TheoremDB Launch


TheoremDB has introduced a shared journal for AI solving mathematical problems. In the early public version of TheoremDB, work on a problem is broken down into attempts, calculations, and partial results. The next agent sees these records before starting new work; formal proofs can be verified with Lean, a program that checks proofs in strict symbolic form.


TheoremDB released an early public version of a shared workspace for machine mathematics on August 9. It stores work on each problem in separate records: formulations, attempts, calculation files, partial results, and proof states. A separate agent session can leave a calculation, partial result, or failed attempt.


In the analysis of Theo-Conjecture, a counterexample saved in the attempt journal forced the system to change the formula. Articles, libraries, and personal notes usually store the outcome; the next agent run starts without the working memory of the previous one. In TheoremDB, an agent first reads the records on the problem, matches its plan with what has already been done, and adds the result to the journal. These records can be read by anyone, and a new entry is linked to the author's account.


The article summarizes the work; the journal keeps separate steps while the problem is still moving. This memory is useful where two attempts can be exactly matched. In formal proof, Lean brings the proof state to a single form and gives it an exact fingerprint. In computational search with predefined boundaries, an agent matches the already checked range, the spent computational budget, and the result: the next agent sees exactly what has already been checked and how many calculations it took.

🔗 Read original →

Cellular Cleanup Molecule Found


A team from Dana-Farber and Harvard described a method to find small molecules that guide the cell's disposal system to a chosen protein in Nature on 5 August. The search among seven E3-ligases and 5,000 compounds led to M12. A cellular enzyme attaches a glutathione molecule to M12, and the resulting product binds the DCAF11 ligase to the DDX18 protein.

Many proteins lack a convenient site for a drug to attach. A molecular glue temporarily connects the target protein to an E3-ligase. The ligase puts a ubiquitin mark on the protein, and the proteasome, the cell's disposal system, breaks it down. For such glues, only a small number of well-studied E3-ligases have been used so far.

The authors started with seven ligases and a library of compounds to find a new pair. They fixed the ligases on particles, added a library of 5,000 compounds, and cellular lysate. Mass spectrometry detected DDX18 in one of the mixtures. Experiments with each fixed ligase indicated DCAF11, and smaller groups of compounds pointed to M12.

The lysate turned out to be part of the search method. With purified DCAF11 and DDX18, no complex formed; in the lysate, it appeared, and after boiling the lysate, it disappeared. The authors found that the GST enzyme attaches glutathione to M12. The resulting GSH-M12 compound is retained in the DCAF11 site and creates a surface for contact with DDX18.

🔗 Read original →

National Geographic Report


National Geographic has released a report on attempts to rejuvenate cells while preserving their specialization. The report showcases partial cellular reprogramming, an effort to restore a young state to old or damaged cells while maintaining their specialization, on different scales: from a single eye to claims of rejuvenating the entire organism.


The clinical route, ER-100, is an experimental gene therapy introduced into one eye, and is currently recruiting people with open-angle glaucoma or ischemic optic neuropathy - damage caused by disrupted blood supply. The ClinicalTrials.gov registry lists up to 18 participants for the study. On June 9, Life Biosciences announced the first administration of ER-100.


The report considers the ongoing trial as one of the routes to cellular rejuvenation. ER-100 is administered in one eye, and the treatment involves a modified virus carrying instructions for three proteins - OCT4, SOX2, and KLF4, or OSK. These proteins regulate which genes are active in the cell, and eight weeks of doxycycline intake enable their production. The first phase of the trial involves doctors selecting the dose and checking safety, including how participants tolerate the treatment.


Their health and vision will be monitored for up to five years. This route grew out of experience in 2020 on mice, where OSK shifted chemical marks on DNA, regulating gene function, to a younger state; stimulated the recovery of neuronal projections after damage; and improved vision in mice with a laboratory model of glaucoma. The eye allows for the first human test of this approach on a limited area, as explained by Sinclear: "The FDA likes it when an intervention can be limited to one isolated area of the body: it's safer."


In the same report, there are also long-term plans. Cellular reprogramming researcher Juan Carlos Izpisua Belmonte talks about creating the foundation for future medicine, and Sinclear describes SL-100 as a program for rejuvenating the entire organism. Alongside the single-eye trial are laboratory projects and claims of intervention in the entire body. Shin'ya Yamanaka links aging to several processes: changes in epigenetic marks regulating gene function; DNA damage; changes in mitochondrial function and the environment around cells. He believes future strategies will likely combine multiple approaches and calls for public discussion of the consequences of significantly extending life, saying "Significant life extension may affect questions of human existence and values deeper than AI."

🔗 Read original →

Math Priority Rule


Tеренс Тао proposes a new rule for determining priority in mathematics, considering the impact of ИИ on the field. He suggests that priority should be given to the person who first publishes a manuscript, provides a clear explanation understandable by colleagues, and includes a formal verification, which will become more common in the future. The release date, according to Тао, should be considered the moment when the last of these materials is ready.


In the past, during his graduate studies, Тао notes that priority was usually determined by the date of the article in a journal. Later, the public timestamp in open archives like arXiv allowed authors to record when their work became available, eliminating concerns that a reviewer might release a similar result earlier. However, Тао writes that ИИ accelerates this race, as authors may announce answers on social media before verification and clear explanation, just to put a timestamp earlier than others.


Тао proposes separating these tasks among different materials: a manuscript to record the result, a journal dialogue with ИИ to show how the solution was found, Lean - a special language in which a computer verifies the logical steps of the proof, to check its formal record, and a lecture or video to explain the idea to people. "Priority is then determined by who first presents a complete set of materials," Тао writes. Early announcement can inform colleagues that the work is still in progress and help competing groups to unite their efforts or negotiate a synchronized publication. Тао also suggests using a cryptographic hash - a computable fingerprint of a file, which allows later verification of whether the published file matches the recorded version.

🔗 Read original →

Biotech Labs Expand


Ginkgo, a biotech company, is building four university labs for remote experiments, while the National Science Foundation has allocated $380 million to 20 teams creating a network of such facilities. On August 5, the company announced new projects for the California Institute of Technology, Massachusetts Institute of Technology, University of Maryland, and Northwestern University.


Three of these projects are part of a four-year federal program, with the Astera Institute promising to add up to $20 million to the funding. Artificial intelligence can propose a biological hypothesis, but it can only be tested through physical experiments, such as those involving cell cultures, bioreactors, or microscopes.


In its August 5 announcement, Ginkgo described four new university projects that will allow researchers to launch such experiments remotely. In the spring, Ginkgo had already launched Cloud Lab, a service where researchers can remotely order experiments according to a standard protocol. Now, the company is building labs at Caltech, MIT, the University of Maryland, and Northwestern, with three projects part of the federal program to create labs accessible for remote process launches, as described in the National Science Foundation publication.

🔗 Read original →

BMS Adopts AI Agent

The pharmaceutical company Bristol Myers Squibb (BMS) plans to deploy Bunsen, an AI agent for molecular calculations, after Schrödinger announced an agreement with BMS on 5 August. The companies will jointly develop Bunsen, and BMS also plans to use RetroSynth, a program for planning the synthesis of selected molecules. The search for a drug begins with the selection of a target protein, a protein in the body that the drug should affect.


The researcher formulates a scientific goal in ordinary language using Bunsen, which translates it into a sequence of calculations on the Schrödinger platform, launches them, and returns an interpretation of the results. This combines the calculations needed for the chemist's task, and physical models evaluate the properties of candidate molecules. After calculations, the selected molecule needs to be assembled from available substances.


RetroSynth searches for and evaluates chains of chemical reactions, as well as checks proposed reactions, so the selection of a molecule is immediately linked to the route by which it can be obtained in the laboratory. On 20 July, BMS reported on the planned merger of two computational clusters, which should provide research teams with shared data and computing power.

🔗 Read original →

David Sinclair Update


Researcher David Sinclair described the search for a single molecule instead of a cellular cocktail on the August 7 Joe Rogan Experience podcast. Computational screening had narrowed down a trillion molecules to 200 candidates, which his team planned to test on human cells. The search for a single molecule grew out of work with OSK, three proteins that reprogram gene function.


In a 2020 article, Sinclair's group had introduced OSK into mouse retina cells and checked if the cells regained younger characteristics, regrew damaged neuronal projections, and improved vision in a glaucoma model. This combination became the starting point for the search for a single molecule that could perform multiple actions in a cell. In conversation with Joe Rogan, Sinclair named the goal more precisely: one molecule should perform three or four actions of the cocktail in a cell.


The computer first narrowed down the enormous number of options to a list that could be tested in the lab by matching potential molecules with ferments - proteins involved in cellular chemical reactions. According to Sinclair, the work took several months, but without computational screening, it would have taken about 160 years. After calculation, 200 substances remained for the next experiment, as cited in Nature Aging, July 2026.

🔗 Read original →

Petter Attia on Peptides


Petter Attia, author and host of The Drive podcast, released a breakdown of peptides on August 10. He suggests evaluating each promise based on the specific molecule, product, dose, indication, and person it is intended for. In the new episode of The Drive, Attia starts by defining a peptide as a short chain of amino acids. This chemical class includes insulin, which has specific indications and clinical data. The market also uses this term to describe substances that are claimed to have energy, healing, and longevity benefits.


Attia evaluates each promise with five questions. Does the substance have a target - a place in the body where it is supposed to act - and a clear chain to the claimed effect? Has this effect been shown in humans? In what form and dose was the substance studied, how does it behave in the body, and what risk does it pose? Is the balance of benefits and risks suitable for a specific person and their goal? Is there already a studied way to treat the same problem? A study evaluates a specific product: its composition, purity, dose, and method of production. Its result applies to the form of the product and situation that was studied.


As an example, Attia breaks down BPC-157, a peptide associated with healing in the wellness market. In his scheme, he relates this substance to means without scientific support and again goes through questions about the target, result in humans, dosing, and risk. In reviews, the action of the substance is mixed with expectations, concomitant treatment, and changes in behavior. A controlled study separates the effect of the substance itself from these factors. "Nature Aging, July 2026" and other studies are not mentioned in this context, but generally, a substance that worked for one disease, in one group of people, or in one dose, cannot be automatically considered to work just as well after changing any of these conditions. For a decision on application, information is needed about the molecule, product, dose, indication - the disease or condition for which it is applied - and person. When at least one of these parts changes, the result of the previous study needs to be verified in a new situation.

🔗 Read original →

Cancer Models Unveiled


Researchers have made available 665 models of tumors from patients and compared them with the original tissue. An international program has collected 665 tumor models grown from tissues of patients with 25 types of cancer. In an article published on August 5, the authors compared them with the original tissue and made the models and data available to other laboratories.

Patient models are cell cultures grown from tumor tissue: three-dimensional organoids and spheroids, as well as cell lines. They help identify genes that tumor cells depend on and test drugs. However, outside the body, cells live in a different environment and may change some properties of the original tumor. Therefore, it is essential to know what the culture has preserved before experimentation.

In 421 pairs of "patient tumor - grown culture," the authors compared key DNA features. In 97.8% of the models, at least two of the four tested features were preserved, including mutations and the number of times specific DNA segments are repeated. In 201 pairs, they also checked methylation - chemical marks on DNA that participate in gene regulation. In 190 of these pairs (95%), the methylation profile of the culture was closer to the profile of the original tumor than would be expected by random pairing.

The state of the cells was more sensitive to growth conditions: in some cultures, the set of active genes changed. In one model of glioblastoma - an aggressive brain tumor - changing the medium in which the cells grow changed this set. When planning an experiment, growth conditions are essential to consider: they can change the state of the cells. 522 models have detailed clinical information. Along with molecular data, they help researchers studying tumor resistance to treatment choose a culture based on the genome, patient history, and growth conditions.

🔗 Read original →

Biotech Expert Weighs In


Жуан Педру де Магальяйнш, a biogerontologist, responded to a National Geographic report on ER-100 testing, where the treatment is being tested in one eye, with plans to potentially rejuvenate the entire organism. He emphasized the need to differentiate results at the cellular, tissue, and organism levels. The current phase 1 trial involves administering ER-100 to one eye to determine the dosage and ensure safety.


In a post on August 9, де Магальяйнш formulated a boundary with the phrase: "cellular rejuvenation does not equal organismal rejuvenation." Cellular reprogramming attempts to restore a more youthful gene expression regime in adult cells while maintaining their specialization. This is achieved by delivering factors that control gene expression to the cell for a limited time. For example, a retinal cell should remain a retinal cell, and a liver cell should remain a liver cell. In a 2020 experiment on mice, researchers introduced three factors, OSK, into ganglion cells in the retina, resulting in more youthful DNA methylation patterns and improved vision in a glaucoma model.


The composition of factors and the duration of their introduction can alter the outcome. In the same study, continuous introduction of the full set of four factors, OSKM, often led to teratomas or death in mice. Therefore, in experiments with local OSK delivery, authors separately checked the frequency of tumors in the retina. The ER-100 trial is testing this approach in humans, with plans to enroll up to 18 adults with visual nerve diseases, using a modified virus to deliver OSK to cells in one eye, followed by 56 days of doxycycline treatment. To apply this method to the entire organism, the delivery, signal duration, and method of preserving cellular specialization will need to be separately determined for multiple tissues, as reported in Nature Aging, July 2026.

🔗 Read original →

Rapamycin Effects

Researchers compared five laboratory lines of fruit flies on two diets in a study published in the GeroScience journal on August 8. At the same dose, the median lifespan - the age to which half of the flies in a group survive - varied from +5.4% to −51.3%. In the wDah line, the treatment slightly extended life on a medium with brewer's yeast and shortened it on a medium with cornmeal and torula yeast.


The study aimed to understand the varying results of rapamycin, which suppresses the TORC1 signaling system that links nutrient availability to cell growth and protein production. A 2010 study found that 50–400 μM of the treatment extended the life of adult flies, with a dose of 200 μM giving the greatest increase in median lifespan in female wDah line flies.


The authors of the new study tested various conditions, including the solvent, light, food freshness, sex, and genotype of the flies. They compared ethanol with DMSO, the solvent used for the treatment, and found that these conditions altered the response strength, but the largest gap appeared when directly comparing the diets.

🔗 Read original →

Robotic Hand


The robotic hand works unlike ordinary mechanical manipulators, almost like a human hand. The main idea is to transfer all motors from the fingers to the forearm and control finger movement through a system of artificial tendons. These tendons pull each joint, opening and closing it, just like our own muscles and ligaments. This makes the fingers very light, fast, and thin enough to react to a quickly flying object.

The hand was able to catch a baseball in flight because the finger mass is minimal, inertia is low, and bending speed is high. The fingers can not only bend but also spread to the sides, allowing the hand to change its grip shape - from cup-shaped for a ball to point-shaped for small objects. The finger position system determines by how much the motors have turned and how taut the tendons are.

This gives precision without the need for a large number of sensors. Magnetic sensors on the joints measure angles with an accuracy of less than one degree, but they serve as a backup control layer. So far, the hand does not use tactile feedback: the throw was pre-programmed, and the hand caught the ball according to the predicted trajectory, not by sensation.

The hand is designed for Phantom humanoids, which Foundation makes for industrial work. Fast, thin, tendon-driven fingers allow such robots to perform tasks that require dexterity, speed, and precise capture - from sorting parts to working with tools. This is a step towards making robotic hands closer to human functionality, but faster and more robust.

🔗 Source: @solid_state_humanity

Captain Saburo Shimada


Captain Saburo Shimada, a veteran of the New Guinea campaign, trained in kendo with a prosthetic hand in 1944. The photograph clearly shows that this is not a modern bionic prosthetic with electric motors and nerve impulse control. It is a mechanical hand prosthesis designed primarily to restore the ability to hold objects and perform certain movements.


The most interesting thing here is the application. Shimada used the prosthesis not only as a replacement for the lost part of the hand in everyday life but also for kendo training. The design allowed him to hold a sword and participate in training. This is a remarkable example of how long people have been trying to solve the problem of limb loss with mechanics, as reported in undisclosed historical archives.


This example highlights the resourcefulness and determination of individuals like Captain Shimada, who have pushed the boundaries of what is possible with prosthetic technology, even in the 1940s.

🔗 Source: @solid_state_humanity

Math Proof Review


Mathematicians are verifying the novelty of Astra model proofs separately from formal verification. On August 6, Scientific American gathered mathematicians' reactions to the ten Astra results presented by OpenAI on August 1. Mathematicians noted that two of the most notable works used ideas from recent literature, and OpenAI later clarified the formulation of problems that had supposedly seen no progress for decades.


In analyzing the manuscripts and Lean code of Astra's ten results, the focus was on the proofs themselves: OpenAI released manuscripts and formalizations in Lean, a language that checks each recorded logical step. This code allows independent verification of whether a theorem follows from the recorded proof. Verifying novelty requires a different reading, where a mathematician must track what previous results the current construction builds upon and what it adds to them.


This difference is evident in the proof of the existence of a non-sofic group, a mathematical object that cannot be approximated by permutations of a finite number of elements. Astra constructed such an object, and Andreas Thom, co-author of one of the precursor works, showed how the new construction connects Gabor Kuhn's 2016 theorem and Kuhn and Thom's 2019 work. Thom called this construction a "creative and at the same time elementary construction", as reported in Scientific American, August 2026.

🔗 Read original →

Brain Chip Breakthrough


The new technology allows for an artificial neural network to be embedded directly into a memory chip, utilizing the device's physics for calculations. This enables the modeling of complex processes, such as the shape and dynamics of the brain's cortex, with a delay of only 2.12 milliseconds. In comparison, conventional GPU systems like the NVIDIA A100 are 50-470 times slower for the same tasks. This level of speed makes real-time "live" simulations possible, where the simulation runs alongside the actual process.

The chip does not shuttle data between memory and processor like in the classical von Neumann architecture. Instead, it computes directly within the memory array, in a so-called computing-in-memory system. The basis for this is the phase-change memristor, an element that changes conductivity depending on the material's state - crystalline or amorphous. Memristors have a physical "flaw" - conductivity drift, where the value gradually shifts over time. However, engineers have turned this defect into a tool: the drift has become the basis for the adaptive integration step, where the physical process within the chip determines how the neural network should update the model's state.

As a result, the system achieves an integration accuracy of 10⁻⁷ at each step and operates in millisecond mode, consuming 12-25 times less energy than previous neurodynamic systems. It is capable of preserving surface topology (e.g., the shape of the brain's cortex without "tears" and artifacts), which is critical for medicine and robotics, as reported in Nature Aging, July 2026. This opens the path to technologies where a computer can track tissue deformations during surgery, control robots with reactions at the level of biological systems, or simulate physics without delays.

🔗 Read original →

Genome Limit Found


The preprint described the limit of genomic instructions: DNA sets the rules for assembling an organism, while precise molecular movements are determined by physics. In late July, a preprint was published on bioRxiv about the limit of genomic instructions. According to the authors' model, DNA and the signals received by cells determine which cells form tissue and how they work, while precise molecular movements are determined by physical processes.

DNA sets the sequence of amino acids in proteins and the rules by which cells turn genes on and off and respond to signals. From these rules, tissue should emerge, where cells and proteins perform their functions. The authors of the preprint ask: is a finite set of instructions enough to separately specify the precise history of each molecule? At the tissue level, its functional state can be described: what types of cells are in it, how many proteins each cell contains, and which genes are turned on.

The authors show the difference on a protein about 300 amino acids long. Its sequence carries about 1,300 bits of information, and the shape of its main chain with an accuracy of up to an angstrom requires about 6,000 bits. Water, electrical charges, and molecular collisions give this chain a specific three-dimensional shape. According to the authors' model, the genome sets the functional state, and physical processes determine its precise molecular details.

The article cites a 2018 article by geneticist Yussi Tlupova, who formulated the same problem: the genome is finite, and there are many more measurable parameters in cellular biochemistry. Part of the order in the cell arises through self-organization, when molecules bind and arrange themselves according to local physical rules. The current work considers at what level of detail such a description ceases to fit in the genome.

In planarians, a failure of the Notum/Wnt signal, which sets the "head-tail" map, shifted this map in old animals, and regeneration temporarily restored fertility. In this experiment, the concerted work of cells in the right place is essential. In the discussion, the authors relate CRISPR gene editing and light control of cells to interventions that change chains of processes in cells and direct the concerted work of their groups. From this model, an engineering question arises: what genes, chemical and electrical signals, and feedback will reassemble tissue with the necessary function?

🔗 Read original →

Heart Transplant Rejection


Researchers studied 62 patients after heart transplants, comparing the usual microscopic biopsy evaluation with a map of gene activity in individual cells. Despite the same degree of rejection, this map showed different combinations of immune cells, vessels, and heart tissue. Some of these combinations were associated with response to treatment and subsequent narrowing of the transplant vessels.

The study, published on August 10, examined 195 tissue fragments with a diameter of 0.6 mm. The authors measured the activity of 477 pre-selected genes in 162,638 cells and distinguished 28 cell types. Spatial transcriptomics shows which of these genes are active in each cell and where the cell is located in the section.

The authors first compared these cell states with conventional histological grades. Then, they compared biopsies before and after anti-rejection therapy and verified some of the differences in an external set of biopsies. The design and samples are available in a public dataset. In 32 patients with cellular rejection, where T-cells play a leading role, 15 had their rejection resolved after therapy, while 17 did not. Before treatment, the second group had higher activity of genes associated with T-cell activation and tissue remodeling.

The authors also tracked the connection between these states and cardiac vasculopathy of the transplant - the gradual narrowing of the donor heart's vessels. In biopsies during and after rejection episodes, they found cell-specific genes associated with its subsequent development, mainly in vascular and supporting tissue cells. In a separate set of hearts with severe vasculopathy, the same genes were expressed in similar cell types, as reported in Nature Aging, July 2026.

🔗 Read original →

Sanger Institute Opens Cancer Organoid Bank


The Wellcome Sanger Institute has published an open collection of organoids of five cancer types in Nature on August 5. Organoids are three-dimensional cultures grown from tissue of a specific tumor. In 162 such cultures, researchers sequentially turned off genes and measured the effect on cell growth. DNA sequencing shows changes in the tumor; the role of these changes in its growth is determined by experiment on a living culture.


The team collected material through a network of five hospitals. Out of 907 samples from 878 donors, they were able to grow 256 renewable cultures of colorectal, esophageal, ovarian, pancreatic, and stomach cancer. For 171 cultures, a matching original tumor was found; in 76 pairs, at least 75% of somatic mutations matched. This match links the experimental result in culture to the tumor from which it grew.


In 162 organoids, the team applied CRISPR-Cas9, a method that turns off a selected gene. If the culture grows worse after this, it means this gene is needed for its cells to grow. Then researchers compared these dependencies with mutations, tumor subtype, and treatment information. This creates a functional map of mutations: which variant of the tumor needs a specific gene. The logic of the map is visible in the example of the KRAS gene, which often mutates in colorectal cancer.

🔗 Read original →

Codebook Project Findings


The Codebook project has identified the preferred short DNA sequences for 177 understudied proteins that regulate gene function. On August 5, an article about the Codebook project was published in Nature. The authors investigated 332 presumed human transcription factors, which are proteins involved in regulating gene function, and obtained motifs for 177 of them.

A motif is a short set of preferences for DNA letters, showing which sequences a protein binds to more readily. For 130 factors, data from cells showed where such preferences are manifested in the genome. In a 2018 catalog, researchers listed 1,639 presumed human transcription factors, with more than a quarter of them having unknown motifs.

Knowing the motif allows researchers to check if replacing one DNA letter changes the binding of a specific protein and then search for its consequences for gene function. To separate laboratory preference of a protein from its behavior in cells, the Codebook team combined several types of experiments. In 4,804 experiments, they studied 393 proteins: 332 candidates and 61 already known factors for control.

In some experiments, the protein selected DNA sequences it bound to, while in others, it was offered fragments of the human genome. The ChIP-seq method showed which DNA fragments the protein was bound to in cultured human HEK293 cells. A motif was considered reliable when a similar result was obtained by at least two methods and its prediction was confirmed by other experiments.

The authors searched for sites where three lines of data converged: the site contained the motif, the protein bound to it in the experiment on genome fragments, and was detected there in the cell. For 85 out of 101 factors with both types of data, at least one such site was found, where the motif was better preserved in mammals than neighboring DNA.

In total, the authors counted 113,577 such conservative sites: 82,760 for Codebook factors and 30,817 for control factors. This map allows formulating a testable question for a DNA variant: which protein can distinguish between two versions of the sequence and where in the genome to look for consequences for gene function. For 2,260 variants that strongly changed the similarity with the motif, the forecast coincided with the measurement of which of the two variants the protein bound to more often in 1,682 cases.

🔗 Read original →

TERRA Model


The TERRA model, trained on 112.6 million cells from 20 human tissues, was introduced in a preprint on bioRxiv on August 4. This model uses spatial measurements of 636 tissue sections to show both gene activity in a cell and its location within an organ.



The model takes into account the active genes of a central cell and up to 10 neighboring cells, ordered by distance, to predict the representation of missing genes. This allows the model to build connected descriptions of genes, individual cells, and their local environment.



The authors tested the model's ability to predict the effects of gene knockout on tissue structure, starting with the kidney. They found that the model correctly predicted changes in the tissue surrounding cells where specific genes were knocked out, including CTLA4 and PDCD1, which are targets of cancer immunotherapy drugs. The results were published in bioRxiv and the model's weights and code are available.

🔗 Read original →

Browser Agent Test


The Claude Code browser agent gathered a basket of parts from supplier Farnell for a physical experiment. On 7 August, blogger Chill Physics Enjoyer gave Claude Code Desktop an article about a scheme for measuring the Boltzmann constant. The agent prepared a prototype board version of the experiment, selecting parts, creating a mass upload file, and filling the Farnell basket.


The experiment to measure the Boltzmann constant uses Johnson noise - random voltage fluctuations in a heated resistor. To assemble such a scheme, it is necessary to translate the article into resistor and capacitor values, microchip cases, and supplier codes. The author had put the experiment on hold, as selecting dozens of positions took a lot of time.


On 7 August, blogger Chill Physics Enjoyer uploaded the article to Claude Code Desktop and asked to prepare a prototype board version of the scheme. The agent found a replacement for a discontinued microchip that amplifies the signal, selected resistors and capacitors from the Farnell catalog of electronic components, and matched them with supplier codes. Then, it created a table for mass upload, sent it to the website, and filled the basket.


The published journal of interaction preserves the path from replacing a part to lines in the supplier's interface. The basket was checked by the author himself. One resistor with a value of 8.45 kOhm was sold in packs of 5000 for £268. Six more positions had to come from a warehouse in the USA and add £15.95 to the delivery. The agent found options from a warehouse in the UK; to remove previous lines, Farnell's guest mode requires account login. Before payment, the author needs to remove seven positions. In this launch, Claude Code prepared a basket with specific codes, and the account owner checked its composition and received a list of lines to remove before payment, as described in the Chill Physics Enjoyer blog.

🔗 Read original →