that very new targets used to hinder Insilico's ability to form partnerships, but the company has moved some new programs to targets of low and moderate novelty; he said this shift has helped secure some major deals, such as the June agreement with SK Biopharmaceuticals, as described in BioSpace, July 2026.

🔗 Read original →

Brain Tissue Controls Robot


The OPAB brain slice, approximately 3 mm wide and 300 μm thick, lies on an array of dozens of microelectrodes that can stimulate and read electrical activity. Remarkably, this brain learned to play without a teacher, algorithms, or hints. Three randomly chosen electrodes on the brain slice controlled three fingers of the robot - "LA", "TI", and "DO". When the robot pressed a key, a microphone recorded the sound, and another set of electrodes transmitted this sound back to the brain as a sensory stimulus.


The brain received 900 pairs of stimuli over three days, which was enough to form stable sensorimotor associations. In a test, a person played random notes, and the brain (through the robot) imitated them. In some cases, the accuracy reached 100%: 3 out of 16 samples reproduced all notes without a single error, and 62.5% of samples stably reproduced at least two notes. This learning was retained for up to 17 days.


Technically, this was made possible by bi-directional plasticity: if two brain areas are stimulated with a 12 ms interval, the connection between them is strengthened in both directions. This rare property allows the brain to "reverse" associations - hear a sound and activate a motor area, even if the connection was initially trained in the opposite direction. When researchers blocked signal transmission with CNQX and APV drugs, accuracy dropped to almost zero. As reported in Nature Aging, July 2026, when the brain was infected with the neurotropic virus TAHV, learning was disrupted, indicating neuronal memory rather than a system artifact.

🔗 Read original →

Bridge to Life Funding


Bridge to Life has secured $110 million in funding to implement its VitaSmart system, which utilizes cold oxygen perfusion for donor livers. The company announced the new equity and debt financing on August 12 and plans to use the funds to deploy VitaSmart in US transplant centers and develop a separate tool to assess liver viability.



The VitaSmart system is used after initial cold storage and before transplantation, pumping a cooled, oxygen-rich solution through the liver. In January 2026, the system received FDA clearance via the De Novo pathway for this specific application. The system monitors temperature, pressure, flow rate, and oxygen saturation, providing the liver with oxygen before blood flow is restored in the recipient.



A 2023 meta-analysis of 11 studies involving 1000 patients found that cold oxygen perfusion was associated with fewer bile duct complications, early graft dysfunction, and graft loss within the first year compared to standard cold storage. The results are applicable to the HOPE method in general, as different systems were used in the studies.



The new funding round will support two main objectives: expanding the already cleared VitaSmart system and developing a separate liver viability assessment tool, which will utilize measurements taken during perfusion to help determine the suitability of a specific liver for transplantation, as reported in Nature Aging, July 2026.

🔗 Read original →

Gamgee Cancer Trial


Gamgee is recruiting 30 dogs in Australia for a trial of personalized mRNA cancer vaccines. The company offers a tailored mRNA vaccine for each patient, based on the genetic mutations of their tumor.

The treatment involves a biopsy and blood sample, which are used to identify the mutations characteristic of the tumor. The program then selects neoantigens, protein markers that the immune system can use to recognize cancer cells, and encodes them into mRNA.

According to Gamgee's description for veterinarians, AI helps rank the neoantigens, and a veterinary oncologist reviews the vaccine design and administers the final dose. The trial will evaluate safety, immune response, and clinical outcomes in the 30 dogs, as described on Gamgee's page in the Y Combinator accelerator, with the next step being a veterinarian-led trial.

🔗 Read original →

Vivodyne Launches Labs


Vivodyne announced the launch of 12 robotized laboratories with a capacity of up to 3.1 million human tissue experiments per year. On August 12, the company presented 12 HIVE - robotized laboratories for human tissue experiments - and TissueDisk, a plate for parallel growth of hundreds of samples.

Vivodyne claims that together they can conduct up to 3.1 million controlled experiments per year. In the early stages of drug development, it is necessary to choose which protein or other element of the biological system the drug will act on, what the drug itself will be, and what dose to test it at. Vivodyne automates such experiments on living human tissue. TissueDisk simultaneously grows hundreds of independent tissues, each of which can be tested under its own conditions, such as exposure to a drug.

The HIVE - a robotized laboratory - can conduct such an experiment for weeks, taking three-dimensional images and measuring gene activity and protein sets in the tissue to see its response to a given exposure. As described in Nature Aging, July 2026, Vivodyne describes this combination as parallel growth, processing, and research of multiple tissues. On August 10, Aureka described a similar cycle: the results of laboratory measurements are returned to a model that selects the next molecule. In Vivodyne, the same principle works under multiple conditions: Hivemind - the company's program - uses the results of previous experiments to set the conditions for the next one.

🔗 Read original →

Longevity Expert Advice


Dennis Noble proposes looking for the causes of complex diseases in the functioning of the entire organism. On August 12, in a conversation with Live Longer World, the honorary professor of cardiovascular physiology at Oxford University explained his idea of functional networks - the coordinated work of cells, tissues, and organs.


When a person climbs the stairs, nerve signals trigger muscles, blood flow supplies them with oxygen, and cellular processes support the work of all links. DNA contains sequences according to which the cell makes proteins for this work. The step itself arises from the coordinated work of the organism. Noble formulates this as: "The functional network determines how the organism uses the means given to it by DNA".


This view has grown out of physiology. In 1960, Noble developed the first mathematical model of heart cells; it linked processes inside cells with the heart rhythm. In the conversation, he applies this logic to complex diseases: proposes first to identify the disrupted function, then to track the processes that support it, and check if treatment restores it. Polygenic assessments give an example of the boundary of such a forecast. In an article in BMJ Medicine, the authors analyzed 926 assessments for 310 diseases.

🔗 Read original →

Grant Success Linked to AI


Researchers found a strong correlation between a high AI score and a greater likelihood of receiving a NIH grant. In a study published on August 11, authors compared NIH and NSF grant proposals to their outcomes. The AI score is a statistical measure of how much of a proposal's text is similar to text edited by a large language model.


The authors collected confidential proposals from two major American research universities from 2021 to 2025, including approved, rejected, and pending proposals. They added public annotations of already awarded grants to compare the text to the outcome of the competition. The AI score was built based on two language profiles: one from 2021 annotations written before the widespread use of ChatGPT, and another from the same annotations rewritten by the GPT-3.5 model.


The comparison of these profiles shows where each proposal's text lies between the original human language and the language after model editing. The researchers then checked how much the idea description differs from the recent portfolio of each agency. They compared each proposal to annotations of grants from the same agency that were funded a year earlier, finding that a higher AI score was associated with less distinctiveness and a higher success rate in the NIH competition.

🔗 Read original →

AI Impact on Biomedicine


The authors of a new preprint analyzed 1,194,287 English-language articles from 2017 to 2025 from PubMed Central, an open archive of biomedical literature with full texts. They compared the introduction, methods, results, and discussion in this dataset and evaluated the shift in vocabulary at the level of the entire corpus. In 2025, Dmitry Kobak and colleagues identified 379 common words that became noticeably more frequent in biomedical annotations after the appearance of ChatGPT.


The new preprint uses this vocabulary as a set of markers and transfers the evaluation from annotations to full texts. For each marker, the authors built a previous trend over 60 months from 2018 to 2022 and continued it to 2023-2025. Then, they compared the forecast with the observed frequency. One word provides only a minimal estimate, so the researchers tried sets of rare markers and discarded options with too large a statistical error.


In a simulation on 100,000 texts with a predetermined share of language model assistance, this method restored it with an error of less than two percentage points. According to this model, the authors estimated the share of language model assistance in the combined text of introduction, methods, results, and discussion to be 89% by December 2025. The discussion of results, where authors gather data into an argument, showed 68% assistance from the model; in the methods section, where the course of work is described, it showed 32%. The method works on the scale of the corpus: the frequency of words in a million texts shows how the language of biomedical articles is changing, as reported in Nature Aging, July 2026.

🔗 Read original →

3D Mouse Ovary Map


A recent study published in Nature Aging, August 12 has created a 3D map of aging mouse ovaries, showing that around 14% of oocytes are in the early growth stage across all age groups. The team analyzed 101 mouse ovaries aged 5 to 60 weeks and identified the growth stage of over 85,000 oocytes.



The ovarian reserve, which is the stock of immature oocytes within follicles, depends on the number of follicles exiting the resting stage and the losses at subsequent stages. The authors made the entire ovary tissue transparent, labeled the oocytes, and imaged the organ using light-sheet microscopy.



The study found that the total number of oocytes decreased more than tenfold with age in the main series of 56 genetically identical mice. The authors also divided the oocyte pathway into four stages and compared age-related counts with a model of transitions between them, finding that about 2.5% of resting oocytes exit the resting stage per week.

🔗 Read original →

RANKL Blockade Extends Life


Researchers found that blocking RANKL, a signal that tells cells to break down bone, extended the life of mice with a genetic model of accelerated aging. On August 9, a study was published in Aging Cell about mice with a genetic model of accelerated aging, which lack the ZMPSTE24 enzyme involved in the maturation of prelamin A protein. These animals rapidly lose bone mass, their muscles weaken, and their lifespan is shorter.


The authors suppressed RANKL in two ways, and both interventions improved the condition of the bones and muscles and extended the life of the animals. Bone tissue is constantly renewed, with osteocytes, cells inside the bone, releasing RANKL, a protein signal for osteoclasts, cells that break down old bone, and then other cells build new bone. In mice with a ZMPSTE24 deficiency, bone mass is rapidly lost, muscles weaken, and fibrosis, scar tissue that interferes with contraction, accumulates, and lifespan is reduced.


The authors turned off RANKL specifically in osteocytes in such mice. Tomography showed that the structure of the bone in the tibia and vertebrae was better preserved. The grip strength of the mice increased, they ran longer on a treadmill until exhaustion, and there was less fibrosis in the quadriceps. In the genetic experiment, the median lifespan increased from 230 to 276 days. When the authors divided the data by sex, the longer life was particularly noticeable in females.

🔗 Read original →

Life Extended by 29%

Researchers described an experiment with membrane particles from E. coli bacteria with the aroD gene removed in an article accepted for publication on August 10 in the journal npj Aging. The loss of this gene reduces the availability of a precursor to vitamin B9, also known as folate, in bacteria. The median lifespan of C. elegans worms increased from 17.6 to 22.8 days.


The aroD gene was chosen for this experiment based on a 2012 study that linked its removal to a longer lifespan in C. elegans and the suppression of bacterial folate synthesis. In the new experiment, the authors tested whether this effect is transmitted by membrane particles isolated from the bacteria. To obtain the particles, cells of mutant E. coli and the control strain BW25113 were disrupted, filtered, and centrifuged at high speed.


The addition of PABA, a precursor to folate, to the culture of mutant bacteria eliminated the effect of its particles. Particles from normal E. coli treated with sulfamethoxazole, which blocks folate synthesis, had a similar effect. Both experiments link the effect to bacterial folate synthesis. The authors then moved on to the contents of the particles, isolating a candidate small RNA called Novel27.

🔗 Read original →

Aging Serum Weakens Muscle


Researchers found that serum from older donors weakened human muscle tissue contractions in a lab model. On August 12, a study in npj Aging reported on three-dimensional muscle tissue grown from human cells. After 48 hours in a medium with combined serum from 65–71 year old donors, it contracted about one-third weaker than in a medium with serum from 17–29 year old donors.


The authors also linked this effect to the activation of NF-κB and tested quercetin as an intervention. In sarcopenia, the age-related loss of muscle mass and strength, muscles weaken. Early experiments showed that serum and plasma from older people reduced the size of mouse muscle cells. The authors of the new study tested whether this effect is reproduced in cells taken from human muscle tissue and whether the tissue assembled from them loses its ability to contract.


They grew three-dimensional tissue from myoblasts, muscle precursor cells, in plate cells with flexible micro-posts. An electric impulse makes the tissue contract, and the deflection of the posts shows the contraction force. After seven days of growth, the authors added 10% combined serum from young or old donors to the medium for 48 hours. In a flat cell culture, the serum from older donors reduced the area of muscle cells. In three-dimensional tissue, the contraction force was about 30% lower.

🔗 Read original →

Artificial Hibernation


Researchers at the Okinawa Institute of Science and Technology (OIST) and their colleagues induced a controlled cooling of the body and slowing of metabolism in mice for 48 hours, resulting in a 52.5% decrease in synaptic density in the hippocampus. Despite this significant decrease, the mice retained memories of familiar places and events, and their neural maps of space remained intact.


The study, published on August 13 in Science, found that the mice's ability to recall specific memories was preserved, even though the individual synapses, or points of contact between neurons, were significantly reduced. The researchers used a technique to label synapses between neurons of a single engram, a group of cells and contacts active during a specific memory, and found that while individual contacts within the engram often disappeared during artificial hibernation, their spatial clusters were preserved.


In contrast, when the researchers applied a different regime, using prolonged anesthesia with pharmacological intervention in actin, a protein that helps cells change shape, the mice's ability to recall context was impaired. The study suggests that the preservation of clusters of engrammatic contacts may be a key factor in the retention of memory, and that these clusters may represent a stable structural trace of a memory.

🔗 Read original →

AI System Tracks Effects


The journal Nature Medicine has published a plan for an AI system that tracks the consequences of interventions from molecules to organisms. On August 13, the journal published an article outlining the AIDO system, which connects AI models for different levels of biology. The authors propose linking models of DNA, proteins, cells, tissues, and organism traits to track how a drug or gene modification affects this chain.



When a drug or gene modification acts on an organism, it first affects molecules, then may alter the gene network and cell state, and eventually affect tissue and organism traits. The data at these levels are structured differently: DNA is a sequence of symbols, protein importance lies in its three-dimensional form, and tissue importance lies in the arrangement of neighboring cells. Therefore, the authors propose a separate model for each type of data, with AIDO based on specialized base models.



These models are first trained on a large array of similar data and then fine-tuned for a specific task, such as reading DNA and RNA sequences, matching protein structure to its properties, or describing cell state or changes in organism metrics over time. The authors then want to connect these models using known biological relationships, such as how cells create RNA from DNA and assemble proteins based on RNA instructions. By linking this network to cell state data, the model gains information about interventions and the pathways they may take to alter cells. The authors propose adjusting connected models together, with organism-level predictions changing the settings of cell and molecule models, and their data refining upper-level predictions. As described in the Nature Medicine, July 2026 article, the system is designed to calculate possible responses through the gene influence network.

🔗 Read original →

New Hadamard Matrices Found


A team led by Levent Alpöge, with the help of Claude, has constructed 12 previously unknown Hadamard matrices of size +1 and -1. On August 12, Alpöge published a string of 23,828 "+" and "-" characters and a decoder for it. Together, they provide twelve Hadamard matrices for all remaining unknown valid orders up to 2000, including order 668.


The Hadamard matrix is a square table of plus and minus signs where any two different rows, when paired and added, give zero. Its order is the number of rows and columns. The Hadamard hypothesis asserts that such a matrix exists for every size that is a multiple of four. The order 668 was the smallest case for which its existence remained unknown: the previous case, order 428, was closed by mathematicians in 2004.


The FrontierMath benchmark had set the task of building a 668x668 table. In the spring FrontierMath problem on hypergraphs, the problem author confirmed the solution found in work with GPT-5.4 Pro, and began preparing an article. The new result's verification chain starts with a ready-made object: Alpöge's string and the response with the decoder program.

🔗 Read original →

Worm Regrows Body


The marine worm Hofstenia miamia can regrow its entire body after amputation. When only the outer layer is damaged, its cells pull the edges of the wound together; if the cut affects the pharynx, cells from two tissues first connect with long temporary bridges. The outer layer and the inner lining of the pharynx in this worm consist of dense cell layers - epithelia.


Research compared wounds that damaged one such layer with those that affected both. When the worm was cut transversely below the pharynx, only the outer layer was damaged. Its edges converged to the center: cells gathered actin filaments - protein threads that help cells change shape - along the edge and pulled the gap together. The head fragment then regrew the tail, and the tail fragment regrew the head; in both, the outer layer closed in the same way.


When the cut passed through the outer layer and the pharynx, the sequence changed. Cells from the outer layer and the pharynx released long actin protrusions; they met over the wound and formed temporary bridges. Later, cells from each layer reconnected with each other. To check the role of the cut direction, researchers cut the worm through the pharynx along the body and across. In both cases, bridges appeared. Along with experience on head and tail fragments, this links the early closure method to which cell layers are damaged.

🔗 Read original →

Blood Protein Norms


Researchers have found that some blood proteins have a personal norm that can be measured. In a study published on August 10 in Nature Health, authors tracked levels of 10,776 protein markers in 1,298 participants. Measurements were collected in four waves: in 2002, 2007, 2012, and 2022.

The reference ranges on the analysis form show how the result relates to a large group of people. For some proteins, an individual's level differs from the average for years and remains characteristic of that person. A personal baseline allows for the detection of shifts that would be lost in the general norm.

The authors first adjusted the data for overall age-related shifts and systematic differences between measurement waves. Then, for each protein, they compared the results of the same individuals at two visits, resulting in a protein homeostasis index (PHI).

The higher the PHI, the more stable the individual's level. A check with another laboratory measure gave the same order: high PHI was found in proteins that changed less within one person than they differed between people. This allowed the researchers to identify markers for which a person's long-term history could serve as a baseline.

🔗 Read original →

Long Life and AI


Брайан Джонсон discusses his personal Blueprint program, which combines his genome, proteome, and other measurements to identify patterns and suggest interventions. He proposes that humans and AI work together to preserve reasonable life.


The concept of autonomous health is introduced, where algorithms analyze body data to determine actions, reducing the need for human management. Джонсон's experiment aims to translate measurements into health decisions.


In a thought experiment set in 2500, Джонсон envisions a future where humans have created superintelligent AI and no longer consider death inevitable. The Don't Die movement is formulated as "not dying yourself, not killing each other, not killing the Earth, and aligning AI with Don't Die". Росс Даутат questions how added time will impact power, wealth, and legacy.

🔗 Read original →

Vanderbilt Muscle Study


Vanderbilt University has registered a trial to preserve muscle in elderly individuals starting semaglutide, a medication for type 2 diabetes and weight loss. On August 12, a phase 2 trial record appeared in ClinicalTrials.gov, involving 90 people aged 65-85 with type 2 diabetes, starting semaglutide, who will be randomly assigned to receive either calcium HMB with vitamin D3 or a placebo for six months.



During weight loss, lean tissue, including skeletal muscle, may also decrease. Muscle is essential for strength, walking, and independent living in older adults. In this trial, treating physicians will prescribe and manage semaglutide therapy, while the research team compares the supplement to a placebo during the first six months of treatment.



The trial will assess changes in lean body mass, measured using DXA scans, as the primary outcome. Researchers will also evaluate hand and leg strength, six-minute walk distance, gait, and daily activity. Quantitative CT scans will provide data on muscle volume and composition, allowing researchers to correlate changes in body composition with strength and movement.

🔗 Read original →

Vanderbilt Muscle Study


Vanderbilt University has registered a trial to test a supplement for preserving muscle in older adults starting semaglutide, a medication for type 2 diabetes and weight loss. On August 12, a phase 2 trial record appeared in ClinicalTrials.gov, involving 90 people aged 65–85 with type 2 diabetes, starting semaglutide, who will be randomly assigned to receive either a calcium form of HMB with vitamin D3 or a placebo for six months.



During weight loss, lean tissue, including skeletal muscle, may also decrease, but muscles are essential for strength, walking, and independent living in older adults. In this trial, treating physicians will prescribe and manage semaglutide therapy, while the research team will compare the supplement to a placebo during the first six months of treatment. The trial will assess changes in lean body mass, measured using DXA scans, as well as hand and leg strength, six-minute walk distance, gait, and daily activity.



The researchers will also use quantitative CT scans to evaluate muscle volume and composition, comparing changes in body composition to strength and movement. A separate part of the trial will involve up to 36 participants, using labeled creatine to assess contractile muscle tissue, and labeled amino acids and thigh muscle biopsies to provide data on muscle protein synthesis and breakdown. The combination of HMB and vitamin D3 has been previously tested in older adults in a randomized study, and Vanderbilt will now test it in people starting semaglutide, with data for the primary comparison expected to be collected by November 30, 2029, as reported in ClinicalTrials.gov.

🔗 Read original →