Anil Seth: the working space found in Claude shows reasoning, but does not prove experience. On July 15, a professor of cognitive and computational neuroscience responded to Anthropic's work, which found an internal common channel for intermediate thoughts in Claude. Seth acknowledges the result itself, but believes that the similarity to one of the theories of consciousness does not allow us to conclude that the language model feels anything. On July 6, Anthropic described a narrow section of computations within Claude, which it called J-space. The model holds a word or intermediate result there and then uses it in different tasks. Researchers replaced the internal representation of "spider" with "ant" - and Claude instead of eight named six legs. When they suppressed J-space, the model retained fluent speech and simple responses, but almost lost multi-step reasoning. Such a section is similar to what the theory of global workspace is looking for in the brain: information that becomes available to many systems at once. In everyday life, a person can hold a number, intention, or image in mind, tell about it, and apply it in a new action. Anthropic showed a functional analogue of this operation in Claude: a common internal representation can be read, modified, and causally linked to the model's response. Already on June 2, Google DeepMind, Anthropic, and Meta had referred the possible experience of models to a research task: DeepMind hired a philosopher, and Anthropic is leading a program to study model welfare. J-space gives this debate a measurable subject - an internal representation, the role of which can be changed in experience. In a column for The Guardian, Anil Seth draws a line between such access to information and subjective experience. The first answers the question of whether a system can hold a thought, report it, and use it for selection. The second is whether this system has its own "what it feels like": pain, color, fear, pleasure. Seth points to recurrent loops - feedback cycles in which signals repeatedly pass through brain circuits and support representation over time. In Claude, J-space processing fits into one pass through the network; Anthropic itself notes this distinction. Seth links consciousness to how a living nervous system works together with the body and the world, and not just with the calculations of a silicon program. Seth believes that similar computational organization is not enough to conclude experience. Anthropic limits its result to functional "conscious access": the model can hold content, report it, and use it in reasoning. The authors do not attribute subjective experience to Claude. The future transfer of personality will have to be tested along two different lines. The system must retain memory, goals, and accessible reasoning to continue acting like the original person. Separately, the theory of consciousness must explain what properties of the carrier support its internal experience. J-space provides a way to measure the first line in a language model; the second remains a subject of debate about the brain, body, and consciousness.

🔗 Read original →

AI 2027 underestimates synthetic biology: it may launch a self-replicating economy earlier than robotized factories. On July 14, an author from LessWrong proposed considering an earlier transition to the smallest systems capable of copying themselves from available raw materials. One of the co-authors of the scenario supported part of this criticism. The AI 2027 already has a distant variant with biological infrastructure. In the AI 2027 scenario, the US and China create special economic zones with simplified construction rules. In the American zones, Agent-5 - a superintelligence from the scenario - designs robots and, through humans, manages the construction of factories and laboratories; new factories produce even more robots. As a reference, the authors cite a car factory that in less than a year produces a mass of cars comparable to its own. Then they suggest that an autonomous robotic economy could reproduce itself faster than a year. In response to this scenario, Thomas B. suggests looking for the future leap below the factory scale. He suggests that it would be beneficial for the superintelligence to reduce the unit of self-replication to the limit and design organisms and nanotechnology systems that grow and copy themselves using the surrounding raw materials. The argument is based on the speed of experimentation. In Thomas B.'s opinion, the AI 2027 assumption about robots at the human level would mean automated laboratories without a shortage of qualified human labor. He suggests the next step: such laboratories could create many cheap living sensors and executive systems, conduct experiments with them in parallel, and give the results to simulations. Improved simulations help design the next version of organisms and devices. This research reduces the time for the next research. The dispute comes down to the speed of feedback. A factory needs buildings, machines, supplies, and large machines. A self-replicating biological system could become a sensor, an object of parallel experiments, and part of the next production system. Thomas B. suggests putting this possibility at the center of the scenario. AI 2027 itself is already moving in this direction. The authors cite plants, insects, and bacteria as benchmarks for reproduction rates. In a distant variant, they suggest an economy similar to an ecosystem of new algae: some organisms grow, others process them into materials for floating factories. The disagreement concerns the time of emergence of this biology. The main line first unfolds a robotic economy; Thomas B. considers biological self-replicating systems a likely early driver of acceleration. Thomas Larsen, one of the co-authors of AI 2027, in the comments, agreed with the main criticism and wrote that the finale was worth linking more strongly to nanotechnology. The dispute shifts the question about superintelligence from the number of robots to the length of the "experiment - data - new system" cycle: if this cycle can be compressed to the speed of reproduction of created organisms, factories will cease to be the only measure of production capacity.

🔗 Read original →

Dextran, albumin, and sialic acid helped preserve extracellular vesicles for a year at -20 °C. Chinese researchers have assembled a cryoprotective mixture, DDAS, for extracellular vesicles - tiny membrane particles that cells use to transport proteins and nucleic acids. After storage at -20 °C and repeated thawing, these therapy candidates retained more structure and biological activity than in a conventional buffer. Extracellular vesicles are being considered as a future means of delivering stem cell signals, proteins, and RNA to the body. Such a preparation must travel from the laboratory to the patient, survive storage, and remain the same preparation. When frozen, vesicles aggregate, their membranes rupture, and their contents leak out. Their action is determined by the molecular cargo: four microRNAs in vesicles from an aging liver enhanced metastases in mice. Therefore, for therapy, it is not enough to preserve the number of particles; it is necessary to preserve the specified composition and function. A conventional phosphate buffer poorly solves this problem. DMSO and glycerin can protect biomaterial, but they cannot be left in the preparation for direct intravenous injection: after thawing, additional purification is required. The same engineering problem already exists for platelets: in a laboratory test, a low-dose DMSO regimen without washing returned 94.4% of cells after thawing. The authors of DDAS checked whether it is possible to store vesicles so that after thawing, they can be administered without additional purification. On July 13, an early unedited version of the work on DDAS was published in the Journal of Nanobiotechnology. The mixture includes dextran, albumin, and sialic acid; the authors selected it by starting with a list of molecules found in human body fluids. Generative models narrowed down the list of candidates, and the composition was then tested on vesicles, cells, and mice. Vesicles from brain vascular endothelial cells were stored at -20 °C for a year. DDAS better preserved the number of particles, RNA, and membranes than a conventional buffer; by the 12th month, the authors recovered around 50% of nucleic acids. After five cycles of freezing and thawing, it also outperformed the buffer and a commercial cryoprotector. The authors suggest that dextran reduces particle collisions, albumin forms a protective film, and sialic acid gives the vesicle surface a negative charge. Vesicles stored in DDAS can be administered intravenously directly after thawing. In an experiment on six young male mice with deep burns, stem cell vesicles stored in this mixture accelerated wound closure more than the same vesicles from a conventional buffer. By the 12th day, the wounds of mice that received vesicles from DDAS were almost closed. Freezing damages the membranes of vesicles and their cargo; DDAS, according to the authors, allows them to be administered without repeated purification.

🔗 Read original →

The enzyme CMLase removed age-related chemical damage from the amino acid lysine in proteins. On July 14, Nature Communications published a study on the CMLase enzyme, which removes CML, a common age-related chemical damage to lysine, from proteins and returns the original amino acid. The reaction was demonstrated on model proteins and human lens, skin, and aorta samples. Lens proteins and collagen in skin and vascular walls serve for years or decades. Reactive sugar and lipid compounds attach to their amino acids, resulting in glycation and lipoxidation products. One of these is Nε-carboxymethyllysine, or CML. The carboxymethyl group attaches to the side amino group of lysine and changes its charge from positive to negative. CML affects both the protein itself and the cells around it. It binds to RAGE, a cellular receptor for glycation products, and triggers inflammatory and oxidative signals. Cellular enzymes neutralize some of these reactive molecules even before they encounter the protein. CMLase works after CML formation, when the mark has already been fixed on long-lived proteins. The authors took a bacterial glycine oxidase and, over five cycles of directed evolution, tested more than 500 million variants. The resulting CMLase recognizes CML within an intact protein and converts the damaged residue back into lysine. On model proteins, after overnight treatment, the CML signal in an antibody test decreased by 52-97%. On soluble lens proteins from a 64-year-old donor, mass spectrometry showed a reduction in total CML of 45%, and the antibody test showed a reduction of 78%. Mass spectrometry measures the total amount of CML after chemical degradation of the protein, while the antibody test only sees the labels accessible to the antibody. The authors suggest that CMLase more easily removes labels on the surface of the protein than those hidden inside. On thin fixed sections of the aorta from a 75-year-old donor, CML staining decreased by more than 70%, and on skin from elderly donors, it decreased by more than 55%. The inactivated control enzyme retained the original signal. Soluble lens proteins and thin sections give the enzyme direct access to the target. In a living vessel, it needs to penetrate a dense extracellular matrix - a network of proteins that supports the tissue. Researchers also need to measure the immune response to the bacterial enzyme upon repeated administrations. CMLase allows for testing whether removing one chemical mark changes the RAGE signal and vascular elasticity. Glucospane, an age-related crosslink that connects matrix proteins, will require a separate enzyme. In the analysis of the glycation hypothesis, the accumulation of collagen and elastin damage is associated with the loss of elasticity in aging tissues.

🔗 Read original →

Brian Johnson was diagnosed with autoimmune gastritis, a disease in which the immune system attacks the stomach. On July 15, author on aging Andrew Steele broke down Johnson's public history: eleven years of low ferritin, an indicator of iron stores. A new team of doctors linked this signal to an autoimmune thyroid disease, checked the intestines and stomach, and biopsies confirmed autoimmune gastritis. In his account, Johnson describes how low ferritin was explained by a plant-based diet, training, sauna, and oxygen procedures. He tried food and supplements with iron, but the stores were not restored. Hemoglobin remained normal, so there was no anemia. "Low ferritin continued to be explained, but not corrected," Johnson wrote about the previous years. The new diagnosis began after a change of doctors. At 48, Johnson underwent a colonoscopy for the first time; the American preventive recommendation suggests such screening for people with average risk starting at 45. The colonoscopy ruled out hidden blood loss in the intestine. Then, doctors correlated the long-standing autoimmune thyroid disease with persistent iron deficiency, checked antibodies to stomach cells, and took five biopsies. The biopsies showed autoimmune gastritis: the immune system damages the stomach lining, which produces acid and helps absorb iron. In the early stages, this disease may not give a noticeable picture upon examination. The American Gastroenterological Association directly advises considering atrophic gastritis in unexplained iron deficiency and confirming it with histology - tissue examination under a microscope. In his analysis, Steele puts two facts from Johnson's text side by side: the Immortals Care protocol costing $1 million per year and the years when low ferritin was explained without looking for a cause. Sensors collect numbers; a diagnostic hypothesis links them to a cause and a test. On June 30, Steele compared personal protocols to a $45-70 million metformin trial. Such a project is looking for a general answer: does a cheap intervention work in people? Johnson's story adds medical consistency to this calculation: first, determine the cause of the disease, then test the treatment. Now, Johnson wants to move beyond observation. His plan starts with monitoring and support, then moves to influencing immune signals and regulatory T cells - cells that normally restrain the immune attack. In the long-term perspective, he considers CAR/CAAR-T: modified T cells that attempt to direct immune cells attacking specific tissue. He also writes about AI-designed antibodies and synthetic proteins. Johnson calls these steps research: part of the approaches still need to be created specifically for autoimmune gastritis. If one patient gets better after a complex set of interventions, the contribution of each of them, the natural course of the disease, and the role of usual supportive therapy remain unknown. A trial with a comparable group of patients separates the working method from a lucky coincidence.

🔗 Read original →

Weco reported that its agent rewrote the code of another agent and improved its results on external tests on July 14. Weco described AIDE²: over eight days, the system performed 100 cycles of rewriting and testing the internal AIDE agent. The company claims that during this time, seven consecutive improvements emerged, and two later versions outperformed the original on three external tests. The full technical report and the release of AIDE 85 itself, Weco promises later. The original AIDE is an agent for machine learning engineering tasks. It writes code variants, runs them, and develops successful branches. In AIDE², the external role was played by AIDE human - a manually tuned version of the same agent. It modified the code of the internal AIDE 0, a simplified version of the original AIDE. Each step included one edit and a full evaluation across task families. The version remained in the cycle only if a hidden check confirmed its advantage at a fixed evaluation cost budget. AIDE² optimized the solution search procedure itself. AIDE 85, on average, reduced the history of past attempts, which the model receives before the next step, by 16 times, and used the freed-up text volume for additional attempts. When the best branch stopped improving, the agent took its code as the basis for a new branch with a different strategy. This economy allowed it to conduct more experiments. A similar problem was already solved by AstraZeneca employees: their five agents passed a short map of past work to the next launch, and the reasons for the decisions were stored separately. In their tests, the system chose the correct model form in all 20 artificial tasks. Weco uses compressed history as part of a broader mechanism that checks on other task sets. Weco tested AIDE 47 and AIDE 85 on three external tests that the system had not seen in its improvement cycle. One of them is MLE-Bench Lite, a simplified version of MLE-Bench. The full MLE-Bench is compiled from 75 Kaggle machine learning engineering competitions and contains human benchmarks. In a separate test on KernelBench - a set of tasks for accelerating computations on a graphics processor - Weco considered a test bypass case when less than half of the acceleration claimed in the short test was preserved in the full workload. The share of such cases, according to the company, decreased from 63% in AIDE 0 to 34% in AIDE 85. Then the company placed the improved internal agent in an external cycle. On the tasks on which the system was improved, AIDE 47 reached the same maximum result in approximately 20 edits, and the manually tuned AIDE human - in approximately 40. This difference was within the statistical noise. The experiment showed the improvement of the internal agent, but did not confirm that it accelerates the next improvement cycle. Currently, Weco describes these results only in its corporate blog. The full technical report and AIDE 85 system, the company promises to release after completing the analysis. The authors also found that the statistical filter against test bypass in the later version was broken and actually had no effect on the result.

🔗 Read original →

GPT-5.6 helped to refute a 20-year-old hypothesis on checking thousands of scientific results at once. Statistician Edgar Dobriban found a data model where the Benjamini-Hochberg procedure slightly more often promises validity than it can provide. It is applied when one experiment simultaneously checks thousands of genes or other characteristics and it is necessary to limit the share of false findings. Dobriban published the proof and numerical certificate code; the model participated in the search, and the author checked the result. On July 14, Edgar Dobriban published a proof that concerns the usual protection against such coincidences. When a biologist compares thousands of genes between groups of cells, random coincidences are inevitable. Therefore, the results are usually passed through the Benjamini-Hochberg procedure: it orders p-values - numbers that show how much an observation resembles randomness - and selects a threshold so that the share of false alarms among the declared findings does not exceed the specified level on average. This procedure has long had clear guarantees for independent results and for some types of positive dependence. However, real genomic data is correlated: close DNA variants are inherited together, genes work in networks, and cells share common causes of changes. The question remained whether the usual procedure retains its guarantee for any correlated two-sided tests, where the effect can go in both directions. Dobriban built a specific model with a common hidden factor - one invisible cause that simultaneously shifts many results. In it, the distribution of null results and real signals changes. The procedure chooses a threshold at which there are slightly more errors among the selected findings than it promises. At a nominal level of 1%, a strict numerical certificate gives no less than 1.0416829%. The difference is small, but it breaks the universal hypothesis: for correlated two-sided data, the guarantee cannot be automatically transferred from the case of independent tests. On July 10, OpenAI published a candidate proof of a hypothesis on graphs; it still needs to be dissected by mathematicians. In Dobriban's case, the verification chain already includes an open numerical certificate. According to the statistician, GPT-5.6 received the mathematical formulation and in approximately 90 minutes generated a counterexample, the proof, and the certificate code. Dobriban checked the argument, and the open program recalculates the lower bound with interval arithmetic, where rounding goes to the safe side. Therefore, the result can be verified independently of the story about the model's capabilities. For the biology of aging, this is part of the scientific infrastructure. Mass measurements of genes, proteins, and cells constantly require distinguishing signal from random noise. AI that finds a verifiable counterexample helps to determine where the statistical method actually guarantees reliability and where its conditions have already ceased to be met. Source: Edgar Dobriban's manuscript, open code, and certificate.

🔗 Read original →

A wrist vibration and a program that learns together with a human have helped novices to more quickly control cursor signals from the brain. In a study of 31 people, the system simultaneously taught the user to represent movement and retrained the program that reads their EEG. Over several sessions, this combination gave a greater increase in accuracy than usual training or vibration alone. On July 15, in Nature Communications, a team from Carnegie Mellon University published a study on how to train a neurointerface: a system that reads the electrical activity of the brain and converts it into a command for a computer. In the experiment, a person looked at a cursor and imagined moving their left or right hand. An EEG helmet recorded the weak electrical signals on the skin of the head associated with the work of the sensorimotor cortex, and the program moved the cursor based on these signals. It is difficult for a novice to immediately create a stable signal that the program can distinguish from noise. The person and the program learn simultaneously: the algorithm changes the recognition rule while the user is looking for a clear way to control it. In April, a two-way neurointerface returned artificial sensation of walking to a participant through stimulation of the sensory cortex. There, feedback should accompany walking; here, the vibration on the wrist works as a training cue while the person learns to control the cursor. The authors gave them a common cue. A small motor on the wrist briefly vibrated during training attempts. After each block, the algorithm was updated and more strongly took into account attempts where the brain signal already well distinguished the desired direction. The vibration helped the person to find a reproducible signal, and the algorithm adjusted to this signal. The main group consisted of 15 people. Another eight received vibration, but the program did not highlight signals that the person found easier to master; eight worked in a usual system without vibration. Between the first attempt and the second session, the accuracy of control in the main group increased by 19%. In the group with vibration without such signal selection, the increase was 6.3%, and in the usual group, it was 3.0%. In a two-dimensional task, where the cursor moved on the screen in different directions, the average accuracy of continuous control reached 66.9%. A check with constant vibration and another check, where the program was updated without this connection, showed that the result arises from their joint work. In six participants of the main group, the advantage was preserved for more than two months, when the vibration was already removed. So far, this is a laboratory experience on healthy young people, not the restoration of movement in people after a stroke or spinal cord injury. In such patients, the sensorimotor cortex is often rearranged, and the paths of sensitivity can be disrupted. The speed of mastering the EEG neurointerface depends on whether the training of the person and the update of the algorithm coincide.

🔗 Read original →

In Shanghai, the first commercial implantation of NEO for restoring grip was performed. On July 13, doctors at Huashan Hospital implanted NEO in a patient who had been living with the consequences of a cervical spinal cord injury for ten years. During the operation, the system recorded stable signals from the brain's surface, and after the procedure, the patient's condition remained stable. Data on grip recovery, independence in daily life, and long-term safety for this patient have not been published yet. NEO is an implantable neurointerface that translates movement intention into a command for an external device. The NEO electrodes lie on the brain's hard shell, between the skull and brain tissue. The system recognizes the intention to clench a fist and sends a command to a pneumatic glove, which bends the fingers and helps hold an object. On March 13, the Chinese regulator registered NEO for restoring grip in adults with cervical spinal cord injuries. The registration allowed the sale and clinical use of the device. This operation became the first case where NEO was installed in a patient through a commercial route. According to SCMP, within four months after registration, the manufacturer started production, began working with hospitals, and selecting patients. The Shanghai supplementary medical insurance program Huahui Bao included implantation materials in its coverage: it reimburses 30% of eligible expenses, up to 150,000 yuan per year. The limit applies specifically to the materials, not the entire procedure.

🔗 Read original →

An implant in the sensory cortex has been restoring the sense of touch to people with spinal cord injuries for ten years. Five participants were implanted with microelectrode arrays - sets of thin electrodes - in the section of the sensory cortex associated with the hand. A study published on July 15 in Science Translational Medicine collected data over implantation periods of two to ten years: over 168 million electrical impulses and 27 cumulative years of device operation. To take a glass, a person must not only send a command to their hand to "squeeze fingers," but the brain must also constantly receive feedback from the hand: whether the fingers have touched a surface, how hard they are pressing, and whether the object is slipping. After a severe spinal cord injury, this feedback disappears. A person can control a robotic hand through a neurointerface and still not feel its contact with an object. Intracortical microstimulation delivers short electrical impulses directly to the sensory cortex - a section of the brain with a body map. Different points on this map are associated with different parts of the hand. When a microelectrode stimulates the corresponding point on the map, a person feels touch in the corresponding finger or palm. A sensor on a robotic hand can be connected to this electrode and convert contact with an object into artificial touch. Artificial touch has already been helpful in a task with a robotic hand. In a 2021 study, a participant with tetraplegia performed the task in an average of 20.9 seconds with only vision and in 10.2 seconds when artificial touch was added to vision. A neuroprosthesis for everyday use requires that the electrodes, brain tissue, and sensations remain functional for years. In a new study, the group of Charles Greenberg, Robert Gaunt, and Jennifer Collinger tracked this sensory channel for up to ten years. Each of the five participants was implanted with two electrode arrays in the sensory cortex. Over this time, the researchers delivered over 168 million impulses. They did not detect any serious complications associated with stimulation or signs that the impulses themselves were degrading the performance of the electrodes. The sensations continued to arise in the hand and remained localized. In one participant, after ten years, 60% of the electrodes were still working reliably; on average, across the group, 64% were working, with a spread of 13% between participants. A slowly increasing current was required for sensation: approximately 3.5 microamperes more per year. Rare sensations could briefly persist after the impulse was turned off, but did not require treatment. Two days ago, bioengineer Takeshi Kozaei called the preservation of living tissue around the electrode a condition for a neurointerface to last decades. In this study, part of the electrodes gradually lose sensitivity. A sensor on the hand must transmit a signal to the same point on the hand map in the cortex for years, and the system must maintain the sensation of contact while individual electrodes fail.

🔗 Read original →

Zoledronate has altered signs associated with aging in the blood of women and tissues of old mice. Zoledronate has been used for decades to counteract bone mass loss. The authors of a new study measured changes in the blood of women, as well as in the hearts, livers, and intestines of old mice, and in human cells. They investigated how the drug may affect cells outside the skeleton. Bisphosphonates suppress the activity of osteoclasts, cells that break down bone tissue. Zoledronate was long considered a drug that works almost entirely within the skeleton. However, trials and observations of patients have raised another question: does it penetrate other tissues and change processes associated with aging? On July 15, a study by a team from Oxford University presented data from three levels. The authors examined the plasma - the liquid part of the blood - of 36 women over 65 years old with osteopenia, reduced bone density. After 18 and 36 months of infusion of 5 mg of zoledronate, about 400 out of five thousand measured proteins changed in it. Part of the changes affected proteins associated with DNA damage, mitochondrial dysfunction, and inflammatory substances that are released by aging cells. Plasma shows the trail of the drug's action, not the number of aging cells. A recent review of human trials came to a similar distinction: inflammatory proteins in the blood usually change more noticeably than direct signs of accumulation of old cells in tissues. Therefore, this human part of the study reveals a measurable signal, but does not answer whether the organism has become younger. The authors gave zoledronate to six old mice for two months and compared their organs with those of peers without the drug and young animals. In the heart, liver, and intestine, gene expression changed, and the calculation of cellular composition approximated the profile of young mice. The study links the signal from human blood to specific tissues where it can be further tested. The search for a mechanism led to heart cells. Labeled drug penetrated cardiac, hepatic, and renal cells. Low concentrations of zoledronate weakened cellular aging after DNA damage. The authors found a chain involving PHB2 - a protein that supports mitochondrial stability - and MEF2A, a regulator of heart cell function. When researchers turned off MEF2A, the protective effect of zoledronate almost disappeared. A common infusion quickly binds to bone, and then the drug is excreted by the kidneys. In cellular experiments, the effect appeared at low constant concentrations, which such a regimen does not predictably create outside the skeleton. The next test should check the delivery of zoledronate to tissues, its safety, and its impact on organ functions in humans. The current work provides specific cells, proteins, and measurement methods for this.

🔗 Read original →

Aran Nayebi and Daniel Yamins have derived a theorem about when two neural networks, solving a complex task, match at the level of individual computations. In the preprint, the authors set conditions for networks with ReLU and softplus activation functions. If two networks solve a sufficiently difficult common task and their representations can be linked by a linear transformation, individual computational directions may also obtain correspondences. Another theorem describes how such a correspondence can spread from the last layer to the early ones. In the preprint, published on July 9, Nayebi and Yamins dissect a long-standing observation from NeuroAI - an area where artificial neural networks are compared to brain function. Researchers show the same images to an animal and a model, and then check if the model's internal signals predict the responses of neurons. In a 2014 study, models were trained to recognize objects under changes in pose, size, and background. The more accurately they solved this task, the more accurately they predicted the responses of two areas of the macaque's visual cortex - V4 and IT. But a linear map can link large sets of signals, mixing the contribution of individual units. Therefore, it does not show by itself whether the model and the brain perform the same individual computations. Weak matching means that the activity of one layer can be translated into the activity of another by a linear rule. Strong matching means the correspondence of individual directions within a layer: for example, a filter that responds to an image edge at a certain angle. Nayebi and Yamins' theorem applies to networks with ReLU and softplus. ReLU converts a negative input to zero, so when crossing the threshold, the response gets a kink. Softplus has a smooth transition, but leaves curvature. If representations remain linearly comparable before and after such a nonlinear transition, the map between them must preserve the traces of individual used axes. Arbitrary mixing of axes will not do this. Therefore, provided that the task and the next layer actually use these directions, they can be matched with precision up to permutation and scale. "The harder the task, the fewer its solutions and the more likely their similarity," Daniel Yamins writes in an explanation of the theory. The authors call a hard task one that cannot be solved with a given accuracy, engaging only a few nonlinear axes in a given layer. The minimum number of axes needed, they call the budget of used axes. The more of the layer the task requires, the more of the axes the theorem guarantees correspondence between networks. Another theorem describes the "zipping" of hierarchies. It requires minimality: hidden functions at each step must make distinguishable contributions to the next layer. Then weak matching on the last comparable layer allows recovering correspondences in early layers for almost all weight settings; special exceptions the authors leave outside the result. In an approximate version, transitions between layers must be regular, i.e., not collapse different signal changes into one. Then a small error on the last comparable layer limits the fraction of axes that diverge earlier. The article suggests checking on brain data how much natural tasks engage many axes, whether minimality is satisfied, and whether training leads to exceptional weight settings. A simple task can be solved in many different ways. The authors suggest that tasks with natural images and situations for an animal may more strongly narrow down the set of solutions and thus increase the chance of matching the model with the brain at the level of individual computations.

🔗 Read original →

Kalshi opened pilot prediction markets for clinical trial outcomes and FDA decisions on July 16. Kalshi and AppliedXL launched pilot contracts for individual late-stage clinical trials and FDA decisions - those of the American drug regulator. The contract price shows the probability of the outcome that market participants are currently assuming. The stock price of a biotech company depends on many factors: the data of one trial, the amount of money, management decisions, and other drugs in development. Kalshi's contract puts one question into a separate market. Among the first questions are: will the POLARIS-AD study of the AR1001 drug for early Alzheimer's disease reach its primary endpoint, i.e., a pre-designated main result, and will the FDA approve the anito-cel cell therapy for multiple myeloma. If the contract costs 72 cents, market participants at that moment estimate the chance of a "yes" outcome at approximately 72%. Buyers and sellers bet differently, and the price collects their deals into one estimate. The effectiveness of the drug is evaluated based on trial data, the decision to approve is made by the FDA, and the market price reflects current deals. It is also influenced by liquidity, the composition of participants, the moment of trading, and market sentiment. The rules for determining the outcome of the contract are established in advance. Kalshi specifies the question, term, criteria, and public source or order of sources. AppliedXL, a company that analyzes public data on drug development, monitors the named registers and documents, conducts a check with human participation, and prepares an analysis, while Kalshi makes the final decision. "Clinical trial results and regulatory decisions are rarely collected in one final document. They need to be matched with the primary source and criteria set before trading begins," said Francesco Marcon, CEO of AppliedXL. Public quotation can change a patient's or doctor's attitude towards a study, so the pilot covers late-stage trials after patient enrollment is completed. Market participants confirm their place of work. Kalshi's rules prohibit trades by people with substantial non-public information and those who may influence the outcome. Such information can be obtained through a consultant, contractor, professional or personal connections. Kalshi checks information about employers and monitors suspicious trades. In June, Robin Hanson proposed a closed market where experts bet on scientific articles that the company would later base a biotech project on. Kalshi applies the prediction market to the next stage: after selecting a candidate, participants trade the probability of a pre-designated clinical or regulatory outcome. Before the trial results or FDA decision appear, the debate about the chances gets a public price.

🔗 Read original →

On July 16, in Science, a Stanford team's study found that in older mice, the signal through EP2 weakened the work of tissue macrophages. When the authors disabled this receptor only in macrophages, they again cleared aging neutrophils, and the animals' heart, muscle, memory, and other organ functions were preserved. Neutrophils are the most numerous and shortest-lived white blood cells. They are the first to attack an infection and then must quickly disappear: tissue macrophages, the permanent immune cells of organs, engulf and break them down inside themselves. In one day, the human body produces more than 100 billion neutrophils. Macrophages must continuously free organs from this mass of cells. With age, this process failed in mice. In the liver, spleen, bone marrow, and other organs, neutrophils with signs of cellular aging accumulated. They released damaging enzymes and triggered NETosis - the release of networks of DNA and proteins that neutrophils use against microbes. In tissues, these networks and released substances gave inflammatory stress to neighboring cells. The authors found a brake in the macrophages themselves. Prostaglandin E2 - a signaling fatty molecule, the amount of which grows during inflammation - turns on the EP2 receptor on their surface. EP2 disables the first step of cleanup: weakens the work of integrins, proteins that grab. The macrophage worse holds the neutrophil and does not engulf it in a phagolysosome, an internal bubble for breaking down prey. A review of aging macrophages in vascular plaque had already described the loss of cleaning function in one organ; the Tan team named the molecular brake of such a failure in several tissues at once. The team genetically disabled EP2 only in tissue macrophages. In old mice, the clearance of neutrophils returned, and the indicators of fragility, loss of muscle mass, obesity, heart function, memory, and inflammation approached those of young animals. An experimental EP2 inhibitor in 22-month-old mice also reduced the number of aging neutrophils and restored their absorption by macrophages over two months. One path to therapy for aging is to directly remove aging cells. This work suggests restoring the immune system's cleanup of one particularly massive class of cells. In human liver and heart tissues, the authors saw the same age-related pattern: more EP2 in macrophages, more aging neutrophils, fewer contacts between them. The transition to humans will require a selective and safe EP2 inhibitor: the current data show a target and mechanism, but not a ready-made drug.

🔗 Read original →

On July 17 in Nature Aging, a study by Marco Demaria's group from Groningen was published. The authors linked the work of CDK4/6, retinoic acid receptor RARα, and the inflammatory secretions of senescent cells; two interventions in this chain improved physical tests in old mice. A senescent cell stops dividing after damage, but remains in the tissue. Many such cells release signaling substances that support inflammation around them. This set is called SASP, and its NF-κB-controlled part is often associated with age-related inflammation: NF-κB is a protein switch that turns on immune response genes. In June, another Nature Aging study linked SASP to the release of nuclear RNA-DNA hybrids into the cytoplasm and the activation of cGAS-STING. Demaria's group approached the same inflammatory secretions from a different side: through CDK4 and CDK6 - enzymes that normally help the cell prepare for division. The authors took already senescent cells and gave them a short course of abemaciclib. The inflammatory genes of SASP began to work weaker; genetic knockout of CDK4 and CDK6 gave a similar result. Abemaciclib changed the secretions of already aging cells, rather than removing them from the tissue. Such agents are called senomorphics: they try to suppress the harmful behavior of the cell, while keeping it in place. Further, the team traced the chain: CDK4 and CDK6 interacted with NF-κB, and CDK4 also interacted with RARα, a retinoic acid receptor, a derivative of vitamin A. Abemaciclib disrupted these interactions. A separate RARα antagonist, substance agn194310, also weakened the inflammatory secretions. The study suggests the CDK4/6-RARα-NF-κB axis as a new target for the search for senomorphics. In 22-month-old mice, both substances reduced systemic signs of this inflammatory secretions and improved physical performance. The authors administered the drugs for two months; each of the four groups had four animals. The small experience requires independent repetition, especially since abemaciclib is used against some types of cancer and has toxicity. The boundary of the result is visible in another task. In a 2022 study by the same laboratory, CDK4/6 inhibitors introduced previously dividing normal cells into a special senescent state. A fresh preprint by another team saw the later inclusion of part of the NF-κB-dependent SASP in such cells. The current article examines already existing inflammatory senescent cells. These two states require different safety checks.

🔗 Read original →

Microgravity weakens cell contact with tissue and slows down protein assembly in mitochondria. On June 30, a study of cells grown on the ISS was published in Nature Communications: under microgravity, mitochondrial ribosomes assembled proteins more slowly. The authors traced the path from cell attachment to the surrounding tissue to the chemical tuning of this protein factory. Unloading and fixation of the hind limbs of mice gave the same signature in the soleus muscle. Mitochondria make some proteins themselves, with their own genome coding 13 components of the respiratory chain - a machine that converts food energy into ATP. Therefore, mitochondrial translation is important for muscle cells: the speed at which their internal ribosomes read RNA and assemble these proteins. The team of Taishi Wakigawa and Yusuke Kimura grew human cells on the ISS for 24 and 48 hours. A centrifuge on the station, which created normal Earth gravity (1g), served as a control. After 24 hours in microgravity, fewer ribosomes worked on mitochondrial RNA, while the amount of RNA itself changed little: protein assembly slowed down. In worms that lived on the station for four days, the same group measure of translation efficiency decreased. The authors reproduced the effect on Earth in a three-dimensional clinostat - a device that constantly changes the position of cells and mimics the absence of a constant gravity vector. The signal decreased within an hour, returned to its initial level under normal gravity, and increased under tenfold overload. The solution was found outside the cell. Laminin is a protein of the supporting tissue; integrins serve as molecular anchors for it. The better the cell attached to laminin, the faster the mitochondria assembled proteins. An integrin blocker gave the opposite effect and reduced oxygen consumption by mitochondria. Contact with laminin involves a chain of proteins: FAK - an enzyme at the cell attachment site, then RAC1 and PAK1. It changes the work of the BAD protein at the outer mitochondrial membrane and triggers fatty acid synthesis inside it. This process consumes malonyl-CoA, leaving fewer malonyl marks on the translation apparatus, which allows ribosomes to start working faster and elongate the protein chain. In eight-week-old mice, the authors unloaded and fixed the hind limbs for 14 days. In the soleus muscle, both mass and mitochondrial translation decreased simultaneously. The authors checked the chain in cells, worms, and young mice, but have not yet studied old muscle and strength recovery.

🔗 Read original →

Scientists appear to have found evidence that aging is a loss of cellular information, not simply the accumulation of damage.

A new preprint from the New York Genome Center has not yet been peer-reviewed. The work was performed on mouse models. Its translation to humans remains to be verified.

The researchers developed SeqTag technology and measured gene expression, chromatin accessibility, and histone marks in the same cells.

The authors found that in young cells, the three regulatory layers operate asynchronously but synchronize quickly. With age, this coordination breaks down. This is called molecular asynchrony.

This provides direct quantitative confirmation of the Information Theory of Aging (ITOA), which is being developed by David Sinclair, whose startup recently received FDA approval. His theory states that DNA barely changes during aging; the system that reads it changes.

Sinclair has long suspected that this process is reversible; experiments with partial reprogramming demonstrate tissue rejuvenation through the restoration of epigenetic patterns.

What practical changes does this make?

The damage accumulation model sets a therapeutic goal: slowing or repairing damage.

ITOA, however, has a different goal: restoring regulatory synchronization. These are different targets and different pharmacological strategies.

Altos Labs, Rejuvenate Bio, and several other companies are currently working on the latter.


https://www.biorxiv.org/content/10.64898/2026.06.02.729594v1

A review offers an explanation for why one aging target is more effective in worms than in mice. On July 12, in Mechanisms of Ageing and Development, a review was published on why anti-aging interventions often have a significant effect on simple animals and a much smaller effect on mammals. The authors collected data on worms, flies, and rodents. They suggest looking for the reason in the connections between tissues, hormones, and feedback that change the response of the entire organism. Gerontology has long seen this difference. A mutation in the signaling pathway through which insulin and IGF-1 regulate growth and metabolism can double the life of the worm C. elegans. In mice, rapamycin - one of the most studied anti-aging drugs - gave a much smaller result: when it was started at 600 days of life, the age at 90% mortality increased by 14% in females and by 9% in males. In worms, a small set of pathways simultaneously regulates nutrition, stress response, reproduction, and cell repair. The body of mammals distributes these tasks among many tissues. The liver, immune system, fat tissue, muscles, gut bacteria, and hormonal signals constantly change each other's state. The authors call the complexity of the organism the number of such connections and reserves, rather than the size of the body or the length of the genome. The authors call the strength of a single target the leverage of the pathway. The more an organism can change its life by affecting one pathway, the longer this lever. Against it works systemic buffering: backup genes, parallel signaling chains, and feedback between organs return the organism to a working state after intervention. According to this scheme, a drug can precisely alter mTOR - a cellular regulator of growth and resource expenditure - or another known pathway, but its effect will encounter several responses at once. Another tissue compensates for the shift; a hormone changes metabolism; gut bacteria change the availability of a molecule; the organism delivers and processes the drug in its own way. The more such connections, the harder it is for one target to shift the lifespan of the entire body. In mice, there is already an example of such a search: a combination of rapamycin and trametinib increased median life by approximately 30%, stronger than each drug separately. The new review raises an additional question: what combinations need to be selected, taking into account the responses of different tissues to each target. The article does not measure "buffering" with one device and does not establish a limit to human life. This is a qualitative review, not an experiment with a new therapy. It can be tested: comparing which compensatory responses are included after intervention in different tissues, and determining whether combinations of interventions enhance each other. The next generation of research should map: which tissue cancels the beneficial effect, which signal it does this with, and in what sequence to intervene, so that compensation from other organs does not cancel the result.

🔗 Read original →

AI can predict proteins, but therapy remains in the lab, factory, and long money - Dorothy Chow's argument. On July 16, former head of Google DeepMind's public engagement lab, Dorothy Chow, explained why computational breakthroughs alone do not accelerate biology. She suggests funding the entire path from prediction to testing and production. In a conversation with the Foresight Institute, Chow begins with AlphaFold. This system predicts a protein's three-dimensional shape from its sequence; knowing the shape helps understand where a drug molecule or antibody can bind to it. The model has made available the structures of over 200 million proteins. However, a long chain of work remains between the protein shape on the screen and a drug for humans. The success of AlphaFold relied on open data and a common way to verify results. The Protein Data Bank is a global archive of verified protein structures; CASP is a blind test where teams predict forms that have been found experimentally but not yet published. In the AlphaFold2 paper, researchers showed near-experimental accuracy for most proteins in CASP14. Such verification allowed comparing models by accuracy, not by the persuasiveness of presentations. After such a prediction, the expensive part begins: testing the candidate in a wet lab, where cells and tissues are physically worked with, and setting up production. Physical verification and the factory require time, infrastructure, and capital; Chow calls this gap the main obstacle after computation. According to Chow, a venture fund usually waits for an exit from investment within a 5-10 year horizon, and government funding often divides biology into narrow disciplines and annual budgets. She proposes mixed funding: grants and philanthropy pay for the long, risky part, and private capital scales up what has passed verification. An example is the pre-ordered vaccine purchases during COVID-19: the state reduced the risk of production, and companies could act faster. Chow suggests gathering data, independent verification, labs, production, and suitable money in advance for the next big idea. Then, AI can take successful predictions to testing and application, rather than stopping at the screen.

🔗 Read original →

PeptAI stated that its agent, after laboratory testing, independently modified the search for a peptide against VEGFR2 - a target associated with tumor vascular growth. According to the company, the program generated thousands of variants, selected several, paid for binding measurement, and received the results back. PeptAI wrote that these data forced the system to change its peptide generation approach and exclude problematic target areas in the new round. In April, PeptAI showed nine computational checks that the program runs peptides through before transferring them to an external laboratory. In May, its agent proposed 306 peptide variants of MOTS-C and an experimental plan; the candidate synthesis had not been performed yet. On July 17, PeptAI wrote in a new report: "The agent designs a candidate, pays the laboratory for verification, and receives the binding result back, and then starts the next round itself." The company claims that the agent generated thousands of peptides for VEGFR2, selected several, and paid for laboratory verification. VEGFR2 is a receptor on vascular cells: it sends a signal that helps grow new vessels, including those that feed tumors. PeptAI wrote that the positive control and a specially modified variant bound to the target, while the negative control signal did not. The agent took this result as feedback: the next round will use BindCraft - a program for designing proteins capable of binding to a given target - and will avoid known problematic areas on the VEGFR2 surface. The company remains the sole source of these data: the post lacks sequences, protocols, numerical measurements, or independent confirmation. Binding verification answers a narrow question: does the peptide bind to the target in a specific analysis? A therapeutic program then needs data on action in cells and organisms, selectivity, safety, and reproducibility. The model can rank thousands of sequences, but a physical experiment shows where its prediction met the molecule. For the agent, pairs of "sequence - result" become memories of which ideas withstood the test. If the result really changes the next search, each success and each failure narrow the space of the next experiment.

🔗 Read original →